add hg code

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 YaboZhang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -1,12 +1,162 @@
----
-title: ControlVideo
-emoji: 😻
-colorFrom: yellow
-colorTo: purple
-sdk: gradio
-sdk_version: 3.36.1
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# ControlVideo
+
+Official pytorch implementation of "ControlVideo: Training-free Controllable Text-to-Video Generation"
+
+[![arXiv](https://img.shields.io/badge/arXiv-2305.13077-b31b1b.svg)](https://arxiv.org/abs/2305.13077)
+![visitors](https://visitor-badge.laobi.icu/badge?page_id=YBYBZhang/ControlVideo)
+[![Replicate](https://replicate.com/cjwbw/controlvideo/badge)](https://replicate.com/cjwbw/controlvideo) 
+
+<p align="center">
+<img src="assets/overview.png" width="1080px"/> 
+<br>
+<em>ControlVideo adapts ControlNet to the video counterpart without any finetuning, aiming to directly inherit its high-quality and consistent generation </em>
+</p>
+
+## News
+* [07/11/2023] Support [ControlNet 1.1](https://github.com/lllyasviel/ControlNet-v1-1-nightly) based version! 
+* [05/28/2023] Thanks [chenxwh](https://github.com/chenxwh), add a [Replicate demo](https://replicate.com/cjwbw/controlvideo)!
+* [05/25/2023] Code [ControlVideo](https://github.com/YBYBZhang/ControlVideo/) released!
+* [05/23/2023] Paper [ControlVideo](https://arxiv.org/abs/2305.13077) released!
+
+## Setup
+
+### 1. Download Weights
+All pre-trained weights are downloaded to `checkpoints/` directory, including the pre-trained weights of [Stable Diffusion v1.5](https://huggingface.co/runwayml/stable-diffusion-v1-5), ControlNet 1.0 conditioned on [canny edges](https://huggingface.co/lllyasviel/sd-controlnet-canny), [depth maps](https://huggingface.co/lllyasviel/sd-controlnet-depth), [human poses](https://huggingface.co/lllyasviel/sd-controlnet-openpose), and ControlNet 1.1 in [here](https://huggingface.co/lllyasviel). 
+The `flownet.pkl` is the weights of [RIFE](https://github.com/megvii-research/ECCV2022-RIFE).
+The final file tree likes:
+
+```none
+checkpoints
+├── stable-diffusion-v1-5
+├── sd-controlnet-canny
+├── sd-controlnet-depth
+├── sd-controlnet-openpose
+├── ...
+├── flownet.pkl
+```
+### 2. Requirements
+
+```shell
+conda create -n controlvideo python=3.10
+conda activate controlvideo
+pip install -r requirements.txt
+```
+Note: `xformers` is recommended to save memory and running time. `controlnet-aux` is updated to version 0.0.6.
+
+## Inference
+
+To perform text-to-video generation, just run this command in `inference.sh`:
+```bash
+python inference.py \
+    --prompt "A striking mallard floats effortlessly on the sparkling pond." \
+    --condition "depth" \
+    --video_path "data/mallard-water.mp4" \
+    --output_path "outputs/" \
+    --video_length 15 \
+    --smoother_steps 19 20 \
+    --width 512 \
+    --height 512 \
+    --frame_rate 2 \
+    --version v10 \
+    # --is_long_video
+```
+where `--video_length` is the length of synthesized video, `--condition` represents the type of structure sequence,
+`--smoother_steps` determines at which timesteps to perform smoothing, `--version` selects the version of ControlNet (e.g., `v10` or `v11`), and `--is_long_video` denotes whether to enable efficient long-video synthesis.
+
+## Visualizations
+
+### ControlVideo on depth maps
+
+<table class="center">
+<tr>
+  <td width=30% align="center"><img src="assets/depth/A_charming_flamingo_gracefully_wanders_in_the_calm_and_serene_water,_its_delicate_neck_curving_into_an_elegant_shape..gif" raw=true></td>
+	<td width=30% align="center"><img src="assets/depth/A_striking_mallard_floats_effortlessly_on_the_sparkling_pond..gif" raw=true></td>
+  <td width=30% align="center"><img src="assets/depth/A_gigantic_yellow_jeep_slowly_turns_on_a_wide,_smooth_road_in_the_city..gif" raw=true></td>
+</tr>
+<tr>
+  <td width=30% align="center">"A charming flamingo gracefully wanders in the calm and serene water, its delicate neck curving into an elegant shape."</td>
+  <td width=30% align="center">"A striking mallard floats effortlessly on the sparkling pond."</td>
+  <td width=30% align="center">"A gigantic yellow jeep slowly turns on a wide, smooth road in the city."</td>
+</tr>
+ <tr>
+	<td width=30% align="center"><img src="assets/depth/A_sleek_boat_glides_effortlessly_through_the_shimmering_river,_van_gogh_style..gif" raw=true></td>
+  <td width=30% align="center"><img src="assets/depth/A_majestic_sailing_boat_cruises_along_the_vast,_azure_sea..gif" raw=true></td>
+	<td width=30% align="center"><img src="assets/depth/A_contented_cow_ambles_across_the_dewy,_verdant_pasture..gif" raw=true></td>
+</tr>
+<tr>
+  <td width=30% align="center">"A sleek boat glides effortlessly through the shimmering river, van gogh style."</td>
+  <td width=30% align="center">"A majestic sailing boat cruises along the vast, azure sea."</td>
+  <td width=30% align="center">"A contented cow ambles across the dewy, verdant pasture."</td>
+</tr>
+</table>
+
+### ControlVideo on canny edges
+
+<table class="center">
+<tr>
+  <td width=30% align="center"><img src="assets/canny/A_young_man_riding_a_sleek,_black_motorbike_through_the_winding_mountain_roads..gif" raw=true></td>
+  <td width=30% align="center"><img src="assets/canny/A_white_swan_moving_on_the_lake,_cartoon_style..gif" raw=true></td>
+	<td width=30% align="center"><img src="assets/canny/A_dusty_old_jeep_was_making_its_way_down_the_winding_forest_road,_creaking_and_groaning_with_each_bump_and_turn..gif" raw=true></td>
+</tr>
+<tr>
+  <td width=30% align="center">"A young man riding a sleek, black motorbike through the winding mountain roads."</td>
+  <td width=30% align="center">"A white swan movingon the lake, cartoon style."</td>
+  <td width=30% align="center">"A dusty old jeep was making its way down the winding forest road, creaking and groaning with each bump and turn."</td>
+</tr>
+ <tr>
+  <td width=30% align="center"><img src="assets/canny/A_shiny_red_jeep_smoothly_turns_on_a_narrow,_winding_road_in_the_mountains..gif" raw=true></td>
+  <td width=30% align="center"><img src="assets/canny/A_majestic_camel_gracefully_strides_across_the_scorching_desert_sands..gif" raw=true></td>
+	<td width=30% align="center"><img src="assets/canny/A_fit_man_is_leisurely_hiking_through_a_lush_and_verdant_forest..gif" raw=true></td>
+</tr>
+<tr>
+  <td width=30% align="center">"A shiny red jeep smoothly turns on a narrow, winding road in the mountains."</td>
+  <td width=30% align="center">"A majestic camel gracefully strides across the scorching desert sands."</td>
+  <td width=30% align="center">"A fit man is leisurely hiking through a lush and verdant forest."</td>
+</tr>
+</table>
+
+
+### ControlVideo on human poses
+
+<table class="center">
+<tr>
+  <td width=25% align="center"><img src="assets/pose/James_bond_moonwalk_on_the_beach,_animation_style.gif" raw=true></td>
+  <td width=25% align="center"><img src="assets/pose/Goku_in_a_mountain_range,_surreal_style..gif" raw=true></td>
+	<td width=25% align="center"><img src="assets/pose/Hulk_is_jumping_on_the_street,_cartoon_style.gif" raw=true></td>
+  <td width=25% align="center"><img src="assets/pose/A_robot_dances_on_a_road,_animation_style.gif" raw=true></td>
+</tr>
+<tr>
+  <td width=25% align="center">"James bond moonwalk on the beach, animation style."</td>
+  <td width=25% align="center">"Goku in a mountain range, surreal style."</td>
+  <td width=25% align="center">"Hulk is jumping on the street, cartoon style."</td>
+  <td width=25% align="center">"A robot dances on a road, animation style."</td>
+</tr></table>
+
+### Long video generation
+
+<table class="center">
+<tr>
+  <td width=60% align="center"><img src="assets/long/A_steamship_on_the_ocean,_at_sunset,_sketch_style.gif" raw=true></td>
+	<td width=40% align="center"><img src="assets/long/Hulk_is_dancing_on_the_beach,_cartoon_style.gif" raw=true></td>
+</tr>
+<tr>
+  <td width=60% align="center">"A steamship on the ocean, at sunset, sketch style."</td>
+  <td width=40% align="center">"Hulk is dancing on the beach, cartoon style."</td>
+</tr>
+</table>
+
+## Citation
+If you make use of our work, please cite our paper.
+```bibtex
+@article{zhang2023controlvideo,
+  title={ControlVideo: Training-free Controllable Text-to-Video Generation},
+  author={Zhang, Yabo and Wei, Yuxiang and Jiang, Dongsheng and Zhang, Xiaopeng and Zuo, Wangmeng and Tian, Qi},
+  journal={arXiv preprint arXiv:2305.13077},
+  year={2023}
+}
+```
+
+## Acknowledgement
+This work repository borrows heavily from [Diffusers](https://github.com/huggingface/diffusers), [ControlNet](https://github.com/lllyasviel/ControlNet), [Tune-A-Video](https://github.com/showlab/Tune-A-Video), and [RIFE](https://github.com/megvii-research/ECCV2022-RIFE).
+
+There are also many interesting works on video generation: [Tune-A-Video](https://github.com/showlab/Tune-A-Video), [Text2Video-Zero](https://github.com/Picsart-AI-Research/Text2Video-Zero), [Follow-Your-Pose](https://github.com/mayuelala/FollowYourPose), [Control-A-Video](https://github.com/Weifeng-Chen/control-a-video), et al.

