Init

README.md

@@ -1,14 +1,12 @@
 ---
-title: Conceptrol
-emoji: 🖼
-colorFrom: purple
+title: PAID
+emoji: 🏢
+colorFrom: pink
 colorTo: red
 sdk: gradio
-sdk_version: 5.0.1
+sdk_version: 4.22.0
 app_file: app.py
 pinned: false
-license: apache-2.0
-short_description: A free lunch eliciting personalized ability of adapters
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -1,154 +1,392 @@
+import os
+
 import gradio as gr
 import numpy as np
-import random
-
-# import spaces #[uncomment to use ZeroGPU]
-from diffusers import DiffusionPipeline
 import torch
+from PIL import Image
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo"  # Replace to the model you would like to use
+from ip_adapter import (
+    ConceptrolIPAdapterPlus,
+    ConceptrolIPAdapterPlusXL,
+)
+from ip_adapter.custom_pipelines import (
+    StableDiffusionCustomPipeline,
+    StableDiffusionXLCustomPipeline,
+)
+from omini_control.conceptrol import Conceptrol
+from omini_control.flux_conceptrol_pipeline import FluxConceptrolPipeline
 
-if torch.cuda.is_available():
-    torch_dtype = torch.float16
-else:
-    torch_dtype = torch.float32
 
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to(device)
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
-
-
-# @spaces.GPU #[uncomment to use ZeroGPU]
-def infer(
-    prompt,
-    negative_prompt,
-    seed,
-    randomize_seed,
-    width,
-    height,
-    guidance_scale,
-    num_inference_steps,
-    progress=gr.Progress(track_tqdm=True),
-):
-    if randomize_seed:
-        seed = random.randint(0, MAX_SEED)
-
-    generator = torch.Generator().manual_seed(seed)
-
-    image = pipe(
-        prompt=prompt,
-        negative_prompt=negative_prompt,
-        guidance_scale=guidance_scale,
-        num_inference_steps=num_inference_steps,
-        width=width,
-        height=height,
-        generator=generator,
-    ).images[0]
-
-    return image, seed
-
-
-examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
-]
-
-css = """
-#col-container {
-    margin: 0 auto;
-    max-width: 640px;
+title = r"""
+<h1 align="center">Conceptrol: Concept Control of Zero-shot Personalized Image Generation</h1>
+"""
+
+description = r"""
+<b>Official 🤗 Gradio demo</b> for <a href='https://github.com/QY-H00/Conceptrol/tree/public' target='_blank'><b>Conceptrol: Concept Control of Zero-shot Personalized Image Generation</b></a>.<br>
+How to use:<br>
+1. Input text prompt, visual specification and the textual concept of the personalized target.
+2. Choose your preferrd base model, the first time for switching might take 30 minutes to download the model.
+3. For each inference, SD-series takes about 10s, SDXL-series takes about 30s, FLUX takes about 50s.
+4. Click the <b>Generate</b> button to enjoy! 😊
+"""
+
+article = r"""
+---
+✒️ **Citation**
+<br>
+If you found this demo/our paper useful, please consider citing:
+```bibtex
+@article{he2025conceptrol,
+  title={Conceptrol: Concept Control of Zero-shot Personalized Image Generation},
+  author={He, Qiyuan and Yao, Angela},
+  journal={arXiv preprint arXiv:2403.17924},
+  year={2024}
 }
+```
+📧 **Contact**
+<br>
+If you have any questions, please feel free to open an issue in our <a href='https://github.com/QY-H00/Conceptrol/tree/public' target='_blank'><b>Github Repo</b></a> or directly reach us out at <b>qhe@u.nus.edu.sg</b>.
 """
 
-with gr.Blocks(css=css) as demo:
-    with gr.Column(elem_id="col-container"):
-        gr.Markdown(" # Text-to-Image Gradio Template")
+MAX_SEED = np.iinfo(np.int32).max
+CACHE_EXAMPLES = False
+USE_TORCH_COMPILE = False
+ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD") == "1"
+PREVIEW_IMAGES = False
 
-        with gr.Row():
-            prompt = gr.Text(
-                label="Prompt",
-                show_label=False,
-                max_lines=1,
-                placeholder="Enter your prompt",
-                container=False,
-            )
+# Default settings
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+adapter_name = "h94/IP-Adapter/models/ip-adapter-plus_sd15.bin"
+pipe = StableDiffusionCustomPipeline.from_pretrained(
+    "SG161222/Realistic_Vision_V5.1_noVAE",
+    torch_dtype=torch.float16,
+    feature_extractor=None,
+    safety_checker=None
+)
+pipeline = ConceptrolIPAdapterPlus(pipe, "", adapter_name, device, num_tokens=16)
 
-            run_button = gr.Button("Run", scale=0, variant="primary")
+def change_model_fn(model_name: str) -> None:
+    global device, pipeline
+    
+    # Clear GPU memory
+    if torch.cuda.is_available():
+        if pipeline is not None:
+            del pipeline
+        torch.cuda.empty_cache()
+        
+    name_mapping = {
+        "SD1.5-512": "stable-diffusion-v1-5/stable-diffusion-v1-5", 
+        "AOM3 (SD-based)": "hogiahien/aom3",
+        "RealVis-v5.1 (SD-based)": "SG161222/Realistic_Vision_V5.1_noVAE",
+        "SDXL-1024": "stabilityai/stable-diffusion-xl-base-1.0",
+        "RealVisXL-v5.0 (SDXL-based)": "SG161222/RealVisXL_V5.0",
+        "Playground-XL-v2 (SDXL-based)": "playgroundai/playground-v2.5-1024px-aesthetic",
+        "Animagine-XL-v4.0 (SDXL-based)": "cagliostrolab/animagine-xl-4.0",
+        "FLUX-schnell": "black-forest-labs/FLUX.1-schnell"
+    }
+    if "XL" in model_name:
+        adapter_name = "h94/IP-Adapter/sdxl_models/ip-adapter-plus_sdxl_vit-h.safetensors"
+        pipe = StableDiffusionXLCustomPipeline.from_pretrained(
+            name_mapping[model_name],
+            # variant="fp16",
+            torch_dtype=torch.float16,
+            feature_extractor=None
+        )
+        pipeline = ConceptrolIPAdapterPlusXL(pipe, "", adapter_name, device, num_tokens=16)
+        globals()["pipeline"] = pipeline
+    
+    elif "FLUX" in model_name:
+        adapter_name = "Yuanshi/OminiControl"
+        pipeline = FluxConceptrolPipeline.from_pretrained(
+            name_mapping[model_name], torch_dtype=torch.bfloat16
+        ).to(device)
+        pipeline.load_lora_weights(
+            adapter_name,
+            weight_name="omini/subject_512.safetensors",
+            adapter_name="subject",
+        )
+        config = {"name": "conceptrol"}
+        conceptrol = Conceptrol(config)
+        pipeline.load_conceptrol(conceptrol)
+        globals()["pipeline"] = pipeline
+        globals()["pipeline"].to(device, dtype=torch.bfloat16)
+    
+    elif "XL" not in model_name and "FLUX" not in model_name:
+        adapter_name = "h94/IP-Adapter/models/ip-adapter-plus_sd15.bin"
+        pipe = StableDiffusionCustomPipeline.from_pretrained(
+            name_mapping[model_name],
+            torch_dtype=torch.float16,
+            feature_extractor=None,
+            safety_checker=None
+        )
+        pipeline = ConceptrolIPAdapterPlus(pipe, "", adapter_name, device, num_tokens=16)
+        globals()["pipeline"] = pipeline
+    else:
+        raise KeyError("Not supported model name!")
 
-        result = gr.Image(label="Result", show_label=False)
 
-        with gr.Accordion("Advanced Settings", open=False):
-            negative_prompt = gr.Text(
-                label="Negative prompt",
-                max_lines=1,
-                placeholder="Enter a negative prompt",
-                visible=False,
-            )
+def save_image(img, index):
+    unique_name = f"{index}.png"
+    img = Image.fromarray(img)
+    img.save(unique_name)
+    return unique_name
 
-            seed = gr.Slider(
-                label="Seed",
-                minimum=0,
-                maximum=MAX_SEED,
-                step=1,
-                value=0,
-            )
 
-            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+def get_example() -> list[list[str | float | int]]:
+    case = [
+        [
+            "A statue is reading the book in the cafe, best quality, high quality",
+            "statue",
+            "deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
+            Image.open("demo/statue.jpg"),
+            50,
+            6.0,
+            1.0,
+            0.2,
+            42,
+            "RealVis-v5.1 (SD-based)"
+        ],
+        [
+            "A hyper-realistic, high-resolution photograph of an astronaut in a meticulously detailed space suit riding a majestic horse across an otherworldly landscape. The image features dynamic lighting, rich textures, and a cinematic atmosphere, capturing every intricate detail in stunning clarity.",
+            "horse",
+            "deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
+            Image.open("demo/horse.jpg"),
+            50,
+            6.0,
+            1.0,
+            0.2,
+            42,
+            "RealVisXL-v5.0 (SDXL-based)"
+        ],
+        [
+            "A man wearing a T-shirt walking on the street",
+            "T-shirt",
+            "",
+            Image.open("demo/t-shirt.jpg"),
+            20,
+            3.5,
+            1.0,
+            0.0,
+            42,
+            "FLUX-schnell"
+        ]
+    ]
+    return case
+
+
+def change_generate_button_fn(enable: int) -> gr.Button:
+    if enable == 0:
+        return gr.Button(interactive=False, value="Switching Model...")
+    else:
+        return gr.Button(interactive=True, value="Generate")
+
+
+def dynamic_gallery_fn():
+    return gr.Image(label="Result", show_label=False)
 
-            with gr.Row():
-                width = gr.Slider(
-                    label="Width",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,  # Replace with defaults that work for your model
-                )
 
