Upload 50 files

.gitattributes

@@ -33,3 +33,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/demo/fix_face_weight.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo/gt_generate_compare.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo/talk_tys_fix_face_post_processing.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo/talk_tys_naive_retarget_post_processing.gif filter=lfs diff=lfs merge=lfs -text
+assets/demo/talk_tys_offset_retarget_post_processing.gif filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,39 @@
+---
+tags:
+  - text-to-image
+  - stable-diffusion
+  - audio-to-video
+license: apache-2.0
+language:
+  - en
+library_name: diffusers
+---
+
+# V-Express Model Card
+
+<div align="center">
+
+[**Project Page**](https://tenvence.github.io/p/v-express/) **|** [**Paper**](https://arxiv.org/abs/2406.02511) **|** [**Code**](https://github.com/tencent-ailab/V-Express)
+
+</div>
+
+---
+
+## Introduction
+
+## Models
+
+### Audio Encoder
+
+- [model_ckpts/wav2vec2-base-960h](https://huggingface.co/tk93/V-Express/tree/main/model_ckpts/wav2vec2-base-960h). (It is also available from the original model card [facebook/wav2vec2-base-960h](https://huggingface.co/facebook/wav2vec2-base-960h))
+
+### Face Analysis
+
+- [model_ckpts/insightface_models/models/buffalo_l](https://huggingface.co/tk93/V-Express/tree/main/model_ckpts/insightface_models/models/buffalo_l). (It is also available from the original repository [insightface/buffalo_l](https://github.com/deepinsight/insightface/releases/download/v0.7/buffalo_l.zip))
+
+### V-Express
+
+- [model_ckpts/sd-vae-ft-mse](https://huggingface.co/tk93/V-Express/tree/main/model_ckpts/sd-vae-ft-mse). VAE encoder. (original model card [stabilityai/sd-vae-ft-mse](https://huggingface.co/stabilityai/sd-vae-ft-mse))
+- [model_ckpts/stable-diffusion-v1-5](https://huggingface.co/tk93/V-Express/tree/main/model_ckpts/stable-diffusion-v1-5). Only the model configuration file for unet is needed here. (original model card [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5))
+- [model_ckpts/v-express](https://huggingface.co/tk93/V-Express/tree/main/model_ckpts/v-express). The video generation model conditional on audio and V-kps we call V-Express.
+- You should download and put all `.bin` model to `model_ckpts/v-express` directory, which includes `audio_projection.bin`, `denoising_unet.bin`, `motion_module.bin`, `reference_net.bin`, and `v_kps_guider.bin`.

SECourses.py

@@ -0,0 +1,393 @@
+import os
+import subprocess
+import gradio as gr
+from retinaface import RetinaFace
+from PIL import Image
+import filetype
+from datetime import datetime
+import re
+import sys
+import torch
+import argparse
+
+import platform, os
+
+def open_folder():
+    open_folder_path = os.path.abspath("outputs")
+    if platform.system() == "Windows":
+        os.startfile(open_folder_path)
+    elif platform.system() == "Linux":
+        os.system(f'xdg-open "{open_folder_path}"')
+
+
+# Get the path to the currently activated Python executable
+python_executable = sys.executable
+
+def display_media(file):
+    # Determine the type of the uploaded file using filetype
+    if file is None:
+       return gr.update(visible=False), gr.update(visible=False)
+    kind = filetype.guess(file.name)
+
+    if kind is None:
+        return gr.update(visible=False), gr.update(visible=False)
+
+    if kind.mime.startswith('video'):
+        return gr.update(value=file.name, visible=True), gr.update(visible=False)
+    elif kind.mime.startswith('audio'):
+        return gr.update(visible=False), gr.update(value=file.name, visible=True)
+    else:
+        return gr.update(visible=False), gr.update(visible=False)
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--share", type=str, default=False, help="Set to True to share the app publicly.")
+args = parser.parse_args()
+
+
+# Function to extract audio from video using FFmpeg
+def extract_audio(video_path, audio_path):
+    command = [python_executable, "-m", "ffmpeg", "-i", video_path, "-vn", "-acodec", "libmp3lame", "-q:a", "2", audio_path]
+    subprocess.call(command)
+
+# Function to convert audio to MP3 using FFmpeg
+def convert_audio_to_mp3(audio_path, mp3_path):
+    command = ["ffmpeg", "-i", audio_path, "-acodec", "libmp3lame", "-q:a", "2", mp3_path]
+    subprocess.call(command)
+
+def crop_and_save_image(image_path, auto_crop, crop_width, crop_height, crop_expansion):
+    cropped_image = auto_crop_image(image_path, crop_expansion, crop_size=(crop_width, crop_height))
+    if cropped_image is not None:
+        cropped_folder = os.path.join("outputs", "cropped_images")
+        os.makedirs(cropped_folder, exist_ok=True)
+        
+        # Get the base name and extension of the image file
+        base_name, extension = os.path.splitext(os.path.basename(image_path))
+        
+        # Initialize the counter for the image number
+        counter = 1
+        
+        # Generate the new image name with the incremented number
+        new_image_name = f"{base_name}_{counter:04d}{extension}"
+        cropped_image_path = os.path.join(cropped_folder, new_image_name)
+        
+        # Check if the image already exists and increment the counter until a unique name is found
+        while os.path.exists(cropped_image_path):
+            counter += 1
+            new_image_name = f"{base_name}_{counter:04d}{extension}"
+            cropped_image_path = os.path.join(cropped_folder, new_image_name)
+        
+        # Save the cropped image with the new name
+        cropped_image.save(cropped_image_path, format='PNG')
+        return cropped_image_path
+    return None
+
+# Function to generate kps sequence and audio from video
+def generate_kps_sequence_and_audio(video_path, kps_sequence_save_path, audio_save_path):
+    command = [python_executable, "scripts/extract_kps_sequence_and_audio.py", "--video_path", video_path, "--kps_sequence_save_path", kps_sequence_save_path, "--audio_save_path", audio_save_path]
+    subprocess.call(command)
+
+def auto_crop_image(image_path, expand_percent, crop_size=(512, 512)):
+    # Check if CUDA is available
+    if torch.cuda.is_available():
+        device = 'cuda'
+        print("Using GPU for RetinaFace detection.")
+    else:
+        device = 'cpu'
+        print("Using CPU for RetinaFace detection.")
+
+    # Load image
+    img = Image.open(image_path)
+
+    # Perform face detection
+    faces = RetinaFace.detect_faces(image_path)
+
+    if not faces:
+        print("No faces detected.")
+        return None
+
+    # Assuming 'faces' is a dictionary of detected faces
+    # Pick the first face detected
+    face = list(faces.values())[0]
+    landmarks = face['landmarks']
+
+    # Extract the landmarks
+    right_eye = landmarks['right_eye']
+    left_eye = landmarks['left_eye']
+    right_mouth = landmarks['mouth_right']
+    left_mouth = landmarks['mouth_left']
+
+    # Calculate the distance between the eyes
+    eye_distance = abs(right_eye[0] - left_eye[0])
+
+    # Estimate the head width and height
+    head_width = eye_distance * 4.5  # Increase the width multiplier
+    head_height = eye_distance * 6.5  # Increase the height multiplier
+
+    # Calculate the center point between the eyes
+    eye_center_x = (right_eye[0] + left_eye[0]) // 2
+    eye_center_y = (right_eye[1] + left_eye[1]) // 2
+
+    # Calculate the top-left and bottom-right coordinates of the assumed head region
+    head_left = max(0, int(eye_center_x - head_width // 2))
+    head_top = max(0, int(eye_center_y - head_height // 2))  # Adjust the top coordinate
+    head_right = min(img.width, int(eye_center_x + head_width // 2))
+    head_bottom = min(img.height, int(eye_center_y + head_height // 2))  # Adjust the bottom coordinate
+
+    # Save the assumed head image
+    assumed_head_img = img.crop((head_left, head_top, head_right, head_bottom))
+    assumed_head_img.save("assumed_head.png", format='PNG')
+
+    # Calculate the expansion in pixels and the new dimensions
+    expanded_w = int(head_width * (1 + expand_percent))
+    expanded_h = int(head_height * (1 + expand_percent))
+
+    # Calculate the top-left and bottom-right points of the expanded box
+    center_x, center_y = head_left + head_width // 2, head_top + head_height // 2
+    left = max(0, center_x - expanded_w // 2)
+    right = min(img.width, center_x + expanded_w // 2)
+    top = max(0, center_y - expanded_h // 2)
+    bottom = min(img.height, center_y + expanded_h // 2)
+
+    # Crop the image with the expanded boundaries
+    cropped_img = img.crop((left, top, right, bottom))
+    cropped_img.save("expanded_face.png", format='PNG')
+
+    # Calculate the aspect ratio of the cropped image
+    cropped_width, cropped_height = cropped_img.size
+    aspect_ratio = cropped_width / cropped_height
+
+    # Calculate the target dimensions based on the desired crop size
+    target_width = crop_size[0]
+    target_height = crop_size[1]
+
+    # Adjust the crop to match the desired aspect ratio
+    if aspect_ratio > target_width / target_height:
+        # Crop from left and right
+        new_width = int(cropped_height * target_width / target_height)
+        left_crop = (cropped_width - new_width) // 2
+        right_crop = left_crop + new_width
+        top_crop = 0
+        bottom_crop = cropped_height
+    else:
+        # Crop from top and bottom
+        new_height = int(cropped_width * target_height / target_width)
+        top_crop = (cropped_height - new_height) // 2
+        bottom_crop = top_crop + new_height
+        left_crop = 0
+        right_crop = cropped_width
+
+    # Crop the image with the adjusted boundaries
+    final_cropped_img = cropped_img.crop((left_crop, top_crop, right_crop, bottom_crop))
+    final_cropped_img.save("final_cropped_img.png", format='PNG')
+
+    # Resize the cropped image to the desired size (512x512 by default) with best quality
+    resized_img = final_cropped_img.resize(crop_size, resample=Image.LANCZOS)
+
+    # Save the resized image as PNG
+    resized_img.save(image_path, format='PNG')
+    return resized_img
+     
+
+def generate_output_video(reference_image_path, audio_path, kps_path, output_path, retarget_strategy, num_inference_steps, reference_attention_weight, audio_attention_weight, auto_crop, crop_width, crop_height, crop_expansion,image_width,image_height, low_vram):
+    print("auto cropping...")
+    if auto_crop:
+        auto_crop_image(reference_image_path,crop_expansion, crop_size=(crop_width, crop_height))
+    
+    print("starting inference...")
+    command = [
+        python_executable, "inference.py",
+        "--reference_image_path", reference_image_path,
+        "--audio_path", audio_path,
+        "--kps_path", kps_path,
+        "--output_path", output_path,
+        "--retarget_strategy", retarget_strategy,
+        "--num_inference_steps", str(num_inference_steps),
+        "--reference_attention_weight", str(reference_attention_weight),
+        "--audio_attention_weight", str(audio_attention_weight),     
+        "--image_width", str(image_width),
+        "--image_height", str(image_height)
+    ]
+
+    if low_vram:  # Add the --save_gpu_memory flag if Low VRAM is checked
+        command.append("--save_gpu_memory")
+    
+    with open("executed_command.txt", "w") as file:
+        file.write(" ".join(command))
+    
+    subprocess.call(command)
+    return output_path, reference_image_path
+
+def sanitize_folder_name(name):
+    # Define a regex pattern to match invalid characters for both Linux and Windows
+    invalid_chars = r'[<>:"/\\|?*\x00-\x1F]'
+    # Replace invalid characters with an underscore
+    sanitized_name = re.sub(invalid_chars, '_', name)
+    return sanitized_name
+
+# Function to handle the input and generate the output
+def process_input(reference_image, target_input, retarget_strategy, num_inference_steps, reference_attention_weight, audio_attention_weight, auto_crop, crop_width, crop_height, crop_expansion,image_width,image_height,low_vram):
+    # Create temp_process directory for intermediate files
+    temp_process_dir = "temp_process"
+    os.makedirs(temp_process_dir, exist_ok=True)
+    
+    input_file_name = os.path.splitext(os.path.basename(reference_image))[0]
+    input_file_name=sanitize_folder_name(input_file_name)
+    timestamp = datetime.now().strftime("%Y%m%d%H%M%S")
+    temp_dir = os.path.join(temp_process_dir, f"{input_file_name}_{timestamp}")
+    os.makedirs(temp_dir, exist_ok=True)
+    
+    kind = filetype.guess(target_input)
+    if not kind:
+        raise ValueError("Cannot determine file type. Please provide a valid video or audio file.")
+    
+    mime_type = kind.mime
+    
+    if mime_type.startswith("video/"):  # Video input
+        audio_path = os.path.join(temp_dir, "target_audio.mp3")
+        kps_path = os.path.join(temp_dir, "kps.pth")
+        print("generating generate_kps_sequence_and_audio...")
+        generate_kps_sequence_and_audio(target_input, kps_path, audio_path)
+    elif mime_type.startswith("audio/"):  # Audio input
+        audio_path = target_input
+        if mime_type != "audio/mpeg":
+            mp3_path = os.path.join(temp_dir, "target_audio_converted.mp3")
+            convert_audio_to_mp3(target_input, mp3_path)
+            audio_path = mp3_path
+        kps_path = ""
+    else:
+        raise ValueError("Unsupported file type. Please provide a video or audio file.")
+    
+    output_dir = "outputs"
+    os.makedirs(output_dir, exist_ok=True)
+    output_file_name = f"{input_file_name}_result_"
+    output_file_name=sanitize_folder_name(output_file_name)
+    output_file_ext = ".mp4"
+    output_file_count = 1
+    while os.path.exists(os.path.join(output_dir, f"{output_file_name}{output_file_count:04d}{output_file_ext}")):
+        output_file_count += 1
+    output_path = os.path.join(output_dir, f"{output_file_name}{output_file_count:04d}{output_file_ext}")
+
+    
+    output_video_path, cropped_image_path = generate_output_video(reference_image, audio_path, kps_path, output_path, retarget_strategy, num_inference_steps, reference_attention_weight, audio_attention_weight, auto_crop,crop_width,crop_height, crop_expansion,image_width,image_height,low_vram)
+    
+    return output_video_path, cropped_image_path
+
+def launch_interface():
+    retarget_strategies = ["fix_face", "no_retarget", "offset_retarget", "naive_retarget"]
+   
+    with gr.Blocks() as demo:
+        gr.Markdown("# Tencent AI Lab - V-Express Image to Animation V4 : https://www.patreon.com/posts/105251204")
+        with gr.Row():          
+            with gr.Column():
+                input_image = gr.Image(label="Reference Image", format="png", type="filepath", height=512)
+                generate_button = gr.Button("Generate Talking Video")
+                low_vram = gr.Checkbox(label="Low VRAM - Greatly reduces VRAM usage but takes longer", value=False,visible=False)
+                crop_button = gr.Button("Crop Image")
+                with gr.Row():
+
+                    with gr.Column(min_width=0):
+                        image_width = gr.Number(label="Target Video Width", value=512)
+
+                    with gr.Column(min_width=0):
+                        image_height = gr.Number(label="Target Video Height", value=512)
+
+                with gr.Row():
+                    with gr.Column(min_width=0):
+                        retarget_strategy = gr.Dropdown(retarget_strategies, label="Retarget Strategy", value="fix_face")
+                    with gr.Column(min_width=0):
+                        inference_steps = gr.Slider(10, 90, step=1, label="Number of Inference Steps", value=30)
+
+                with gr.Row():
+                    with gr.Column(min_width=0):
+                        reference_attention = gr.Slider(0.80, 1.1, step=0.01, label="Reference Attention Weight", value=0.95)
+                    with gr.Column(min_width=0):
+                        audio_attention = gr.Slider(1.0, 5.0, step=0.1, label="Audio Attention Weight", value=3.0)             
+
+                with gr.Row(visible=True) as crop_size_row:
+                    with gr.Column(min_width=0):
+                        auto_crop = gr.Checkbox(label="Auto Crop Image", value=True)
+                    with gr.Column(min_width=0):
+                        crop_expansion = gr.Slider(0.0, 1.0, step=0.01, label="Face Focus Expansion Percent", value=0.15)
+                with gr.Row():
+                    with gr.Column(min_width=0):
+                        crop_width = gr.Number(label="Crop Width", value=512)
+                    with gr.Column(min_width=0):
+                        crop_height = gr.Number(label="Crop Height", value=512)
+        
+            with gr.Column():
+                input_video = gr.File(
+                    label="Target Input (Image or Video)",
+                    type="filepath",
+                    file_count="single",
+                    file_types=[
+                        ".mp4", ".avi", ".mov", ".wmv", ".flv", ".mkv", ".webm",  # Video extensions
+                        ".3gp", ".m4v", ".mpg", ".mpeg", ".m2v", ".m4v", ".mts",  # More video extensions
+                        ".mp3", ".wav", ".aac", ".flac", ".m4a", ".wma", ".ogg"   # Audio extensions
+                    ],
+                    height=512        )
+                video_output = gr.Video(visible=False)
+                audio_output = gr.Audio(visible=False)
+    
+                input_video.change(display_media, inputs=input_video, outputs=[video_output, audio_output])
+                btn_open_outputs = gr.Button("Open Outputs Folder")
+                btn_open_outputs.click(fn=open_folder)
+                gr.Markdown("""
+
+                            Retarget Strategies
+
+                            Only target audio : fix_face
+
+                            Input picture and target video (same person - best practice) select : no_retarget
+
+                            Input picture and target video (different person) select : offset_retarget or naive_retarget
+
+                            Please look examples in Tests folder to see which settings you like most. I feel like offset_retarget is best
+
+                            You can turn up reference_attention_weight to make the model maintain higher character consistency, and turn down audio_attention_weight to reduce mouth artifacts. E.g. setting both values to 1.0
+                            """)
+
+
+
+            with gr.Column():
+                output_video = gr.Video(label="Generated Video", height=512)
+                output_image = gr.Image(label="Cropped Image")
+        
+        
+        generate_button.click(
+            fn=process_input,
+            inputs=[
+                input_image,
+                input_video,
+                retarget_strategy,
+                inference_steps,
+                reference_attention,
+                audio_attention,
+                auto_crop,
+                crop_width,
+                crop_height,
+                crop_expansion,
+                image_width,
+                image_height,
+                low_vram
+            ],
+            outputs=[output_video, output_image]
+        )
+
+        crop_button.click(
+            fn=crop_and_save_image,
+            inputs=[
+                input_image,
+                auto_crop,
+                crop_width,
+                crop_height,
+                crop_expansion
+            ],
+            outputs=output_image
+        )
+    
+    demo.queue()
+    demo.launch(inbrowser=True,share=args.share)
+
+# Run the Gradio interface
+launch_interface()