app.py

@@ -0,0 +1,348 @@
+import gradio as gr
+import os
+import shutil
+import subprocess
+import cv2
+import numpy as np
+import math
+
+from huggingface_hub import snapshot_download
+
+os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
+
+
+model_ids = [
+    'runwayml/stable-diffusion-v1-5',
+    'lllyasviel/sd-controlnet-depth', 
+    'lllyasviel/sd-controlnet-canny', 
+    'lllyasviel/sd-controlnet-openpose',
+    "lllyasviel/control_v11p_sd15_softedge",
+    "lllyasviel/control_v11p_sd15_scribble",
+    "lllyasviel/control_v11p_sd15_lineart_anime",
+    "lllyasviel/control_v11p_sd15_lineart",
+    "lllyasviel/control_v11f1p_sd15_depth",
+    "lllyasviel/control_v11p_sd15_canny",
+    "lllyasviel/control_v11p_sd15_openpose",
+    "lllyasviel/control_v11p_sd15_normalbae"
+]
+
+
+for model_id in model_ids:
+    model_name = model_id.split('/')[-1]
+    snapshot_download(model_id, cache_dir=f'checkpoints/{model_name}')
+
+def load_model(model_id):
+    local_dir = f'checkpoints/stable-diffusion-v1-5'
+    # Check if the directory exists
+    if os.path.exists(local_dir):
+        # Delete the directory if it exists
+        shutil.rmtree(local_dir)
+
+    model_name = model_id.split('/')[-1]
+    snapshot_download(model_id, local_dir=f'checkpoints/{model_name}')
+    os.rename(f'checkpoints/{model_name}', f'checkpoints/stable-diffusion-v1-5')
+    return "model loaded"
+
+def get_frame_count(filepath):
+    if filepath is not None:
+        video = cv2.VideoCapture(filepath) 
+        frame_count = int(video.get(cv2.CAP_PROP_FRAME_COUNT))
+    
+        video.release()
+
+        # LIMITS
+        if frame_count > 100 :
+            frame_count = 100 # limit to 100 frames to avoid cuDNN errors
+
+        return gr.update(maximum=frame_count)
+
+    else:
+        return gr.update(value=1, maximum=100 )
+
+def get_video_dimension(filepath):
+    video = cv2.VideoCapture(filepath)
+    width = int(video.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(video.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    fps = int(video.get(cv2.CAP_PROP_FPS))
+    frame_count = int(video.get(cv2.CAP_PROP_FRAME_COUNT))
+    video.release()
+    return width, height, fps, frame_count
+
+def resize_video(input_vid, output_vid, width, height, fps):
+    print(f"RESIZING ...")
+    # Open the input video file
+    video = cv2.VideoCapture(input_vid)
+
+    # Create a VideoWriter object to write the resized video
+    fourcc = cv2.VideoWriter_fourcc(*'mp4v')  # Codec for the output video
+    output_video = cv2.VideoWriter(output_vid, fourcc, fps, (width, height))
+
+    while True:
+        # Read a frame from the input video
+        ret, frame = video.read()
+        if not ret:
+            break
+
+        # Resize the frame to the desired dimensions
+        resized_frame = cv2.resize(frame, (width, height))
+
+        # Write the resized frame to the output video file
+        output_video.write(resized_frame)
+
+    # Release the video objects
+    video.release()
+    output_video.release()
+    print(f"RESIZE VIDEO DONE!")
+    return output_vid
+
+def make_nearest_multiple_of_32(number):
+    remainder = number % 32
+    if remainder <= 16:
+        number -= remainder
+    else:
+        number += 32 - remainder
+    return number 
+
+def change_video_fps(input_path):
+    print(f"CHANGING FIANL OUTPUT FPS")
+    cap = cv2.VideoCapture(input_path)
+    # Check if the final file already exists
+    if os.path.exists('output_video.mp4'):
+        # Delete the existing file
+        os.remove('output_video.mp4')
+    output_path = 'output_video.mp4'
+    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
+    output_fps = 12
+    output_size = (int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)), int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)))
+    out = cv2.VideoWriter(output_path, fourcc, output_fps, output_size)
+    
+    frame_count = 0
+    while cap.isOpened():
+        ret, frame = cap.read()
+        if not ret:
+            break
+        
+        # Write the current frame to the output video multiple times to increase the frame rate
+        for _ in range(output_fps // 8):
+            out.write(frame)
+        
+        frame_count += 1
+        print(f'Processed frame {frame_count}')
+
+    cap.release()
+    out.release()
+    cv2.destroyAllWindows()
+
+    return 'output_video.mp4'
+
+def run_inference(prompt, video_path, version_condition, video_length, seed):
+    
+    seed = math.floor(seed)
+    o_width = get_video_dimension(video_path)[0]
+    o_height = get_video_dimension(video_path)[1]
+    version, condition = version_condition.split("+")
+
+    # Prepare dimensions
+    if o_width > 512 :
+        # Calculate the new height while maintaining the aspect ratio
+        n_height = int(o_height / o_width * 512)
+        n_width = 512
+    else:
+        n_height = o_height
+        n_width = o_width
+
+    # Make sure new dimensions are multipe of 32
+    r_width = make_nearest_multiple_of_32(n_width)
+    r_height = make_nearest_multiple_of_32(n_height)
+    print(f"multiple of 32 sizes : {r_width}x{r_height}")
+
+    # Get FPS of original video input
+    original_fps = get_video_dimension(video_path)[2] 
+    if original_fps > 12 :
+        print(f"FPS is too high: {original_fps}")          
+        target_fps = 12
+    else : 
+        target_fps = original_fps
+    print(f"NEW INPUT FPS: {target_fps}, NEW LENGTH: {video_length}")
+    
+    # Check if the resized file already exists
+    if os.path.exists('resized.mp4'):
+        # Delete the existing file
+        os.remove('resized.mp4')
+    
+    resized = resize_video(video_path, 'resized.mp4', r_width, r_height, target_fps)
+    resized_video_fcount = get_video_dimension(resized)[3]
+    print(f"RESIZED VIDEO FRAME COUNT: {resized_video_fcount}")
+
+    # Make sure new total frame count is enough to handle chosen video length
+    if video_length > resized_video_fcount :
+        video_length = resized_video_fcount
+    # video_length = int((target_fps * video_length) / original_fps)
+    
+    output_path = 'output/'
+    os.makedirs(output_path, exist_ok=True)
+            
+    # Check if the file already exists
+    if os.path.exists(os.path.join(output_path, f"result.mp4")):
+        # Delete the existing file
+        os.remove(os.path.join(output_path, f"result.mp4"))
+
+    print(f"RUNNING INFERENCE ...")
+    if video_length > 16:
+        command = f"python inference.py --prompt '{prompt}' --condition '{condition}' --video_path '{resized}' --output_path '{output_path}' --temp_video_name 'result' --width {r_width} --height {r_height} --seed {seed} --video_length {video_length} --smoother_steps 19 20 --version {version} --is_long_video"
+    else:
+        command = f"python inference.py --prompt '{prompt}' --condition '{condition}' --video_path '{resized}' --output_path '{output_path}' --temp_video_name 'result'  --width {r_width} --height {r_height} --seed {seed} --video_length {video_length} --smoother_steps 19 20 --version {version} "
+    
+    try:
+        subprocess.run(command, shell=True)
+    except cuda.Error as e:
+        return f"CUDA Error: {e}", None
+    except RuntimeError as e:
+        return f"Runtime Error: {e}", None
+
+    # Construct the video path
+    video_path_output = os.path.join(output_path, f"result.mp4")
+
+    # Resize to original video input size
+    #o_width = get_video_dimension(video_path)[0]
+    #o_height = get_video_dimension(video_path)[1]
+    #resize_video(video_path_output, 'resized_final.mp4', o_width, o_height, target_fps)
+
+    # Check generated video FPS
+    gen_fps = get_video_dimension(video_path_output)[2] 
+    print(f"GEN VIDEO FPS: {gen_fps}")
+    final = change_video_fps(video_path_output)
+    print(f"FINISHED !")
+    
+    return final
+    # return final, gr.Group.update(visible=True)
+ 
+
+css="""
+#col-container {max-width: 810px; margin-left: auto; margin-right: auto;}
+.animate-spin {
+  animation: spin 1s linear infinite;
+}
+@keyframes spin {
+  from {
+      transform: rotate(0deg);
+  }
+  to {
+      transform: rotate(360deg);
+  }
+}
+#share-btn-container {
+  display: flex; 
+  padding-left: 0.5rem !important; 
+  padding-right: 0.5rem !important; 
+  background-color: #000000; 
+  justify-content: center; 
+  align-items: center; 
+  border-radius: 9999px !important; 
+  max-width: 13rem;
+}
+#share-btn-container:hover {
+  background-color: #060606;
+}
+#share-btn {
+  all: initial; 
+  color: #ffffff;
+  font-weight: 600; 
+  cursor:pointer; 
+  font-family: 'IBM Plex Sans', sans-serif; 
+  margin-left: 0.5rem !important; 
+  padding-top: 0.5rem !important; 
+  padding-bottom: 0.5rem !important;
+  right:0;
+}
+#share-btn * {
+  all: unset;
+}
+#share-btn-container div:nth-child(-n+2){
+  width: auto !important;
+  min-height: 0px !important;
+}
+#share-btn-container .wrap {
+  display: none !important;
+}
+#share-btn-container.hidden {
+  display: none!important;
+}
+img[src*='#center'] { 
+    display: block;
+    margin: auto;
+}
+"""
+with gr.Blocks(css=css) as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.Markdown("""
+            <h1 style="text-align: center;">ControlVideo: Training-free Controllable Text-to-Video Generation</h1>
+            <p style="text-align: center;">
+                [<a href="https://arxiv.org/abs/2305.13077" style="color:blue;">arXiv</a>] 
+                [<a href="https://github.com/YBYBZhang/ControlVideo" style="color:blue;">GitHub</a>]
+            </p>
+            <p style="text-align: center;"> ControlVideo adapts ControlNet to the video counterpart without any finetuning, aiming to directly inherit its high-quality and consistent generation. </p>            
+        """)
+        
+        with gr.Column():
+            with gr.Row():
+                video_path = gr.Video(label="Input video", source="upload", type="filepath", visible=True, elem_id="video-in")
+                video_res = gr.Video(label="result", elem_id="video-out")
+                
+                # with gr.Column():
+                #     video_res = gr.Video(label="result", elem_id="video-out")
+                #     with gr.Group(elem_id="share-btn-container", visible=False) as share_group:
+                #         community_icon = gr.HTML(community_icon_html)
+                #         loading_icon = gr.HTML(loading_icon_html)
+                #         share_button = gr.Button("Share to community", elem_id="share-btn")
+            with gr.Row():
+                chosen_model = gr.Dropdown(label="Diffusion model (*1.5)", choices=['runwayml/stable-diffusion-v1-5','nitrosocke/Ghibli-Diffusion'], value="runwayml/stable-diffusion-v1-5", allow_custom_value=True)
+                model_status = gr.Textbox(label="status")
+                load_model_btn = gr.Button("load model (optional)")
+            prompt = gr.Textbox(label="prompt", info="If you loaded a custom model, do not forget to include Prompt trigger", elem_id="prompt-in")
+            with gr.Column():
+                video_length = gr.Slider(label="Video length", info="How many frames do you want to process ? For demo purpose, max is set to 24", minimum=1, maximum=12, step=1, value=2)
+                with gr.Row():
+                    # version = gr.Dropdown(label="ControlNet version", choices=["v10", "v11"], value="v10")
+                    version_condition = gr.Dropdown(label="ControlNet version + Condition", 
+                                                    choices=["v10+depth_midas", "v10+canny", "v10+openpose", "v11+softedge_pidinet", "v11+softedge_pidsafe",
+                                                    "v11+softedge_hed", "v11+softedge_hedsafe", "v11+scribble_hed", "v11+scribble_pidinet", "v11+lineart_anime",
+                                                    "v11+lineart_coarse", "v11+lineart_realistic", "v11+depth_midas", "v11+depth_leres", "v11+depth_leres++", 
+                                                    "v11+depth_zoe", "v11+canny", "v11+openpose", "v11+openpose_face", "v11+openpose_faceonly", "v11+openpose_full", 
+                                                    "v11+openpose_hand", "v11+normal_bae"], value="v10+depth_midas")
+                    seed = gr.Number(label="seed", value=42)
+            submit_btn = gr.Button("Submit")
+        
+            
+            gr.Examples(
+                examples=[["James bond moonwalks on the beach.", "./data/moonwalk.mp4", 'v10+openpose', 15, 42],
+                          ["A striking mallard floats effortlessly on the sparkling pond.", "./data/mallard-water.mp4", "v11+depth_midas", 15, 42]],
+                fn=run_inference,
+                inputs=[prompt,
+                         video_path,
+                         version_condition,
+                         video_length,
+                         seed,
+                        ],
+                # outputs=[video_res, share_group],
+                outputs=video_res,
+                cache_examples=False
+            )
+                
+    # share_button.click(None, [], [], _js=share_js)
+    load_model_btn.click(fn=load_model, inputs=[chosen_model], outputs=[model_status], queue=False)
+    video_path.change(fn=get_frame_count,
+                      inputs=[video_path],
+                      outputs=[video_length],
+                      queue=False
+                     )
+    submit_btn.click(fn=run_inference, 
+                     inputs=[prompt,
+                             video_path,
+                             version_condition,
+                             video_length,
+                             seed,
+                            ],
+                    outputs=video_res)
+
+demo.queue(max_size=12).launch()

cog.yaml

@@ -0,0 +1,53 @@
+# Configuration for Cog ⚙️
+# Reference: https://github.com/replicate/cog/blob/main/docs/yaml.md
+
+build:
+  gpu: true
+  cuda: "11.6"
+  system_packages:
+    - "libgl1-mesa-glx"
+    - "libglib2.0-0"
+  python_version: "3.8"
+  python_packages:
+   - "accelerate==0.17.1"
+   - "addict==2.4.0"
+   - "basicsr==1.4.2"
+   - "bitsandbytes==0.35.4"
+   - "clip==0.2.0"
+   - "cmake==3.25.2"
+   - "controlnet-aux==0.0.4"
+   - "decord==0.6.0"
+   - "deepspeed==0.8.0"
+   - "diffusers==0.14.0"
+   - "easydict==1.10"
+   - "einops==0.6.0"
+   - "ffmpy==0.3.0"
+   - "ftfy==6.1.1"
+   - "imageio==2.25.1"
+   - "imageio-ffmpeg==0.4.8"
+   - "moviepy==1.0.3"
+   - "numpy==1.24.2"
+   - "omegaconf==2.3.0"
+   - "opencv-python==4.7.0.68"
+   - "pandas==1.5.3"
+   - "pillow==9.4.0"
+   - "scikit-image==0.19.3"
+   - "scipy==1.10.1"
+   - "tensorboard==2.12.0"
+   - "tensorboard-data-server==0.7.0"
+   - "tensorboard-plugin-wit==1.8.1"
+   - "termcolor==2.2.0"
+   - "thinc==8.1.10"
+   - "timm==0.6.12"
+   - "tokenizers==0.13.2"
+   - "torch==1.13.1"
+   - "torchvision==0.14.1"
+   - "tqdm==4.64.1"
+   - "transformers==4.26.1"
+   - "wandb==0.13.10"
+   - "xformers==0.0.16"
+   - "positional_encodings==6.0.1"
+   - "mediapipe==0.10.0"
+   - "triton==2.0.0.post1"
+
+predict: "predict.py:Predictor"