-                height = gr.Slider(
-                    label="Height",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,  # Replace with defaults that work for your model
+@torch.no_grad()
+def generate(
+    prompt="a statue is reading the book in the cafe",
+    subject="cat",
+    negative_prompt="",
+    image=None,
+    num_inference_steps=20,
+    guidance_scale=3.5,
+    condition_scale=1.0,
+    control_guidance_start=0.0,
+    seed=0,
+    model_name="RealVis-v5.1 (SD-based)"
+) -> np.ndarray:
+    global pipeline
+    change_model_fn(model_name)
+    if isinstance(pipeline, FluxConceptrolPipeline):
+        images = pipeline(
+            prompt=prompt,
+            image=image,
+            subject=subject,
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+            condition_scale=condition_scale,
+            control_guidance_start=control_guidance_start,
+            height=512,
+            width=512,
+            seed=seed,
+        ).images[0]
+    elif isinstance(pipeline, ConceptrolIPAdapterPlus) or isinstance(pipeline, ConceptrolIPAdapterPlusXL):
+        images = pipeline.generate(
+            prompt=prompt,
+            pil_images=[image],
+            subjects=[subject],
+            num_samples=1,
+            num_inference_steps=50,
+            scale=condition_scale,
+            negative_prompt=negative_prompt,
+            control_guidance_start=control_guidance_start,
+            seed=seed,
+        )[0]
+    else:
+        raise TypeError("Unsupported Pipeline")
+
+    return images
+
+with gr.Blocks(css="style.css") as demo:
+    gr.Markdown(title)
+    gr.Markdown(description)
+    with gr.Row(elem_classes="grid-container"):
+        with gr.Group():
+            with gr.Row(elem_classes="flex-grow"):
+                with gr.Column(elem_classes="grid-item"):  # 左侧列
+                    prompt = gr.Text(
+                        label="Prompt",
+                        max_lines=3,
+                        placeholder="Enter the Descriptive Prompt",
+                        interactive=True,
+                        value="A statue is reading the book in the cafe, best quality, high quality",
+                    )
+                    textual_concept = gr.Text(
+                        label="Textual Concept",
+                        max_lines=3,
+                        placeholder="Enter the Textual Concept required customization",
+                        interactive=True,
+                        value="statue",
+                    )
+                    negative_prompt = gr.Text(
+                        label="Negative prompt",
+                        max_lines=3,
+                        placeholder="Enter a Negative Prompt",
+                        interactive=True,
+                        value="deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality"
+                    )
+        
+        with gr.Row(elem_classes="flex-grow"):
+            image_prompt = gr.Image(
+                    label="Reference Image for customization",
+                    interactive=True,
+                    height=280
                 )
+                
 
-            with gr.Row():
-                guidance_scale = gr.Slider(
-                    label="Guidance scale",
+        with gr.Group():
+            with gr.Column(elem_classes="grid-item"):  # 右侧列
+                with gr.Row(elem_classes="flex-grow"):
+                    
+                    with gr.Group():
+                        # result = gr.Gallery(label="Result", show_label=False, rows=1, columns=1)
+                        result = gr.Image(label="Result", show_label=False, height=238, width=256)
+                        generate_button = gr.Button(value="Generate", variant="primary")
+
+    with gr.Accordion("Advanced options", open=True):
+        with gr.Row():
+            with gr.Column():
+                # with gr.Row(elem_classes="flex-grow"):
+                model_choice = gr.Dropdown(
+                    [
+                        "AOM3 (SD-based)",
+                        "SD1.5-512",
+                        "RealVis-v5.1 (SD-based)",
+                        "SDXL-1024", 
+                        "RealVisXL-v5.0 (SDXL-based)",
+                        "Animagine-XL-v4.0 (SDXL-based)",
+                        "FLUX-schnell"
+                    ],
+                    label="Model",
+                    value="RealVis-v5.1 (SD-based)",
+                    interactive=True,
+                    info="XL-Series takes longer time and FLUX takes even more",
+                )
+                condition_scale = gr.Slider(
+                    label="Condition Scale of Reference Image",
+                    minimum=0.4,
+                    maximum=1.5,
+                    step=0.05,
+                    value=1.0,
+                    interactive=True,
+                )
+                warmup_ratio = gr.Slider(
+                    label="Warmup Ratio",
                     minimum=0.0,
-                    maximum=10.0,
-                    step=0.1,
-                    value=0.0,  # Replace with defaults that work for your model
+                    maximum=1,
+                    step=0.05,
+                    value=0.2,
+                    interactive=True,
                 )
-
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=50,
-                    step=1,
-                    value=2,  # Replace with defaults that work for your model
+                guidance_scale = gr.Slider(
+                    label="Guidance Scale",
+                    minimum=0,
+                    maximum=10,
+                    step=0.1,
+                    value=5.0,
+                    interactive=True,
                 )
+        num_inference_steps = gr.Slider(
+            label="Inference Steps",
+            minimum=10,
+            maximum=50,
+            step=1,
+            value=50,
+            interactive=True,
+        )
+        with gr.Column():
+            seed = gr.Slider(
+                label="Seed",
+                minimum=0,
+                maximum=MAX_SEED,
+                step=1,
+                value=0,
+            )
 
-        gr.Examples(examples=examples, inputs=[prompt])
-    gr.on(
-        triggers=[run_button.click, prompt.submit],
-        fn=infer,
+    gr.Examples(
+        examples=get_example(),
         inputs=[
             prompt,
+            textual_concept,
             negative_prompt,
-            seed,
-            randomize_seed,
-            width,
-            height,
-            guidance_scale,
+            image_prompt,
             num_inference_steps,
+            guidance_scale,
+            condition_scale,
+            warmup_ratio,
+            seed,
+            model_choice
         ],
-        outputs=[result, seed],
+        cache_examples=CACHE_EXAMPLES,
+    )
+
+    # model_choice.change(
+    #     fn=change_generate_button_fn,
+    #     inputs=gr.Number(0, visible=False),
+    #     outputs=generate_button,
+    # )
+    
+    # .then(fn=change_model_fn, inputs=model_choice).then(
+    #     fn=change_generate_button_fn,
+    #     inputs=gr.Number(1, visible=False),
+    #     outputs=generate_button,
+    # )
+
+    inputs = [
+        prompt,
+        textual_concept,
+        negative_prompt,
+        image_prompt,
+        num_inference_steps,
+        guidance_scale,
+        condition_scale,
+        warmup_ratio,
+        seed,
+        model_choice
+    ]
+    generate_button.click(
+        fn=dynamic_gallery_fn,
+        outputs=result,
+    ).then(
+        fn=generate,
+        inputs=inputs,
+        outputs=result,
     )
+    gr.Markdown(article)
 
-if __name__ == "__main__":
-    demo.launch()
+demo.launch()

ip_adapter/__init__.py

@@ -0,0 +1,23 @@
+from .ip_adapter import (
+    IPAdapter,
+    IPAdapterFull,
+    IPAdapterPlus,
+    IPAdapterPlusXL,
+    IPAdapterXL,
+    ConceptrolIPAdapter,
+    ConceptrolIPAdapterPlus,
+    ConceptrolIPAdapterPlusXL,
+    ConceptrolIPAdapterXL,
+)
+
+__all__ = [
+    "IPAdapter",
+    "IPAdapterPlus",
+    "IPAdapterPlusXL",
+    "IPAdapterXL",
+    "IPAdapterFull",
+    "ConceptrolIPAdapter",
+    "ConceptrolIPAdapterPlus",
+    "ConceptrolIPAdapterXL",
+    "ConceptrolIPAdapterPlusXL",
+]