executed_command.txt

@@ -0,0 +1 @@
+R:\V_Express_Installers_V6\V-Express\venv\Scripts\python.exe inference.py --reference_image_path C:\Users\King\AppData\Local\Temp\gradio\e57288fe0b82a668f63c481f82c0fc65eaa84618\Biden_Photo_Big.png --audio_path temp_process\Biden_Photo_Big_20240620112047\target_audio.mp3 --kps_path temp_process\Biden_Photo_Big_20240620112047\kps.pth --output_path outputs\Biden_Photo_Big_result_0006.mp4 --retarget_strategy offset_retarget --num_inference_steps 30 --reference_attention_weight 0.95 --audio_attention_weight 3 --image_width 512 --image_height 512

inference.py

@@ -0,0 +1,291 @@
+import argparse
+import os
+import time
+
+import accelerate
+import cv2
+import numpy as np
+import torch
+import torchaudio.functional
+import torchvision.io
+from PIL import Image
+from diffusers import AutoencoderKL, DDIMScheduler
+from diffusers.utils.import_utils import is_xformers_available
+from insightface.app import FaceAnalysis
+from omegaconf import OmegaConf
+from transformers import Wav2Vec2Model, Wav2Vec2Processor
+
+from modules import UNet2DConditionModel, UNet3DConditionModel, VKpsGuider, AudioProjection
+from pipelines import VExpressPipeline
+from pipelines.utils import draw_kps_image, save_video
+from pipelines.utils import retarget_kps
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--unet_config_path', type=str, default='./model_ckpts/stable-diffusion-v1-5/unet/config.json')
+    parser.add_argument('--vae_path', type=str, default='./model_ckpts/sd-vae-ft-mse/')
+    parser.add_argument('--audio_encoder_path', type=str, default='./model_ckpts/wav2vec2-base-960h/')
+    parser.add_argument('--insightface_model_path', type=str, default='./model_ckpts/insightface_models/')
+
+    parser.add_argument('--denoising_unet_path', type=str, default='./model_ckpts/v-express/denoising_unet.bin')
+    parser.add_argument('--reference_net_path', type=str, default='./model_ckpts/v-express/reference_net.bin')
+    parser.add_argument('--v_kps_guider_path', type=str, default='./model_ckpts/v-express/v_kps_guider.bin')
+    parser.add_argument('--audio_projection_path', type=str, default='./model_ckpts/v-express/audio_projection.bin')
+    parser.add_argument('--motion_module_path', type=str, default='./model_ckpts/v-express/motion_module.bin')
+
+    parser.add_argument('--retarget_strategy', type=str, default='fix_face',
+                        help='{fix_face, no_retarget, offset_retarget, naive_retarget}')
+
+    parser.add_argument('--dtype', type=str, default='fp16')
+    parser.add_argument('--device', type=str, default='cuda')
+    parser.add_argument('--gpu_id', type=int, default=0)
+    parser.add_argument('--do_multi_devices_inference', action='store_true')
+    parser.add_argument('--save_gpu_memory', action='store_true')
+
+    parser.add_argument('--num_pad_audio_frames', type=int, default=2)
+    parser.add_argument('--standard_audio_sampling_rate', type=int, default=16000)
+
+    parser.add_argument('--reference_image_path', type=str, default='./test_samples/emo/talk_emotion/ref.jpg')
+    parser.add_argument('--audio_path', type=str, default='./test_samples/emo/talk_emotion/aud.mp3')
+    parser.add_argument('--kps_path', type=str, default='./test_samples/emo/talk_emotion/kps.pth')
+    parser.add_argument('--output_path', type=str, default='./output/emo/talk_emotion.mp4')
+
+    parser.add_argument('--image_width', type=int, default=512)
+    parser.add_argument('--image_height', type=int, default=512)
+    parser.add_argument('--fps', type=float, default=30.0)
+    parser.add_argument('--seed', type=int, default=42)
+    parser.add_argument('--num_inference_steps', type=int, default=25)
+    parser.add_argument('--guidance_scale', type=float, default=3.5)
+    parser.add_argument('--context_frames', type=int, default=12)
+    parser.add_argument('--context_overlap', type=int, default=4)
+    parser.add_argument('--reference_attention_weight', default=0.95, type=float)
+    parser.add_argument('--audio_attention_weight', default=3., type=float)
+
+    args = parser.parse_args()
+
+    return args
+
+
+def load_reference_net(unet_config_path, reference_net_path, dtype, device):
+    reference_net = UNet2DConditionModel.from_config(unet_config_path).to(dtype=dtype, device=device)
+    reference_net.load_state_dict(torch.load(reference_net_path, map_location="cpu"), strict=False)
+    print(f'Loaded weights of Reference Net from {reference_net_path}.')
+    return reference_net
+
+
+def load_denoising_unet(inf_config_path, unet_config_path, denoising_unet_path, motion_module_path, dtype, device):
+    inference_config = OmegaConf.load(inf_config_path)
+    denoising_unet = UNet3DConditionModel.from_config_2d(
+        unet_config_path,
+        unet_additional_kwargs=inference_config.unet_additional_kwargs,
+    ).to(dtype=dtype, device=device)
+    denoising_unet.load_state_dict(torch.load(denoising_unet_path, map_location="cpu"), strict=False)
+    print(f'Loaded weights of Denoising U-Net from {denoising_unet_path}.')
+
+    denoising_unet.load_state_dict(torch.load(motion_module_path, map_location="cpu"), strict=False)
+    print(f'Loaded weights of Denoising U-Net Motion Module from {motion_module_path}.')
+
+    return denoising_unet
+
+
+def load_v_kps_guider(v_kps_guider_path, dtype, device):
+    v_kps_guider = VKpsGuider(320, block_out_channels=(16, 32, 96, 256)).to(dtype=dtype, device=device)
+    v_kps_guider.load_state_dict(torch.load(v_kps_guider_path, map_location="cpu"))
+    print(f'Loaded weights of V-Kps Guider from {v_kps_guider_path}.')
+    return v_kps_guider
+
+
+def load_audio_projection(
+        audio_projection_path,
+        dtype,
+        device,
+        inp_dim: int,
+        mid_dim: int,
+        out_dim: int,
+        inp_seq_len: int,
+        out_seq_len: int,
+):
+    audio_projection = AudioProjection(
+        dim=mid_dim,
+        depth=4,
+        dim_head=64,
+        heads=12,
+        num_queries=out_seq_len,
+        embedding_dim=inp_dim,
+        output_dim=out_dim,
+        ff_mult=4,
+        max_seq_len=inp_seq_len,
+    ).to(dtype=dtype, device=device)
+    audio_projection.load_state_dict(torch.load(audio_projection_path, map_location='cpu'))
+    print(f'Loaded weights of Audio Projection from {audio_projection_path}.')
+    return audio_projection
+
+
+def get_scheduler(inference_config_path):
+    inference_config = OmegaConf.load(inference_config_path)
+    scheduler_kwargs = OmegaConf.to_container(inference_config.noise_scheduler_kwargs)
+    scheduler = DDIMScheduler(**scheduler_kwargs)
+    return scheduler
+
+
+def main():
+    args = parse_args()
+    start_time = time.time()
+
+    if not args.do_multi_devices_inference:
+        # TODO
+        accelerator = None
+        device = torch.device(f'{args.device}:{args.gpu_id}' if args.device == 'cuda' else args.device)
+    else:
+        accelerator = accelerate.Accelerator()
+        device = torch.device(f'cuda:{accelerator.process_index}')
+    dtype = torch.float16 if args.dtype == 'fp16' else torch.float32
+
+    vae_path = args.vae_path
+    audio_encoder_path = args.audio_encoder_path
+
+    vae = AutoencoderKL.from_pretrained(vae_path).to(dtype=dtype, device=device)
+    audio_encoder = Wav2Vec2Model.from_pretrained(audio_encoder_path).to(dtype=dtype, device=device)
+    audio_processor = Wav2Vec2Processor.from_pretrained(audio_encoder_path)
+
+    unet_config_path = args.unet_config_path
+    reference_net_path = args.reference_net_path
+    denoising_unet_path = args.denoising_unet_path
+    v_kps_guider_path = args.v_kps_guider_path
+    audio_projection_path = args.audio_projection_path
+    motion_module_path = args.motion_module_path
+
+    inference_config_path = './inference_v2.yaml'
+    scheduler = get_scheduler(inference_config_path)
+    reference_net = load_reference_net(unet_config_path, reference_net_path, dtype, device)
+    denoising_unet = load_denoising_unet(
+        inference_config_path, unet_config_path, denoising_unet_path, motion_module_path,
+        dtype, device
+    )
+    v_kps_guider = load_v_kps_guider(v_kps_guider_path, dtype, device)
+    audio_projection = load_audio_projection(
+        audio_projection_path,
+        dtype,
+        device,
+        inp_dim=denoising_unet.config.cross_attention_dim,
+        mid_dim=denoising_unet.config.cross_attention_dim,
+        out_dim=denoising_unet.config.cross_attention_dim,
+        inp_seq_len=2 * (2 * args.num_pad_audio_frames + 1),
+        out_seq_len=2 * args.num_pad_audio_frames + 1,
+    )
+
+    if is_xformers_available():
+        reference_net.enable_xformers_memory_efficient_attention()
+        denoising_unet.enable_xformers_memory_efficient_attention()
+    else:
+        raise ValueError("xformers is not available. Make sure it is installed correctly")
+
+    generator = torch.manual_seed(args.seed)
+    pipeline = VExpressPipeline(
+        vae=vae,
+        reference_net=reference_net,
+        denoising_unet=denoising_unet,
+        v_kps_guider=v_kps_guider,
+        audio_processor=audio_processor,
+        audio_encoder=audio_encoder,
+        audio_projection=audio_projection,
+        scheduler=scheduler,
+    ).to(dtype=dtype, device=device)
+
+    app = FaceAnalysis(
+        providers=['CUDAExecutionProvider' if args.device == 'cuda' else 'CPUExecutionProvider'],
+        provider_options=[{'device_id': args.gpu_id}] if args.device == 'cuda' else [],
+        root=args.insightface_model_path,
+    )
+    app.prepare(ctx_id=0, det_size=(args.image_height, args.image_width))
+
+    reference_image = Image.open(args.reference_image_path).convert('RGB')
+    reference_image = reference_image.resize((args.image_height, args.image_width))
+
+    reference_image_for_kps = cv2.imread(args.reference_image_path)
+    reference_image_for_kps = cv2.resize(reference_image_for_kps, (args.image_width, args.image_height))
+    reference_kps = app.get(reference_image_for_kps)[0].kps[:3]
+    if args.save_gpu_memory:
+        del app
+    torch.cuda.empty_cache()
+
+    _, audio_waveform, meta_info = torchvision.io.read_video(os.path.join(os.path.dirname(args.audio_path), os.path.basename(args.audio_path)), pts_unit='sec')
+    audio_sampling_rate = meta_info['audio_fps']
+    print(f'Length of audio is {audio_waveform.shape[1]} with the sampling rate of {audio_sampling_rate}.')
+    if audio_sampling_rate != args.standard_audio_sampling_rate:
+        audio_waveform = torchaudio.functional.resample(
+            audio_waveform,
+            orig_freq=audio_sampling_rate,
+            new_freq=args.standard_audio_sampling_rate,
+        )
+    audio_waveform = audio_waveform.mean(dim=0)
+
+    duration = audio_waveform.shape[0] / args.standard_audio_sampling_rate
+    init_video_length = int(duration * args.fps)
+    num_contexts = np.around((init_video_length + args.context_overlap) / args.context_frames)
+    video_length = int(num_contexts * args.context_frames - args.context_overlap)
+    fps = video_length / duration
+    print(f'The corresponding video length is {video_length}.')
+
+    kps_sequence = None
+    if args.kps_path != "":
+        assert os.path.exists(args.kps_path), f'{args.kps_path} does not exist'
+        kps_sequence = torch.tensor(torch.load(args.kps_path))  # [len, 3, 2]
+        print(f'The original length of kps sequence is {kps_sequence.shape[0]}.')
+
+        if kps_sequence.shape[0] > video_length:
+            kps_sequence = kps_sequence[:video_length, :, :]
+
+        kps_sequence = torch.nn.functional.interpolate(kps_sequence.permute(1, 2, 0), size=video_length, mode='linear')
+        kps_sequence = kps_sequence.permute(2, 0, 1)
+        print(f'The interpolated length of kps sequence is {kps_sequence.shape[0]}.')
+
+    retarget_strategy = args.retarget_strategy
+    if retarget_strategy == 'fix_face':
+        kps_sequence = torch.tensor([reference_kps] * video_length)
+    elif retarget_strategy == 'no_retarget':
+        kps_sequence = kps_sequence
+    elif retarget_strategy == 'offset_retarget':
+        kps_sequence = retarget_kps(reference_kps, kps_sequence, only_offset=True)
+    elif retarget_strategy == 'naive_retarget':
+        kps_sequence = retarget_kps(reference_kps, kps_sequence, only_offset=False)
+    else:
+        raise ValueError(f'The retarget strategy {retarget_strategy} is not supported.')
+
+    kps_images = []
+    for i in range(video_length):
+        kps_image = draw_kps_image(args.image_height, args.image_width, kps_sequence[i])
+        kps_images.append(Image.fromarray(kps_image))
+
+    video_tensor = pipeline(
+        reference_image=reference_image,
+        kps_images=kps_images,
+        audio_waveform=audio_waveform,
+        width=args.image_width,
+        height=args.image_height,
+        video_length=video_length,
+        num_inference_steps=args.num_inference_steps,
+        guidance_scale=args.guidance_scale,
+        context_frames=args.context_frames,
+        context_overlap=args.context_overlap,
+        reference_attention_weight=args.reference_attention_weight,
+        audio_attention_weight=args.audio_attention_weight,
+        num_pad_audio_frames=args.num_pad_audio_frames,
+        generator=generator,
+        do_multi_devices_inference=args.do_multi_devices_inference,
+        save_gpu_memory=args.save_gpu_memory,
+    )
+
+    if accelerator is None or accelerator.is_main_process:
+        save_video(video_tensor, args.audio_path, args.output_path, device, fps)
+        consumed_time = time.time() - start_time
+        generation_fps = video_tensor.shape[2] / consumed_time
+        print(f'The generated video has been saved at {args.output_path}. '
+              f'The generation time is {consumed_time:.1f} seconds. '
+              f'The generation FPS is {generation_fps:.2f}.')
+
+
+if __name__ == '__main__':
+    main()