inference.py

@@ -0,0 +1,153 @@
+import os
+import numpy as np
+import argparse
+import imageio
+import torch
+
+from einops import rearrange
+from diffusers import DDIMScheduler, AutoencoderKL
+from transformers import CLIPTextModel, CLIPTokenizer
+
+import torchvision
+from controlnet_aux.processor import Processor
+
+from models.pipeline_controlvideo import ControlVideoPipeline
+from models.util import save_videos_grid, read_video
+from models.unet import UNet3DConditionModel
+from models.controlnet import ControlNetModel3D
+from models.RIFE.IFNet_HDv3 import IFNet
+
+
+device = "cuda"
+sd_path = "checkpoints/stable-diffusion-v1-5"
+inter_path = "checkpoints/flownet.pkl"
+controlnet_dict_version = {
+    "v10":{
+        "openpose": "checkpoints/sd-controlnet-openpose",
+        "depth_midas": "checkpoints/sd-controlnet-depth",
+        "canny": "checkpoints/sd-controlnet-canny",
+    },
+    "v11": {
+    "softedge_pidinet": "checkpoints/control_v11p_sd15_softedge",
+    "softedge_pidsafe": "checkpoints/control_v11p_sd15_softedge",
+    "softedge_hed": "checkpoints/control_v11p_sd15_softedge",
+    "softedge_hedsafe": "checkpoints/control_v11p_sd15_softedge",
+    "scribble_hed": "checkpoints/control_v11p_sd15_scribble",
+    "scribble_pidinet": "checkpoints/control_v11p_sd15_scribble",
+    "lineart_anime": "checkpoints/control_v11p_sd15_lineart_anime",
+    "lineart_coarse": "checkpoints/control_v11p_sd15_lineart",
+    "lineart_realistic": "checkpoints/control_v11p_sd15_lineart",
+    "depth_midas": "checkpoints/control_v11f1p_sd15_depth",
+    "depth_leres": "checkpoints/control_v11f1p_sd15_depth",
+    "depth_leres++": "checkpoints/control_v11f1p_sd15_depth",
+    "depth_zoe": "checkpoints/control_v11f1p_sd15_depth",
+    "canny": "checkpoints/control_v11p_sd15_canny",
+    "openpose": "checkpoints/control_v11p_sd15_openpose",
+    "openpose_face": "checkpoints/control_v11p_sd15_openpose",
+    "openpose_faceonly": "checkpoints/control_v11p_sd15_openpose",
+    "openpose_full": "checkpoints/control_v11p_sd15_openpose",
+    "openpose_hand": "checkpoints/control_v11p_sd15_openpose",
+    "normal_bae": "checkpoints/control_v11p_sd15_normalbae"
+    }
+}
+# load processor from processor_id
+# options are:
+# ["canny", "depth_leres", "depth_leres++", "depth_midas", "depth_zoe", "lineart_anime",
+#  "lineart_coarse", "lineart_realistic", "mediapipe_face", "mlsd", "normal_bae", "normal_midas",
+#  "openpose", "openpose_face", "openpose_faceonly", "openpose_full", "openpose_hand",
+#  "scribble_hed, "scribble_pidinet", "shuffle", "softedge_hed", "softedge_hedsafe",
+#  "softedge_pidinet", "softedge_pidsafe"]
+
+POS_PROMPT = " ,best quality, extremely detailed, HD, ultra-realistic, 8K, HQ, masterpiece, trending on artstation, art, smooth"
+NEG_PROMPT = "longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic"
+
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--prompt", type=str, required=True, help="Text description of target video")
+    parser.add_argument("--video_path", type=str, required=True, help="Path to a source video")
+    parser.add_argument("--output_path", type=str, default="./outputs", help="Directory of output")
+    parser.add_argument("--condition", type=str, default="depth", help="Condition of structure sequence")
+    parser.add_argument("--video_length", type=int, default=15, help="Length of synthesized video")
+    parser.add_argument("--height", type=int, default=512, help="Height of synthesized video, and should be a multiple of 32")
+    parser.add_argument("--width", type=int, default=512, help="Width of synthesized video, and should be a multiple of 32")
+    parser.add_argument("--smoother_steps", nargs='+', default=[19, 20], type=int, help="Timesteps at which using interleaved-frame smoother")
+    parser.add_argument("--is_long_video", action='store_true', help="Whether to use hierarchical sampler to produce long video")
+    parser.add_argument("--seed", type=int, default=42, help="Random seed of generator")
+    parser.add_argument("--version", type=str, default='v10', choices=["v10", "v11"], help="Version of ControlNet")
+    parser.add_argument("--frame_rate", type=int, default=None, help="The frame rate of loading input video. Default rate is computed according to video length.")
+    parser.add_argument("--temp_video_name", type=str, default=None, help="Default video name")
+    
+    args = parser.parse_args()
+    return args
+
+if __name__ == "__main__":
+    args = get_args()
+    os.makedirs(args.output_path, exist_ok=True)
+    
+    # Height and width should be a multiple of 32
+    args.height = (args.height // 32) * 32    
+    args.width = (args.width // 32) * 32    
+
+    processor = Processor(args.condition)
+    controlnet_dict = controlnet_dict_version[args.version]
+    
+    tokenizer = CLIPTokenizer.from_pretrained(sd_path, subfolder="tokenizer")
+    text_encoder = CLIPTextModel.from_pretrained(sd_path, subfolder="text_encoder").to(dtype=torch.float16)
+    vae = AutoencoderKL.from_pretrained(sd_path, subfolder="vae").to(dtype=torch.float16)
+    unet = UNet3DConditionModel.from_pretrained_2d(sd_path, subfolder="unet").to(dtype=torch.float16)
+    controlnet = ControlNetModel3D.from_pretrained_2d(controlnet_dict[args.condition]).to(dtype=torch.float16)
+    interpolater = IFNet(ckpt_path=inter_path).to(dtype=torch.float16)
+    scheduler=DDIMScheduler.from_pretrained(sd_path, subfolder="scheduler")
+
+    pipe = ControlVideoPipeline(
+            vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet,
+            controlnet=controlnet, interpolater=interpolater, scheduler=scheduler,
+        )
+    pipe.enable_vae_slicing()
+    pipe.enable_xformers_memory_efficient_attention()
+    pipe.to(device)
+
+    generator = torch.Generator(device="cuda")
+    generator.manual_seed(args.seed)
+
+    # Step 1. Read a video
+    video = read_video(video_path=args.video_path, video_length=args.video_length, width=args.width, height=args.height, frame_rate=args.frame_rate)
+
+    # Save source video
+    original_pixels = rearrange(video, "(b f) c h w -> b c f h w", b=1)
+    save_videos_grid(original_pixels, os.path.join(args.output_path, "source_video.mp4"), rescale=True)
+
+
+    # Step 2. Parse a video to conditional frames
+    t2i_transform = torchvision.transforms.ToPILImage()
+    pil_annotation = []
+    for frame in video:
+        pil_frame = t2i_transform(frame)
+        pil_annotation.append(processor(pil_frame, to_pil=True))
+
+    # Save condition video
+    video_cond = [np.array(p).astype(np.uint8) for p in pil_annotation]
+    imageio.mimsave(os.path.join(args.output_path, f"{args.condition}_condition.mp4"), video_cond, fps=8)
+
+    # Reduce memory (optional)
+    del processor; torch.cuda.empty_cache()
+
+    # Step 3. inference
+
+    if args.is_long_video:
+        window_size = int(np.sqrt(args.video_length))
+        sample = pipe.generate_long_video(args.prompt + POS_PROMPT, video_length=args.video_length, frames=pil_annotation, 
+                    num_inference_steps=50, smooth_steps=args.smoother_steps, window_size=window_size,
+                    generator=generator, guidance_scale=12.5, negative_prompt=NEG_PROMPT,
+                    width=args.width, height=args.height
+                ).videos
+    else:
+        sample = pipe(args.prompt + POS_PROMPT, video_length=args.video_length, frames=pil_annotation, 
+                    num_inference_steps=50, smooth_steps=args.smoother_steps,
+                    generator=generator, guidance_scale=12.5, negative_prompt=NEG_PROMPT,
+                    width=args.width, height=args.height
+                ).videos
+    args.temp_video_name = args.prompt if args.temp_video_name is None else args.temp_video_name
+    save_videos_grid(sample, f"{args.output_path}/{args.temp_video_name}.mp4")

inference.sh

@@ -0,0 +1,12 @@
+python inference.py \
+    --prompt "A striking mallard floats effortlessly on the sparkling pond." \
+    --condition "depth_midas" \
+    --video_path "data/mallard-water.mp4" \
+    --output_path "outputs/" \
+    --video_length 15 \
+    --smoother_steps 19 20 \
+    --width 512 \
+    --height 512 \
+    --frame_rate 2 \
+    --version v10 \
+    # --is_long_video

models/RIFE/IFNet_HDv3.py

@@ -0,0 +1,130 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers import ModelMixin
+
+from .warplayer import warp
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=True),        
+        nn.PReLU(out_planes)
+    )
+
+def conv_bn(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=False),
+        nn.BatchNorm2d(out_planes),
+        nn.PReLU(out_planes)
+    )
+
+def convert(param):
+        return {
+            k.replace("module.", ""): v
+            for k, v in param.items()
+            if "module." in k
+        }
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, c=64):
+        super(IFBlock, self).__init__()
+        self.conv0 = nn.Sequential(
+            conv(in_planes, c//2, 3, 2, 1),
+            conv(c//2, c, 3, 2, 1),
+            )
+        self.convblock0 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.convblock1 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.convblock2 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.convblock3 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.conv1 = nn.Sequential(
+            nn.ConvTranspose2d(c, c//2, 4, 2, 1),
+            nn.PReLU(c//2),
+            nn.ConvTranspose2d(c//2, 4, 4, 2, 1),
+        )
+        self.conv2 = nn.Sequential(
+            nn.ConvTranspose2d(c, c//2, 4, 2, 1),
+            nn.PReLU(c//2),
+            nn.ConvTranspose2d(c//2, 1, 4, 2, 1),
+        )
+
+    def forward(self, x, flow, scale=1):
+        x = F.interpolate(x, scale_factor= 1. / scale, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+        flow = F.interpolate(flow, scale_factor= 1. / scale, mode="bilinear", align_corners=False, recompute_scale_factor=False) * 1. / scale
+        feat = self.conv0(torch.cat((x, flow), 1))
+        feat = self.convblock0(feat) + feat
+        feat = self.convblock1(feat) + feat
+        feat = self.convblock2(feat) + feat
+        feat = self.convblock3(feat) + feat        
+        flow = self.conv1(feat)
+        mask = self.conv2(feat)
+        flow = F.interpolate(flow, scale_factor=scale, mode="bilinear", align_corners=False, recompute_scale_factor=False) * scale
+        mask = F.interpolate(mask, scale_factor=scale, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+        return flow, mask
+        
+class IFNet(ModelMixin):
+    def __init__(self, ckpt_path="checkpoints/flownet.pkl"):
+        super(IFNet, self).__init__()
+        self.block0 = IFBlock(7+4, c=90)
+        self.block1 = IFBlock(7+4, c=90)
+        self.block2 = IFBlock(7+4, c=90)
+        self.block_tea = IFBlock(10+4, c=90)
+        if ckpt_path is not None:
+            self.load_state_dict(convert(torch.load(ckpt_path, map_location ='cpu')))
+        
+    def inference(self, img0, img1, scale=1.0):
+        imgs = torch.cat((img0, img1), 1)
+        scale_list = [4/scale, 2/scale, 1/scale]
+        flow, mask, merged = self.forward(imgs, scale_list)
+        return merged[2]
+
+    def forward(self, x, scale_list=[4, 2, 1], training=False):
+        if training == False:
+            channel = x.shape[1] // 2
+            img0 = x[:, :channel]
+            img1 = x[:, channel:]
+        flow_list = []
+        merged = []
+        mask_list = []
+        warped_img0 = img0
+        warped_img1 = img1
+        flow = (x[:, :4]).detach() * 0
+        mask = (x[:, :1]).detach() * 0
+        loss_cons = 0
+        block = [self.block0, self.block1, self.block2]
+        for i in range(3):
+            f0, m0 = block[i](torch.cat((warped_img0[:, :3], warped_img1[:, :3], mask), 1), flow, scale=scale_list[i])
+            f1, m1 = block[i](torch.cat((warped_img1[:, :3], warped_img0[:, :3], -mask), 1), torch.cat((flow[:, 2:4], flow[:, :2]), 1), scale=scale_list[i])
+            flow = flow + (f0 + torch.cat((f1[:, 2:4], f1[:, :2]), 1)) / 2
+            mask = mask + (m0 + (-m1)) / 2
+            mask_list.append(mask)
+            flow_list.append(flow)
+            warped_img0 = warp(img0, flow[:, :2])
+            warped_img1 = warp(img1, flow[:, 2:4])
+            merged.append((warped_img0, warped_img1))
+        '''
+        c0 = self.contextnet(img0, flow[:, :2])
+        c1 = self.contextnet(img1, flow[:, 2:4])
+        tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1)
+        res = tmp[:, 1:4] * 2 - 1
+        '''
+        for i in range(3):
+            mask_list[i] = torch.sigmoid(mask_list[i])
+            merged[i] = merged[i][0] * mask_list[i] + merged[i][1] * (1 - mask_list[i])
+            # merged[i] = torch.clamp(merged[i] + res, 0, 1)        
+        return flow_list, mask_list[2], merged

models/RIFE/warplayer.py

@@ -0,0 +1,22 @@
+import torch
+import torch.nn as nn
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+backwarp_tenGrid = {}
+
+
+def warp(tenInput, tenFlow):
+    k = (str(tenFlow.device), str(tenFlow.size()))
+    if k not in backwarp_tenGrid:
+        tenHorizontal = torch.linspace(-1.0, 1.0, tenFlow.shape[3], device=device).view(
+            1, 1, 1, tenFlow.shape[3]).expand(tenFlow.shape[0], -1, tenFlow.shape[2], -1)
+        tenVertical = torch.linspace(-1.0, 1.0, tenFlow.shape[2], device=device).view(
+            1, 1, tenFlow.shape[2], 1).expand(tenFlow.shape[0], -1, -1, tenFlow.shape[3])
+        backwarp_tenGrid[k] = torch.cat(
+            [tenHorizontal, tenVertical], 1).to(device)
+
+    tenFlow = torch.cat([tenFlow[:, 0:1, :, :] / ((tenInput.shape[3] - 1.0) / 2.0),
+                         tenFlow[:, 1:2, :, :] / ((tenInput.shape[2] - 1.0) / 2.0)], 1)
+
+    g = (backwarp_tenGrid[k] + tenFlow).permute(0, 2, 3, 1).to(dtype=tenInput.dtype)
+    return torch.nn.functional.grid_sample(input=tenInput, grid=g, mode='bilinear', padding_mode='border', align_corners=True)