ip_adapter/attention_processor.py

@@ -0,0 +1,948 @@
+# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# Global Variable
+global_concept_mask = []
+attn_mask_logs = {}
+text_attn_map_logs = {}
+image_attn_map_logs = {}
+
+
+class AttnProcessor(nn.Module):
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        *args,
+        **kwargs,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape
+            if encoder_hidden_states is None
+            else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(
+            attention_mask, sequence_length, batch_size
+        )
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(
+                encoder_hidden_states
+            )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor(nn.Module):
+    r"""
+    Attention processor for IP-Adapater.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.to_k_ip = nn.Linear(
+            cross_attention_dim or hidden_size, hidden_size, bias=False
+        )
+        self.to_v_ip = nn.Linear(
+            cross_attention_dim or hidden_size, hidden_size, bias=False
+        )
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        *args,
+        **kwargs,
+    ):
+        global global_concept_mask
+        global attn_mask_logs
+        global text_attn_map_logs
+        global image_attn_map_logs
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape
+            if encoder_hidden_states is None
+            else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(
+            attention_mask, sequence_length, batch_size
+        )
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(
+                    encoder_hidden_states
+                )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        ip_key = attn.head_to_batch_dim(ip_key)
+        ip_value = attn.head_to_batch_dim(ip_value)
+
+        ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+        self.attn_map = ip_attention_probs
+        ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+        ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class ConceptrolAttnProcessor(nn.Module):
+    r"""
+    Attention processor for IP-Adapater.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(
+        self,
+        hidden_size,
+        cross_attention_dim=None,
+        scale=1.0,
+        num_tokens=4,
+        textual_concept_idxs=None,
+        name=None,
+        global_masking=False,
+        adaptive_scale_mask=False,
+        concept_mask_layer=None,
+    ):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.textual_concept_idxs = textual_concept_idxs
+        self.name = name
+
+        self.to_k_ip = nn.Linear(
+            cross_attention_dim or hidden_size, hidden_size, bias=False
+        )
+        self.to_v_ip = nn.Linear(
+            cross_attention_dim or hidden_size, hidden_size, bias=False
+        )
+
+        self.global_masking = global_masking
+        self.adaptive_scale_mask = adaptive_scale_mask
+
+        if concept_mask_layer is None:
+            concept_mask_layer = [
+                "mid_block.attentions.0.transformer_blocks.0.attn2.processor"
+            ]  # For SD
+            print("Warning: Using default concept mask layer for SD. For SDXL, use 'up_blocks.0.attentions.1.transformer_blocks.5.attn2.processor'")
+            # concept_mask_layer = ['up_blocks.0.attentions.1.transformer_blocks.1.attn2.processor'] # For SDXL
+        self.concept_mask_layer = concept_mask_layer
+
+    def set_global_view(self, attn_procs):
+        self.attn_procs = attn_procs
+        # print(self.name, self.attn_procs.keys())
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        *args,
+        **kwargs,
+    ):
+        global global_concept_mask
+        global attn_mask_logs
+
+        if self.textual_concept_idxs is None:
+            raise ValueError(
+                "textual_concept_idxs should be provided for ConceptrolAttnProcessor"
+            )
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape
+            if encoder_hidden_states is None
+            else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(
+            attention_mask, sequence_length, batch_size
+        )
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = 77  # Both SD and SDXL use 77 as length of text tokens
+            encoder_hidden_states, ip_hidden_states_cat = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            num_concepts = ip_hidden_states_cat.shape[1] // self.num_tokens
+            ip_hidden_states_list = torch.chunk(
+                ip_hidden_states_cat, num_concepts, dim=1
+            )
+            assert len(ip_hidden_states_list) == len(
+                self.textual_concept_idxs
+            ), f"register_idxs should have the same length as the number of concepts, but got {len(ip_hidden_states_list)} and {len(self.textual_concept_idxs)}"
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(
+                    encoder_hidden_states
+                )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)  # [16, 4096, 40]
+        key = attn.head_to_batch_dim(key)  # [16, 77, 40]
+        value = attn.head_to_batch_dim(value)  # [16, 77, 40]
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        concept_mask_layer = self.concept_mask_layer
+        if len(global_concept_mask) == 0:
+            global_concept_mask = [None for _ in range(len(self.textual_concept_idxs))]
+        for i in range(len(self.textual_concept_idxs)):
+            ip_hidden_states = ip_hidden_states_list[i]
+            textual_concept_start_idx, textual_concept_end_idx = (
+                self.textual_concept_idxs[i]
+            )
+            ip_key = self.to_k_ip(ip_hidden_states)
+            ip_value = self.to_v_ip(ip_hidden_states)
+
+            ip_key = attn.head_to_batch_dim(ip_key)  # [16, 4, 40]
+            ip_value = attn.head_to_batch_dim(ip_value)  # [16, 4, 40]
+
+            # attention_probs: [20/40, 4096, 77]
+
+            ip_attention_mask = attention_probs[
+                :, :, textual_concept_start_idx:textual_concept_end_idx
+            ]  # [16, 4096, T]
+            ip_attention_mask = torch.mean(
+                ip_attention_mask, dim=-1, keepdim=True
+            )  # [16, 4096, 1]
+            ip_attention_mask = attn.batch_to_head_dim(
+                ip_attention_mask
+            )  # [2, 4096, 8]
+            ip_attention_mask = torch.mean(
+                ip_attention_mask, dim=-1, keepdim=True
+            )  # [2, 4096, 1]
+
+            ip_attention_mask = ip_attention_mask / (
+                torch.amax(ip_attention_mask, dim=-2, keepdim=True) + 1e-6
+            )
+
+            ip_attention_mask = ip_attention_mask[1:2]  # (use the classifier one)
+
+            # Visualization
+            if self.name not in attn_mask_logs:
+                attn_mask_logs[self.name] = []
+                text_attn_map_logs[self.name] = []
+                image_attn_map_logs[self.name] = []
+            attn_mask_logs[self.name].append(
+                ip_attention_mask.detach().cpu().numpy()[0, :, 0]
+            )
+            text_attn_map_logs[self.name].append(
+                ip_attention_mask.detach().cpu().numpy()[0, :, 0]
+            )
+
+            if self.global_masking and (
+                self.name == concept_mask_layer[0]
+            ):
+                global_concept_mask[i] = ip_attention_mask
+
+            if (
+                self.global_masking
+                and self.name != concept_mask_layer[0]
+                and global_concept_mask[i] is not None
+            ):
+                original_dim = int(global_concept_mask[i].shape[1] ** 0.5)
+                target_dim = int(hidden_states.shape[1] ** 0.5)
+                global_concept_mask_2d = global_concept_mask[i].view(
+                    global_concept_mask[i].shape[0], 1, original_dim, original_dim
+                )
+                resized_global_concept_mask_2d = F.interpolate(
+                    global_concept_mask_2d,
+                    size=(target_dim, target_dim),
+                    mode="nearest",
+                )
+                resized_global_concept_mask = resized_global_concept_mask_2d.view(
+                    global_concept_mask[i].shape[0], -1, 1
+                )
+                ip_attention_mask = resized_global_concept_mask
+
+            ip_attention_probs = attn.get_attention_scores(
+                query, ip_key, None
+            )  # [16, 4096, 4]
+
+            #  Visualization
+            ip_attention_map = attention_probs[:, :, 15:16]  # [16, 4096, T]
+            ip_attention_map = torch.mean(
+                ip_attention_map, dim=-1, keepdim=True
+            )  # [16, 4096, 1]
+            ip_attention_map = torch.mean(
+                ip_attention_map, dim=-1, keepdim=True
+            )  # [16, 4096, 1]
+            ip_attention_map = attn.batch_to_head_dim(ip_attention_map)  # [2, 4096, 8]
+            ip_attention_map = torch.mean(
+                ip_attention_map, dim=-1, keepdim=True
+            )  # [2, 4096, 1]
+            ip_attention_map = ip_attention_map / (
+                torch.amax(ip_attention_map, dim=-2, keepdim=True) + 1e-6
+            )
+            ip_attention_map = ip_attention_map[1:2]  # (use the classifier one)
+            image_attn_map_logs[self.name].append(
+                ip_attention_map.detach().cpu().numpy()[0, :, 0]
+            )
+
+            ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)  # [16, 4096, 40]
+            ip_hidden_states = attn.batch_to_head_dim(
+                ip_hidden_states
+            )  # [2, 4096, 320]
+            ip_hidden_states = ip_hidden_states * ip_attention_mask
+
+            if self.adaptive_scale_mask:
+                raise ValueError("adaptive_scale_mask is deprecated already")
+
+            hidden_states += self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        *args,
+        **kwargs,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape
+            if encoder_hidden_states is None
+            else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(
+                attention_mask, sequence_length, batch_size
+            )
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(
+                batch_size, attn.heads, -1, attention_mask.shape[-1]
+            )
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(
+                encoder_hidden_states
+            )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(
+            batch_size, -1, attn.heads * head_dim
+        )
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.to_k_ip = nn.Linear(
+            cross_attention_dim or hidden_size, hidden_size, bias=False
+        )
+        self.to_v_ip = nn.Linear(
+            cross_attention_dim or hidden_size, hidden_size, bias=False
+        )
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        *args,
+        **kwargs,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape
+            if encoder_hidden_states is None
+            else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(
+                attention_mask, sequence_length, batch_size
+            )
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(
+                batch_size, attn.heads, -1, attention_mask.shape[-1]
+            )
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(
+                    encoder_hidden_states
+                )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(
+            batch_size, -1, attn.heads * head_dim
+        )
+        hidden_states = hidden_states.to(query.dtype)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        ip_hidden_states = F.scaled_dot_product_attention(
+            query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+        )
+        with torch.no_grad():
+            self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
+            # print(self.attn_map.shape)
+
+        ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(
+            batch_size, -1, attn.heads * head_dim
+        )
+        ip_hidden_states = ip_hidden_states.to(query.dtype)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+## for controlnet
+class CNAttnProcessor:
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(self, num_tokens=4):
+        self.num_tokens = num_tokens
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        *args,
+        **kwargs,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape
+            if encoder_hidden_states is None
+            else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(
+            attention_mask, sequence_length, batch_size
+        )
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(
+                    encoder_hidden_states
+                )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class CNAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self, num_tokens=4):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+        self.num_tokens = num_tokens
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        *args,
+        **kwargs,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape
+            if encoder_hidden_states is None
+            else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(
+                attention_mask, sequence_length, batch_size
+            )
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(
+                batch_size, attn.heads, -1, attention_mask.shape[-1]
+            )
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(
+                    encoder_hidden_states
+                )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(
+            batch_size, -1, attn.heads * head_dim
+        )
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states