inference_v2.yaml

@@ -0,0 +1,35 @@
+unet_additional_kwargs:
+  use_inflated_groupnorm: true
+  unet_use_cross_frame_attention: false
+  unet_use_temporal_attention: false
+  use_motion_module: true
+  motion_module_resolutions:
+    - 1
+    - 2
+    - 4
+    - 8
+  motion_module_mid_block: true
+  motion_module_decoder_only: false
+  motion_module_type: Vanilla
+  motion_module_kwargs:
+    num_attention_heads: 8
+    num_transformer_block: 1
+    attention_block_types:
+      - Temporal_Self
+      - Temporal_Self
+    temporal_position_encoding: true
+    temporal_position_encoding_max_len: 32
+    temporal_attention_dim_div: 1
+
+noise_scheduler_kwargs:
+  beta_start: 0.00085
+  beta_end: 0.012
+  beta_schedule: "scaled_linear"
+  clip_sample: false
+  steps_offset: 1
+  ### Zero-SNR params
+  prediction_type: "v_prediction"
+  rescale_betas_zero_snr: True
+  timestep_spacing: "trailing"
+
+sampler: DDIM

modules/__init__.py

@@ -0,0 +1,5 @@
+from .unet_2d_condition import UNet2DConditionModel
+from .unet_3d import UNet3DConditionModel
+from .v_kps_guider import VKpsGuider
+from .audio_projection import AudioProjection
+from .mutual_self_attention import ReferenceAttentionControl

modules/attention.py

@@ -0,0 +1,626 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.models.attention import AdaLayerNorm, AdaLayerNormZero, Attention, FeedForward, GatedSelfAttentionDense
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from einops import rearrange
+from torch import nn
+
+
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(
+                dim, max_seq_length=num_positional_embeddings
+            )
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(
+                    dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+                )
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim
+                if not double_self_attention
+                else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(
+                dim, cross_attention_dim, num_attention_heads, attention_head_dim
+            )
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = (
+            cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        )
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states
+            if self.only_cross_attention
+            else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = (
+                norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+            )
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class TemporalBasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
+        self.unet_use_temporal_attention = unet_use_temporal_attention
+
+        # old self attention layer for only self-attention
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+        self.norm1 = (
+            AdaLayerNorm(dim, num_embeds_ada_norm)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(dim)
+        )
+
+        # new self attention layer for reference features
+        self.attn1_5 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+        self.norm1_5 = (
+            AdaLayerNorm(dim, num_embeds_ada_norm)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(dim)
+        )
+
+        # Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn2 = None
+
+        if cross_attention_dim is not None:
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+        else:
+            self.norm2 = None
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+        self.use_ada_layer_norm_zero = False
+
+        # Temp-Attn
+        assert unet_use_temporal_attention is not None
+        if unet_use_temporal_attention:
+            self.attn_temp = Attention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+            nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
+            self.norm_temp = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        timestep=None,
+        attention_mask=None,
+        video_length=None,
+    ):
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep)
+            if self.use_ada_layer_norm
+            else self.norm1(hidden_states)
+        )
+
+        if self.unet_use_cross_frame_attention:
+            hidden_states = (
+                self.attn1(
+                    norm_hidden_states,
+                    attention_mask=attention_mask,
+                    video_length=video_length,
+                )
+                + hidden_states
+            )
+        else:
+            hidden_states = (
+                self.attn1(norm_hidden_states, attention_mask=attention_mask)
+                + hidden_states
+            )
+
+        norm_hidden_states = (
+            self.norm1_5(hidden_states, timestep)
+            if self.use_ada_layer_norm
+            else self.norm1_5(hidden_states)
+        )
+
+        if self.unet_use_cross_frame_attention:
+            hidden_states = (
+                self.attn1_5(
+                    norm_hidden_states,
+                    attention_mask=attention_mask,
+                    video_length=video_length,
+                )
+                + hidden_states
+            )
+        else:
+            hidden_states = (
+                self.attn1_5(norm_hidden_states, attention_mask=attention_mask)
+                + hidden_states
+            )
+
+        if self.attn2 is not None:
+            # Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm2(hidden_states)
+            )
+            hidden_states = (
+                self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                )
+                + hidden_states
+            )
+
+        # Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        # Temporal-Attention
+        if self.unet_use_temporal_attention:
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+            norm_hidden_states = (
+                self.norm_temp(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm_temp(hidden_states)
+            )
+            hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states
+
+class TemporalBasicTransformerBlockOld(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
+        self.unet_use_temporal_attention = unet_use_temporal_attention
+
+        # SC-Attn
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+        self.norm1 = (
+            AdaLayerNorm(dim, num_embeds_ada_norm)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(dim)
+        )
+
+        # Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn2 = None
+
+        if cross_attention_dim is not None:
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+        else:
+            self.norm2 = None
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+        self.use_ada_layer_norm_zero = False
+
+        # Temp-Attn
+        assert unet_use_temporal_attention is not None
+        if unet_use_temporal_attention:
+            self.attn_temp = Attention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+            nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
+            self.norm_temp = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        timestep=None,
+        attention_mask=None,
+        video_length=None,
+    ):
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep)
+            if self.use_ada_layer_norm
+            else self.norm1(hidden_states)
+        )
+
+        if self.unet_use_cross_frame_attention:
+            hidden_states = (
+                self.attn1(
+                    norm_hidden_states,
+                    attention_mask=attention_mask,
+                    video_length=video_length,
+                )
+                + hidden_states
+            )
+        else:
+            hidden_states = (
+                self.attn1(norm_hidden_states, attention_mask=attention_mask)
+                + hidden_states
+            )
+
+        if self.attn2 is not None:
+            # Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm2(hidden_states)
+            )
+            hidden_states = (
+                self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                )
+                + hidden_states
+            )
+
+        # Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        # Temporal-Attention
+        if self.unet_use_temporal_attention:
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+            norm_hidden_states = (
+                self.norm_temp(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm_temp(hidden_states)
+            )
+            hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

modules/audio_projection.py

@@ -0,0 +1,150 @@
+import math
+
+import torch
+import torch.nn as nn
+from diffusers.models.modeling_utils import ModelMixin
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+def masked_mean(t, *, dim, mask=None):
+    if mask is None:
+        return t.mean(dim=dim)
+
+    denom = mask.sum(dim=dim, keepdim=True)
+    mask = rearrange(mask, "b n -> b n 1")
+    masked_t = t.masked_fill(~mask, 0.0)
+
+    return masked_t.sum(dim=dim) / denom.clamp(min=1e-5)
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head ** -0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+class AudioProjection(ModelMixin):
+    def __init__(
+            self,
+            dim=1024,
+            depth=8,
+            dim_head=64,
+            heads=16,
+            num_queries=8,
+            embedding_dim=768,
+            output_dim=1024,
+            ff_mult=4,
+            max_seq_len: int = 257,
+            num_latents_mean_pooled: int = 0,
+    ):
+        super().__init__()
+
+        self.pos_emb = nn.Embedding(max_seq_len, embedding_dim)
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim ** 0.5)
+
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.to_latents_from_mean_pooled_seq = (
+            nn.Sequential(
+                nn.LayerNorm(dim),
+                nn.Linear(dim, dim * num_latents_mean_pooled),
+                Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled),
+            )
+            if num_latents_mean_pooled > 0
+            else None
+        )
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(nn.ModuleList([
+                PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                FeedForward(dim=dim, mult=ff_mult),
+            ]))
+
+    def forward(self, x):
+        if self.pos_emb is not None:
+            n, device = x.shape[1], x.device
+            pos_emb = self.pos_emb(torch.arange(n, device=device))
+            x = x + pos_emb
+
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        if self.to_latents_from_mean_pooled_seq:
+            meanpooled_seq = masked_mean(x, dim=1, mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool))
+            meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq)
+            latents = torch.cat((meanpooled_latents, latents), dim=-2)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)

modules/motion_module.py

@@ -0,0 +1,388 @@
+# Adapt from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/motion_module.py
+import math
+from dataclasses import dataclass
+from typing import Callable, Optional
+
+import torch
+from diffusers.models.attention import FeedForward
+from diffusers.models.attention_processor import Attention, AttnProcessor
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from einops import rearrange, repeat
+from torch import nn
+
+
+def zero_module(module):
+    # Zero out the parameters of a module and return it.
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+@dataclass
+class TemporalTransformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+def get_motion_module(in_channels, motion_module_type: str, motion_module_kwargs: dict):
+    if motion_module_type == "Vanilla":
+        return VanillaTemporalModule(
+            in_channels=in_channels,
+            **motion_module_kwargs,
+        )
+    else:
+        raise ValueError
+
+
+class VanillaTemporalModule(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads=8,
+        num_transformer_block=2,
+        attention_block_types=("Temporal_Self", "Temporal_Self"),
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        temporal_attention_dim_div=1,
+        zero_initialize=True,
+    ):
+        super().__init__()
+
+        self.temporal_transformer = TemporalTransformer3DModel(
+            in_channels=in_channels,
+            num_attention_heads=num_attention_heads,
+            attention_head_dim=in_channels
+            // num_attention_heads
+            // temporal_attention_dim_div,
+            num_layers=num_transformer_block,
+            attention_block_types=attention_block_types,
+            cross_frame_attention_mode=cross_frame_attention_mode,
+            temporal_position_encoding=temporal_position_encoding,
+            temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+        )
+
+        if zero_initialize:
+            self.temporal_transformer.proj_out = zero_module(
+                self.temporal_transformer.proj_out
+            )
+
+    def forward(
+        self,
+        input_tensor,
+        temb,
+        encoder_hidden_states,
+        attention_mask=None,
+        anchor_frame_idx=None,
+    ):
+        hidden_states = input_tensor
+        hidden_states = self.temporal_transformer(
+            hidden_states, encoder_hidden_states, attention_mask
+        )
+
+        output = hidden_states
+        return output
+
+
+class TemporalTransformer3DModel(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads,
+        attention_head_dim,
+        num_layers,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
+        )
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalTransformerBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    attention_block_types=attention_block_types,
+                    dropout=dropout,
+                    norm_num_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+                for d in range(num_layers)
+            ]
+        )
+        self.proj_out = nn.Linear(inner_dim, in_channels)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        assert (
+            hidden_states.dim() == 5
+        ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+            batch, height * weight, inner_dim
+        )
+        hidden_states = self.proj_in(hidden_states)
+
+        # Transformer Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                video_length=video_length,
+            )
+
+        # output
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = (
+            hidden_states.reshape(batch, height, weight, inner_dim)
+            .permute(0, 3, 1, 2)
+            .contiguous()
+        )
+
+        output = hidden_states + residual
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+
+        return output
+
+
+class TemporalTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_attention_heads,
+        attention_head_dim,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        attention_blocks = []
+        norms = []
+
+        for block_name in attention_block_types:
+            attention_blocks.append(
+                VersatileAttention(
+                    attention_mode=block_name.split("_")[0],
+                    cross_attention_dim=cross_attention_dim
+                    if block_name.endswith("_Cross")
+                    else None,
+                    query_dim=dim,
+                    heads=num_attention_heads,
+                    dim_head=attention_head_dim,
+                    dropout=dropout,
+                    bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+            )
+            norms.append(nn.LayerNorm(dim))
+
+        self.attention_blocks = nn.ModuleList(attention_blocks)
+        self.norms = nn.ModuleList(norms)
+
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.ff_norm = nn.LayerNorm(dim)
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        video_length=None,
+    ):
+        for attention_block, norm in zip(self.attention_blocks, self.norms):
+            norm_hidden_states = norm(hidden_states)
+            hidden_states = (
+                attention_block(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states
+                    if attention_block.is_cross_attention
+                    else None,
+                    video_length=video_length,
+                )
+                + hidden_states
+            )
+
+        hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states
+
+        output = hidden_states
+        return output
+
+
+class PositionalEncoding(nn.Module):
+    def __init__(self, d_model, dropout=0.0, max_len=24):
+        super().__init__()
+        self.dropout = nn.Dropout(p=dropout)
+        position = torch.arange(max_len).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
+        )
+        pe = torch.zeros(1, max_len, d_model)
+        pe[0, :, 0::2] = torch.sin(position * div_term)
+        pe[0, :, 1::2] = torch.cos(position * div_term)
+        self.register_buffer("pe", pe)
+
+    def forward(self, x):
+        x = x + self.pe[:, : x.size(1)]
+        return self.dropout(x)
+
+
+class VersatileAttention(Attention):
+    def __init__(
+        self,
+        attention_mode=None,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        *args,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        assert attention_mode == "Temporal"
+
+        self.attention_mode = attention_mode
+        self.is_cross_attention = kwargs["cross_attention_dim"] is not None
+
+        self.pos_encoder = (
+            PositionalEncoding(
+                kwargs["query_dim"],
+                dropout=0.0,
+                max_len=temporal_position_encoding_max_len,
+            )
+            if (temporal_position_encoding and attention_mode == "Temporal")
+            else None
+        )
+
+    def extra_repr(self):
+        return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"
+
+    def set_use_memory_efficient_attention_xformers(
+        self,
+        use_memory_efficient_attention_xformers: bool,
+        attention_op: Optional[Callable] = None,
+    ):
+        if use_memory_efficient_attention_xformers:
+            if not is_xformers_available():
+                raise ModuleNotFoundError(
+                    (
+                        "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                        " xformers"
+                    ),
+                    name="xformers",
+                )
+            elif not torch.cuda.is_available():
+                raise ValueError(
+                    "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
+                    " only available for GPU "
+                )
+            else:
+                try:
+                    # Make sure we can run the memory efficient attention
+                    _ = xformers.ops.memory_efficient_attention(
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                    )
+                except Exception as e:
+                    raise e
+
+            # XFormersAttnProcessor corrupts video generation and work with Pytorch 1.13.
+            # Pytorch 2.0.1 AttnProcessor works the same as XFormersAttnProcessor in Pytorch 1.13.
+            # You don't need XFormersAttnProcessor here.
+            # processor = XFormersAttnProcessor(
+            #     attention_op=attention_op,
+            # )
+            processor = AttnProcessor()
+        else:
+            processor = AttnProcessor()
+
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        video_length=None,
+        **cross_attention_kwargs,
+    ):
+        if self.attention_mode == "Temporal":
+            d = hidden_states.shape[1]  # d means HxW
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+
+            if self.pos_encoder is not None:
+                hidden_states = self.pos_encoder(hidden_states)
+
+            encoder_hidden_states = (
+                repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d)
+                if encoder_hidden_states is not None
+                else encoder_hidden_states
+            )
+
+        else:
+            raise NotImplementedError
+
+        hidden_states = self.processor(
+            self,
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+        if self.attention_mode == "Temporal":
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