models/attention.py

@@ -0,0 +1,478 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from dataclasses import dataclass
+from typing import Optional, Callable
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn
+from positional_encodings.torch_encodings import PositionalEncoding2D
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers import ModelMixin
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from diffusers.models.attention import CrossAttention, FeedForward, AdaLayerNorm
+from einops import rearrange, repeat
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    @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,
+    ):
+        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(
+            [
+                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,
+                    upcast_attention=upcast_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)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True, \
+                inter_frame=False):
+        # 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")
+        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 block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                timestep=timestep,
+                video_length=video_length,
+                inter_frame=inter_frame
+            )
+
+        # 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)
+
+
+class BasicTransformerBlock(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,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+
+        # Fully
+        self.attn1 = FullyFrameAttention(
+            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,
+        )
+
+        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 = CrossAttention(
+                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)
+
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None):
+        if not is_xformers_available():
+            print("Here is how to install it")
+            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
+            self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+            if self.attn2 is not None:
+                self.attn2._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+
+    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, attention_mask=None, video_length=None, \
+                inter_frame=False):
+        # SparseCausal-Attention
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
+        )
+
+        if self.only_cross_attention:
+            hidden_states = (
+                self.attn1(norm_hidden_states, encoder_hidden_states, attention_mask=attention_mask, inter_frame=inter_frame) + hidden_states
+            )
+        else:
+            hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask, video_length=video_length, inter_frame=inter_frame) + 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
+
+        return hidden_states
+
+class FullyFrameAttention(nn.Module):
+    r"""
+    A cross attention layer.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        cross_attention_dim (`int`, *optional*):
+            The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
+        heads (`int`,  *optional*, defaults to 8): The number of heads to use for multi-head attention.
+        dim_head (`int`,  *optional*, defaults to 64): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        bias (`bool`, *optional*, defaults to False):
+            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
+    """
+
+    def __init__(
+        self,
+        query_dim: int,
+        cross_attention_dim: Optional[int] = None,
+        heads: int = 8,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias=False,
+        upcast_attention: bool = False,
+        upcast_softmax: bool = False,
+        added_kv_proj_dim: Optional[int] = None,
+        norm_num_groups: Optional[int] = None,
+    ):
+        super().__init__()
+        inner_dim = dim_head * heads
+        cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+
+        self.scale = dim_head**-0.5
+
+        self.heads = heads
+        # for slice_size > 0 the attention score computation
+        # is split across the batch axis to save memory
+        # You can set slice_size with `set_attention_slice`
+        self.sliceable_head_dim = heads
+        self._slice_size = None
+        self._use_memory_efficient_attention_xformers = False
+        self.added_kv_proj_dim = added_kv_proj_dim
+
+        if norm_num_groups is not None:
+            self.group_norm = nn.GroupNorm(num_channels=inner_dim, num_groups=norm_num_groups, eps=1e-5, affine=True)
+        else:
+            self.group_norm = None
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=bias)
+        self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
+        self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
+
+        if self.added_kv_proj_dim is not None:
+            self.add_k_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
+            self.add_v_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
+
+        self.to_out = nn.ModuleList([])
+        self.to_out.append(nn.Linear(inner_dim, query_dim))
+        self.to_out.append(nn.Dropout(dropout))
+
+    def reshape_heads_to_batch_dim(self, tensor):
+        batch_size, seq_len, dim = tensor.shape
+        head_size = self.heads
+        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
+        return tensor
+
+    def reshape_batch_dim_to_heads(self, tensor):
+        batch_size, seq_len, dim = tensor.shape
+        head_size = self.heads
+        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
+        return tensor
+
+    def set_attention_slice(self, slice_size):
+        if slice_size is not None and slice_size > self.sliceable_head_dim:
+            raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")
+
+        self._slice_size = slice_size
+    
+    def _attention(self, query, key, value, attention_mask=None):
+        if self.upcast_attention:
+            query = query.float()
+            key = key.float()
+
+        attention_scores = torch.baddbmm(
+            torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device),
+            query,
+            key.transpose(-1, -2),
+            beta=0,
+            alpha=self.scale,
+        )
+        if attention_mask is not None:
+            attention_scores = attention_scores + attention_mask
+
+        if self.upcast_softmax:
+            attention_scores = attention_scores.float()
+
+        attention_probs = attention_scores.softmax(dim=-1)
+
+        # cast back to the original dtype
+        attention_probs = attention_probs.to(value.dtype)
+
+        # compute attention output
+        hidden_states = torch.bmm(attention_probs, value)
+
+        # reshape hidden_states
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+    def _sliced_attention(self, query, key, value, sequence_length, dim, attention_mask):
+        batch_size_attention = query.shape[0]
+        hidden_states = torch.zeros(
+            (batch_size_attention, sequence_length, dim // self.heads), device=query.device, dtype=query.dtype
+        )
+        slice_size = self._slice_size if self._slice_size is not None else hidden_states.shape[0]
+        for i in range(hidden_states.shape[0] // slice_size):
+            start_idx = i * slice_size
+            end_idx = (i + 1) * slice_size
+
+            query_slice = query[start_idx:end_idx]
+            key_slice = key[start_idx:end_idx]
+
+            if self.upcast_attention:
+                query_slice = query_slice.float()
+                key_slice = key_slice.float()
+
+            attn_slice = torch.baddbmm(
+                torch.empty(slice_size, query.shape[1], key.shape[1], dtype=query_slice.dtype, device=query.device),
+                query_slice,
+                key_slice.transpose(-1, -2),
+                beta=0,
+                alpha=self.scale,
+            )
+
+            if attention_mask is not None:
+                attn_slice = attn_slice + attention_mask[start_idx:end_idx]
+
+            if self.upcast_softmax:
+                attn_slice = attn_slice.float()
+
+            attn_slice = attn_slice.softmax(dim=-1)
+
+            # cast back to the original dtype
+            attn_slice = attn_slice.to(value.dtype)
+            attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
+
+            hidden_states[start_idx:end_idx] = attn_slice
+
+        # reshape hidden_states
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+    def _memory_efficient_attention_xformers(self, query, key, value, attention_mask):
+        # TODO attention_mask
+        query = query.contiguous()
+        key = key.contiguous()
+        value = value.contiguous()
+        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, video_length=None, inter_frame=False):
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        encoder_hidden_states = encoder_hidden_states
+
+        if self.group_norm is not None:
+            hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = self.to_q(hidden_states)  # (bf) x d(hw) x c
+        dim = query.shape[-1]
+
+        # All frames
+        query = rearrange(query, "(b f) d c -> b (f d) c", f=video_length)
+        
+        query = self.reshape_heads_to_batch_dim(query)
+
+        if self.added_kv_proj_dim is not None:
+            raise NotImplementedError
+
+        encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
+        key = self.to_k(encoder_hidden_states)
+        value = self.to_v(encoder_hidden_states)
+
+        if inter_frame:
+            key = rearrange(key, "(b f) d c -> b f d c", f=video_length)[:, [0, -1]]
+            value = rearrange(value, "(b f) d c -> b f d c", f=video_length)[:, [0, -1]]
+            key = rearrange(key, "b f d c -> b (f d) c",)
+            value = rearrange(value, "b f d c -> b (f d) c")
+        else:
+            # All frames
+            key = rearrange(key, "(b f) d c -> b (f d) c", f=video_length)
+            value = rearrange(value, "(b f) d c -> b (f d) c", f=video_length)
+
+        key = self.reshape_heads_to_batch_dim(key)
+        value = self.reshape_heads_to_batch_dim(value)
+
+        if attention_mask is not None:
+            if attention_mask.shape[-1] != query.shape[1]:
+                target_length = query.shape[1]
+                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
+                attention_mask = attention_mask.repeat_interleave(self.heads, dim=0)
+
+        # attention, what we cannot get enough of
+        if self._use_memory_efficient_attention_xformers:
+            hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
+            # Some versions of xformers return output in fp32, cast it back to the dtype of the input
+            hidden_states = hidden_states.to(query.dtype)
+        else:
+            if self._slice_size is None or query.shape[0] // self._slice_size == 1:
+                hidden_states = self._attention(query, key, value, attention_mask)
+            else:
+                hidden_states = self._sliced_attention(query, key, value, sequence_length, dim, attention_mask)
+
+        # linear proj
+        hidden_states = self.to_out[0](hidden_states)
+
+        # dropout
+        hidden_states = self.to_out[1](hidden_states)
+
+        # All frames
+        hidden_states = rearrange(hidden_states, "b (f d) c -> (b f) d c", f=video_length)
+        return hidden_states

models/controlnet.py

@@ -0,0 +1,605 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+import os
+import json
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from diffusers import ModelMixin
+from .controlnet_unet_blocks import (
+    CrossAttnDownBlock3D,
+    DownBlock3D,
+    UNetMidBlock3DCrossAttn,
+    get_down_block,
+)
+from .resnet import InflatedConv3d
+
+from diffusers.models.unet_2d_condition import UNet2DConditionModel
+from diffusers.models.cross_attention import AttnProcessor
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class ControlNetOutput(BaseOutput):
+    down_block_res_samples: Tuple[torch.Tensor]
+    mid_block_res_sample: torch.Tensor
+
+
+class ControlNetConditioningEmbedding(nn.Module):
+    """
+    Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
+    [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
+    training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
+    convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
+    (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
+    model) to encode image-space conditions ... into feature maps ..."
+    """
+
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        conditioning_channels: int = 3,
+        block_out_channels: Tuple[int] = (16, 32, 96, 256),
+    ):
+        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
+
+
+class ControlNetModel3D(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 4,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ),
+        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: Optional[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",
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int]] = (16, 32, 96, 256),
+    ):
+        super().__init__()
+
+        # Check inputs
+        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(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}."
+            )
+
+        # input
+        conv_in_kernel = 3
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = InflatedConv3d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        # time
+        time_embed_dim = 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]
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+        )
+
+        # 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)
+        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)
+        else:
+            self.class_embedding = None
+
+        # control net conditioning embedding
+        self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
+            conditioning_embedding_channels=block_out_channels[0],
+            block_out_channels=conditioning_embedding_out_channels,
+        )
+
+        self.down_blocks = nn.ModuleList([])
+        self.controlnet_down_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]
+
+        controlnet_block = InflatedConv3d(output_channel, output_channel, kernel_size=1)
+        controlnet_block = zero_module(controlnet_block)
+        self.controlnet_down_blocks.append(controlnet_block)
+
+        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,
+                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,
+            )
+            self.down_blocks.append(down_block)
+
+            for _ in range(layers_per_block):
+                controlnet_block = InflatedConv3d(output_channel, output_channel, kernel_size=1)
+                controlnet_block = zero_module(controlnet_block)
+                self.controlnet_down_blocks.append(controlnet_block)
+
+            if not is_final_block:
+                controlnet_block = InflatedConv3d(output_channel, output_channel, kernel_size=1)
+                controlnet_block = zero_module(controlnet_block)
+                self.controlnet_down_blocks.append(controlnet_block)
+
+        # mid
+        mid_block_channel = block_out_channels[-1]
+
+        controlnet_block = InflatedConv3d(mid_block_channel, mid_block_channel, kernel_size=1)
+        controlnet_block = zero_module(controlnet_block)
+        self.controlnet_mid_block = controlnet_block
+
+        # mid
+        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,
+        )
+
+    @classmethod
+    def from_unet(
+        cls,
+        unet: UNet2DConditionModel,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int]] = (16, 32, 96, 256),
+        load_weights_from_unet: bool = True,
+    ):
+        r"""
+        Instantiate Controlnet class from UNet2DConditionModel.
+
+        Parameters:
+            unet (`UNet2DConditionModel`):
+                UNet model which weights are copied to the ControlNet. Note that all configuration options are also
+                copied where applicable.
+        """
+        controlnet = cls(
+            in_channels=unet.config.in_channels,
+            flip_sin_to_cos=unet.config.flip_sin_to_cos,
+            freq_shift=unet.config.freq_shift,
+            down_block_types=unet.config.down_block_types,
+            only_cross_attention=unet.config.only_cross_attention,
+            block_out_channels=unet.config.block_out_channels,
+            layers_per_block=unet.config.layers_per_block,
+            downsample_padding=unet.config.downsample_padding,
+            mid_block_scale_factor=unet.config.mid_block_scale_factor,
+            act_fn=unet.config.act_fn,
+            norm_num_groups=unet.config.norm_num_groups,
+            norm_eps=unet.config.norm_eps,
+            cross_attention_dim=unet.config.cross_attention_dim,
+            attention_head_dim=unet.config.attention_head_dim,
+            use_linear_projection=unet.config.use_linear_projection,
+            class_embed_type=unet.config.class_embed_type,
+            num_class_embeds=unet.config.num_class_embeds,
+            upcast_attention=unet.config.upcast_attention,
+            resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
+            projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
+            controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
+            conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+        )
+
+        if load_weights_from_unet:
+            controlnet.conv_in.load_state_dict(unet.conv_in.state_dict())
+            controlnet.time_proj.load_state_dict(unet.time_proj.state_dict())
+            controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
+
+            if controlnet.class_embedding:
+                controlnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
+
+            controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict())
+            controlnet.mid_block.load_state_dict(unet.mid_block.state_dict())
+
+        return controlnet
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttnProcessor]:
+        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, AttnProcessor]):
+            if hasattr(module, "set_processor"):
+                processors[f"{name}.processor"] = module.processor
+
+            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
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttnProcessor, Dict[str, AttnProcessor]]):
+        r"""
+        Parameters:
+            `processor (`dict` of `AttnProcessor` or `AttnProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                of **all** `Attention` layers.
+            In case `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():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+    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"`, maximum 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_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 isinstance(module, (CrossAttnDownBlock3D, DownBlock3D)):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        controlnet_cond: torch.FloatTensor,
+        conditioning_scale: float = 1.0,
+        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,
+        return_dict: bool = True,
+    ) -> Union[ControlNetOutput, Tuple]:
+        # check channel order
+        channel_order = self.config.controlnet_conditioning_channel_order
+
+        if channel_order == "rgb":
+            # in rgb order by default
+            ...
+        elif channel_order == "bgr":
+            controlnet_cond = torch.flip(controlnet_cond, dims=[1])
+        else:
+            raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
+
+        # 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)
+
+        # 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=self.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+
+        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
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        controlnet_cond = self.controlnet_cond_embedding(controlnet_cond)
+
+        sample += controlnet_cond
+
+        # 3. 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,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=attention_mask,
+                cross_attention_kwargs=cross_attention_kwargs,
+            )
+
+        # 5. Control net blocks
+
+        controlnet_down_block_res_samples = ()
+
+        for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
+            down_block_res_sample = controlnet_block(down_block_res_sample)
+            controlnet_down_block_res_samples += (down_block_res_sample,)
+
+        down_block_res_samples = controlnet_down_block_res_samples
+
+        mid_block_res_sample = self.controlnet_mid_block(sample)
+
+        # 6. scaling
+        down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
+        mid_block_res_sample *= conditioning_scale
+
+        if not return_dict:
+            return (down_block_res_samples, mid_block_res_sample)
+
+        return ControlNetOutput(
+            down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
+        )
+
+    @classmethod
+    def from_pretrained_2d(cls, pretrained_model_path, control_path=None):
+        config_file = os.path.join(pretrained_model_path, 'config.json')
+        if not os.path.isfile(config_file):
+            raise RuntimeError(f"{config_file} does not exist")
+        with open(config_file, "r") as f:
+            config = json.load(f)
+        config["_class_name"] = cls.__name__
+        config["down_block_types"] = [
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D"
+        ]
+
+        from diffusers.utils import WEIGHTS_NAME
+        model = cls.from_config(config)
+        if control_path is None:
+            model_file = os.path.join(pretrained_model_path, WEIGHTS_NAME)
+            state_dict = torch.load(model_file, map_location="cpu")
+        else:
+            model_file = control_path
+            state_dict = torch.load(model_file, map_location="cpu")
+            state_dict = {k[14:]: state_dict[k] for k in state_dict.keys()}
+            
+        
+        for k, v in model.state_dict().items():
+            if '_temp.' in k:
+                state_dict.update({k: v})
+        model.load_state_dict(state_dict)
+
+        return model
+
+def zero_module(module):
+    for p in module.parameters():
+        nn.init.zeros_(p)
+    return module