ip_adapter/custom_pipelines.py

@@ -0,0 +1,805 @@
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from diffusers import StableDiffusionPipeline, StableDiffusionXLPipeline
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.image_processor import PipelineImageInput
+from diffusers.pipelines.stable_diffusion.pipeline_output import (
+    StableDiffusionPipelineOutput,
+)
+from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
+from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl import (
+    rescale_noise_cfg,
+)
+
+from .attention_processor import IPAttnProcessor, ConceptrolAttnProcessor
+from . import attention_processor
+
+
+class StableDiffusionXLCustomPipeline(StableDiffusionXLPipeline):
+    def set_scale(self, scale):
+        for attn_processor in self.unet.attn_processors.values():
+            if isinstance(attn_processor, (IPAttnProcessor, ConceptrolAttnProcessor)):
+                attn_processor.scale = scale
+
+    @torch.no_grad()
+    def __call__(  # noqa: C901
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 6.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        num_images_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,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        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,
+        guidance_rescale: float = 0.0,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Optional[Tuple[int, int]] = None,
+        negative_original_size: Optional[Tuple[int, int]] = None,
+        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+        negative_target_size: Optional[Tuple[int, int]] = None,
+        control_guidance_start: float = 0.0,
+        control_guidance_end: float = 1.0,
+    ):
+        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.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            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.
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. As a result, the returned sample will
+                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                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. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            num_images_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.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled 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_xl.StableDiffusionXLPipelineOutput`] 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.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.7):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                For most cases, `target_size` should be set to the desired height and width of the generated image. If
+                not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
+                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            control_guidance_start (`float`, *optional*, defaults to 0.0):
+                The percentage of total steps at which the ControlNet starts applying.
+            control_guidance_end (`float`, *optional*, defaults to 1.0):
+                The percentage of total steps at which the ControlNet stops applying.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        attention_processor.attn_mask_logs = {}
+        attention_processor.image_attn_map_logs = {}
+        attention_processor.text_attn_map_logs = {}
+        # Clear the global concept mask
+        attention_processor.global_concept_mask = []
+
+        # 0. Default height and width to unet
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        )
+
+        # 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
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None)
+            if cross_attention_kwargs is not None
+            else None
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. 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)
+
+        # 7. Prepare added time ids & embeddings
+        add_text_embeds = pooled_prompt_embeds
+        if self.text_encoder_2 is None:
+            text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+        else:
+            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+        add_time_ids = self._get_add_time_ids(
+            original_size,
+            crops_coords_top_left,
+            target_size,
+            dtype=prompt_embeds.dtype,
+            text_encoder_projection_dim=text_encoder_projection_dim,
+        )
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+                text_encoder_projection_dim=text_encoder_projection_dim,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
+
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat(
+                [negative_pooled_prompt_embeds, add_text_embeds], dim=0
+            )
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(
+            batch_size * num_images_per_prompt, 1
+        )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(
+            len(timesteps) - num_inference_steps * self.scheduler.order, 0
+        )
+
+        # 7.1 Apply denoising_end
+        if (
+            denoising_end is not None
+            and isinstance(denoising_end, float)
+            and denoising_end > 0
+            and denoising_end < 1
+        ):
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (denoising_end * self.scheduler.config.num_train_timesteps)
+                )
+            )
+            num_inference_steps = len(
+                list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))
+            )
+            timesteps = timesteps[:num_inference_steps]
+
+        # get init conditioning scale
+        for attn_processor in self.unet.attn_processors.values():
+            if isinstance(attn_processor, (IPAttnProcessor, ConceptrolAttnProcessor)):
+                conditioning_scale = attn_processor.scale
+                break
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if (i / len(timesteps) < control_guidance_start) or (
+                    (i + 1) / len(timesteps) > control_guidance_end
+                ):
+                    self.set_scale(0.0)
+                else:
+                    self.set_scale(conditioning_scale)
+
+                # 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
+                )
+
+                # predict the noise residual
+                added_cond_kwargs = {
+                    "text_embeds": add_text_embeds,
+                    "time_ids": add_time_ids,
+                }
+
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # 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
+                    )
+
+                if do_classifier_free_guidance and guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(
+                        noise_pred, noise_pred_text, guidance_rescale=guidance_rescale
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs, return_dict=False
+                )[0]
+
+                # 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 not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = (
+                self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+            )
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(
+                    next(iter(self.vae.post_quant_conv.parameters())).dtype
+                )
+
+            image = self.vae.decode(
+                latents / self.vae.config.scaling_factor, return_dict=False
+            )[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+        else:
+            image = latents
+
+        if output_type != "latent":
+            # apply watermark if available
+            if self.watermark is not None:
+                image = self.watermark.apply_watermark(image)
+
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusionXLPipelineOutput(images=image)
+
+
+class StableDiffusionCustomPipeline(StableDiffusionPipeline):
+    def set_scale(self, scale):
+        for attn_processor in self.unet.attn_processors.values():
+            if isinstance(attn_processor, (IPAttnProcessor, ConceptrolAttnProcessor)):
+                attn_processor.scale = scale
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[
+            Union[
+                Callable[[int, int, Dict], None],
+                PipelineCallback,
+                MultiPipelineCallbacks,
+            ]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        control_guidance_start: float = 0.0,
+        control_guidance_end: float = 1.0,
+        **kwargs,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            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.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+            num_images_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) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *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 is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
+                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
+                contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
+                provided, embeddings are computed from the `ip_adapter_image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.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.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
+                using zero terminal SNR.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list with the generated images and the
+                second element is a list of `bool`s indicating whether the corresponding generated image contains
+                "not-safe-for-work" (nsfw) content.
+        """
+
+        attention_processor.attn_mask_logs = {}
+        attention_processor.image_attn_map_logs = {}
+        attention_processor.text_attn_map_logs = {}
+        attention_processor.global_concept_mask = []
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+        # to deal with lora scaling and other possible forward hooks
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
+
+        # 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
+
+        # 3. Encode input prompt
+        lora_scale = (
+            self.cross_attention_kwargs.get("scale", None)
+            if self.cross_attention_kwargs is not None
+            else None
+        )
+
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            self.do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=lora_scale,
+            clip_skip=self.clip_skip,
+        )
+
+        # 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
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+            image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. 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)
+
+        # 6.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = (
+            {"image_embeds": image_embeds}
+            if (ip_adapter_image is not None or ip_adapter_image_embeds is not None)
+            else None
+        )
+
+        # 6.2 Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(
+                batch_size * num_images_per_prompt
+            )
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        # get init conditioning scale
+        for attn_processor in self.unet.attn_processors.values():
+            if isinstance(attn_processor, (ConceptrolAttnProcessor, IPAttnProcessor)):
+                conditioning_scale = attn_processor.scale
+                break
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if (i / len(timesteps) < control_guidance_start) or (
+                    (i + 1) / len(timesteps) > control_guidance_end
+                ):
+                    self.set_scale(0.0)
+                else:
+                    self.set_scale(conditioning_scale)
+
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = (
+                    torch.cat([latents] * 2)
+                    if self.do_classifier_free_guidance
+                    else latents
+                )
+                latent_model_input = self.scheduler.scale_model_input(
+                    latent_model_input, t
+                )
+
+                # predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (
+                        noise_pred_text - noise_pred_uncond
+                    )
+
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(
+                        noise_pred,
+                        noise_pred_text,
+                        guidance_rescale=self.guidance_rescale,
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                results = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs, return_dict=False
+                )
+                latents = results[0]
+                # pred_original = results[1]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop(
+                        "negative_prompt_embeds", negative_prompt_embeds
+                    )
+
+                # 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:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if output_type != "latent":
+            image = self.vae.decode(
+                latents / self.vae.config.scaling_factor,
+                return_dict=False,
+                generator=generator,
+            )[0]
+            image, has_nsfw_concept = self.run_safety_checker(
+                image, device, prompt_embeds.dtype
+            )
+        else:
+            image = latents
+            has_nsfw_concept = None
+
+        if has_nsfw_concept is None:
+            do_denormalize = [True] * image.shape[0]
+        else:
+            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+        image = self.image_processor.postprocess(
+            image, output_type=output_type, do_denormalize=do_denormalize
+        )
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(
+            images=image, nsfw_content_detected=has_nsfw_concept
+        )