modules/mutual_self_attention.py

@@ -0,0 +1,376 @@
+# Adapted from https://github.com/magic-research/magic-animate/blob/main/magicanimate/models/mutual_self_attention.py
+from typing import Any, Dict, Optional
+
+import torch
+from einops import rearrange
+
+from .attention import BasicTransformerBlock
+from .attention import TemporalBasicTransformerBlock
+
+
+def torch_dfs(model: torch.nn.Module):
+    result = [model]
+    for child in model.children():
+        result += torch_dfs(child)
+    return result
+
+
+class ReferenceAttentionControl:
+    def __init__(
+            self,
+            unet,
+            mode="write",
+            do_classifier_free_guidance=False,
+            attention_auto_machine_weight=float("inf"),
+            gn_auto_machine_weight=1.0,
+            style_fidelity=1.0,
+            reference_attn=True,
+            reference_adain=False,
+            fusion_blocks="midup",
+            batch_size=1,
+            reference_attention_weight=1.,
+            audio_attention_weight=1.,
+    ) -> None:
+        # 10. Modify self attention and group norm
+        self.unet = unet
+        assert mode in ["read", "write"]
+        assert fusion_blocks in ["midup", "full"]
+        self.reference_attn = reference_attn
+        self.reference_adain = reference_adain
+        self.fusion_blocks = fusion_blocks
+        self.reference_attention_weight = reference_attention_weight
+        self.audio_attention_weight = audio_attention_weight
+        self.register_reference_hooks(
+            mode,
+            do_classifier_free_guidance,
+            attention_auto_machine_weight,
+            gn_auto_machine_weight,
+            style_fidelity,
+            reference_attn,
+            reference_adain,
+            fusion_blocks,
+            batch_size=batch_size,
+        )
+
+    def register_reference_hooks(
+            self,
+            mode,
+            do_classifier_free_guidance,
+            attention_auto_machine_weight,
+            gn_auto_machine_weight,
+            style_fidelity,
+            reference_attn,
+            reference_adain,
+            dtype=torch.float16,
+            batch_size=1,
+            num_images_per_prompt=1,
+            device=torch.device("cpu"),
+            fusion_blocks="midup",
+    ):
+        MODE = mode
+        do_classifier_free_guidance = do_classifier_free_guidance
+        attention_auto_machine_weight = attention_auto_machine_weight
+        gn_auto_machine_weight = gn_auto_machine_weight
+        style_fidelity = style_fidelity
+        reference_attn = reference_attn
+        reference_adain = reference_adain
+        fusion_blocks = fusion_blocks
+        num_images_per_prompt = num_images_per_prompt
+        reference_attention_weight = self.reference_attention_weight
+        audio_attention_weight = self.audio_attention_weight
+        dtype = dtype
+        if do_classifier_free_guidance:
+            uc_mask = (
+                torch.Tensor(
+                    [1] * batch_size * num_images_per_prompt * 16
+                    + [0] * batch_size * num_images_per_prompt * 16
+                )
+                .to(device)
+                .bool()
+            )
+        else:
+            uc_mask = (
+                torch.Tensor([0] * batch_size * num_images_per_prompt * 2)
+                .to(device)
+                .bool()
+            )
+
+        def hacked_basic_transformer_inner_forward(
+                self,
+                hidden_states: torch.FloatTensor,
+                attention_mask: Optional[torch.FloatTensor] = None,
+                encoder_hidden_states: Optional[torch.FloatTensor] = None,
+                encoder_attention_mask: Optional[torch.FloatTensor] = None,
+                timestep: Optional[torch.LongTensor] = None,
+                cross_attention_kwargs: Dict[str, Any] = None,
+                class_labels: Optional[torch.LongTensor] = None,
+                video_length=None,
+        ):
+            if self.use_ada_layer_norm:  # False
+                norm_hidden_states = self.norm1(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero:
+                (
+                    norm_hidden_states,
+                    gate_msa,
+                    shift_mlp,
+                    scale_mlp,
+                    gate_mlp,
+                ) = self.norm1(
+                    hidden_states,
+                    timestep,
+                    class_labels,
+                    hidden_dtype=hidden_states.dtype,
+                )
+            else:
+                norm_hidden_states = self.norm1(hidden_states)
+
+            # 1. Self-Attention
+            # self.only_cross_attention = False
+            cross_attention_kwargs = (
+                cross_attention_kwargs if cross_attention_kwargs is not None else {}
+            )
+            if self.only_cross_attention:
+                attn_output = self.attn1(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states
+                    if self.only_cross_attention
+                    else None,
+                    attention_mask=attention_mask,
+                    **cross_attention_kwargs,
+                )
+            else:
+                if MODE == "write":
+                    attn_output = self.attn1(
+                        norm_hidden_states,
+                        encoder_hidden_states=encoder_hidden_states
+                        if self.only_cross_attention
+                        else None,
+                        attention_mask=attention_mask,
+                        **cross_attention_kwargs,
+                    )
+
+                    if self.use_ada_layer_norm_zero:
+                        attn_output = gate_msa.unsqueeze(1) * attn_output
+                    hidden_states = attn_output + hidden_states
+
+                    if self.attn2 is not None:
+                        norm_hidden_states = (
+                            self.norm2(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm2(hidden_states)
+                        )
+                        self.bank.append(norm_hidden_states.clone())
+
+                        # 2. Cross-Attention
+                        attn_output = self.attn2(
+                            norm_hidden_states,
+                            encoder_hidden_states=encoder_hidden_states,
+                            attention_mask=encoder_attention_mask,
+                            **cross_attention_kwargs,
+                        )
+                        hidden_states = attn_output + hidden_states
+
+                if MODE == "read":
+                    hidden_states = (
+                            self.attn1(
+                                norm_hidden_states,
+                                encoder_hidden_states=norm_hidden_states,
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states
+                    )
+
+                    if self.use_ada_layer_norm:  # False
+                        norm_hidden_states = self.norm1_5(hidden_states, timestep)
+                    elif self.use_ada_layer_norm_zero:
+                        (
+                            norm_hidden_states,
+                            gate_msa,
+                            shift_mlp,
+                            scale_mlp,
+                            gate_mlp,
+                        ) = self.norm1_5(
+                            hidden_states,
+                            timestep,
+                            class_labels,
+                            hidden_dtype=hidden_states.dtype,
+                        )
+                    else:
+                        norm_hidden_states = self.norm1_5(hidden_states)
+
+                    bank_fea = []
+                    for d in self.bank:
+                        if len(d.shape) == 3:
+                            d = d.unsqueeze(1).repeat(1, video_length, 1, 1)
+                        bank_fea.append(rearrange(d, "b t l c -> (b t) l c"))
+
+                    attn_hidden_states = self.attn1_5(
+                        norm_hidden_states,
+                        encoder_hidden_states=bank_fea[0],
+                        attention_mask=attention_mask,
+                    )
+
+                    if reference_attention_weight != 1.:
+                        attn_hidden_states *= reference_attention_weight
+
+                    hidden_states = (attn_hidden_states + hidden_states)
+
+                    # self.bank.clear()
+                    if self.attn2 is not None:
+                        # Cross-Attention
+                        norm_hidden_states = (
+                            self.norm2(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm2(hidden_states)
+                        )
+
+                        attn_hidden_states = self.attn2(
+                            norm_hidden_states,
+                            encoder_hidden_states=encoder_hidden_states,
+                            attention_mask=attention_mask,
+                        )
+
+                        if audio_attention_weight != 1.:
+                            attn_hidden_states *= audio_attention_weight
+
+                        hidden_states = (attn_hidden_states + hidden_states)
+
+                    # Feed-forward
+                    hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+                    # Temporal-Attention
+                    if self.unet_use_temporal_attention:
+                        d = hidden_states.shape[1]
+                        hidden_states = rearrange(
+                            hidden_states, "(b f) d c -> (b d) f c", f=video_length
+                        )
+                        norm_hidden_states = (
+                            self.norm_temp(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm_temp(hidden_states)
+                        )
+                        hidden_states = (
+                                self.attn_temp(norm_hidden_states) + hidden_states
+                        )
+                        hidden_states = rearrange(
+                            hidden_states, "(b d) f c -> (b f) d c", d=d
+                        )
+
+                    return hidden_states
+
+            # 3. Feed-forward
+            norm_hidden_states = self.norm3(hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                norm_hidden_states = (
+                        norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+                )
+
+            ff_output = self.ff(norm_hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+            hidden_states = ff_output + hidden_states
+
+            return hidden_states
+
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                attn_modules = [
+                    module
+                    for module in (
+                            torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                       or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                       or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            attn_modules = sorted(
+                attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+
+            for i, module in enumerate(attn_modules):
+                module._original_inner_forward = module.forward
+                if isinstance(module, BasicTransformerBlock):
+                    module.forward = hacked_basic_transformer_inner_forward.__get__(
+                        module, BasicTransformerBlock
+                    )
+                if isinstance(module, TemporalBasicTransformerBlock):
+                    module.forward = hacked_basic_transformer_inner_forward.__get__(
+                        module, TemporalBasicTransformerBlock
+                    )
+
+                module.bank = []
+                module.attn_weight = float(i) / float(len(attn_modules))
+
+    def update(
+            self,
+            writer,
+            do_classifier_free_guidance=True,
+            dtype=torch.float16,
+    ):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks))
+                    if isinstance(module, TemporalBasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in (torch_dfs(writer.unet.mid_block) + torch_dfs(writer.unet.up_blocks))
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, TemporalBasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in torch_dfs(writer.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            writer_attn_modules = sorted(
+                writer_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            for r, w in zip(reader_attn_modules, writer_attn_modules):
+                if do_classifier_free_guidance:
+                    r.bank = [torch.cat([torch.zeros_like(v), v]).to(dtype) for v in w.bank]
+                else:
+                    r.bank = [v.clone().to(dtype) for v in w.bank]
+
+    def clear(self):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (
+                            torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                       or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                       or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            for r in reader_attn_modules:
+                r.bank.clear()

modules/resnet.py

@@ -0,0 +1,256 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/resnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+
+class InflatedConv3d(nn.Conv2d):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class InflatedGroupNorm(nn.GroupNorm):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class Upsample3D(nn.Module):
+    def __init__(
+            self,
+            channels,
+            use_conv=False,
+            use_conv_transpose=False,
+            out_channels=None,
+            name="conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        conv = None
+        if use_conv_transpose:
+            raise NotImplementedError
+        elif use_conv:
+            self.conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, hidden_states, output_size=None):
+        assert hidden_states.shape[1] == self.channels
+
+        if self.use_conv_transpose:
+            raise NotImplementedError
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if output_size is None:
+            hidden_states = F.interpolate(
+                hidden_states, scale_factor=[1.0, 2.0, 2.0], mode="nearest"
+            )
+        else:
+            hidden_states = F.interpolate(
+                hidden_states, size=output_size, mode="nearest"
+            )
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        # if self.use_conv:
+        #     if self.name == "conv":
+        #         hidden_states = self.conv(hidden_states)
+        #     else:
+        #         hidden_states = self.Conv2d_0(hidden_states)
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class Downsample3D(nn.Module):
+    def __init__(
+            self, channels, use_conv=False, out_channels=None, padding=1, name="conv"
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if use_conv:
+            self.conv = InflatedConv3d(
+                self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            raise NotImplementedError
+
+    def forward(self, hidden_states):
+        assert hidden_states.shape[1] == self.channels
+        if self.use_conv and self.padding == 0:
+            raise NotImplementedError
+
+        assert hidden_states.shape[1] == self.channels
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class ResnetBlock3D(nn.Module):
+    def __init__(
+            self,
+            *,
+            in_channels,
+            out_channels=None,
+            conv_shortcut=False,
+            dropout=0.0,
+            temb_channels=512,
+            groups=32,
+            groups_out=None,
+            pre_norm=True,
+            eps=1e-6,
+            non_linearity="swish",
+            time_embedding_norm="default",
+            output_scale_factor=1.0,
+            use_in_shortcut=None,
+            use_inflated_groupnorm=None,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.pre_norm = True
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.time_embedding_norm = time_embedding_norm
+        self.output_scale_factor = output_scale_factor
+
+        if groups_out is None:
+            groups_out = groups
+
+        assert use_inflated_groupnorm != None
+        if use_inflated_groupnorm:
+            self.norm1 = InflatedGroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True
+            )
+        else:
+            self.norm1 = torch.nn.GroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True
+            )
+
+        self.conv1 = InflatedConv3d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                time_emb_proj_out_channels = out_channels
+            elif self.time_embedding_norm == "scale_shift":
+                time_emb_proj_out_channels = out_channels * 2
+            else:
+                raise ValueError(
+                    f"unknown time_embedding_norm : {self.time_embedding_norm} "
+                )
+
+            self.time_emb_proj = torch.nn.Linear(
+                temb_channels, time_emb_proj_out_channels
+            )
+        else:
+            self.time_emb_proj = None
+
+        if use_inflated_groupnorm:
+            self.norm2 = InflatedGroupNorm(
+                num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
+            )
+        else:
+            self.norm2 = torch.nn.GroupNorm(
+                num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
+            )
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = InflatedConv3d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+        if non_linearity == "swish":
+            self.nonlinearity = lambda x: F.silu(x)
+        elif non_linearity == "mish":
+            self.nonlinearity = Mish()
+        elif non_linearity == "silu":
+            self.nonlinearity = nn.SiLU()
+
+        self.use_in_shortcut = (
+            self.in_channels != self.out_channels
+            if use_in_shortcut is None
+            else use_in_shortcut
+        )
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = InflatedConv3d(
+                in_channels, out_channels, kernel_size=1, stride=1, padding=0
+            )
+
+    def forward(self, input_tensor, temb):
+        hidden_states = input_tensor
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            temb = self.time_emb_proj(self.nonlinearity(temb))
+            if len(temb.shape) == 2:
+                temb = temb[:, :, None, None, None]
+            elif len(temb.shape) == 3:
+                temb = temb[:, :, :, None, None].permute(0, 2, 1, 3, 4)
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+class Mish(torch.nn.Module):
+    def forward(self, hidden_states):
+        return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))