models/controlnet_attention.py

@@ -0,0 +1,483 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from dataclasses import dataclass
+from typing import Optional, Callable
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn
+from positional_encodings.torch_encodings import PositionalEncoding2D
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers import ModelMixin
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from diffusers.models.attention import CrossAttention, FeedForward, AdaLayerNorm
+from einops import rearrange, repeat
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    @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,
+    ):
+        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(
+            [
+                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,
+                    upcast_attention=upcast_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)
+
+    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")
+        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 block in 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)
+
+
+class BasicTransformerBlock(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,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+
+        # Individual-Attn
+        self.attn1 = IndividualAttention(
+            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,
+        )
+        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 = CrossAttention(
+                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.norm_temp = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None):
+        if not is_xformers_available():
+            print("Here is how to install it")
+            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
+            self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+            if self.attn2 is not None:
+                self.attn2._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+            # self.attn_temp._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+
+    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, attention_mask=None, video_length=None):
+        # Individual-Attention
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
+        )
+
+        if self.only_cross_attention:
+            hidden_states = (
+                self.attn1(norm_hidden_states, encoder_hidden_states, attention_mask=attention_mask) + hidden_states
+            )
+        else:
+            hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask, video_length=video_length) + 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
+        # 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 IndividualAttention(nn.Module):
+    r"""
+    A cross attention layer.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        cross_attention_dim (`int`, *optional*):
+            The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
+        heads (`int`,  *optional*, defaults to 8): The number of heads to use for multi-head attention.
+        dim_head (`int`,  *optional*, defaults to 64): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        bias (`bool`, *optional*, defaults to False):
+            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
+    """
+
+    def __init__(
+        self,
+        query_dim: int,
+        cross_attention_dim: Optional[int] = None,
+        heads: int = 8,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias=False,
+        upcast_attention: bool = False,
+        upcast_softmax: bool = False,
+        added_kv_proj_dim: Optional[int] = None,
+        norm_num_groups: Optional[int] = None,
+    ):
+        super().__init__()
+        inner_dim = dim_head * heads
+        cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+
+        self.scale = dim_head**-0.5
+
+        self.heads = heads
+        # for slice_size > 0 the attention score computation
+        # is split across the batch axis to save memory
+        # You can set slice_size with `set_attention_slice`
+        self.sliceable_head_dim = heads
+        self._slice_size = None
+        self._use_memory_efficient_attention_xformers = False
+        self.added_kv_proj_dim = added_kv_proj_dim
+
+        if norm_num_groups is not None:
+            self.group_norm = nn.GroupNorm(num_channels=inner_dim, num_groups=norm_num_groups, eps=1e-5, affine=True)
+        else:
+            self.group_norm = None
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=bias)
+        self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
+        self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
+
+        if self.added_kv_proj_dim is not None:
+            self.add_k_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
+            self.add_v_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
+
+        self.to_out = nn.ModuleList([])
+        self.to_out.append(nn.Linear(inner_dim, query_dim))
+        self.to_out.append(nn.Dropout(dropout))
+
+    def reshape_heads_to_batch_dim(self, tensor):
+        batch_size, seq_len, dim = tensor.shape
+        head_size = self.heads
+        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
+        return tensor
+
+    def reshape_batch_dim_to_heads(self, tensor):
+        batch_size, seq_len, dim = tensor.shape
+        head_size = self.heads
+        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
+        return tensor
+
+    def set_attention_slice(self, slice_size):
+        if slice_size is not None and slice_size > self.sliceable_head_dim:
+            raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")
+
+        self._slice_size = slice_size
+    
+    def _attention(self, query, key, value, attention_mask=None):
+        if self.upcast_attention:
+            query = query.float()
+            key = key.float()
+
+        attention_scores = torch.baddbmm(
+            torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device),
+            query,
+            key.transpose(-1, -2),
+            beta=0,
+            alpha=self.scale,
+        )
+
+        if attention_mask is not None:
+            attention_scores = attention_scores + attention_mask
+
+        if self.upcast_softmax:
+            attention_scores = attention_scores.float()
+
+        attention_probs = attention_scores.softmax(dim=-1)
+
+        # cast back to the original dtype
+        attention_probs = attention_probs.to(value.dtype)
+
+        # compute attention output
+        hidden_states = torch.bmm(attention_probs, value)
+
+        # reshape hidden_states
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+    def _sliced_attention(self, query, key, value, sequence_length, dim, attention_mask):
+        batch_size_attention = query.shape[0]
+        hidden_states = torch.zeros(
+            (batch_size_attention, sequence_length, dim // self.heads), device=query.device, dtype=query.dtype
+        )
+        slice_size = self._slice_size if self._slice_size is not None else hidden_states.shape[0]
+        for i in range(hidden_states.shape[0] // slice_size):
+            start_idx = i * slice_size
+            end_idx = (i + 1) * slice_size
+
+            query_slice = query[start_idx:end_idx]
+            key_slice = key[start_idx:end_idx]
+
+            if self.upcast_attention:
+                query_slice = query_slice.float()
+                key_slice = key_slice.float()
+
+            attn_slice = torch.baddbmm(
+                torch.empty(slice_size, query.shape[1], key.shape[1], dtype=query_slice.dtype, device=query.device),
+                query_slice,
+                key_slice.transpose(-1, -2),
+                beta=0,
+                alpha=self.scale,
+            )
+
+            if attention_mask is not None:
+                attn_slice = attn_slice + attention_mask[start_idx:end_idx]
+
+            if self.upcast_softmax:
+                attn_slice = attn_slice.float()
+
+            attn_slice = attn_slice.softmax(dim=-1)
+
+            # cast back to the original dtype
+            attn_slice = attn_slice.to(value.dtype)
+            attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
+
+            hidden_states[start_idx:end_idx] = attn_slice
+
+        # reshape hidden_states
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+    def _memory_efficient_attention_xformers(self, query, key, value, attention_mask):
+        # TODO attention_mask
+        query = query.contiguous()
+        key = key.contiguous()
+        value = value.contiguous()
+        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, video_length=None):
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        encoder_hidden_states = encoder_hidden_states
+
+        if self.group_norm is not None:
+            hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = self.to_q(hidden_states)  # (bf) x d(hw) x c
+        dim = query.shape[-1]
+
+        query = self.reshape_heads_to_batch_dim(query)
+
+        if self.added_kv_proj_dim is not None:
+            raise NotImplementedError
+
+        encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
+        key = self.to_k(encoder_hidden_states)
+        value = self.to_v(encoder_hidden_states)
+
+        curr_frame_index = torch.arange(video_length)
+
+        key = rearrange(key, "(b f) d c -> b f d c", f=video_length)
+
+        key = key[:, curr_frame_index]
+        key = rearrange(key, "b f d c -> (b f) d c")
+
+        value = rearrange(value, "(b f) d c -> b f d c", f=video_length)
+
+        value = value[:, curr_frame_index]
+        value = rearrange(value, "b f d c -> (b f) d c")
+
+        key = self.reshape_heads_to_batch_dim(key)
+        value = self.reshape_heads_to_batch_dim(value)
+
+        if attention_mask is not None:
+            if attention_mask.shape[-1] != query.shape[1]:
+                target_length = query.shape[1]
+                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
+                attention_mask = attention_mask.repeat_interleave(self.heads, dim=0)
+
+        # attention, what we cannot get enough of
+        if self._use_memory_efficient_attention_xformers:
+            hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
+            # Some versions of xformers return output in fp32, cast it back to the dtype of the input
+            hidden_states = hidden_states.to(query.dtype)
+        else:
+            if self._slice_size is None or query.shape[0] // self._slice_size == 1:
+                hidden_states = self._attention(query, key, value, attention_mask)
+            else:
+                hidden_states = self._sliced_attention(query, key, value, sequence_length, dim, attention_mask)
+
+        # linear proj
+        hidden_states = self.to_out[0](hidden_states)
+
+        # dropout
+        hidden_states = self.to_out[1](hidden_states)
+        return hidden_states

models/controlnet_unet_blocks.py

@@ -0,0 +1,589 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+
+import torch
+from torch import nn
+
+from .controlnet_attention import Transformer3DModel
+from .resnet import Downsample3D, ResnetBlock3D, Upsample3D
+
+
+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",
+):
+    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,
+        )
+    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,
+        )
+    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",
+):
+    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,
+        )
+    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,
+        )
+    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,
+    ):
+        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,
+            )
+        ]
+        attentions = []
+
+        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,
+                )
+            )
+            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,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None, cross_attention_kwargs=None):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
+            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,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        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,
+                )
+            )
+            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,
+                )
+            )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        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,cross_attention_kwargs=None):
+        output_states = ()
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            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]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
+
+            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,
+    ):
+        super().__init__()
+        resnets = []
+
+        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,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        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):
+        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
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+            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,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        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,
+                )
+            )
+            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,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        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,
+        cross_attention_kwargs=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]
+            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 = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                )[0]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
+
+        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,
+    ):
+        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(
+                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,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        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):
+        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]
+            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
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states