ip_adapter/ip_adapter.py

@@ -0,0 +1,1043 @@
+import os
+from typing import List
+
+import torch
+from diffusers.pipelines.controlnet import MultiControlNetModel
+from PIL import Image
+from safetensors import safe_open
+from transformers import (
+    CLIPImageProcessor,
+    CLIPVisionModelWithProjection,
+    CLIPTokenizer,
+)
+
+from .attention_processor import (
+    AttnProcessor,
+    CNAttnProcessor,
+    IPAttnProcessor,
+    ConceptrolAttnProcessor,
+)
+from .resampler import Resampler
+from .utils import get_generator
+from huggingface_hub import hf_hub_download
+
+SD_CONCEPT_LAYER = ["up_blocks.1.attentions.0.transformer_blocks.0.attn2.processor"]
+SDXL_CONCEPT_LAYER = ["up_blocks.0.attentions.1.transformer_blocks.3.attn2.processor"]
+
+
+class ImageProjModel(torch.nn.Module):
+    """Projection Model"""
+
+    def __init__(
+        self,
+        cross_attention_dim=1024,
+        clip_embeddings_dim=1024,
+        clip_extra_context_tokens=4,
+    ):
+        super().__init__()
+
+        self.generator = None
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = torch.nn.Linear(
+            clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim
+        )
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+class MLPProjModel(torch.nn.Module):
+    """SD model with image prompt"""
+
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024):
+        super().__init__()
+
+        self.proj = torch.nn.Sequential(
+            torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim),
+            torch.nn.GELU(),
+            torch.nn.Linear(clip_embeddings_dim, cross_attention_dim),
+            torch.nn.LayerNorm(cross_attention_dim),
+        )
+
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class IPAdapter:
+    def __init__(self, sd_pipe, image_encoder_path, ip_ckpt, device, num_tokens=4):
+        self.device = device
+        self.image_encoder_path = image_encoder_path
+        self.ip_ckpt = ip_ckpt
+        self.num_tokens = num_tokens
+
+        self.pipe = sd_pipe.to(self.device)
+        self.set_ip_adapter()
+
+        # load image encoder
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+            "h94/IP-Adapter",
+            subfolder="models/image_encoder",
+            torch_dtype=torch.float16,
+        ).to(self.device, dtype=torch.float16)
+        self.clip_image_processor = CLIPImageProcessor()
+        # image proj model
+        self.image_proj_model = self.init_proj()
+
+        self.load_ip_adapter()
+
+    def init_proj(self):
+        image_proj_model = ImageProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.projection_dim,
+            clip_extra_context_tokens=self.num_tokens,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    def set_ip_adapter(self):
+        unet = self.pipe.unet
+        attn_procs = {}
+        for name in unet.attn_processors.keys():  # noqa: SIM118
+            cross_attention_dim = (
+                None
+                if name.endswith("attn1.processor")
+                else unet.config.cross_attention_dim
+            )
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor()
+            else:
+                attn_procs[name] = IPAttnProcessor(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                    scale=1.0,
+                    num_tokens=self.num_tokens,
+                ).to(self.device, dtype=torch.float16)
+        unet.set_attn_processor(attn_procs)
+        if hasattr(self.pipe, "controlnet"):
+            if isinstance(self.pipe.controlnet, MultiControlNetModel):
+                for controlnet in self.pipe.controlnet.nets:
+                    controlnet.set_attn_processor(
+                        CNAttnProcessor(num_tokens=self.num_tokens)
+                    )
+            else:
+                self.pipe.controlnet.set_attn_processor(
+                    CNAttnProcessor(num_tokens=self.num_tokens)
+                )
+
+    def load_ip_adapter(self):
+        if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
+            state_dict = {"image_proj": {}, "ip_adapter": {}}
+            with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
+                for key in f.keys():  # noqa: SIM118
+                    if key.startswith("image_proj."):
+                        state_dict["image_proj"][key.replace("image_proj.", "")] = (
+                            f.get_tensor(key)
+                        )
+                    elif key.startswith("ip_adapter."):
+                        state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = (
+                            f.get_tensor(key)
+                        )
+        else:
+            state_dict = torch.load(self.ip_ckpt, map_location="cpu")
+        self.image_proj_model.load_state_dict(state_dict["image_proj"])
+        ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
+        ip_layers.load_state_dict(state_dict["ip_adapter"])
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
+        if pil_image is not None:
+            if isinstance(pil_image, Image.Image):
+                pil_image = [pil_image]
+            clip_image = self.clip_image_processor(
+                images=pil_image, return_tensors="pt"
+            ).pixel_values
+            clip_image_embeds = self.image_encoder(
+                clip_image.to(self.device, dtype=torch.float16)
+            ).image_embeds
+        else:
+            clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_image_prompt_embeds = self.image_proj_model(
+            torch.zeros_like(clip_image_embeds)
+        )
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def set_scale(self, scale):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, IPAttnProcessor):
+                attn_processor.scale = scale
+
+    def generate(
+        self,
+        pil_images=None,
+        clip_image_embeds=None,
+        prompt=None,
+        negative_prompt=None,
+        scale=1.0,
+        num_samples=1,
+        guidance_scale=7.5,
+        num_inference_steps=30,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1 if pil_images is not None else clip_image_embeds.size(0)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = (
+                "monochrome, lowres, bad anatomy, worst quality, low quality"
+            )
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds_list = []
+        uncond_image_prompt_embeds_list = []
+        for pil_image in pil_images:
+            image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
+                pil_image=pil_image, clip_image_embeds=clip_image_embeds
+            )
+            bs_embed, seq_len, _ = image_prompt_embeds.shape
+            image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+            image_prompt_embeds = image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
+                1, num_samples, 1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            image_prompt_embeds_list.append(image_prompt_embeds)
+            uncond_image_prompt_embeds_list.append(uncond_image_prompt_embeds)
+
+        with torch.inference_mode():
+            prompt_embeds_, negative_prompt_embeds_ = self.pipe.encode_prompt(
+                prompt,
+                device=self.device,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat(
+                [prompt_embeds_, *image_prompt_embeds_list], dim=1
+            )
+            negative_prompt_embeds = torch.cat(
+                [negative_prompt_embeds_, *uncond_image_prompt_embeds_list], dim=1
+            )
+
+        # generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            # generator=generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class ConceptrolIPAdapter:
+    def __init__(
+        self,
+        sd_pipe,
+        image_encoder_path,
+        ip_ckpt,
+        device,
+        num_tokens=4,
+        global_masking=False,
+        adaptive_scale_mask=False,
+    ):
+        self.device = device
+        self.image_encoder_path = image_encoder_path
+        self.ip_ckpt = ip_ckpt
+        self.num_tokens = num_tokens
+
+        self.pipe = sd_pipe.to(self.device)
+        self.set_ip_adapter(global_masking, adaptive_scale_mask)
+
+        # load image encoder
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+            "h94/IP-Adapter",
+            subfolder="models/image_encoder",
+            torch_dtype=torch.float16,
+        ).to(self.device, dtype=torch.float16)
+        self.tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+        self.clip_image_processor = CLIPImageProcessor()
+        # image proj model
+        self.image_proj_model = self.init_proj()
+
+        self.load_ip_adapter()
+
+    def init_proj(self):
+        image_proj_model = ImageProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.projection_dim,
+            clip_extra_context_tokens=self.num_tokens,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    def set_ip_adapter(self, global_masking, adaptive_scale_mask):
+        unet = self.pipe.unet
+        attn_procs = {}
+        for name in unet.attn_processors.keys():  # noqa: SIM118
+            cross_attention_dim = (
+                None
+                if name.endswith("attn1.processor")
+                else unet.config.cross_attention_dim
+            )
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor()
+            else:
+                attn_procs[name] = ConceptrolAttnProcessor(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                    scale=1.0,
+                    num_tokens=self.num_tokens,
+                    name=name,
+                    global_masking=global_masking,
+                    adaptive_scale_mask=adaptive_scale_mask,
+                    concept_mask_layer=SD_CONCEPT_LAYER,
+                ).to(self.device, dtype=torch.float16)
+        unet.set_attn_processor(attn_procs)
+        for name in unet.attn_processors.keys():  # noqa: SIM118
+            cross_attention_dim = (
+                None
+                if name.endswith("attn1.processor")
+                else unet.config.cross_attention_dim
+            )
+            if cross_attention_dim is not None:
+                unet.attn_processors[name].set_global_view(unet.attn_processors)
+        if hasattr(self.pipe, "controlnet"):
+            if isinstance(self.pipe.controlnet, MultiControlNetModel):
+                for controlnet in self.pipe.controlnet.nets:
+                    controlnet.set_attn_processor(
+                        CNAttnProcessor(num_tokens=self.num_tokens)
+                    )
+            else:
+                self.pipe.controlnet.set_attn_processor(
+                    CNAttnProcessor(num_tokens=self.num_tokens)
+                )
+
+    def load_ip_adapter(self):
+        ckpt_path = self.ip_ckpt
+        # If the checkpoint path is not an existing file and is not a full URL,
+        # assume it's a Huggingface repository specification.
+        if not os.path.exists(self.ip_ckpt) and not self.ip_ckpt.startswith("http"):
+            # If a colon is present, use it to split repo_id and filename.
+            if ":" in self.ip_ckpt:
+                repo_id, filename = self.ip_ckpt.split(":", 1)
+            else:
+                parts = self.ip_ckpt.split('/')
+                if len(parts) > 2:
+                    # For example, "h94/IP-Adapter/models/ip-adapter-plus_sd15.bin"
+                    # repo_id becomes "h94/IP-Adapter" and filename "models/ip-adapter-plus_sd15.bin".
+                    repo_id = '/'.join(parts[:2])
+                    filename = '/'.join(parts[2:])
+                else:
+                    repo_id = self.ip_ckpt
+                    filename = "models/ip-adapter-plus_sd15.bin"  # default filename if not specified
+            ckpt_path = hf_hub_download(repo_id=repo_id, filename=filename)
+        
+        # Load the state dictionary from the checkpoint file.
+        if os.path.splitext(ckpt_path)[-1] == ".safetensors":
+            state_dict = {"image_proj": {}, "ip_adapter": {}}
+            with safe_open(ckpt_path, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    if key.startswith("image_proj."):
+                        state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                    elif key.startswith("ip_adapter."):
+                        state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+        else:
+            state_dict = torch.load(ckpt_path, map_location="cpu")
+        
+        # Load the state dictionaries into the corresponding models.
+        self.image_proj_model.load_state_dict(state_dict["image_proj"])
+        ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
+        ip_layers.load_state_dict(state_dict["ip_adapter"])
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
+        if pil_image is not None:
+            if isinstance(pil_image, Image.Image):
+                pil_image = [pil_image]
+            clip_image = self.clip_image_processor(
+                images=pil_image, return_tensors="pt"
+            ).pixel_values
+            clip_image_embeds = self.image_encoder(
+                clip_image.to(self.device, dtype=torch.float16)
+            ).image_embeds
+        else:
+            clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_image_prompt_embeds = self.image_proj_model(
+            torch.zeros_like(clip_image_embeds)
+        )
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def set_scale(self, scale):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, ConceptrolAttnProcessor):
+                attn_processor.scale = scale
+
+    def load_textual_concept(self, prompt, subjects):
+        tokens = self.tokenizer.tokenize(prompt)
+        textual_concept_idxs = []
+        offset = 1 # TODO: change back to 1 if not true
+
+        for subject in subjects:
+            subject_tokens = self.tokenizer.tokenize(subject)
+            start_idx = tokens.index(subject_tokens[0]) + offset
+            end_idx = tokens.index(subject_tokens[-1]) + offset
+            textual_concept_idxs.append((start_idx, end_idx + 1))
+            print("Locate:", subject, start_idx, end_idx + 1)
+
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, ConceptrolAttnProcessor):
+                attn_processor.textual_concept_idxs = textual_concept_idxs
+
+    def generate(
+        self,
+        pil_images=None,
+        clip_image_embeds=None,
+        prompt=None,
+        negative_prompt=None,
+        scale=1.0,
+        num_samples=1,
+        seed=42,
+        subjects=None,
+        guidance_scale=7.5,
+        num_inference_steps=30,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1  # not support multiple prompts
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = (
+                "monochrome, lowres, bad anatomy, worst quality, low quality"
+            )
+
+        if subjects:
+            self.load_textual_concept(prompt, subjects)
+        else:
+            raise ValueError("Subjects must be provided")
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds_list = []
+        uncond_image_prompt_embeds_list = []
+        for pil_image in pil_images:
+            image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
+                pil_image=pil_image, clip_image_embeds=clip_image_embeds
+            )
+            bs_embed, seq_len, _ = image_prompt_embeds.shape
+            image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+            image_prompt_embeds = image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
+                1, num_samples, 1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            image_prompt_embeds_list.append(image_prompt_embeds)
+            uncond_image_prompt_embeds_list.append(uncond_image_prompt_embeds)
+
+        with torch.inference_mode():
+            prompt_embeds_, negative_prompt_embeds_ = self.pipe.encode_prompt(
+                prompt,
+                device=self.device,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat(
+                [prompt_embeds_, *image_prompt_embeds_list], dim=1
+            )
+            negative_prompt_embeds = torch.cat(
+                [negative_prompt_embeds_, *uncond_image_prompt_embeds_list], dim=1
+            )
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class IPAdapterXL(IPAdapter):
+    """SDXL"""
+
+    def generate(
+        self,
+        pil_images,
+        prompt=None,
+        negative_prompt=None,
+        scale=1.0,
+        num_samples=1,
+        seed=None,
+        num_inference_steps=30,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1  # not support multiple prompts
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = (
+                "monochrome, lowres, bad anatomy, worst quality, low quality"
+            )
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds_list = []
+        uncond_image_prompt_embeds_list = []
+        for pil_image in pil_images:
+            image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
+                pil_image=pil_image
+            )
+            bs_embed, seq_len, _ = image_prompt_embeds.shape
+            image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+            image_prompt_embeds = image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
+                1, num_samples, 1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            image_prompt_embeds_list.append(image_prompt_embeds)
+            uncond_image_prompt_embeds_list.append(uncond_image_prompt_embeds)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, *image_prompt_embeds_list], dim=1)
+            negative_prompt_embeds = torch.cat(
+                [negative_prompt_embeds, *uncond_image_prompt_embeds_list], dim=1
+            )
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class ConceptrolIPAdapterXL(ConceptrolIPAdapter):
+    """SDXL"""
+
+    def set_ip_adapter(self, global_masking, adaptive_scale_mask):
+        unet = self.pipe.unet
+        attn_procs = {}
+        for name in unet.attn_processors.keys():  # noqa: SIM118
+            cross_attention_dim = (
+                None
+                if name.endswith("attn1.processor")
+                else unet.config.cross_attention_dim
+            )
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor()
+            else:
+                attn_procs[name] = ConceptrolAttnProcessor(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                    scale=1.0,
+                    num_tokens=self.num_tokens,
+                    name=name,
+                    global_masking=global_masking,
+                    adaptive_scale_mask=adaptive_scale_mask,
+                    concept_mask_layer=SDXL_CONCEPT_LAYER,
+                ).to(self.device, dtype=torch.float16)
+        unet.set_attn_processor(attn_procs)
+        for name in unet.attn_processors.keys():  # noqa: SIM118
+            cross_attention_dim = (
+                None
+                if name.endswith("attn1.processor")
+                else unet.config.cross_attention_dim
+            )
+            if cross_attention_dim is not None:
+                unet.attn_processors[name].set_global_view(unet.attn_processors)
+        if hasattr(self.pipe, "controlnet"):
+            if isinstance(self.pipe.controlnet, MultiControlNetModel):
+                for controlnet in self.pipe.controlnet.nets:
+                    controlnet.set_attn_processor(
+                        CNAttnProcessor(num_tokens=self.num_tokens)
+                    )
+            else:
+                self.pipe.controlnet.set_attn_processor(
+                    CNAttnProcessor(num_tokens=self.num_tokens)
+                )
+
+    def generate(
+        self,
+        pil_images=None,
+        prompt=None,
+        negative_prompt=None,
+        subjects=None,
+        scale=1.0,
+        num_samples=1,
+        num_inference_steps=30,
+        seed=None,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1  # not support multiple prompts
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = (
+                "monochrome, lowres, bad anatomy, worst quality, low quality"
+            )
+
+        if subjects:
+            self.load_textual_concept(prompt, subjects)
+        else:
+            raise ValueError("Subjects must be provided")
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds_list = []
+        uncond_image_prompt_embeds_list = []
+        for pil_image in pil_images:
+            image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
+                pil_image=pil_image
+            )
+            bs_embed, seq_len, _ = image_prompt_embeds.shape
+            image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+            image_prompt_embeds = image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
+                1, num_samples, 1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            image_prompt_embeds_list.append(image_prompt_embeds)
+            uncond_image_prompt_embeds_list.append(uncond_image_prompt_embeds)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, *image_prompt_embeds_list], dim=1)
+            negative_prompt_embeds = torch.cat(
+                [negative_prompt_embeds, *uncond_image_prompt_embeds_list], dim=1
+            )
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class IPAdapterPlus(IPAdapter):
+    """IP-Adapter with fine-grained features"""
+
+    def init_proj(self):
+        image_proj_model = Resampler(
+            dim=self.pipe.unet.config.cross_attention_dim,
+            depth=4,
+            dim_head=64,
+            heads=12,
+            num_queries=self.num_tokens,
+            embedding_dim=self.image_encoder.config.hidden_size,
+            output_dim=self.pipe.unet.config.cross_attention_dim,
+            ff_mult=4,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self.clip_image_processor(
+            images=pil_image, return_tensors="pt"
+        ).pixel_values
+        clip_image = clip_image.to(self.device, dtype=torch.float16)
+        clip_image_embeds = self.image_encoder(
+            clip_image, output_hidden_states=True
+        ).hidden_states[-2]
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = self.image_encoder(
+            torch.zeros_like(clip_image), output_hidden_states=True
+        ).hidden_states[-2]
+        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+class ConceptrolIPAdapterPlus(ConceptrolIPAdapter):
+    """IP-Adapter with fine-grained features"""
+
+    def init_proj(self):
+        image_proj_model = Resampler(
+            dim=self.pipe.unet.config.cross_attention_dim,
+            depth=4,
+            dim_head=64,
+            heads=12,
+            num_queries=self.num_tokens,
+            embedding_dim=self.image_encoder.config.hidden_size,
+            output_dim=self.pipe.unet.config.cross_attention_dim,
+            ff_mult=4,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self.clip_image_processor(
+            images=pil_image, return_tensors="pt"
+        ).pixel_values
+        clip_image = clip_image.to(self.device, dtype=torch.float16)
+        clip_image_embeds = self.image_encoder(
+            clip_image, output_hidden_states=True
+        ).hidden_states[-2]
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = self.image_encoder(
+            torch.zeros_like(clip_image), output_hidden_states=True
+        ).hidden_states[-2]
+        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+class IPAdapterFull(IPAdapterPlus):
+    """IP-Adapter with full features"""
+
+    def init_proj(self):
+        image_proj_model = MLPProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.hidden_size,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+
+class IPAdapterPlusXL(IPAdapter):
+    """SDXL"""
+
+    def init_proj(self):
+        image_proj_model = Resampler(
+            dim=1280,
+            depth=4,
+            dim_head=64,
+            heads=20,
+            num_queries=self.num_tokens,
+            embedding_dim=self.image_encoder.config.hidden_size,
+            output_dim=self.pipe.unet.config.cross_attention_dim,
+            ff_mult=4,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self.clip_image_processor(
+            images=pil_image, return_tensors="pt"
+        ).pixel_values
+        clip_image = clip_image.to(self.device, dtype=torch.float16)
+        clip_image_embeds = self.image_encoder(
+            clip_image, output_hidden_states=True
+        ).hidden_states[-2]
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = self.image_encoder(
+            torch.zeros_like(clip_image), output_hidden_states=True
+        ).hidden_states[-2]
+        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def generate(
+        self,
+        pil_images=None,
+        prompt=None,
+        negative_prompt=None,
+        scale=1.0,
+        num_samples=1,
+        seed=42,
+        num_inference_steps=30,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1  # not support multiple prompts
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = (
+                "monochrome, lowres, bad anatomy, worst quality, low quality"
+            )
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds_list = []
+        uncond_image_prompt_embeds_list = []
+        for pil_image in pil_images:
+            image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
+                pil_image=pil_image
+            )
+            bs_embed, seq_len, _ = image_prompt_embeds.shape
+            image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+            image_prompt_embeds = image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
+                1, num_samples, 1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            image_prompt_embeds_list.append(image_prompt_embeds)
+            uncond_image_prompt_embeds_list.append(uncond_image_prompt_embeds)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, *image_prompt_embeds_list], dim=1)
+            negative_prompt_embeds = torch.cat(
+                [negative_prompt_embeds, *uncond_image_prompt_embeds_list], dim=1
+            )
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class ConceptrolIPAdapterPlusXL(ConceptrolIPAdapterXL):
+    """SDXL"""
+
+    def init_proj(self):
+        image_proj_model = Resampler(
+            dim=1280,
+            depth=4,
+            dim_head=64,
+            heads=20,
+            num_queries=self.num_tokens,
+            embedding_dim=self.image_encoder.config.hidden_size,
+            output_dim=self.pipe.unet.config.cross_attention_dim,
+            ff_mult=4,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self.clip_image_processor(
+            images=pil_image, return_tensors="pt"
+        ).pixel_values
+        clip_image = clip_image.to(self.device, dtype=torch.float16)
+        clip_image_embeds = self.image_encoder(
+            clip_image, output_hidden_states=True
+        ).hidden_states[-2]
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = self.image_encoder(
+            torch.zeros_like(clip_image), output_hidden_states=True
+        ).hidden_states[-2]
+        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def generate(
+        self,
+        pil_images=None,
+        prompt=None,
+        negative_prompt=None,
+        scale=1.0,
+        subjects=None,
+        num_samples=1,
+        seed=42,
+        num_inference_steps=30,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1  # not support multiple prompts
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = (
+                "monochrome, lowres, bad anatomy, worst quality, low quality"
+            )
+
+        if subjects:
+            self.load_textual_concept(prompt, subjects)
+        else:
+            raise ValueError("Subjects must be provided")
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds_list = []
+        uncond_image_prompt_embeds_list = []
+        for pil_image in pil_images:
+            image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
+                pil_image=pil_image
+            )
+            bs_embed, seq_len, _ = image_prompt_embeds.shape
+            image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+            image_prompt_embeds = image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(
+                1, num_samples, 1
+            )
+            uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(
+                bs_embed * num_samples, seq_len, -1
+            )
+            image_prompt_embeds_list.append(image_prompt_embeds)
+            uncond_image_prompt_embeds_list.append(uncond_image_prompt_embeds)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, *image_prompt_embeds_list], dim=1)
+            negative_prompt_embeds = torch.cat(
+                [negative_prompt_embeds, *uncond_image_prompt_embeds_list], dim=1
+            )
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images