modules/transformer_2d.py

@@ -0,0 +1,401 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformer_2d.py
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+
+try:
+    from diffusers.models.embeddings import CaptionProjection
+except:
+    from diffusers.models.embeddings import PixArtAlphaTextProjection as CaptionProjection
+
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
+from torch import nn
+
+from .attention import BasicTransformerBlock
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+    ref_feature: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate(
+                "norm_type!=num_embeds_ada_norm",
+                "1.0.0",
+                deprecation_message,
+                standard_warn=False,
+            )
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif (
+            not self.is_input_continuous
+            and not self.is_input_vectorized
+            and not self.is_input_patches
+        ):
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups,
+            num_channels=in_channels,
+            eps=1e-6,
+            affine=True,
+        )
+        if use_linear_projection:
+            self.proj_in = linear_cls(in_channels, inner_dim)
+        else:
+            self.proj_in = conv_cls(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        # TODO: should use out_channels for continuous projections
+        if use_linear_projection:
+            self.proj_out = linear_cls(inner_dim, in_channels)
+        else:
+            self.proj_out = conv_cls(
+                inner_dim, in_channels, kernel_size=1, stride=1, padding=0
+            )
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(
+                inner_dim, use_additional_conditions=self.use_additional_conditions
+            )
+
+        self.caption_projection = None
+        if caption_channels is not None:
+            self.caption_projection = CaptionProjection(
+                in_features=caption_channels, hidden_size=inner_dim
+            )
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(hidden_states.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 1. Input
+        batch, _, height, width = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(
+                batch_size, -1, hidden_states.shape[-1]
+            )
+
+        ref_feature = hidden_states.reshape(batch, height, width, inner_dim)
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+
+            output = hidden_states + residual
+        if not return_dict:
+            return (output, ref_feature)
+
+        return Transformer2DModelOutput(sample=output, ref_feature=ref_feature)

modules/transformer_3d.py

@@ -0,0 +1,169 @@
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models import ModelMixin
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from einops import rearrange, repeat
+from torch import nn
+
+from .attention import TemporalBasicTransformerBlock
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
+        )
+        if use_linear_projection:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_in = nn.Conv2d(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+
+        # Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalBasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        if use_linear_projection:
+            self.proj_out = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_out = nn.Conv2d(
+                inner_dim, in_channels, kernel_size=1, stride=1, padding=0
+            )
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        timestep=None,
+        return_dict: bool = True,
+    ):
+        # Input
+        assert (
+            hidden_states.dim() == 5
+        ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+        if encoder_hidden_states.shape[0] != hidden_states.shape[0]:
+            encoder_hidden_states = repeat(
+                encoder_hidden_states, "b n c -> (b f) n c", f=video_length
+            )
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * weight, inner_dim
+            )
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * weight, inner_dim
+            )
+            hidden_states = self.proj_in(hidden_states)
+
+        # Blocks
+        for i, block in enumerate(self.transformer_blocks):
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                timestep=timestep,
+                video_length=video_length,
+            )
+
+        # Output
+        if not self.use_linear_projection:
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim)
+                .permute(0, 3, 1, 2)
+                .contiguous()
+            )
+            hidden_states = self.proj_out(hidden_states)
+        else:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim)
+                .permute(0, 3, 1, 2)
+                .contiguous()
+            )
+
+        output = hidden_states + residual
+
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+        if not return_dict:
+            return (output,)
+
+        return Transformer3DModelOutput(sample=output)

modules/unet_2d_blocks.py

@@ -0,0 +1,1072 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import Attention
+from diffusers.models.dual_transformer_2d import DualTransformer2DModel
+from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
+from diffusers.utils import is_torch_version, logging
+from diffusers.utils.torch_utils import apply_freeu
+from torch import nn
+
+from .transformer_2d import Transformer2DModel
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def get_down_block(
+    down_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    temb_channels: int,
+    add_downsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    downsample_padding: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    downsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+):
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    down_block_type = (
+        down_block_type[7:]
+        if down_block_type.startswith("UNetRes")
+        else down_block_type
+    )
+    if down_block_type == "DownBlock2D":
+        return DownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "CrossAttnDownBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnDownBlock2D"
+            )
+        return CrossAttnDownBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    prev_output_channel: int,
+    temb_channels: int,
+    add_upsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    resolution_idx: Optional[int] = None,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    upsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+) -> nn.Module:
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    up_block_type = (
+        up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    )
+    if up_block_type == "UpBlock2D":
+        return UpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif up_block_type == "CrossAttnUpBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnUpBlock2D"
+            )
+        return CrossAttnUpBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class AutoencoderTinyBlock(nn.Module):
+    """
+    Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
+    blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        out_channels (`int`): The number of output channels.
+        act_fn (`str`):
+            ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
+
+    Returns:
+        `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
+        `out_channels`.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, act_fn: str):
+        super().__init__()
+        act_fn = get_activation(act_fn)
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+        )
+        self.skip = (
+            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+        self.fuse = nn.ReLU()
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        return self.fuse(self.conv(x) + self.skip(x))
+
+
+class UNetMidBlock2D(nn.Module):
+    """
+    A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        temb_channels (`int`): The number of temporal embedding channels.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
+        num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
+        resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
+        resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
+            The type of normalization to apply to the time embeddings. This can help to improve the performance of the
+            model on tasks with long-range temporal dependencies.
+        resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
+        resnet_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use in the group normalization layers of the resnet blocks.
+        attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
+        resnet_pre_norm (`bool`, *optional*, defaults to `True`):
+            Whether to use pre-normalization for the resnet blocks.
+        add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
+        attention_head_dim (`int`, *optional*, defaults to 1):
+            Dimension of a single attention head. The number of attention heads is determined based on this value and
+            the number of input channels.
+        output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
+
+    Returns:
+        `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
+        in_channels, height, width)`.
+
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        attn_groups: Optional[int] = None,
+        resnet_pre_norm: bool = True,
+        add_attention: bool = True,
+        attention_head_dim: int = 1,
+        output_scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+        self.add_attention = add_attention
+
+        if attn_groups is None:
+            attn_groups = (
+                resnet_groups if resnet_time_scale_shift == "default" else None
+            )
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        if attention_head_dim is None:
+            logger.warn(
+                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+            )
+            attention_head_dim = in_channels
+
+        for _ in range(num_layers):
+            if self.add_attention:
+                attentions.append(
+                    Attention(
+                        in_channels,
+                        heads=in_channels // attention_head_dim,
+                        dim_head=attention_head_dim,
+                        rescale_output_factor=output_scale_factor,
+                        eps=resnet_eps,
+                        norm_num_groups=attn_groups,
+                        spatial_norm_dim=temb_channels
+                        if resnet_time_scale_shift == "spatial"
+                        else None,
+                        residual_connection=True,
+                        bias=True,
+                        upcast_softmax=True,
+                        _from_deprecated_attn_block=True,
+                    )
+                )
+            else:
+                attentions.append(None)
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None
+    ) -> torch.FloatTensor:
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                hidden_states = attn(hidden_states, temb=temb)
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        output_scale_factor: float = 1.0,
+        cross_attention_dim: int = 1280,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+
+        # support for variable transformer layers per block
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for i in range(num_layers):
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        downsample_padding: int = 1,
+        add_downsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        additional_residuals: Optional[torch.FloatTensor] = None,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        blocks = list(zip(self.resnets, self.attentions))
+
+        for i, (resnet, attn) in enumerate(blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+            # apply additional residuals to the output of the last pair of resnet and attention blocks
+            if i == len(blocks) - 1 and additional_residuals is not None:
+                hidden_states = hidden_states + additional_residuals
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_downsample: bool = True,
+        downsample_padding: int = 1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        scale: float = 1.0,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        upsample_size: Optional[int] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(
+                    hidden_states, upsample_size, scale=lora_scale
+                )
+
+        return hidden_states
+
+
+class UpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        upsample_size: Optional[int] = None,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+        return hidden_states

modules/unet_2d_condition.py

@@ -0,0 +1,1312 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+try:
+    from diffusers.models.embeddings import PositionNet
+except:
+    from diffusers.models.embeddings import GLIGENTextBoundingboxProjection as PositionNet
+
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    deprecate,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+
+from .unet_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+    ref_features: Tuple[torch.FloatTensor] = None
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+       reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+        *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+        *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+            self,
+            sample_size: Optional[int] = None,
+            in_channels: int = 4,
+            out_channels: int = 4,
+            center_input_sample: bool = False,
+            flip_sin_to_cos: bool = True,
+            freq_shift: int = 0,
+            down_block_types: Tuple[str] = (
+                    "CrossAttnDownBlock2D",
+                    "CrossAttnDownBlock2D",
+                    "CrossAttnDownBlock2D",
+                    "DownBlock2D",
+            ),
+            mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+            up_block_types: Tuple[str] = (
+                    "UpBlock2D",
+                    "CrossAttnUpBlock2D",
+                    "CrossAttnUpBlock2D",
+                    "CrossAttnUpBlock2D",
+            ),
+            only_cross_attention: Union[bool, Tuple[bool]] = False,
+            block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+            layers_per_block: Union[int, Tuple[int]] = 2,
+            downsample_padding: int = 1,
+            mid_block_scale_factor: float = 1,
+            dropout: float = 0.0,
+            act_fn: str = "silu",
+            norm_num_groups: Optional[int] = 32,
+            norm_eps: float = 1e-5,
+            cross_attention_dim: Union[int, Tuple[int]] = 1280,
+            transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+            reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+            encoder_hid_dim: Optional[int] = None,
+            encoder_hid_dim_type: Optional[str] = None,
+            attention_head_dim: Union[int, Tuple[int]] = 8,
+            num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+            dual_cross_attention: bool = False,
+            use_linear_projection: bool = False,
+            class_embed_type: Optional[str] = None,
+            addition_embed_type: Optional[str] = None,
+            addition_time_embed_dim: Optional[int] = None,
+            num_class_embeds: Optional[int] = None,
+            upcast_attention: bool = False,
+            resnet_time_scale_shift: str = "default",
+            resnet_skip_time_act: bool = False,
+            resnet_out_scale_factor: int = 1.0,
+            time_embedding_type: str = "positional",
+            time_embedding_dim: Optional[int] = None,
+            time_embedding_act_fn: Optional[str] = None,
+            timestep_post_act: Optional[str] = None,
+            time_cond_proj_dim: Optional[int] = None,
+            conv_in_kernel: int = 3,
+            conv_out_kernel: int = 3,
+            projection_class_embeddings_input_dim: Optional[int] = None,
+            attention_type: str = "default",
+            class_embeddings_concat: bool = False,
+            mid_block_only_cross_attention: Optional[bool] = None,
+            cross_attention_norm: Optional[str] = None,
+            addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(
+                only_cross_attention
+        ) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
+                down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(
+                down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(
+                down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(
+                down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if (
+                isinstance(transformer_layers_per_block, list)
+                and reverse_transformer_layers_per_block is None
+        ):
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError(
+                        "Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet."
+                    )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=conv_in_kernel,
+            padding=conv_in_padding,
+        )
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(
+                    f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}."
+                )
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2,
+                set_W_to_weight=False,
+                log=False,
+                flip_sin_to_cos=flip_sin_to_cos,
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(
+                block_out_channels[0], flip_sin_to_cos, freq_shift
+            )
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info(
+                "encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined."
+            )
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(
+                timestep_input_dim, time_embed_dim, act_fn=act_fn
+            )
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim,
+                time_embed_dim,
+                num_heads=addition_embed_type_num_heads,
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim,
+                image_embed_dim=cross_attention_dim,
+                time_embed_dim=time_embed_dim,
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(
+                addition_time_embed_dim, flip_sin_to_cos, freq_shift
+            )
+            self.add_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type is not None:
+            raise ValueError(
+                f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'."
+            )
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(
+                down_block_types
+            )
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i]
+                if attention_head_dim[i] is not None
+                else output_channel,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            raise NotImplementedError(f"Unsupport mid_block_type: {mid_block_type}")
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[
+                min(i + 1, len(block_out_channels) - 1)
+            ]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i]
+                if attention_head_dim[i] is not None
+                else output_channel,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+        self.conv_norm_out = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        self.conv_out = nn.Conv2d(
+            block_out_channels[0],
+            out_channels,
+            kernel_size=conv_out_kernel,
+            padding=conv_out_padding,
+        )
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(
+                    cross_attention_dim, list
+            ):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = PositionNet(
+                positive_len=positive_len,
+                out_dim=cross_attention_dim,
+                feature_type=feature_type,
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+                name: str,
+                module: torch.nn.Module,
+                processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(
+                    return_deprecated_lora=True
+                )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(
+            self,
+            processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
+            _remove_lora=False,
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(
+                        processor.pop(f"{name}.processor"), _remove_lora=_remove_lora
+                    )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(
+                proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
+                for proc in self.attn_processors.values()
+        ):
+            processor = AttnAddedKVProcessor()
+        elif all(
+                proc.__class__ in CROSS_ATTENTION_PROCESSORS
+                for proc in self.attn_processors.values()
+        ):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = (
+            num_sliceable_layers * [slice_size]
+            if not isinstance(slice_size, list)
+            else slice_size
+        )
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(
+                module: torch.nn.Module, slice_size: List[int]
+        ):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1, s2, b1, b2):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if (
+                        hasattr(upsample_block, k)
+                        or getattr(upsample_block, k, None) is not None
+                ):
+                    setattr(upsample_block, k, None)
+
+    def forward(
+            self,
+            sample: torch.FloatTensor,
+            timestep: Union[torch.Tensor, float, int],
+            encoder_hidden_states: torch.Tensor,
+            class_labels: Optional[torch.Tensor] = None,
+            timestep_cond: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+            added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+            down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+            mid_block_additional_residual: Optional[torch.Tensor] = None,
+            down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+            encoder_attention_mask: Optional[torch.Tensor] = None,
+            return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+                example from ControlNet side model(s)
+            mid_block_additional_residual (`torch.Tensor`, *optional*):
+                additional residual to be added to UNet mid block output, for example from ControlNet side model
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2 ** self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (
+                                             1 - encoder_attention_mask.to(sample.dtype)
+                                     ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError(
+                    "class_labels should be provided when num_class_embeds > 0"
+                )
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if (
+                    "image_embeds" not in added_cond_kwargs
+                    or "hint" not in added_cond_kwargs
+            ):
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if (
+                self.encoder_hid_proj is not None
+                and self.config.encoder_hid_dim_type == "text_proj"
+        ):
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif (
+                self.encoder_hid_proj is not None
+                and self.config.encoder_hid_dim_type == "text_image_proj"
+        ):
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(
+                encoder_hidden_states, image_embeds
+            )
+        elif (
+                self.encoder_hid_proj is not None
+                and self.config.encoder_hid_dim_type == "image_proj"
+        ):
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif (
+                self.encoder_hid_proj is not None
+                and self.config.encoder_hid_dim_type == "ip_image_proj"
+        ):
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds).to(
+                encoder_hidden_states.dtype
+            )
+            encoder_hidden_states = torch.cat(
+                [encoder_hidden_states, image_embeds], dim=1
+            )
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if (
+                cross_attention_kwargs is not None
+                and cross_attention_kwargs.get("gligen", None) is not None
+        ):
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {
+                "objs": self.position_net(**gligen_args)
+            }
+
+        # 3. down
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = (
+                mid_block_additional_residual is not None
+                and down_block_additional_residuals is not None
+        )
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if (
+                not is_adapter
+                and mid_block_additional_residual is None
+                and down_block_additional_residuals is not None
+        ):
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        tot_referece_features = ()
+        for downsample_block in self.down_blocks:
+            if (
+                    hasattr(downsample_block, "has_cross_attention")
+                    and downsample_block.has_cross_attention
+            ):
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals[
+                        "additional_residuals"
+                    ] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample, temb=emb, scale=lora_scale
+                )
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                    down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = (
+                        down_block_res_sample + down_block_additional_residual
+                )
+                new_down_block_res_samples = new_down_block_res_samples + (
+                    down_block_res_sample,
+                )
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if (
+                    hasattr(self.mid_block, "has_cross_attention")
+                    and self.mid_block.has_cross_attention
+            ):
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                    is_adapter
+                    and len(down_intrablock_additional_residuals) > 0
+                    and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets):]
+            down_block_res_samples = down_block_res_samples[
+                                     : -len(upsample_block.resnets)
+                                     ]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if (
+                    hasattr(upsample_block, "has_cross_attention")
+                    and upsample_block.has_cross_attention
+            ):
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    scale=lora_scale,
+                )
+
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