models/pipeline_controlvideo.py

@@ -0,0 +1,1351 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import inspect
+import os
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+from dataclasses import dataclass
+
+import numpy as np
+import PIL.Image
+import torch
+from torch import nn
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from diffusers.models import AutoencoderKL
+from .controlnet import ControlNetOutput
+from diffusers import ModelMixin
+from diffusers.schedulers import DDIMScheduler
+from diffusers.utils import (
+    PIL_INTERPOLATION,
+    is_accelerate_available,
+    is_accelerate_version,
+    logging,
+    randn_tensor,
+    BaseOutput
+)
+from diffusers.pipeline_utils import DiffusionPipeline
+
+from einops import rearrange
+
+from .unet import UNet3DConditionModel
+from .controlnet import ControlNetModel3D
+from .RIFE.IFNet_HDv3 import IFNet
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class ControlVideoPipelineOutput(BaseOutput):
+    videos: Union[torch.Tensor, np.ndarray]
+
+
+class MultiControlNetModel3D(ModelMixin):
+    r"""
+    Multiple `ControlNetModel` wrapper class for Multi-ControlNet
+
+    This module is a wrapper for multiple instances of the `ControlNetModel`. The `forward()` API is designed to be
+    compatible with `ControlNetModel`.
+
+    Args:
+        controlnets (`List[ControlNetModel]`):
+            Provides additional conditioning to the unet during the denoising process. You must set multiple
+            `ControlNetModel` as a list.
+    """
+
+    def __init__(self, controlnets: Union[List[ControlNetModel3D], Tuple[ControlNetModel3D]]):
+        super().__init__()
+        self.nets = nn.ModuleList(controlnets)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        controlnet_cond: List[List[torch.tensor]],
+        conditioning_scale: List[float],
+        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,
+        return_dict: bool = True,
+    ) -> Union[ControlNetOutput, Tuple]:
+        for i, (image, scale, controlnet) in enumerate(zip(controlnet_cond, conditioning_scale, self.nets)):
+            down_samples, mid_sample = controlnet(
+                sample,
+                timestep,
+                encoder_hidden_states,
+                torch.cat(image, dim=0),
+                scale,
+                class_labels,
+                timestep_cond,
+                attention_mask,
+                cross_attention_kwargs,
+                return_dict,
+            )
+
+            # merge samples
+            if i == 0:
+                down_block_res_samples, mid_block_res_sample = down_samples, mid_sample
+            else:
+                down_block_res_samples = [
+                    samples_prev + samples_curr
+                    for samples_prev, samples_curr in zip(down_block_res_samples, down_samples)
+                ]
+                mid_block_res_sample += mid_sample
+
+        return down_block_res_samples, mid_block_res_sample
+
+
+class ControlVideoPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-video generation using Stable Diffusion with ControlNet guidance.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
+            Provides additional conditioning to the unet during the denoising process. If you set multiple ControlNets
+            as a list, the outputs from each ControlNet are added together to create one combined additional
+            conditioning.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPImageProcessor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+    _optional_components = ["safety_checker", "feature_extractor"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet3DConditionModel,
+        controlnet: Union[ControlNetModel3D, List[ControlNetModel3D], Tuple[ControlNetModel3D], MultiControlNetModel3D],
+        scheduler: DDIMScheduler,
+        interpolater: IFNet,
+    ):
+        super().__init__()
+
+        if isinstance(controlnet, (list, tuple)):
+            controlnet = MultiControlNetModel3D(controlnet)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            controlnet=controlnet,
+            scheduler=scheduler,
+            interpolater=interpolater,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae, controlnet, and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        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, self.controlnet]:
+            cpu_offload(cpu_offloaded_model, device)
+
+        if self.safety_checker is not None:
+            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
+
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        if self.safety_checker is not None:
+            # the safety checker can offload the vae again
+            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+        # control net hook has be manually offloaded as it alternates with unet
+        cpu_offload_with_hook(self.controlnet, device)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if 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
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+    def _encode_prompt(
+        self,
+        prompt,
+        device,
+        num_videos_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+             prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_videos_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+                Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+        """
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            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]
+                )
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+                )
+
+            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
+
+            prompt_embeds = self.text_encoder(
+                text_input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            prompt_embeds = prompt_embeds[0]
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            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 = prompt_embeds.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
+
+            negative_prompt_embeds = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            negative_prompt_embeds = negative_prompt_embeds[0]
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.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
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        return prompt_embeds
+
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+    def decode_latents(self, latents, return_tensor=False):
+        video_length = latents.shape[2]
+        latents = 1 / 0.18215 * latents
+        latents = rearrange(latents, "b c f h w -> (b f) c h w")
+        video = self.vae.decode(latents).sample
+        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
+        video = (video / 2 + 0.5).clamp(0, 1)
+        if return_tensor:
+            return video
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        video = video.cpu().float().numpy()
+        return video
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    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 check_inputs(
+        self,
+        prompt,
+        # image,
+        height,
+        width,
+        callback_steps,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        controlnet_conditioning_scale=1.0,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        # Check `image`
+
+        # if isinstance(self.controlnet, ControlNetModel):
+        #     self.check_image(image, prompt, prompt_embeds)
+        # elif isinstance(self.controlnet, MultiControlNetModel):
+        #     if not isinstance(image, list):
+        #         raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+        #     if len(image) != len(self.controlnet.nets):
+        #         raise ValueError(
+        #             "For multiple controlnets: `image` must have the same length as the number of controlnets."
+        #         )
+
+        #     for image_ in image:
+        #         self.check_image(image_, prompt, prompt_embeds)
+        # else:
+        #     assert False
+
+        # Check `controlnet_conditioning_scale`
+
+        if isinstance(self.controlnet, ControlNetModel3D):
+            if not isinstance(controlnet_conditioning_scale, float):
+                raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+        elif isinstance(self.controlnet, MultiControlNetModel3D):
+            if isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+                self.controlnet.nets
+            ):
+                raise ValueError(
+                    "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+                    " the same length as the number of controlnets"
+                )
+        else:
+            assert False
+
+    def check_image(self, image, prompt, prompt_embeds):
+        image_is_pil = isinstance(image, PIL.Image.Image)
+        image_is_tensor = isinstance(image, torch.Tensor)
+        image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+        image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+
+        if not image_is_pil and not image_is_tensor and not image_is_pil_list and not image_is_tensor_list:
+            raise TypeError(
+                "image must be passed and be one of PIL image, torch tensor, list of PIL images, or list of torch tensors"
+            )
+
+        if image_is_pil:
+            image_batch_size = 1
+        elif image_is_tensor:
+            image_batch_size = image.shape[0]
+        elif image_is_pil_list:
+            image_batch_size = len(image)
+        elif image_is_tensor_list:
+            image_batch_size = len(image)
+
+        if prompt is not None and isinstance(prompt, str):
+            prompt_batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            prompt_batch_size = len(prompt)
+        elif prompt_embeds is not None:
+            prompt_batch_size = prompt_embeds.shape[0]
+
+        if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+            raise ValueError(
+                f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+            )
+
+    def prepare_image(
+        self, image, width, height, batch_size, num_videos_per_prompt, device, dtype, do_classifier_free_guidance
+    ):
+        if not isinstance(image, torch.Tensor):
+            if isinstance(image, PIL.Image.Image):
+                image = [image]
+
+            if isinstance(image[0], PIL.Image.Image):
+                images = []
+
+                for image_ in image:
+                    image_ = image_.convert("RGB")
+                    image_ = image_.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])
+                    image_ = np.array(image_)
+                    image_ = image_[None, :]
+                    images.append(image_)
+
+                image = images
+
+                image = np.concatenate(image, axis=0)
+                image = np.array(image).astype(np.float32) / 255.0
+                image = image.transpose(0, 3, 1, 2)
+                image = torch.from_numpy(image)
+            elif isinstance(image[0], torch.Tensor):
+                image = torch.cat(image, dim=0)
+
+        image_batch_size = image.shape[0]
+
+        if image_batch_size == 1:
+            repeat_by = batch_size
+        else:
+            # image batch size is the same as prompt batch size
+            repeat_by = num_videos_per_prompt
+
+        image = image.repeat_interleave(repeat_by, dim=0)
+
+        image = image.to(device=device, dtype=dtype)
+
+        if do_classifier_free_guidance:
+            image = torch.cat([image] * 2)
+
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, video_length, height, width, dtype, \
+                    device, generator, latents=None, same_frame_noise=True):
+        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:
+            if same_frame_noise:
+                shape = (batch_size, num_channels_latents, 1, height // self.vae_scale_factor, width // self.vae_scale_factor)
+                latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+                latents = latents.repeat(1, 1, video_length, 1, 1)
+            else:
+                shape = (batch_size, num_channels_latents, video_length, height // self.vae_scale_factor, width // self.vae_scale_factor)
+                latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            shape = (batch_size, num_channels_latents, video_length, height // self.vae_scale_factor, width // self.vae_scale_factor)
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            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 _default_height_width(self, height, width, image):
+        # NOTE: It is possible that a list of images have different
+        # dimensions for each image, so just checking the first image
+        # is not _exactly_ correct, but it is simple.
+        while isinstance(image, list):
+            image = image[0]
+
+        if height is None:
+            if isinstance(image, PIL.Image.Image):
+                height = image.height
+            elif isinstance(image, torch.Tensor):
+                height = image.shape[3]
+
+            height = (height // 8) * 8  # round down to nearest multiple of 8
+
+        if width is None:
+            if isinstance(image, PIL.Image.Image):
+                width = image.width
+            elif isinstance(image, torch.Tensor):
+                width = image.shape[2]
+
+            width = (width // 8) * 8  # round down to nearest multiple of 8
+
+        return height, width
+
+    # override DiffusionPipeline
+    def save_pretrained(
+        self,
+        save_directory: Union[str, os.PathLike],
+        safe_serialization: bool = False,
+        variant: Optional[str] = None,
+    ):
+        if isinstance(self.controlnet, ControlNetModel3D):
+            super().save_pretrained(save_directory, safe_serialization, variant)
+        else:
+            raise NotImplementedError("Currently, the `save_pretrained()` is not implemented for Multi-ControlNet.")
+    
+    def get_alpha_prev(self, timestep):
+        prev_timestep = timestep - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps
+        alpha_prod_t_prev = self.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.scheduler.final_alpha_cumprod
+        return alpha_prod_t_prev
+
+    def get_slide_window_indices(self, video_length, window_size):
+        assert window_size >=3 
+        key_frame_indices = np.arange(0, video_length, window_size-1).tolist()
+
+        # Append last index
+        if key_frame_indices[-1] != (video_length-1):
+            key_frame_indices.append(video_length-1)
+        
+        slices = np.split(np.arange(video_length), key_frame_indices)
+        inter_frame_list = []
+        for s in slices:
+            if len(s) < 2:
+                continue
+            inter_frame_list.append(s[1:].tolist())
+        return key_frame_indices, inter_frame_list
+        
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        video_length: Optional[int] = 1,
+        frames: Union[List[torch.FloatTensor], List[PIL.Image.Image], List[List[torch.FloatTensor]], List[List[PIL.Image.Image]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_videos_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "tensor",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+        smooth_steps: List = [19, 20],
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            frames (`List[torch.FloatTensor]`, `List[PIL.Image.Image]`,
+                    `List[List[torch.FloatTensor]]`, or `List[List[PIL.Image.Image]]`):
+                The ControlVideo input condition. ControlVideo uses this input condition to generate guidance to Unet. If
+                the type is specified as `Torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can
+                also be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If
+                height and/or width are passed, `image` is resized according to them. If multiple ControlNets are
+                specified in init, images must be passed as a list such that each element of the list can be correctly
+                batched for input to a single controlnet.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+                Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added
+                to the residual in the original unet. If multiple ControlNets are specified in init, you can set the
+                corresponding scale as a list.
+            smooth_steps (`List[int]`):
+                Perform smoother on predicted RGB frames at these timesteps. 
+
+        Examples:
+
+        Returns:
+            [`ControlVideoPipelineOutput`] or `tuple`:
+            [`ControlVideoPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+        # 0. Default height and width to unet
+        height, width = self._default_height_width(height, width, frames)
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            controlnet_conditioning_scale,
+        )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        if isinstance(self.controlnet, MultiControlNetModel3D) and isinstance(controlnet_conditioning_scale, float):
+            controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(self.controlnet.nets)
+
+        # 3. Encode input prompt
+        prompt_embeds = self._encode_prompt(
+            prompt,
+            device,
+            num_videos_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+        )
+
+        # 4. Prepare image
+        if isinstance(self.controlnet, ControlNetModel3D):
+            images = []
+            for i_img in frames:
+                i_img = self.prepare_image(
+                    image=i_img,
+                    width=width,
+                    height=height,
+                    batch_size=batch_size * num_videos_per_prompt,
+                    num_videos_per_prompt=num_videos_per_prompt,
+                    device=device,
+                    dtype=self.controlnet.dtype,
+                    do_classifier_free_guidance=do_classifier_free_guidance,
+                )
+                images.append(i_img)
+            frames = torch.stack(images, dim=2)  # b x c x f x h x w
+        elif isinstance(self.controlnet, MultiControlNetModel3D):
+            images = []
+            for i_img in frames:
+                i_images = []
+                for ii_img in i_img:
+                    ii_img = self.prepare_image(
+                        image=ii_img,
+                        width=width,
+                        height=height,
+                        batch_size=batch_size * num_videos_per_prompt,
+                        num_videos_per_prompt=num_videos_per_prompt,
+                        device=device,
+                        dtype=self.controlnet.dtype,
+                        do_classifier_free_guidance=do_classifier_free_guidance,
+                    )
+
+                    i_images.append(ii_img)
+                images.append(torch.stack(i_images, dim=2))
+            frames = images
+        else:
+            assert False
+
+        # 5. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 6. Prepare latent variables
+        num_channels_latents = self.unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            video_length,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+            same_frame_noise=True,
+        )
+
+        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+
+        # Prepare video indices if performing smoothing
+        if len(smooth_steps) > 0:
+            video_indices = np.arange(video_length)
+            zero_indices = video_indices[0::2]
+            one_indices = video_indices[1::2]
+
+        # 8. Denoising loop
+        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):
+                torch.cuda.empty_cache()
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # controlnet(s) inference
+                down_block_res_samples, mid_block_res_sample = self.controlnet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    controlnet_cond=frames,
+                    conditioning_scale=controlnet_conditioning_scale,
+                    return_dict=False,
+                )
+                # predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    down_block_additional_residuals=down_block_res_samples,
+                    mid_block_additional_residual=mid_block_res_sample,
+                ).sample
+
+                # perform guidance
+                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)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                step_dict = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs)
+                latents = step_dict.prev_sample
+                pred_original_sample = step_dict.pred_original_sample
+                
+                # Smooth videos
+                if (num_inference_steps - i) in smooth_steps:
+                    pred_video = self.decode_latents(pred_original_sample, return_tensor=True)  # b c f h w
+                    pred_video = rearrange(pred_video, "b c f h w -> b f c h w")
+                    for b_i in range(len(pred_video)):
+                        if i % 2 == 0:
+                            for v_i in range(len(zero_indices)-1):
+                                s_frame = pred_video[b_i][zero_indices[v_i]].unsqueeze(0)
+                                e_frame = pred_video[b_i][zero_indices[v_i+1]].unsqueeze(0)
+                                pred_video[b_i][one_indices[v_i]] = self.interpolater.inference(s_frame, e_frame)[0]
+                        else:
+                            if video_length % 2 == 1:
+                                tmp_one_indices = [0] + one_indices.tolist() + [video_length-1]
+                            else:
+                                tmp_one_indices = [0] + one_indices.tolist()
+
+                            for v_i in range(len(tmp_one_indices)-1):
+                                s_frame = pred_video[b_i][tmp_one_indices[v_i]].unsqueeze(0)
+                                e_frame = pred_video[b_i][tmp_one_indices[v_i+1]].unsqueeze(0)
+                                pred_video[b_i][zero_indices[v_i]] = self.interpolater.inference(s_frame, e_frame)[0]
+                    pred_video = rearrange(pred_video, "b f c h w -> (b f) c h w")
+                    pred_video = 2.0 * pred_video - 1.0
+                    # ori_pred_original_sample = pred_original_sample
+                    pred_original_sample = self.vae.encode(pred_video).latent_dist.sample(generator)
+                    pred_original_sample *= self.vae.config.scaling_factor
+                    pred_original_sample = rearrange(pred_original_sample, "(b f) c h w -> b c f h w", f=video_length)
+                    
+                    # predict xt-1 with smoothed x0
+                    alpha_prod_t_prev =self.get_alpha_prev(t)
+                    # preserve more details
+                    # pred_original_sample = ori_pred_original_sample * alpha_prod_t_prev + (1 - alpha_prod_t_prev) * pred_original_sample
+                    # compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+                    pred_sample_direction = (1 - alpha_prod_t_prev) ** (0.5) * noise_pred
+                    # compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+                    latents = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+                    
+                    
+                # call the callback, if provided
+                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:
+                        callback(i, t, latents)
+
+        # If we do sequential model offloading, let's offload unet and controlnet
+        # manually for max memory savings
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.unet.to("cpu")
+            self.controlnet.to("cpu")
+            torch.cuda.empty_cache()
+        # Post-processing
+        video = self.decode_latents(latents)
+
+        # Convert to tensor
+        if output_type == "tensor":
+            video = torch.from_numpy(video)
+
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.final_offload_hook.offload()
+
+        if not return_dict:
+            return video
+
+        return ControlVideoPipelineOutput(videos=video)
+
+    @torch.no_grad()
+    def generate_long_video(
+        self,
+        prompt: Union[str, List[str]] = None,
+        video_length: Optional[int] = 1,
+        frames: Union[List[torch.FloatTensor], List[PIL.Image.Image], List[List[torch.FloatTensor]], List[List[PIL.Image.Image]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_videos_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "tensor",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+        smooth_steps: List = [19, 20],
+        window_size: int = 8,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            frames (`List[torch.FloatTensor]`, `List[PIL.Image.Image]`,
+                    `List[List[torch.FloatTensor]]`, or `List[List[PIL.Image.Image]]`):
+                The ControlVideo input condition. ControlVideo uses this input condition to generate guidance to Unet. If
+                the type is specified as `Torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can
+                also be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If
+                height and/or width are passed, `image` is resized according to them. If multiple ControlNets are
+                specified in init, images must be passed as a list such that each element of the list can be correctly
+                batched for input to a single controlnet.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+                Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added
+                to the residual in the original unet. If multiple ControlNets are specified in init, you can set the
+                corresponding scale as a list.
+            smooth_steps (`List[int]`):
+                Perform smoother on predicted RGB frames at these timesteps. 
+            window_size ('int'):
+                The length of each short clip.
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+        # 0. Default height and width to unet
+        height, width = self._default_height_width(height, width, frames)
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            controlnet_conditioning_scale,
+        )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        if isinstance(self.controlnet, MultiControlNetModel3D) and isinstance(controlnet_conditioning_scale, float):
+            controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(self.controlnet.nets)
+
+        # 3. Encode input prompt
+        prompt_embeds = self._encode_prompt(
+            prompt,
+            device,
+            num_videos_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+        )
+
+        # 4. Prepare image
+        if isinstance(self.controlnet, ControlNetModel3D):
+            images = []
+            for i_img in frames:
+                i_img = self.prepare_image(
+                    image=i_img,
+                    width=width,
+                    height=height,
+                    batch_size=batch_size * num_videos_per_prompt,
+                    num_videos_per_prompt=num_videos_per_prompt,
+                    device=device,
+                    dtype=self.controlnet.dtype,
+                    do_classifier_free_guidance=do_classifier_free_guidance,
+                )
+                images.append(i_img)
+            frames = torch.stack(images, dim=2)  # b x c x f x h x w
+        elif isinstance(self.controlnet, MultiControlNetModel3D):
+            images = []
+            for i_img in frames:
+                i_images = []
+                for ii_img in i_img:
+                    ii_img = self.prepare_image(
+                        image=ii_img,
+                        width=width,
+                        height=height,
+                        batch_size=batch_size * num_videos_per_prompt,
+                        num_videos_per_prompt=num_videos_per_prompt,
+                        device=device,
+                        dtype=self.controlnet.dtype,
+                        do_classifier_free_guidance=do_classifier_free_guidance,
+                    )
+
+                    i_images.append(ii_img)
+                images.append(torch.stack(i_images, dim=2))
+            frames = images
+        else:
+            assert False
+
+        # 5. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 6. Prepare latent variables
+        num_channels_latents = self.unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            video_length,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+            same_frame_noise=True,
+        )
+
+        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # Prepare indices of key frames and interval frames
+        key_frame_indices, inter_frame_list = self.get_slide_window_indices(video_length, window_size)
+
+        # Prepare video indices if performing smoothing
+        if len(smooth_steps) > 0:
+            video_indices = np.arange(video_length)
+            zero_indices = video_indices[0::2]
+            one_indices = video_indices[1::2]
+
+        # 8. Denoising loop
+        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):
+                torch.cuda.empty_cache()
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+                noise_pred = torch.zeros_like(latents)
+                pred_original_sample = torch.zeros_like(latents)
+                
+                # 8.1 Key frames
+                # controlnet(s) inference
+                key_down_block_res_samples, key_mid_block_res_sample = self.controlnet(
+                    latent_model_input[:, :, key_frame_indices],
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    controlnet_cond=frames[:, :, key_frame_indices],
+                    conditioning_scale=controlnet_conditioning_scale,
+                    return_dict=False,
+                )
+                # predict the noise residual
+                key_noise_pred = self.unet(
+                    latent_model_input[:, :, key_frame_indices],
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    down_block_additional_residuals=key_down_block_res_samples,
+                    mid_block_additional_residual=key_mid_block_res_sample,
+                    inter_frame=False,
+                ).sample
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = key_noise_pred.chunk(2)
+                    noise_pred[:, :, key_frame_indices] = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                key_step_dict = self.scheduler.step(noise_pred[:, :, key_frame_indices], t, latents[:, :, key_frame_indices], **extra_step_kwargs)
+                latents[:, :, key_frame_indices] = key_step_dict.prev_sample
+                pred_original_sample[:, :, key_frame_indices] = key_step_dict.pred_original_sample
+
+                # 8.2 compute interval frames
+                for f_i, frame_ids in enumerate(inter_frame_list):
+                    input_frame_ids = key_frame_indices[f_i:f_i+2] + frame_ids
+                    # controlnet(s) inference
+                    inter_down_block_res_samples, inter_mid_block_res_sample = self.controlnet(
+                        latent_model_input[:, :, input_frame_ids],
+                        t,
+                        encoder_hidden_states=prompt_embeds,
+                        controlnet_cond=frames[:, :, input_frame_ids],
+                        conditioning_scale=controlnet_conditioning_scale,
+                        return_dict=False,
+                    )
+                    # predict the noise residual
+                    inter_noise_pred = self.unet(
+                        latent_model_input[:, :, input_frame_ids],
+                        t,
+                        encoder_hidden_states=prompt_embeds,
+                        cross_attention_kwargs=cross_attention_kwargs,
+                        down_block_additional_residuals=inter_down_block_res_samples,
+                        mid_block_additional_residual=inter_mid_block_res_sample,
+                        inter_frame=True,
+                    ).sample
+
+                    # perform guidance
+                    if do_classifier_free_guidance:
+                        noise_pred_uncond, noise_pred_text = inter_noise_pred[:, :, 2:].chunk(2)
+                        noise_pred[:, :, frame_ids] = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                    # compute the previous noisy sample x_t -> x_t-1
+                    step_dict = self.scheduler.step(noise_pred[:, :, frame_ids], t, latents[:, :, frame_ids], **extra_step_kwargs)
+                    latents[:, :, frame_ids] = step_dict.prev_sample
+                    pred_original_sample[:, :, frame_ids] = step_dict.pred_original_sample
+                
+                # Smooth videos
+                if (num_inference_steps - i) in smooth_steps:
+                    pred_video = self.decode_latents(pred_original_sample, return_tensor=True)  # b c f h w
+                    pred_video = rearrange(pred_video, "b c f h w -> b f c h w")
+                    for b_i in range(len(pred_video)):
+                        if i % 2 == 0:
+                            for v_i in range(len(zero_indices)-1):
+                                s_frame = pred_video[b_i][zero_indices[v_i]].unsqueeze(0)
+                                e_frame = pred_video[b_i][zero_indices[v_i+1]].unsqueeze(0)
+                                pred_video[b_i][one_indices[v_i]] = self.interpolater.inference(s_frame, e_frame)[0]
+                        else:
+                            if video_length % 2 == 1:
+                                tmp_one_indices = [0] + one_indices.tolist() + [video_length-1]
+                            else:
+                                tmp_one_indices = [0] + one_indices.tolist()
+                            for v_i in range(len(tmp_one_indices)-1):
+                                s_frame = pred_video[b_i][tmp_one_indices[v_i]].unsqueeze(0)
+                                e_frame = pred_video[b_i][tmp_one_indices[v_i+1]].unsqueeze(0)
+                                pred_video[b_i][zero_indices[v_i]] = self.interpolater.inference(s_frame, e_frame)[0]
+                    pred_video = rearrange(pred_video, "b f c h w -> (b f) c h w")
+                    pred_video = 2.0 * pred_video - 1.0
+                    for v_i in range(len(pred_video)):
+                        pred_original_sample[:, :, v_i] = self.vae.encode(pred_video[v_i:v_i+1]).latent_dist.sample(generator)
+                        pred_original_sample[:, :, v_i] *= self.vae.config.scaling_factor
+
+                    
+                    # predict xt-1 with smoothed x0
+                    alpha_prod_t_prev =self.get_alpha_prev(t)
+                    # preserve more details
+                    pred_sample_direction = (1 - alpha_prod_t_prev) ** (0.5) * noise_pred
+                    # compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+                    latents = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+                    
+                    
+                # call the callback, if provided
+                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:
+                        callback(i, t, latents)
+
+        # If we do sequential model offloading, let's offload unet and controlnet
+        # manually for max memory savings
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.unet.to("cpu")
+            self.controlnet.to("cpu")
+            torch.cuda.empty_cache()
+        # Post-processing
+        video = self.decode_latents(latents)
+
+        # Convert to tensor
+        if output_type == "tensor":
+            video = torch.from_numpy(video)
+
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.final_offload_hook.offload()
+
+        if not return_dict:
+            return video
+
+        return ControlVideoPipelineOutput(videos=video)