ip_adapter/resampler.py

@@ -0,0 +1,247 @@
+# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py
+
+import math
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(
+            -2, -1
+        )  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+        max_seq_len: int = 257,  # CLIP tokens + CLS token
+        apply_pos_emb: bool = False,
+        num_latents_mean_pooled: int = 0,  # number of latents derived from mean pooled representation of the sequence
+    ):
+        super().__init__()
+        self.pos_emb = (
+            nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None
+        )
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.to_latents_from_mean_pooled_seq = (
+            nn.Sequential(
+                nn.LayerNorm(dim),
+                nn.Linear(dim, dim * num_latents_mean_pooled),
+                Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled),
+            )
+            if num_latents_mean_pooled > 0
+            else None
+        )
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        if self.pos_emb is not None:
+            n, device = x.shape[1], x.device
+            pos_emb = self.pos_emb(torch.arange(n, device=device))
+            x = x + pos_emb
+
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        if self.to_latents_from_mean_pooled_seq:
+            meanpooled_seq = masked_mean(
+                x,
+                dim=1,
+                mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool),
+            )
+            meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq)
+            latents = torch.cat((meanpooled_latents, latents), dim=-2)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
+class ResamplerZeroInOut(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+        max_seq_len: int = 257,  # CLIP tokens + CLS token
+        apply_pos_emb: bool = False,
+        num_latents_mean_pooled: int = 0,  # number of latents derived from mean pooled representation of the sequence
+    ):
+        super().__init__()
+        self.pos_emb = (
+            nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None
+        )
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
+        self.proj_in_zero = nn.Linear(embedding_dim, embedding_dim, bias=False)
+        self.proj_in = nn.Linear(embedding_dim, dim)
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.proj_out_zero = nn.Linear(output_dim, output_dim, bias=False)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        nn.init.zeros_(self.proj_in_zero.weight)
+        nn.init.zeros_(self.proj_out_zero.weight)
+
+        self.to_latents_from_mean_pooled_seq = (
+            nn.Sequential(
+                nn.LayerNorm(dim),
+                nn.Linear(dim, dim * num_latents_mean_pooled),
+                Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled),
+            )
+            if num_latents_mean_pooled > 0
+            else None
+        )
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        if self.pos_emb is not None:
+            n, device = x.shape[1], x.device
+            pos_emb = self.pos_emb(torch.arange(n, device=device))
+            x = x + pos_emb
+
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in_zero(x)
+        x = self.proj_in(x)
+
+        if self.to_latents_from_mean_pooled_seq:
+            meanpooled_seq = masked_mean(
+                x,
+                dim=1,
+                mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool),
+            )
+            meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq)
+            latents = torch.cat((meanpooled_latents, latents), dim=-2)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        latents = self.proj_out_zero(latents)
+        return self.norm_out(latents)
+
+
+def masked_mean(t, *, dim, mask=None):
+    if mask is None:
+        return t.mean(dim=dim)
+
+    denom = mask.sum(dim=dim, keepdim=True)
+    mask = rearrange(mask, "b n -> b n 1")
+    masked_t = t.masked_fill(~mask, 0.0)
+
+    return masked_t.sum(dim=dim) / denom.clamp(min=1e-5)