modules/unet_3d.py

@@ -0,0 +1,698 @@
+# Adapted from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/unet_blocks.py
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from os import PathLike
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.attention_processor import AttentionProcessor
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import SAFETENSORS_WEIGHTS_NAME, WEIGHTS_NAME, BaseOutput, logging
+from safetensors.torch import load_file
+
+from .resnet import InflatedConv3d, InflatedGroupNorm
+from .unet_3d_blocks import UNetMidBlock3DCrossAttn, get_down_block, get_up_block
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet3DConditionOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+class UNet3DConditionModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+            self,
+            sample_size: Optional[int] = None,
+            in_channels: int = 4,
+            out_channels: int = 4,
+            center_input_sample: bool = False,
+            flip_sin_to_cos: bool = True,
+            freq_shift: int = 0,
+            down_block_types: Tuple[str] = (
+                    "CrossAttnDownBlock3D",
+                    "CrossAttnDownBlock3D",
+                    "CrossAttnDownBlock3D",
+                    "DownBlock3D",
+            ),
+            mid_block_type: str = "UNetMidBlock3DCrossAttn",
+            up_block_types: Tuple[str] = (
+                    "UpBlock3D",
+                    "CrossAttnUpBlock3D",
+                    "CrossAttnUpBlock3D",
+                    "CrossAttnUpBlock3D",
+            ),
+            only_cross_attention: Union[bool, Tuple[bool]] = False,
+            block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+            layers_per_block: int = 2,
+            downsample_padding: int = 1,
+            mid_block_scale_factor: float = 1,
+            act_fn: str = "silu",
+            norm_num_groups: int = 32,
+            norm_eps: float = 1e-5,
+            cross_attention_dim: int = 1280,
+            attention_head_dim: Union[int, Tuple[int]] = 8,
+            dual_cross_attention: bool = False,
+            use_linear_projection: bool = False,
+            class_embed_type: Optional[str] = None,
+            num_class_embeds: Optional[int] = None,
+            upcast_attention: bool = False,
+            resnet_time_scale_shift: str = "default",
+            use_inflated_groupnorm=False,
+            # Additional
+            use_motion_module=False,
+            motion_module_resolutions=(1, 2, 4, 8),
+            motion_module_mid_block=False,
+            motion_module_decoder_only=False,
+            motion_module_type=None,
+            motion_module_kwargs={},
+            unet_use_cross_frame_attention=None,
+            unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # input
+        self.conv_in = InflatedConv3d(
+            in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1)
+        )
+
+        # time
+        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            res = 2 ** i
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module
+                                  and (res in motion_module_resolutions)
+                                  and (not motion_module_decoder_only),
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock3DCrossAttn":
+            self.mid_block = UNetMidBlock3DCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module and motion_module_mid_block,
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+            )
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the videos
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_attention_head_dim = list(reversed(attention_head_dim))
+        only_cross_attention = list(reversed(only_cross_attention))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            res = 2 ** (3 - i)
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[
+                min(i + 1, len(block_out_channels) - 1)
+            ]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=reversed_attention_head_dim[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module
+                                  and (res in motion_module_resolutions),
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if use_inflated_groupnorm:
+            self.conv_norm_out = InflatedGroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        else:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        self.conv_act = nn.SiLU()
+        self.conv_out = InflatedConv3d(
+            block_out_channels[0], out_channels, kernel_size=3, padding=1
+        )
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+                name: str,
+                module: torch.nn.Module,
+                processors: Dict[str, AttentionProcessor],
+        ):
+            # if hasattr(module, "set_processor"):
+            #     processors[f"{name}.processor"] = module.processor
+
+            if hasattr(module, "get_processor") or hasattr(module, "set_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_slicable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_slicable_dims(module)
+
+        num_slicable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_slicable_layers * [1]
+
+        slice_size = (
+            num_slicable_layers * [slice_size]
+            if not isinstance(slice_size, list)
+            else slice_size
+        )
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(
+                module: torch.nn.Module, slice_size: List[int]
+        ):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(
+            self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_attn_processor(name, module, processor)
+
+    def forward(
+            self,
+            sample: torch.FloatTensor,
+            timestep: Union[torch.Tensor, float, int],
+            encoder_hidden_states: torch.Tensor,
+            class_labels: Optional[torch.Tensor] = None,
+            kps_features: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+            mid_block_additional_residual: Optional[torch.Tensor] = None,
+            return_dict: bool = True,
+    ) -> Union[UNet3DConditionOutput, Tuple]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
+            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
+            encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2 ** self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=self.dtype)
+        emb = self.time_embedding(t_emb)
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError(
+                    "class_labels should be provided when num_class_embeds > 0"
+                )
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        # pre-process
+        sample = self.conv_in(sample)
+        if kps_features is not None:
+            sample = sample + kps_features
+
+        # down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if (
+                    hasattr(downsample_block, "has_cross_attention")
+                    and downsample_block.has_cross_attention
+            ):
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                )
+
+            down_block_res_samples += res_samples
+
+        if down_block_additional_residuals is not None:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                    down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = (
+                        down_block_res_sample + down_block_additional_residual
+                )
+                new_down_block_res_samples += (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # mid
+        sample = self.mid_block(
+            sample,
+            emb,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+        )
+
+        if mid_block_additional_residual is not None:
+            sample = sample + mid_block_additional_residual
+
+        # up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets):]
+            down_block_res_samples = down_block_res_samples[
+                                     : -len(upsample_block.resnets)
+                                     ]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if (
+                    hasattr(upsample_block, "has_cross_attention")
+                    and upsample_block.has_cross_attention
+            ):
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    encoder_hidden_states=encoder_hidden_states,
+                )
+
+        # post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet3DConditionOutput(sample=sample)
+
+    @classmethod
+    def from_pretrained_2d(
+            cls,
+            pretrained_model_path: PathLike,
+            motion_module_path: PathLike,
+            subfolder=None,
+            unet_additional_kwargs=None,
+            mm_zero_proj_out=False,
+    ):
+        pretrained_model_path = Path(pretrained_model_path)
+        motion_module_path = Path(motion_module_path)
+        if subfolder is not None:
+            pretrained_model_path = pretrained_model_path.joinpath(subfolder)
+        logger.info(
+            f"loaded temporal unet's pretrained weights from {pretrained_model_path} ..."
+        )
+
+        config_file = pretrained_model_path / "config.json"
+        if not (config_file.exists() and config_file.is_file()):
+            raise RuntimeError(f"{config_file} does not exist or is not a file")
+
+        unet_config = cls.load_config(config_file)
+        unet_config["_class_name"] = cls.__name__
+        unet_config["down_block_types"] = [
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ]
+        unet_config["up_block_types"] = [
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ]
+        unet_config["mid_block_type"] = "UNetMidBlock3DCrossAttn"
+
+        model = cls.from_config(unet_config, **unet_additional_kwargs)
+        # load the vanilla weights
+        if pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME).exists():
+            logger.debug(
+                f"loading safeTensors weights from {pretrained_model_path} ..."
+            )
+            state_dict = load_file(
+                pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME), device="cpu"
+            )
+
+        elif pretrained_model_path.joinpath(WEIGHTS_NAME).exists():
+            logger.debug(f"loading weights from {pretrained_model_path} ...")
+            state_dict = torch.load(
+                pretrained_model_path.joinpath(WEIGHTS_NAME),
+                map_location="cpu",
+                weights_only=True,
+            )
+        else:
+            raise FileNotFoundError(f"no weights file found in {pretrained_model_path}")
+
+        # load the motion module weights
+        if motion_module_path.exists() and motion_module_path.is_file():
+            if motion_module_path.suffix.lower() in [".pth", ".pt", ".ckpt", ".bin"]:
+                logger.info(f"Load motion module params from {motion_module_path}")
+                motion_state_dict = torch.load(
+                    motion_module_path, map_location="cpu", weights_only=True
+                )
+            elif motion_module_path.suffix.lower() == ".safetensors":
+                motion_state_dict = load_file(motion_module_path, device="cpu")
+            else:
+                raise RuntimeError(
+                    f"unknown file format for motion module weights: {motion_module_path.suffix}"
+                )
+            if mm_zero_proj_out:
+                logger.info(f"Zero initialize proj_out layers in motion module...")
+                new_motion_state_dict = OrderedDict()
+                for k in motion_state_dict:
+                    if "proj_out" in k:
+                        continue
+                    new_motion_state_dict[k] = motion_state_dict[k]
+                motion_state_dict = new_motion_state_dict
+
+            # merge the state dicts
+            state_dict.update(motion_state_dict)
+
+        # load the weights into the model
+        m, u = model.load_state_dict(state_dict, strict=False)
+        logger.debug(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
+
+        params = [
+            p.numel() if "temporal" in n else 0 for n, p in model.named_parameters()
+        ]
+        logger.info(f"Loaded {sum(params) / 1e6}M-parameter motion module")
+
+        return model
+
+    @classmethod
+    def from_config_2d(
+            cls,
+            unet_config_path: PathLike,
+            unet_additional_kwargs=None,
+    ):
+        config_file = unet_config_path
+
+        unet_config = cls.load_config(config_file)
+        unet_config["_class_name"] = cls.__name__
+        unet_config["down_block_types"] = [
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ]
+        unet_config["up_block_types"] = [
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ]
+        unet_config["mid_block_type"] = "UNetMidBlock3DCrossAttn"
+
+        model = cls.from_config(unet_config, **unet_additional_kwargs)
+        return model

modules/unet_3d_blocks.py

@@ -0,0 +1,862 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+
+import pdb
+
+import torch
+from torch import nn
+
+from .motion_module import get_motion_module
+
+# from .motion_module import get_motion_module
+from .resnet import Downsample3D, ResnetBlock3D, Upsample3D
+from .transformer_3d import Transformer3DModel
+
+
+def get_down_block(
+    down_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    temb_channels,
+    add_downsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    downsample_padding=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_type=None,
+    motion_module_kwargs=None,
+):
+    down_block_type = (
+        down_block_type[7:]
+        if down_block_type.startswith("UNetRes")
+        else down_block_type
+    )
+    if down_block_type == "DownBlock3D":
+        return DownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    elif down_block_type == "CrossAttnDownBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnDownBlock3D"
+            )
+        return CrossAttnDownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    prev_output_channel,
+    temb_channels,
+    add_upsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_type=None,
+    motion_module_kwargs=None,
+):
+    up_block_type = (
+        up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    )
+    if up_block_type == "UpBlock3D":
+        return UpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    elif up_block_type == "CrossAttnUpBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnUpBlock3D"
+            )
+        return CrossAttnUpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlock3DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        output_scale_factor=1.0,
+        cross_attention_dim=1280,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock3D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+            )
+        ]
+        attentions = []
+        motion_modules = []
+
+        for _ in range(num_layers):
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=in_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=in_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+    def forward(
+        self,
+        hidden_states,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+    ):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet, motion_module in zip(
+            self.attentions, self.resnets[1:], self.motion_modules
+        ):
+            hidden_states = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+            ).sample
+            hidden_states = (
+                motion_module(
+                    hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                )
+                if motion_module is not None
+                else hidden_states
+            )
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        downsample_padding=1,
+        add_downsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+    ):
+        output_states = ()
+
+        for i, (resnet, attn, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.motion_modules)
+        ):
+            # self.gradient_checkpointing = False
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                )[0]
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                ).sample
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_downsample=True,
+        downsample_padding=1,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        # use_motion_module = False
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
+        output_states = ()
+
+        for resnet, motion_module in zip(self.resnets, self.motion_modules):
+            # print(f"DownBlock3D {self.gradient_checkpointing = }")
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_motion_module=None,
+        use_inflated_groupnorm=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        res_hidden_states_tuple,
+        temb=None,
+        encoder_hidden_states=None,
+        upsample_size=None,
+        attention_mask=None,
+    ):
+        for i, (resnet, attn, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.motion_modules)
+        ):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                ).sample
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                ).sample
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states
+
+
+class UpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        # use_motion_module = False
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        res_hidden_states_tuple,
+        temb=None,
+        upsample_size=None,
+        encoder_hidden_states=None,
+    ):
+        for resnet, motion_module in zip(self.resnets, self.motion_modules):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            # print(f"UpBlock3D {self.gradient_checkpointing = }")
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states

modules/v_kps_guider.py

@@ -0,0 +1,45 @@
+from typing import Tuple
+
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.models.modeling_utils import ModelMixin
+from .motion_module import zero_module
+from .resnet import InflatedConv3d
+
+
+class VKpsGuider(ModelMixin):
+    def __init__(
+            self,
+            conditioning_embedding_channels: int,
+            conditioning_channels: int = 3,
+            block_out_channels: Tuple[int] = (16, 32, 64, 128),
+    ):
+        super().__init__()
+        self.conv_in = InflatedConv3d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
+
+        self.blocks = nn.ModuleList([])
+
+        for i in range(len(block_out_channels) - 1):
+            channel_in = block_out_channels[i]
+            channel_out = block_out_channels[i + 1]
+            self.blocks.append(InflatedConv3d(channel_in, channel_in, kernel_size=3, padding=1))
+            self.blocks.append(InflatedConv3d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
+
+        self.conv_out = zero_module(InflatedConv3d(
+            block_out_channels[-1],
+            conditioning_embedding_channels,
+            kernel_size=3,
+            padding=1,
+        ))
+
+    def forward(self, conditioning):
+        embedding = self.conv_in(conditioning)
+        embedding = F.silu(embedding)
+
+        for block in self.blocks:
+            embedding = block(embedding)
+            embedding = F.silu(embedding)
+
+        embedding = self.conv_out(embedding)
+
+        return embedding