models/resnet.py

@@ -0,0 +1,217 @@
+# 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 TemporalConv1d(nn.Conv1d):
+    def forward(self, x):
+        b, c, f, h, w = x.shape
+        y = rearrange(x.clone(), "b c f h w -> (b h w) c f")
+        y = super().forward(y)
+        y = rearrange(y, "(b h w) c f -> b c f h w", b=b, h=h, w=w)
+        return y
+
+
+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:
+            conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
+
+        if name == "conv":
+            self.conv = conv
+        else:
+            self.Conv2d_0 = conv
+
+    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)
+
+        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:
+            conv = InflatedConv3d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
+        else:
+            raise NotImplementedError
+
+        if name == "conv":
+            self.Conv2d_0 = conv
+            self.conv = conv
+        elif name == "Conv2d_0":
+            self.conv = conv
+        else:
+            self.conv = conv
+
+    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,
+    ):
+        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
+
+        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
+
+        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))[:, :, None, None, None]
+
+        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))

models/unet.py

@@ -0,0 +1,472 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import os
+import json
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers import ModelMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from .unet_blocks import (
+    CrossAttnDownBlock3D,
+    CrossAttnUpBlock3D,
+    DownBlock3D,
+    UNetMidBlock3DCrossAttn,
+    UpBlock3D,
+    get_down_block,
+    get_up_block,
+)
+from .resnet import InflatedConv3d
+
+
+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",
+    ):
+        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):
+            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,
+            )
+            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,
+            )
+        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):
+            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,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        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)
+
+    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 isinstance(module, (CrossAttnDownBlock3D, DownBlock3D, CrossAttnUpBlock3D, UpBlock3D)):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        cross_attention_kwargs = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        inter_frame = False,
+    ) -> 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)
+
+        # 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,
+                    inter_frame=inter_frame
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+            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_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,
+            inter_frame=inter_frame
+
+        )
+
+        if mid_block_additional_residual is not None:
+            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,
+                    inter_frame=inter_frame
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
+                )
+        # 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, subfolder=None):
+        if subfolder is not None:
+            pretrained_model_path = os.path.join(pretrained_model_path, subfolder)
+
+        config_file = os.path.join(pretrained_model_path, 'config.json')
+        if not os.path.isfile(config_file):
+            raise RuntimeError(f"{config_file} does not exist")
+        with open(config_file, "r") as f:
+            config = json.load(f)
+        config["_class_name"] = cls.__name__
+        config["down_block_types"] = [
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D"
+        ]
+        config["up_block_types"] = [
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D"
+        ]
+
+        from diffusers.utils import WEIGHTS_NAME
+        model = cls.from_config(config)
+        model_file = os.path.join(pretrained_model_path, WEIGHTS_NAME)
+        if not os.path.isfile(model_file):
+            raise RuntimeError(f"{model_file} does not exist")
+        state_dict = torch.load(model_file, map_location="cpu")
+        # for k, v in model.state_dict().items():
+        #     if '_temp.' in k:
+        #         state_dict.update({k: v})
+        model.load_state_dict(state_dict, strict=False)
+
+        return model