ip_adapter/utils.py

@@ -0,0 +1,140 @@
+import numpy as np
+import torch
+import torch.nn.functional as F
+from PIL import Image
+
+# global variable
+raw_attn_maps = {}
+raw_ip_attn_maps = {}
+attn_maps = {}
+ip_attn_maps = {}
+
+
+def hook_fn(name):
+    def forward_hook(module, input, output):
+        if hasattr(module.processor, "attn_map"):
+            if name not in raw_attn_maps:
+                raw_attn_maps[name] = []
+            if name not in raw_ip_attn_maps:
+                raw_ip_attn_maps[name] = []
+            raw_attn_maps[name].append(module.processor.attn_map)
+            raw_ip_attn_maps[name].append(module.processor.ip_attn_map)
+            del module.processor.attn_map
+            del module.processor.ip_attn_map
+
+    return forward_hook
+
+
+def post_process_attn_maps():
+    global raw_attn_maps, raw_ip_attn_maps, attn_maps, ip_attn_maps
+    attn_maps = [
+        dict(zip(raw_attn_maps.keys(), values))
+        for values in zip(*raw_attn_maps.values())
+    ]
+    ip_attn_maps = [
+        dict(zip(raw_ip_attn_maps.keys(), values))
+        for values in zip(*raw_ip_attn_maps.values())
+    ]
+
+    return attn_maps, ip_attn_maps
+
+
+def register_cross_attention_hook(unet):
+    for name, module in unet.named_modules():
+        if name.split(".")[-1].startswith("attn2"):
+            module.register_forward_hook(hook_fn(name))
+
+    return unet
+
+
+def upscale(attn_map, target_size):
+    attn_map = torch.mean(attn_map, dim=0)
+    attn_map = attn_map.permute(1, 0)
+    temp_size = None
+
+    for i in range(0, 5):
+        scale = 2**i
+        if (target_size[0] // scale) * (target_size[1] // scale) == attn_map.shape[
+            1
+        ] * 64:
+            temp_size = (target_size[0] // (scale * 8), target_size[1] // (scale * 8))
+            break
+
+    assert temp_size is not None, "temp_size cannot is None"
+
+    attn_map = attn_map.view(attn_map.shape[0], *temp_size)
+
+    attn_map = F.interpolate(
+        attn_map.unsqueeze(0).to(dtype=torch.float32),
+        size=target_size,
+        mode="bilinear",
+        align_corners=False,
+    )[0]
+
+    attn_map = torch.softmax(attn_map, dim=0)
+    return attn_map
+
+
+def get_net_attn_map(
+    image_size, batch_size=2, instance_or_negative=False, detach=True, step=-1
+):
+
+    idx = 0 if instance_or_negative else 1
+    net_attn_maps = []
+    net_ip_attn_maps = []
+
+    for _, attn_map in attn_maps[step].items():
+        attn_map = attn_map.cpu() if detach else attn_map
+        attn_map = torch.chunk(attn_map, batch_size)[
+            idx
+        ].squeeze()  # get the attention map of text
+        attn_map = upscale(attn_map, image_size)
+        net_attn_maps.append(attn_map)
+
+    net_attn_maps = torch.mean(torch.stack(net_attn_maps, dim=0), dim=0)
+
+    for _, attn_map in ip_attn_maps[step].items():
+        attn_map = attn_map.cpu() if detach else attn_map
+        attn_map = torch.chunk(attn_map, batch_size)[
+            idx
+        ].squeeze()  # get the attention map of text
+        attn_map = upscale(attn_map, image_size)
+        net_ip_attn_maps.append(attn_map)
+
+    net_ip_attn_maps = torch.mean(torch.stack(net_ip_attn_maps, dim=0), dim=0)
+
+    return net_attn_maps, net_ip_attn_maps
+
+
+def attnmaps2images(net_attn_maps):
+    images = []
+
+    for attn_map in net_attn_maps:
+        attn_map = attn_map.cpu().numpy()
+        normalized_attn_map = (
+            (attn_map - np.min(attn_map)) / (np.max(attn_map) - np.min(attn_map)) * 255
+        )
+        normalized_attn_map = normalized_attn_map.astype(np.uint8)
+        image = Image.fromarray(normalized_attn_map)
+        images.append(image)
+
+    return images
+
+
+def is_torch2_available():
+    return hasattr(F, "scaled_dot_product_attention")
+
+
+def get_generator(seed, device):
+
+    if seed is not None:
+        if isinstance(seed, list):
+            generator = [
+                torch.Generator(device).manual_seed(seed_item) for seed_item in seed
+            ]
+        else:
+            generator = torch.Generator(device).manual_seed(seed)
+    else:
+        generator = None
+
+    return generator