pipelines/__init__.py

@@ -0,0 +1 @@
+from .v_express_pipeline import VExpressPipeline

pipelines/context.py

@@ -0,0 +1,78 @@
+# TODO: Adapted from cli
+from typing import Callable, List, Optional
+
+import numpy as np
+
+
+def ordered_halving(val):
+    bin_str = f"{val:064b}"
+    bin_flip = bin_str[::-1]
+    as_int = int(bin_flip, 2)
+
+    return as_int / (1 << 64)
+
+
+def uniform(
+    step: int = ...,
+    num_frames: int = ...,
+    context_size: Optional[int] = None,
+    context_stride: int = 3,
+    context_overlap: int = 4,
+    closed_loop: bool = True,
+):
+    if num_frames <= context_size:
+        yield list(range(num_frames))
+        return
+
+    context_stride = min(
+        context_stride, int(np.ceil(np.log2(num_frames / context_size))) + 1
+    )
+
+    for context_step in 1 << np.arange(context_stride):
+        pad = int(round(num_frames * ordered_halving(step)))
+        for j in range(
+            int(ordered_halving(step) * context_step) + pad,
+            num_frames + pad + (0 if closed_loop else -context_overlap),
+            (context_size * context_step - context_overlap),
+        ):
+            next_itr = []
+            for e in range(j, j + context_size * context_step, context_step):
+                if e >= num_frames:
+                    e = num_frames - 2 - e % num_frames
+                next_itr.append(e)
+
+            yield next_itr
+
+
+def get_context_scheduler(name: str) -> Callable:
+    if name == "uniform":
+        return uniform
+    else:
+        raise ValueError(f"Unknown context_overlap policy {name}")
+
+
+def get_total_steps(
+    scheduler,
+    timesteps: List[int],
+    num_steps: Optional[int] = None,
+    num_frames: int = ...,
+    context_size: Optional[int] = None,
+    context_stride: int = 3,
+    context_overlap: int = 4,
+    closed_loop: bool = True,
+):
+    return sum(
+        len(
+            list(
+                scheduler(
+                    i,
+                    num_steps,
+                    num_frames,
+                    context_size,
+                    context_stride,
+                    context_overlap,
+                )
+            )
+        )
+        for i in range(len(timesteps))
+    )

pipelines/utils.py

@@ -0,0 +1,218 @@
+import math
+import os
+import pathlib
+
+import cv2
+import numpy as np
+import torch
+import torch.nn.functional as func
+import tqdm
+from imageio_ffmpeg import get_ffmpeg_exe
+
+tensor_interpolation = None
+
+
+def get_tensor_interpolation_method():
+    return tensor_interpolation
+
+
+def set_tensor_interpolation_method(is_slerp):
+    global tensor_interpolation
+    tensor_interpolation = slerp if is_slerp else linear
+
+
+def linear(v1, v2, t):
+    return (1.0 - t) * v1 + t * v2
+
+
+def slerp(v0: torch.Tensor, v1: torch.Tensor, t: float, DOT_THRESHOLD: float = 0.9995) -> torch.Tensor:
+    u0 = v0 / v0.norm()
+    u1 = v1 / v1.norm()
+    dot = (u0 * u1).sum()
+    if dot.abs() > DOT_THRESHOLD:
+        # logger.info(f'warning: v0 and v1 close to parallel, using linear interpolation instead.')
+        return (1.0 - t) * v0 + t * v1
+    omega = dot.acos()
+    return (((1.0 - t) * omega).sin() * v0 + (t * omega).sin() * v1) / omega.sin()
+
+
+def draw_kps_image(height, width, kps, color_list=[(255, 0, 0), (0, 255, 0), (0, 0, 255)]):
+    stick_width = 4
+    limb_seq = np.array([[0, 2], [1, 2]])
+    kps = np.array(kps)
+
+    canvas = np.zeros((height, width, 3), dtype=np.uint8)
+
+    for i in range(len(limb_seq)):
+        index = limb_seq[i]
+        color = color_list[index[0]]
+
+        x = kps[index][:, 0]
+        y = kps[index][:, 1]
+        length = ((x[0] - x[1]) ** 2 + (y[0] - y[1]) ** 2) ** 0.5
+        angle = int(math.degrees(math.atan2(y[0] - y[1], x[0] - x[1])))
+        polygon = cv2.ellipse2Poly((int(np.mean(x)), int(np.mean(y))), (int(length / 2), stick_width), angle, 0, 360, 1)
+        cv2.fillConvexPoly(canvas, polygon, [int(float(c) * 0.6) for c in color])
+
+    for idx_kp, kp in enumerate(kps):
+        color = color_list[idx_kp]
+        x, y = kp
+        cv2.circle(canvas, (int(x), int(y)), 4, color, -1)
+
+    return canvas
+
+
+
+import os
+import pathlib
+import shutil
+import cv2
+import numpy as np
+from scipy.ndimage.filters import median_filter
+
+def get_ffmpeg_exe():
+    if os.name == 'nt':  # Windows
+        return 'ffmpeg'
+    else:  # Ubuntu and other Unix-based systems
+        return 'ffmpeg'
+
+
+def median_filter_3d(video_tensor, kernel_size, device):
+    _, video_length, height, width = video_tensor.shape
+
+    pad_size = kernel_size // 2
+    video_tensor = func.pad(video_tensor, (pad_size, pad_size, pad_size, pad_size, pad_size, pad_size), mode='reflect')
+
+    filtered_video_tensor = []
+    for i in tqdm.tqdm(range(video_length), desc='Median Filtering'):
+        video_segment = video_tensor[:, i:i + kernel_size, ...].to(device)
+        video_segment = video_segment.unfold(dimension=2, size=kernel_size, step=1)
+        video_segment = video_segment.unfold(dimension=3, size=kernel_size, step=1)
+        video_segment = video_segment.permute(0, 2, 3, 1, 4, 5).reshape(3, height, width, -1)
+        filtered_video_frame = torch.median(video_segment, dim=-1)[0]
+        filtered_video_tensor.append(filtered_video_frame.cpu())
+    filtered_video_tensor = torch.stack(filtered_video_tensor, dim=1)
+    return filtered_video_tensor
+
+
+def save_video(video_tensor, audio_path, output_path, device, fps=30.0):
+    pathlib.Path(output_path).parent.mkdir(exist_ok=True, parents=True)
+
+    video_tensor = video_tensor[0, ...]
+    _, num_frames, height, width = video_tensor.shape
+
+    video_tensor = median_filter_3d(video_tensor, kernel_size=3, device=device)
+    video_tensor = video_tensor.permute(1, 2, 3, 0)
+    video_frames = (video_tensor * 255).numpy().astype(np.uint8)
+
+    output_name = pathlib.Path(output_path).stem
+    temp_output_path = output_path.replace(output_name, output_name + '-temp')
+    video_writer = cv2.VideoWriter(temp_output_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (width, height))
+
+    for i in tqdm.tqdm(range(num_frames), 'Writing frames into file'):
+        frame_image = video_frames[i, ...]
+        frame_image = cv2.cvtColor(frame_image, cv2.COLOR_RGB2BGR)
+        video_writer.write(frame_image)
+    video_writer.release()
+
+    cmd = (f'{get_ffmpeg_exe()} -i "{temp_output_path}" -i "{audio_path}" '
+           f'-map 0:v -map 1:a -c:v h264 -shortest -y "{output_path}" -loglevel quiet')
+    os.system(cmd)
+
+    os.remove(temp_output_path)
+
+
+
+def compute_dist(x1, y1, x2, y2):
+    return math.sqrt((x1 - x2) ** 2 + (y1 - y2) ** 2)
+
+
+def compute_ratio(kps):
+    l_eye_x, l_eye_y = kps[0][0], kps[0][1]
+    r_eye_x, r_eye_y = kps[1][0], kps[1][1]
+    nose_x, nose_y = kps[2][0], kps[2][1]
+    d_left = compute_dist(l_eye_x, l_eye_y, nose_x, nose_y)
+    d_right = compute_dist(r_eye_x, r_eye_y, nose_x, nose_y)
+    ratio = d_left / (d_right + 1e-6)
+    return ratio
+
+
+def point_to_line_dist(point, line_points):
+    point = np.array(point)
+    line_points = np.array(line_points)
+    line_vec = line_points[1] - line_points[0]
+    point_vec = point - line_points[0]
+    line_norm = line_vec / np.sqrt(np.sum(line_vec ** 2))
+    point_vec_scaled = point_vec * 1.0 / np.sqrt(np.sum(line_vec ** 2))
+    t = np.dot(line_norm, point_vec_scaled)
+    if t < 0.0:
+        t = 0.0
+    elif t > 1.0:
+        t = 1.0
+    nearest = line_points[0] + t * line_vec
+    dist = np.sqrt(np.sum((point - nearest) ** 2))
+    return dist
+
+
+def get_face_size(kps):
+    # 0: left eye, 1: right eye, 2: nose
+    A = kps[0, :]
+    B = kps[1, :]
+    C = kps[2, :]
+
+    AB_dist = math.sqrt((A[0] - B[0]) ** 2 + (A[1] - B[1]) ** 2)
+    C_AB_dist = point_to_line_dist(C, [A, B])
+    return AB_dist, C_AB_dist
+
+
+def get_rescale_params(kps_ref, kps_target):
+    kps_ref = np.array(kps_ref)
+    kps_target = np.array(kps_target)
+
+    ref_AB_dist, ref_C_AB_dist = get_face_size(kps_ref)
+    target_AB_dist, target_C_AB_dist = get_face_size(kps_target)
+
+    scale_width = ref_AB_dist / target_AB_dist
+    scale_height = ref_C_AB_dist / target_C_AB_dist
+
+    return scale_width, scale_height
+
+
+def retarget_kps(ref_kps, tgt_kps_list, only_offset=True):
+    ref_kps = np.array(ref_kps)
+    tgt_kps_list = np.array(tgt_kps_list)
+
+    ref_ratio = compute_ratio(ref_kps)
+
+    ratio_delta = 10000
+    selected_tgt_kps_idx = None
+    for idx, tgt_kps in enumerate(tgt_kps_list):
+        tgt_ratio = compute_ratio(tgt_kps)
+        if math.fabs(tgt_ratio - ref_ratio) < ratio_delta:
+            selected_tgt_kps_idx = idx
+            ratio_delta = tgt_ratio
+
+    scale_width, scale_height = get_rescale_params(
+        kps_ref=ref_kps,
+        kps_target=tgt_kps_list[selected_tgt_kps_idx],
+    )
+
+    rescaled_tgt_kps_list = np.array(tgt_kps_list)
+    rescaled_tgt_kps_list[:, :, 0] *= scale_width
+    rescaled_tgt_kps_list[:, :, 1] *= scale_height
+
+    if only_offset:
+        nose_offset = rescaled_tgt_kps_list[:, 2, :] - rescaled_tgt_kps_list[0, 2, :]
+        nose_offset = nose_offset[:, np.newaxis, :]
+        ref_kps_repeat = np.tile(ref_kps, (tgt_kps_list.shape[0], 1, 1))
+
+        ref_kps_repeat[:, :, :] -= (nose_offset / 2.0)
+        rescaled_tgt_kps_list = ref_kps_repeat
+    else:
+        nose_offset_x = rescaled_tgt_kps_list[0, 2, 0] - ref_kps[2][0]
+        nose_offset_y = rescaled_tgt_kps_list[0, 2, 1] - ref_kps[2][1]
+
+        rescaled_tgt_kps_list[:, :, 0] -= nose_offset_x
+        rescaled_tgt_kps_list[:, :, 1] -= nose_offset_y
+
+    return rescaled_tgt_kps_list