models/unet_blocks.py

@@ -0,0 +1,588 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+
+import torch
+from torch import nn
+
+from .attention import Transformer3DModel
+from .resnet import Downsample3D, ResnetBlock3D, Upsample3D
+
+
+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",
+):
+    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,
+        )
+    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,
+        )
+    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",
+):
+    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,
+        )
+    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,
+        )
+    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,
+    ):
+        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,
+            )
+        ]
+        attentions = []
+
+        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,
+                )
+            )
+            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,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None, inter_frame=False):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states, inter_frame=inter_frame).sample
+            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,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        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,
+                )
+            )
+            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,
+                )
+            )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        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, inter_frame=False):
+        output_states = ()
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None, inter_frame=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict, inter_frame=inter_frame)
+                        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, inter_frame=inter_frame),
+                    hidden_states,
+                    encoder_hidden_states,
+                )[0]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states, inter_frame=inter_frame).sample
+
+            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,
+    ):
+        super().__init__()
+        resnets = []
+
+        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,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        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):
+        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
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+            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,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        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,
+                )
+            )
+            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,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        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, 
+        inter_frame=False
+    ):
+        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]
+            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, inter_frame=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict, inter_frame=inter_frame)
+                        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, inter_frame=inter_frame),
+                    hidden_states,
+                    encoder_hidden_states,
+                )[0]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states, inter_frame=inter_frame).sample
+
+        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,
+    ):
+        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(
+                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,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        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):
+        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]
+            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
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states

models/util.py

@@ -0,0 +1,122 @@
+import os
+import imageio
+import numpy as np
+from typing import Union
+import decord
+decord.bridge.set_bridge('torch')
+import torch
+import torchvision
+import PIL
+from typing import List
+from tqdm import tqdm
+from einops import rearrange
+
+from controlnet_aux import CannyDetector
+
+def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=4, fps=8):
+    videos = rearrange(videos, "b c t h w -> t b c h w")
+    outputs = []
+    for x in videos:
+        x = torchvision.utils.make_grid(x, nrow=n_rows)
+        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+        if rescale:
+            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
+        x = (x * 255).numpy().astype(np.uint8)
+        outputs.append(x)
+
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    imageio.mimsave(path, outputs, fps=fps)
+
+def save_videos_grid_pil(videos: List[PIL.Image.Image], path: str, rescale=False, n_rows=4, fps=8):
+    videos = rearrange(videos, "b c t h w -> t b c h w")
+    outputs = []
+    for x in videos:
+        x = torchvision.utils.make_grid(x, nrow=n_rows)
+        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+        if rescale:
+            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
+        x = (x * 255).numpy().astype(np.uint8)
+        outputs.append(x)
+
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    imageio.mimsave(path, outputs, fps=fps)
+
+def read_video(video_path, video_length, width=512, height=512, frame_rate=None):
+    vr = decord.VideoReader(video_path, width=width, height=height)
+    if frame_rate is None:
+        frame_rate = max(1, len(vr) // video_length)
+    sample_index = list(range(0, len(vr), frame_rate))[:video_length]
+    video = vr.get_batch(sample_index)
+    video = rearrange(video, "f h w c -> f c h w")
+    video = (video / 127.5 - 1.0)
+    return video
+
+
+def get_annotation(video, annotator):
+    t2i_transform = torchvision.transforms.ToPILImage()
+    annotation = []
+    for frame in video:
+        pil_frame = t2i_transform(frame)
+        if isinstance(annotator, CannyDetector):
+            annotation.append(annotator(pil_frame, low_threshold=100, high_threshold=200))
+        else:
+            annotation.append(annotator(pil_frame))
+    return annotation
+
+# DDIM Inversion
+@torch.no_grad()
+def init_prompt(prompt, pipeline):
+    uncond_input = pipeline.tokenizer(
+        [""], padding="max_length", max_length=pipeline.tokenizer.model_max_length,
+        return_tensors="pt"
+    )
+    uncond_embeddings = pipeline.text_encoder(uncond_input.input_ids.to(pipeline.device))[0]
+    text_input = pipeline.tokenizer(
+        [prompt],
+        padding="max_length",
+        max_length=pipeline.tokenizer.model_max_length,
+        truncation=True,
+        return_tensors="pt",
+    )
+    text_embeddings = pipeline.text_encoder(text_input.input_ids.to(pipeline.device))[0]
+    context = torch.cat([uncond_embeddings, text_embeddings])
+
+    return context
+
+
+def next_step(model_output: Union[torch.FloatTensor, np.ndarray], timestep: int,
+              sample: Union[torch.FloatTensor, np.ndarray], ddim_scheduler):
+    timestep, next_timestep = min(
+        timestep - ddim_scheduler.config.num_train_timesteps // ddim_scheduler.num_inference_steps, 999), timestep
+    alpha_prod_t = ddim_scheduler.alphas_cumprod[timestep] if timestep >= 0 else ddim_scheduler.final_alpha_cumprod
+    alpha_prod_t_next = ddim_scheduler.alphas_cumprod[next_timestep]
+    beta_prod_t = 1 - alpha_prod_t
+    next_original_sample = (sample - beta_prod_t ** 0.5 * model_output) / alpha_prod_t ** 0.5
+    next_sample_direction = (1 - alpha_prod_t_next) ** 0.5 * model_output
+    next_sample = alpha_prod_t_next ** 0.5 * next_original_sample + next_sample_direction
+    return next_sample
+
+
+def get_noise_pred_single(latents, t, context, unet):
+    noise_pred = unet(latents, t, encoder_hidden_states=context)["sample"]
+    return noise_pred
+
+
+@torch.no_grad()
+def ddim_loop(pipeline, ddim_scheduler, latent, num_inv_steps, prompt):
+    context = init_prompt(prompt, pipeline)
+    uncond_embeddings, cond_embeddings = context.chunk(2)
+    all_latent = [latent]
+    latent = latent.clone().detach()
+    for i in tqdm(range(num_inv_steps)):
+        t = ddim_scheduler.timesteps[len(ddim_scheduler.timesteps) - i - 1]
+        noise_pred = get_noise_pred_single(latent, t, cond_embeddings, pipeline.unet)
+        latent = next_step(noise_pred, t, latent, ddim_scheduler)
+        all_latent.append(latent)
+    return all_latent
+
+
+@torch.no_grad()
+def ddim_inversion(pipeline, ddim_scheduler, video_latent, num_inv_steps, prompt=""):
+    ddim_latents = ddim_loop(pipeline, ddim_scheduler, video_latent, num_inv_steps, prompt)
+    return ddim_latents

predict.py

@@ -0,0 +1,164 @@
+# Prediction interface for Cog ⚙️
+# https://github.com/replicate/cog/blob/main/docs/python.md
+import os
+import numpy as np
+import argparse
+import imageio
+import torch
+
+from einops import rearrange
+from diffusers import DDIMScheduler, AutoencoderKL
+from transformers import CLIPTextModel, CLIPTokenizer
+import controlnet_aux
+from controlnet_aux import OpenposeDetector, CannyDetector, MidasDetector
+
+from models.pipeline_controlvideo import ControlVideoPipeline
+from models.util import save_videos_grid, read_video, get_annotation
+from models.unet import UNet3DConditionModel
+from models.controlnet import ControlNetModel3D
+from models.RIFE.IFNet_HDv3 import IFNet
+from cog import BasePredictor, Input, Path
+
+
+sd_path = "checkpoints/stable-diffusion-v1-5"
+inter_path = "checkpoints/flownet.pkl"
+controlnet_dict = {
+    "pose": "checkpoints/sd-controlnet-openpose",
+    "depth": "checkpoints/sd-controlnet-depth",
+    "canny": "checkpoints/sd-controlnet-canny",
+}
+
+controlnet_parser_dict = {
+    "pose": OpenposeDetector,
+    "depth": MidasDetector,
+    "canny": CannyDetector,
+}
+
+POS_PROMPT = " ,best quality, extremely detailed, HD, ultra-realistic, 8K, HQ, masterpiece, trending on artstation, art, smooth"
+NEG_PROMPT = "longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic"
+
+
+class Predictor(BasePredictor):
+    def setup(self):
+        """Load the model into memory to make running multiple predictions efficient"""
+
+        self.tokenizer = CLIPTokenizer.from_pretrained(sd_path, subfolder="tokenizer")
+        self.text_encoder = CLIPTextModel.from_pretrained(
+            sd_path, subfolder="text_encoder"
+        ).to(dtype=torch.float16)
+        self.vae = AutoencoderKL.from_pretrained(sd_path, subfolder="vae").to(
+            dtype=torch.float16
+        )
+        self.unet = UNet3DConditionModel.from_pretrained_2d(
+            sd_path, subfolder="unet"
+        ).to(dtype=torch.float16)
+        self.interpolater = IFNet(ckpt_path=inter_path).to(dtype=torch.float16)
+        self.scheduler = DDIMScheduler.from_pretrained(sd_path, subfolder="scheduler")
+        self.controlnet = {
+            k: ControlNetModel3D.from_pretrained_2d(controlnet_dict[k]).to(
+                dtype=torch.float16
+            )
+            for k in ["depth", "canny", "pose"]
+        }
+        self.annotator = {k: controlnet_parser_dict[k]() for k in ["depth", "canny"]}
+        self.annotator["pose"] = controlnet_parser_dict["pose"].from_pretrained(
+            "lllyasviel/ControlNet", cache_dir="checkpoints"
+        )
+
+    def predict(
+        self,
+        prompt: str = Input(
+            description="Text description of target video",
+            default="A striking mallard floats effortlessly on the sparkling pond.",
+        ),
+        video_path: Path = Input(description="source video"),
+        condition: str = Input(
+            default="depth",
+            choices=["depth", "canny", "pose"],
+            description="Condition of structure sequence",
+        ),
+        video_length: int = Input(
+            default=15, description="Length of synthesized video"
+        ),
+        smoother_steps: str = Input(
+            default="19, 20",
+            description="Timesteps at which using interleaved-frame smoother, separate with comma",
+        ),
+        is_long_video: bool = Input(
+            default=False,
+            description="Whether to use hierarchical sampler to produce long video",
+        ),
+        num_inference_steps: int = Input(
+            description="Number of denoising steps", default=50
+        ),
+        guidance_scale: float = Input(
+            description="Scale for classifier-free guidance", ge=1, le=20, default=12.5
+        ),
+        seed: str = Input(
+            default=None, description="Random seed. Leave blank to randomize the seed"
+        ),
+    ) -> Path:
+        """Run a single prediction on the model"""
+        if seed is None:
+            seed = int.from_bytes(os.urandom(2), "big")
+        else:
+            seed = int(seed)
+        print(f"Using seed: {seed}")
+
+        generator = torch.Generator(device="cuda")
+        generator.manual_seed(seed)
+
+        pipe = ControlVideoPipeline(
+            vae=self.vae,
+            text_encoder=self.text_encoder,
+            tokenizer=self.tokenizer,
+            unet=self.unet,
+            controlnet=self.controlnet[condition],
+            interpolater=self.interpolater,
+            scheduler=self.scheduler,
+        )
+
+        pipe.enable_vae_slicing()
+        pipe.enable_xformers_memory_efficient_attention()
+        pipe.to("cuda")
+
+        # Step 1. Read a video
+        video = read_video(video_path=str(video_path), video_length=video_length)
+
+        # Step 2. Parse a video to conditional frames
+        pil_annotation = get_annotation(video, self.annotator[condition])
+
+        # Step 3. inference
+        smoother_steps = [int(s) for s in smoother_steps.split(",")]
+
+        if is_long_video:
+            window_size = int(np.sqrt(video_length))
+            sample = pipe.generate_long_video(
+                prompt + POS_PROMPT,
+                video_length=video_length,
+                frames=pil_annotation,
+                num_inference_steps=num_inference_steps,
+                smooth_steps=smoother_steps,
+                window_size=window_size,
+                generator=generator,
+                guidance_scale=guidance_scale,
+                negative_prompt=NEG_PROMPT,
+            ).videos
+        else:
+            sample = pipe(
+                prompt + POS_PROMPT,
+                video_length=video_length,
+                frames=pil_annotation,
+                num_inference_steps=num_inference_steps,
+                smooth_steps=smoother_steps,
+                generator=generator,
+                guidance_scale=guidance_scale,
+                negative_prompt=NEG_PROMPT,
+            ).videos
+
+        out_path = "/tmp/out.mp4"
+        save_videos_grid(sample, out_path)
+        del pipe
+        torch.cuda.empty_cache()
+
+        return Path(out_path)

requirements.txt

@@ -0,0 +1,38 @@
+accelerate==0.17.1
+addict==2.4.0
+basicsr==1.4.2
+bitsandbytes==0.35.4
+clip==1.0
+cmake==3.25.2
+controlnet-aux==0.0.6
+decord==0.6.0
+deepspeed==0.8.0
+diffusers==0.14.0
+easydict==1.10
+einops==0.6.0
+ffmpy==0.3.0
+ftfy==6.1.1
+imageio==2.25.1
+imageio-ffmpeg==0.4.8
+moviepy==1.0.3
+numpy==1.24.2
+omegaconf==2.3.0
+opencv-python==4.7.0.68
+pandas==1.5.3
+pillow==9.4.0
+scikit-image==0.19.3
+scipy==1.10.1
+tensorboard==2.12.0
+tensorboard-data-server==0.7.0
+tensorboard-plugin-wit==1.8.1
+termcolor==2.2.0
+thinc==8.1.10
+timm==0.6.12
+tokenizers==0.13.2
+torch==1.13.1+cu116
+torchvision==0.14.1+cu116
+tqdm==4.64.1
+transformers==4.26.1
+wandb==0.13.10
+xformers==0.0.16
+modelcards