pipelines/v_express_pipeline.py

@@ -0,0 +1,586 @@
+# Adapted from https://github.com/magic-research/magic-animate/blob/main/magicanimate/pipelines/pipeline_animation.py
+import inspect
+import math
+from typing import Callable, List, Optional, Union
+
+import torch
+from diffusers import DiffusionPipeline
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import (
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerAncestralDiscreteScheduler,
+    EulerDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+)
+from diffusers.utils import is_accelerate_available
+from diffusers.utils.torch_utils import randn_tensor
+from einops import rearrange
+from tqdm import tqdm
+from transformers import CLIPImageProcessor
+
+from modules import ReferenceAttentionControl
+from .context import get_context_scheduler
+
+
+def retrieve_timesteps(
+        scheduler,
+        num_inference_steps: Optional[int] = None,
+        device: Optional[Union[str, torch.device]] = None,
+        timesteps: Optional[List[int]] = None,
+        **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used,
+            `timesteps` must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+                timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+                must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class VExpressPipeline(DiffusionPipeline):
+    _optional_components = []
+
+    def __init__(
+            self,
+            vae,
+            reference_net,
+            denoising_unet,
+            v_kps_guider,
+            audio_processor,
+            audio_encoder,
+            audio_projection,
+            scheduler: Union[
+                DDIMScheduler,
+                PNDMScheduler,
+                LMSDiscreteScheduler,
+                EulerDiscreteScheduler,
+                EulerAncestralDiscreteScheduler,
+                DPMSolverMultistepScheduler,
+            ],
+            image_proj_model=None,
+            tokenizer=None,
+            text_encoder=None,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            reference_net=reference_net,
+            denoising_unet=denoising_unet,
+            v_kps_guider=v_kps_guider,
+            audio_processor=audio_processor,
+            audio_encoder=audio_encoder,
+            audio_projection=audio_projection,
+            scheduler=scheduler,
+            image_proj_model=image_proj_model,
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.clip_image_processor = CLIPImageProcessor()
+        self.reference_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True
+        )
+        self.condition_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor,
+            do_convert_rgb=True,
+            do_normalize=False,
+        )
+
+    def enable_vae_slicing(self):
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    def _execution_device(self):
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                    hasattr(module, "_hf_hook")
+                    and hasattr(module._hf_hook, "execution_device")
+                    and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    @torch.no_grad()
+    def decode_latents(self, latents):
+        video_length = latents.shape[2]
+        latents = 1 / 0.18215 * latents
+        latents = rearrange(latents, "b c f h w -> (b f) c h w")
+        video = []
+        for frame_idx in tqdm(range(latents.shape[0]), desc='Decoding latents into frames'):
+            image = self.vae.decode(latents[frame_idx: frame_idx + 1].to(self.vae.device)).sample
+            image = (image / 2 + 0.5).clamp(0, 1)
+            image = image.cpu().float()
+            video.append(image)
+        video = torch.cat(video)
+        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
+
+        return video
+
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def prepare_latents(
+            self,
+            batch_size,
+            num_channels_latents,
+            width,
+            height,
+            video_length,
+            dtype,
+            device,
+            generator,
+            latents=None
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            video_length,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(
+                shape, generator=generator, device=device, dtype=dtype
+            )
+
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def _encode_prompt(
+            self,
+            prompt,
+            device,
+            num_videos_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+    ):
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(
+            prompt, padding="longest", return_tensors="pt"
+        ).input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                text_input_ids, untruncated_ids
+        ):
+            removed_text = self.tokenizer.batch_decode(
+                untruncated_ids[:, self.tokenizer.model_max_length - 1: -1]
+            )
+
+        if (
+                hasattr(self.text_encoder.config, "use_attention_mask")
+                and self.text_encoder.config.use_attention_mask
+        ):
+            attention_mask = text_inputs.attention_mask.to(device)
+        else:
+            attention_mask = None
+
+        text_embeddings = self.text_encoder(
+            text_input_ids.to(device),
+            attention_mask=attention_mask,
+        )
+        text_embeddings = text_embeddings[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_videos_per_prompt, 1)
+        text_embeddings = text_embeddings.view(
+            bs_embed * num_videos_per_prompt, seq_len, -1
+        )
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if (
+                    hasattr(self.text_encoder.config, "use_attention_mask")
+                    and self.text_encoder.config.use_attention_mask
+            ):
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            uncond_embeddings = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            uncond_embeddings = uncond_embeddings[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_videos_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(
+                batch_size * num_videos_per_prompt, seq_len, -1
+            )
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        return text_embeddings
+
+    def get_timesteps(self, num_inference_steps, strength, device):
+        # get the original timestep using init_timestep
+        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+        t_start = max(num_inference_steps - init_timestep, 0)
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order:]
+
+        return timesteps, num_inference_steps - t_start
+
+    def prepare_reference_latent(self, reference_image, height, width):
+        reference_image_tensor = self.reference_image_processor.preprocess(reference_image, height=height, width=width)
+        reference_image_tensor = reference_image_tensor.to(dtype=self.dtype, device=self.device)
+        reference_image_latents = self.vae.encode(reference_image_tensor).latent_dist.mean
+        reference_image_latents = reference_image_latents * 0.18215
+        return reference_image_latents
+
+    def prepare_kps_feature(self, kps_images, height, width, do_classifier_free_guidance):
+        kps_image_tensors = []
+        for idx, kps_image in enumerate(kps_images):
+            kps_image_tensor = self.condition_image_processor.preprocess(kps_image, height=height, width=width)
+            kps_image_tensor = kps_image_tensor.unsqueeze(2)  # [bs, c, 1, h, w]
+            kps_image_tensors.append(kps_image_tensor)
+        kps_images_tensor = torch.cat(kps_image_tensors, dim=2)  # [bs, c, t, h, w]
+
+        bs = 16
+        num_forward = math.ceil(kps_images_tensor.shape[2] / bs)
+        kps_feature = []
+        for i in range(num_forward):
+            tensor = kps_images_tensor[:, :, i * bs:(i + 1) * bs, ...].to(device=self.device, dtype=self.dtype)
+            feature = self.v_kps_guider(tensor).cpu()
+            kps_feature.append(feature)
+            torch.cuda.empty_cache()
+        kps_feature = torch.cat(kps_feature, dim=2)
+
+        if do_classifier_free_guidance:
+            uc_kps_feature = torch.zeros_like(kps_feature)
+            kps_feature = torch.cat([uc_kps_feature, kps_feature], dim=0)
+
+        return kps_feature
+
+    def prepare_audio_embeddings(self, audio_waveform, video_length, num_pad_audio_frames, do_classifier_free_guidance):
+        audio_waveform = self.audio_processor(audio_waveform, return_tensors="pt", sampling_rate=16000)['input_values']
+        audio_waveform = audio_waveform.to(self.device, self.dtype)
+        audio_embeddings = self.audio_encoder(audio_waveform).last_hidden_state  # [1, num_embeds, d]
+
+        audio_embeddings = torch.nn.functional.interpolate(
+            audio_embeddings.permute(0, 2, 1),
+            size=2 * video_length,
+            mode='linear',
+        )[0, :, :].permute(1, 0)  # [2*vid_len, dim]
+
+        audio_embeddings = torch.cat([
+            torch.zeros_like(audio_embeddings)[:2 * num_pad_audio_frames, :],
+            audio_embeddings,
+            torch.zeros_like(audio_embeddings)[:2 * num_pad_audio_frames, :],
+        ], dim=0)  # [2*num_pad+2*vid_len+2*num_pad, dim]
+
+        frame_audio_embeddings = []
+        for frame_idx in range(video_length):
+            start_sample = frame_idx
+            end_sample = frame_idx + 2 * num_pad_audio_frames
+
+            frame_audio_embedding = audio_embeddings[2 * start_sample:2 * (end_sample + 1), :]  # [2*num_pad+1, dim]
+            frame_audio_embeddings.append(frame_audio_embedding)
+        audio_embeddings = torch.stack(frame_audio_embeddings, dim=0)  # [vid_len, 2*num_pad+1, dim]
+
+        audio_embeddings = self.audio_projection(audio_embeddings).unsqueeze(0)
+        if do_classifier_free_guidance:
+            uc_audio_embeddings = torch.zeros_like(audio_embeddings)
+            audio_embeddings = torch.cat([uc_audio_embeddings, audio_embeddings], dim=0)
+        return audio_embeddings
+
+    @torch.no_grad()
+    def __call__(
+            self,
+            reference_image,
+            kps_images,
+            audio_waveform,
+            width,
+            height,
+            video_length,
+            num_inference_steps,
+            guidance_scale,
+            strength=1.,
+            num_images_per_prompt=1,
+            eta: float = 0.0,
+            generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+            output_type: Optional[str] = "tensor",
+            return_dict: bool = True,
+            callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+            callback_steps: Optional[int] = 1,
+            context_schedule="uniform",
+            context_frames=24,
+            context_overlap=4,
+            reference_attention_weight=1.,
+            audio_attention_weight=1.,
+            num_pad_audio_frames=2,
+            do_multi_devices_inference=False,
+            save_gpu_memory=False,
+            **kwargs,
+    ):
+        # Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        device = self._execution_device
+
+        do_classifier_free_guidance = guidance_scale > 1.0
+        batch_size = 1
+
+        # Prepare timesteps
+        timesteps = None
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+
+        reference_control_writer = ReferenceAttentionControl(
+            self.reference_net,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            mode="write",
+            batch_size=batch_size,
+            fusion_blocks="full",
+        )
+        reference_control_reader = ReferenceAttentionControl(
+            self.denoising_unet,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            mode="read",
+            batch_size=batch_size,
+            fusion_blocks="full",
+            reference_attention_weight=reference_attention_weight,
+            audio_attention_weight=audio_attention_weight,
+        )
+
+        num_channels_latents = self.denoising_unet.in_channels
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        reference_image_latents = self.prepare_reference_latent(reference_image, height, width)
+        kps_feature = self.prepare_kps_feature(kps_images, height, width, do_classifier_free_guidance)
+        if save_gpu_memory:
+            del self.v_kps_guider
+        torch.cuda.empty_cache()
+        audio_embeddings = self.prepare_audio_embeddings(
+            audio_waveform,
+            video_length,
+            num_pad_audio_frames,
+            do_classifier_free_guidance,
+        )
+        if save_gpu_memory:
+            del self.audio_processor, self.audio_encoder, self.audio_projection
+        torch.cuda.empty_cache()
+
+        context_scheduler = get_context_scheduler(context_schedule)
+        context_queue = list(
+            context_scheduler(
+                step=0,
+                num_frames=video_length,
+                context_size=context_frames,
+                context_stride=1,
+                context_overlap=context_overlap,
+                closed_loop=False,
+            )
+        )
+
+        num_frame_context = torch.zeros(video_length, device=device, dtype=torch.long)
+        for context in context_queue:
+            num_frame_context[context] += 1
+
+        encoder_hidden_states = torch.zeros((1, 1, 768), dtype=self.dtype, device=self.device)
+        self.reference_net(
+            reference_image_latents,
+            timestep=0,
+            encoder_hidden_states=encoder_hidden_states,
+            return_dict=False,
+        )
+        reference_control_reader.update(reference_control_writer, do_classifier_free_guidance)
+        if save_gpu_memory:
+            del self.reference_net
+        torch.cuda.empty_cache()
+
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            width,
+            height,
+            video_length,
+            self.dtype,
+            torch.device('cpu'),
+            generator,
+        )
+
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                context_counter = torch.zeros(video_length, device=device, dtype=torch.long)
+                noise_preds = [None] * video_length
+                for context_idx, context in enumerate(context_queue):
+                    latent_kps_feature = kps_feature[:, :, context].to(device, self.dtype)
+
+                    latent_audio_embeddings = audio_embeddings[:, context, ...]
+                    _, _, num_tokens, dim = latent_audio_embeddings.shape
+                    latent_audio_embeddings = latent_audio_embeddings.reshape(-1, num_tokens, dim)
+
+                    input_latents = latents[:, :, context, ...].to(device)
+                    input_latents = input_latents.repeat(2 if do_classifier_free_guidance else 1, 1, 1, 1, 1)
+                    input_latents = self.scheduler.scale_model_input(input_latents, t)
+                    noise_pred = self.denoising_unet(
+                        input_latents,
+                        t,
+                        encoder_hidden_states=latent_audio_embeddings.reshape(-1, num_tokens, dim),
+                        kps_features=latent_kps_feature,
+                        return_dict=False,
+                    )[0]
+                    if do_classifier_free_guidance:
+                        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                    context_counter[context] += 1
+                    noise_pred /= num_frame_context[context][None, None, :, None, None]
+                    step_frame_ids = []
+                    step_noise_preds = []
+                    for latent_idx, frame_idx in enumerate(context):
+                        if noise_preds[frame_idx] is None:
+                            noise_preds[frame_idx] = noise_pred[:, :, latent_idx, ...]
+                        else:
+                            noise_preds[frame_idx] += noise_pred[:, :, latent_idx, ...]
+                        if context_counter[frame_idx] == num_frame_context[frame_idx]:
+                            step_frame_ids.append(frame_idx)
+                            step_noise_preds.append(noise_preds[frame_idx])
+                            noise_preds[frame_idx] = None
+                    step_noise_preds = torch.stack(step_noise_preds, dim=2)
+                    output_latents = self.scheduler.step(
+                        step_noise_preds,
+                        t,
+                        latents[:, :, step_frame_ids, ...].to(device),
+                        **extra_step_kwargs,
+                    ).prev_sample
+                    latents[:, :, step_frame_ids, ...] = output_latents.cpu()
+
+                    progress_bar.set_description(
+                        f'Denoising Step Index: {i + 1} / {len(timesteps)}, '
+                        f'Context Index: {context_idx + 1} / {len(context_queue)}'
+                    )
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        reference_control_reader.clear()
+        reference_control_writer.clear()
+
+        video_tensor = self.decode_latents(latents)
+        return video_tensor

requirements.txt

@@ -0,0 +1,21 @@
+diffusers==0.24.0
+imageio-ffmpeg==0.4.9
+insightface==0.7.3
+omegaconf==2.2.3
+onnx==1.14.0
+onnxruntime-gpu==1.16.3
+safetensors==0.4.2
+torch==2.0.1
+torchaudio==2.0.2
+torchvision==0.15.2
+transformers==4.41.1
+einops==0.4.1
+tqdm==4.66.4
+xformers==0.0.22
+av==11.0.0
+gradio
+retina-face
+tf-keras
+filetype
+bitsandbytes==0.43.0
+accelerate

scripts/crop.py

@@ -0,0 +1,108 @@
+from retinaface import RetinaFace
+from PIL import Image
+import torch
+
+ 
+
+def auto_crop_image(image_path=r"F:\V_Express_V1\Material\Biden_Photo_Big.png", expand_percent=0.15, crop_size=(512, 512)):
+    # Check if CUDA is available
+    if torch.cuda.is_available():
+        device = 'cuda'
+        print("Using GPU for RetinaFace detection.")
+    else:
+        device = 'cpu'
+        print("Using CPU for RetinaFace detection.")
+
+    # Load image
+    img = Image.open(image_path)
+
+    # Perform face detection
+    faces = RetinaFace.detect_faces(image_path)
+
+    if not faces:
+        print("No faces detected.")
+        return None
+
+    # Assuming 'faces' is a dictionary of detected faces
+    # Pick the first face detected
+    face = list(faces.values())[0]
+    landmarks = face['landmarks']
+
+    # Extract the landmarks
+    right_eye = landmarks['right_eye']
+    left_eye = landmarks['left_eye']
+    right_mouth = landmarks['mouth_right']
+    left_mouth = landmarks['mouth_left']
+
+    # Calculate the distance between the eyes
+    eye_distance = abs(right_eye[0] - left_eye[0])
+
+    # Estimate the head width and height
+    head_width = eye_distance * 4.5  # Increase the width multiplier
+    head_height = eye_distance * 6.5  # Increase the height multiplier
+
+    # Calculate the center point between the eyes
+    eye_center_x = (right_eye[0] + left_eye[0]) // 2
+    eye_center_y = (right_eye[1] + left_eye[1]) // 2
+
+    # Calculate the top-left and bottom-right coordinates of the assumed head region
+    head_left = max(0, int(eye_center_x - head_width // 2))
+    head_top = max(0, int(eye_center_y - head_height // 2))  # Adjust the top coordinate
+    head_right = min(img.width, int(eye_center_x + head_width // 2))
+    head_bottom = min(img.height, int(eye_center_y + head_height // 2))  # Adjust the bottom coordinate
+
+    # Save the assumed head image
+    assumed_head_img = img.crop((head_left, head_top, head_right, head_bottom))
+    assumed_head_img.save("assumed_head.png", format='PNG')
+
+    # Calculate the expansion in pixels and the new dimensions
+    expanded_w = int(head_width * (1 + expand_percent))
+    expanded_h = int(head_height * (1 + expand_percent))
+
+    # Calculate the top-left and bottom-right points of the expanded box
+    center_x, center_y = head_left + head_width // 2, head_top + head_height // 2
+    left = max(0, center_x - expanded_w // 2)
+    right = min(img.width, center_x + expanded_w // 2)
+    top = max(0, center_y - expanded_h // 2)
+    bottom = min(img.height, center_y + expanded_h // 2)
+
+    # Crop the image with the expanded boundaries
+    cropped_img = img.crop((left, top, right, bottom))
+    cropped_img.save("expanded_face.png", format='PNG')
+
+    # Calculate the aspect ratio of the cropped image
+    cropped_width, cropped_height = cropped_img.size
+    aspect_ratio = cropped_width / cropped_height
+
+    # Calculate the target dimensions based on the desired crop size
+    target_width = crop_size[0]
+    target_height = crop_size[1]
+
+    # Adjust the crop to match the desired aspect ratio
+    if aspect_ratio > target_width / target_height:
+        # Crop from left and right
+        new_width = int(cropped_height * target_width / target_height)
+        left_crop = (cropped_width - new_width) // 2
+        right_crop = left_crop + new_width
+        top_crop = 0
+        bottom_crop = cropped_height
+    else:
+        # Crop from top and bottom
+        new_height = int(cropped_width * target_height / target_width)
+        top_crop = (cropped_height - new_height) // 2
+        bottom_crop = top_crop + new_height
+        left_crop = 0
+        right_crop = cropped_width
+
+    # Crop the image with the adjusted boundaries
+    final_cropped_img = cropped_img.crop((left_crop, top_crop, right_crop, bottom_crop))
+    final_cropped_img.save("final_cropped_img.png", format='PNG')
+
+    # Resize the cropped image to the desired size (512x512 by default) with best quality
+    resized_img = final_cropped_img.resize(crop_size, resample=Image.LANCZOS)
+
+    # Save the resized image as PNG
+    resized_img_path = image_path.rsplit('.', 1)[0] + '_cropped.png'  # Change file name to avoid overwriting
+    resized_img.save("resized_img.png", format='PNG')
+
+auto_crop_image()