omini_control/block.py

@@ -0,0 +1,354 @@
+import torch
+from typing import Optional, Dict, Any
+from diffusers.models.attention_processor import Attention, F
+from .lora_controller import enable_lora
+from .conceptrol import Conceptrol
+
+
+def attn_forward(
+    attn: Attention,
+    hidden_states: torch.FloatTensor,
+    encoder_hidden_states: torch.FloatTensor = None,
+    condition_latents: torch.FloatTensor = None,
+    attention_mask: Optional[torch.FloatTensor] = None,
+    image_rotary_emb: Optional[torch.Tensor] = None,
+    cond_rotary_emb: Optional[torch.Tensor] = None,
+    model_config: Optional[Dict[str, Any]] = {},
+    conceptrol: Conceptrol = None,
+) -> torch.FloatTensor:
+    global attn_maps
+    batch_size, _, _ = (
+        hidden_states.shape
+        if encoder_hidden_states is None
+        else encoder_hidden_states.shape
+    )
+
+    with enable_lora(
+        (attn.to_q, attn.to_k, attn.to_v), model_config.get("latent_lora", False)
+    ):
+        # `sample` projections.
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+    inner_dim = key.shape[-1]
+    head_dim = inner_dim // attn.heads
+
+    query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+    key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+    value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+    if attn.norm_q is not None:
+        query = attn.norm_q(query)
+    if attn.norm_k is not None:
+        key = attn.norm_k(key)
+
+    # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
+    if encoder_hidden_states is not None:
+        # `context` projections.
+        encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+        encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
+            batch_size, -1, attn.heads, head_dim
+        ).transpose(1, 2)
+        encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
+            batch_size, -1, attn.heads, head_dim
+        ).transpose(1, 2)
+        encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
+            batch_size, -1, attn.heads, head_dim
+        ).transpose(1, 2)
+
+        if attn.norm_added_q is not None:
+            encoder_hidden_states_query_proj = attn.norm_added_q(
+                encoder_hidden_states_query_proj
+            )
+        if attn.norm_added_k is not None:
+            encoder_hidden_states_key_proj = attn.norm_added_k(
+                encoder_hidden_states_key_proj
+            )
+
+        # attention
+        query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
+        key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
+        value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
+
+    if image_rotary_emb is not None:
+        from diffusers.models.embeddings import apply_rotary_emb
+
+        query = apply_rotary_emb(query, image_rotary_emb)
+        key = apply_rotary_emb(key, image_rotary_emb)
+
+    if condition_latents is not None:
+        cond_query = attn.to_q(condition_latents)
+        cond_key = attn.to_k(condition_latents)
+        cond_value = attn.to_v(condition_latents)
+
+        cond_query = cond_query.view(batch_size, -1, attn.heads, head_dim).transpose(
+            1, 2
+        )
+        cond_key = cond_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        cond_value = cond_value.view(batch_size, -1, attn.heads, head_dim).transpose(
+            1, 2
+        )
+        if attn.norm_q is not None:
+            cond_query = attn.norm_q(cond_query)
+        if attn.norm_k is not None:
+            cond_key = attn.norm_k(cond_key)
+
+    if cond_rotary_emb is not None:
+        cond_query = apply_rotary_emb(cond_query, cond_rotary_emb)
+        cond_key = apply_rotary_emb(cond_key, cond_rotary_emb)
+
+    if condition_latents is not None:
+        query = torch.cat([query, cond_query], dim=2)
+        key = torch.cat([key, cond_key], dim=2)
+        value = torch.cat([value, cond_value], dim=2)
+
+    if not model_config.get("union_cond_attn", True):
+        # If we don't want to use the union condition attention, we need to mask the attention
+        # between the hidden states and the condition latents
+        attention_mask = torch.ones(
+            query.shape[2], key.shape[2], device=query.device, dtype=torch.bool
+        )
+        condition_n = cond_query.shape[2]
+        attention_mask[-condition_n:, :-condition_n] = False
+        attention_mask[:-condition_n, -condition_n:] = False
+    if hasattr(attn, "c_factor"):
+        attention_mask = torch.zeros(
+            query.shape[2], key.shape[2], device=query.device, dtype=query.dtype
+        )
+        condition_n = cond_query.shape[2]
+        bias = torch.log(attn.c_factor[0])
+        attention_mask[-condition_n:, :-condition_n] = bias
+        attention_mask[:-condition_n, -condition_n:] = bias
+
+    if conceptrol is None:
+        print(
+            "Conceptrol using this stuff indicates that the implementation is problematic"
+        )
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, dropout_p=0.0, is_causal=False, attn_mask=attention_mask
+        )
+        hidden_states = hidden_states.transpose(1, 2).reshape(
+            batch_size, -1, attn.heads * head_dim
+        )
+        hidden_states = hidden_states.to(query.dtype)
+    else:
+        conceptrolled_attention_probs = conceptrol(
+            query, key, attention_mask, c_factor=attn.c_factor
+        )
+        hidden_states = conceptrolled_attention_probs @ value
+        hidden_states = hidden_states.transpose(1, 2).reshape(
+            batch_size, -1, attn.heads * head_dim
+        )
+        hidden_states = hidden_states.to(query.dtype)
+
+    if encoder_hidden_states is not None:
+        if condition_latents is not None:
+            encoder_hidden_states, hidden_states, condition_latents = (
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+                hidden_states[
+                    :, encoder_hidden_states.shape[1] : -condition_latents.shape[1]
+                ],
+                hidden_states[:, -condition_latents.shape[1] :],
+            )
+        else:
+            encoder_hidden_states, hidden_states = (
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+                hidden_states[:, encoder_hidden_states.shape[1] :],
+            )
+
+        with enable_lora((attn.to_out[0],), model_config.get("latent_lora", False)):
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            hidden_states = attn.to_out[1](hidden_states)
+        encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        if condition_latents is not None:
+            condition_latents = attn.to_out[0](condition_latents)
+            condition_latents = attn.to_out[1](condition_latents)
+
+        return (
+            (hidden_states, encoder_hidden_states, condition_latents)
+            if condition_latents is not None
+            else (hidden_states, encoder_hidden_states)
+        )
+    elif condition_latents is not None:
+        # if there are condition_latents, we need to separate the hidden_states and the condition_latents
+        hidden_states, condition_latents = (
+            hidden_states[:, : -condition_latents.shape[1]],
+            hidden_states[:, -condition_latents.shape[1] :],
+        )
+        return hidden_states, condition_latents
+    else:
+        return hidden_states
+
+
+def block_forward(
+    self,
+    hidden_states: torch.FloatTensor,
+    encoder_hidden_states: torch.FloatTensor,
+    condition_latents: torch.FloatTensor,
+    temb: torch.FloatTensor,
+    cond_temb: torch.FloatTensor,
+    cond_rotary_emb=None,
+    image_rotary_emb=None,
+    model_config: Optional[Dict[str, Any]] = {},
+    conceptrol: Conceptrol = None,
+):
+    use_cond = condition_latents is not None
+    with enable_lora((self.norm1.linear,), model_config.get("latent_lora", False)):
+        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+            hidden_states, emb=temb
+        )
+
+    norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = (
+        self.norm1_context(encoder_hidden_states, emb=temb)
+    )
+
+    if use_cond:
+        (
+            norm_condition_latents,
+            cond_gate_msa,
+            cond_shift_mlp,
+            cond_scale_mlp,
+            cond_gate_mlp,
+        ) = self.norm1(condition_latents, emb=cond_temb)
+
+    # Attention.
+    result = attn_forward(
+        self.attn,
+        model_config=model_config,
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=norm_encoder_hidden_states,
+        condition_latents=norm_condition_latents if use_cond else None,
+        image_rotary_emb=image_rotary_emb,
+        cond_rotary_emb=cond_rotary_emb if use_cond else None,
+        conceptrol=conceptrol if use_cond else None,
+    )
+    attn_output, context_attn_output = result[:2]
+    cond_attn_output = result[2] if use_cond else None
+
+    # Process attention outputs for the `hidden_states`.
+    # 1. hidden_states
+    attn_output = gate_msa.unsqueeze(1) * attn_output
+    hidden_states = hidden_states + attn_output
+    # 2. encoder_hidden_states
+    context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
+    encoder_hidden_states = encoder_hidden_states + context_attn_output
+    # 3. condition_latents
+    if use_cond:
+        cond_attn_output = cond_gate_msa.unsqueeze(1) * cond_attn_output
+        condition_latents = condition_latents + cond_attn_output
+        if model_config.get("add_cond_attn", False):
+            hidden_states += cond_attn_output
+
+    # LayerNorm + MLP.
+    # 1. hidden_states
+    norm_hidden_states = self.norm2(hidden_states)
+    norm_hidden_states = (
+        norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+    )
+    # 2. encoder_hidden_states
+    norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+    norm_encoder_hidden_states = (
+        norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+    )
+    # 3. condition_latents
+    if use_cond:
+        norm_condition_latents = self.norm2(condition_latents)
+        norm_condition_latents = (
+            norm_condition_latents * (1 + cond_scale_mlp[:, None])
+            + cond_shift_mlp[:, None]
+        )
+
+    # Feed-forward.
+    with enable_lora((self.ff.net[2],), model_config.get("latent_lora", False)):
+        # 1. hidden_states
+        ff_output = self.ff(norm_hidden_states)
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+    # 2. encoder_hidden_states
+    context_ff_output = self.ff_context(norm_encoder_hidden_states)
+    context_ff_output = c_gate_mlp.unsqueeze(1) * context_ff_output
+    # 3. condition_latents
+    if use_cond:
+        cond_ff_output = self.ff(norm_condition_latents)
+        cond_ff_output = cond_gate_mlp.unsqueeze(1) * cond_ff_output
+
+    # Process feed-forward outputs.
+    hidden_states = hidden_states + ff_output
+    encoder_hidden_states = encoder_hidden_states + context_ff_output
+    if use_cond:
+        condition_latents = condition_latents + cond_ff_output
+
+    # Clip to avoid overflow.
+    if encoder_hidden_states.dtype == torch.float16:
+        encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
+
+    return encoder_hidden_states, hidden_states, condition_latents if use_cond else None
+
+
+def single_block_forward(
+    self,
+    hidden_states: torch.FloatTensor,
+    temb: torch.FloatTensor,
+    image_rotary_emb=None,
+    condition_latents: torch.FloatTensor = None,
+    cond_temb: torch.FloatTensor = None,
+    cond_rotary_emb=None,
+    model_config: Optional[Dict[str, Any]] = {},
+    conceptrol: Conceptrol = None,
+):
+
+    using_cond = condition_latents is not None
+    residual = hidden_states
+    with enable_lora(
+        (
+            self.norm.linear,
+            self.proj_mlp,
+        ),
+        model_config.get("latent_lora", False),
+    ):
+        norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
+        mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
+    if using_cond:
+        residual_cond = condition_latents
+        norm_condition_latents, cond_gate = self.norm(condition_latents, emb=cond_temb)
+        mlp_cond_hidden_states = self.act_mlp(self.proj_mlp(norm_condition_latents))
+
+    attn_output = attn_forward(
+        self.attn,
+        model_config=model_config,
+        hidden_states=norm_hidden_states,
+        image_rotary_emb=image_rotary_emb,
+        **(
+            {
+                "condition_latents": norm_condition_latents,
+                "cond_rotary_emb": cond_rotary_emb if using_cond else None,
+                "conceptrol": conceptrol if using_cond else None,
+            }
+            if using_cond
+            else {}
+        ),
+    )
+    if using_cond:
+        attn_output, cond_attn_output = attn_output
+
+    with enable_lora((self.proj_out,), model_config.get("latent_lora", False)):
+        hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
+        gate = gate.unsqueeze(1)
+        hidden_states = gate * self.proj_out(hidden_states)
+        hidden_states = residual + hidden_states
+    if using_cond:
+        condition_latents = torch.cat([cond_attn_output, mlp_cond_hidden_states], dim=2)
+        cond_gate = cond_gate.unsqueeze(1)
+        condition_latents = cond_gate * self.proj_out(condition_latents)
+        condition_latents = residual_cond + condition_latents
+
+    if hidden_states.dtype == torch.float16:
+        hidden_states = hidden_states.clip(-65504, 65504)
+
+    return hidden_states if not using_cond else (hidden_states, condition_latents)

omini_control/conceptrol.py

@@ -0,0 +1,208 @@
+import os
+import math
+import numpy as np
+import matplotlib.pyplot as plt
+import torch
+
+
+class Conceptrol:
+    def __init__(self, config):
+        if "name" not in config:
+            raise KeyError("name has to be provided as 'conceptrol' or 'ominicontrol'")
+
+        name = config["name"]
+        if name not in ["conceptrol", "ominicontrol"]:
+            raise ValueError(
+                f"Name must be one of ['conceptrol', 'ominicontrol'], got {name}"
+            )
+
+        try:
+            log_attn_map = config["log_attn_map"]
+        except KeyError:
+            log_attn_map = False
+
+        # static
+        self.NUM_BLOCKS = 19  # this is fixed for FLUX
+        self.M = 512  # num of text tokens, fixed for FLUX
+        self.N = 1024  # num of latent / image condtion tokens, fixed for FLUX
+        self.EP = -10e6
+        self.CONCEPT_BLOCK_IDX = 18
+
+        # fixed during one generation
+        self.name = name
+
+        # variable during one generation
+        self.textual_concept_mask = None
+        self.forward_count = 0
+
+        # log out for visualization
+        if log_attn_map:
+            self.attn_maps = {"latent_to_concept": [], "latent_to_image": []}
+
+    def __call__(
+        self,
+        query: torch.FloatTensor,
+        key: torch.FloatTensor,
+        attention_mask: torch.Tensor,
+        c_factor: float = 1.0,
+    ) -> torch.Tensor:
+
+        if not hasattr(self, "textual_concept_idx"):
+            raise AttributeError(
+                "textual_concept_idx must be registered before calling Conceptrol"
+            )
+
+        # Skip computation for ominicontrol
+        if self.name == "ominicontrol":
+            scale_factor = 1 / math.sqrt(query.size(-1))
+            attention_weight = (
+                query @ key.transpose(-2, -1) * scale_factor + attention_mask
+            )
+            attention_probs = torch.softmax(
+                attention_weight, dim=-1
+            )  # [B, H, M+2N, M+2N]
+            return attention_probs
+
+        if not self.textual_concept_idx[0] < self.textual_concept_idx[1]:
+            raise ValueError(
+                f"register_idx[0] must be less than register_idx[1], "
+                f"got {self.textual_concept_idx[0]} >= {self.textual_concept_idx[1]}"
+            )
+
+        ### Reset attention mask predefined in ominicontrol
+        attention_mask = torch.zeros_like(attention_mask)
+        bias = torch.log(c_factor[0])
+        # attention of image condition to latent
+        attention_mask[-self.N :, self.M : -self.N] = bias
+        # attention of latent to image condition
+        attention_mask[self.M : -self.N, -self.N :] = bias
+
+        # attention of textual concept to image condition
+        attention_mask[
+            self.textual_concept_idx[0] : self.textual_concept_idx[1], -self.N :
+        ] = bias
+        # attention of other words to image condition (set as negative inf)
+        attention_mask[: self.textual_concept_idx[0], -self.N :] = self.EP
+        attention_mask[self.textual_concept_idx[1] : self.M, -self.N :] = self.EP
+
+        # If there is no textual_concept_mask, it means currently in layers previous to the first concept-specific block
+        if self.textual_concept_mask is None:
+            self.textual_concept_mask = (
+                torch.zeros_like(attention_mask).unsqueeze(0).unsqueeze(0)
+            )
+
+        ### Compute attention
+        scale_factor = 1 / math.sqrt(query.size(-1))
+        attention_weight = (
+            query @ key.transpose(-2, -1) * scale_factor
+            + attention_mask
+            + self.textual_concept_mask
+        )
+        # [B, H, M+2N, M+2N]
+        attention_probs = torch.softmax(attention_weight, dim=-1)
+
+        ### Extract textual concept mask if it's concept-specific block
+        is_concept_block = (
+            self.forward_count % self.NUM_BLOCKS == self.CONCEPT_BLOCK_IDX
+        )
+        if is_concept_block:
+            # Shape: [B, H, N, S], where S is the token numbers of the subject
+            textual_concept_mask_local = attention_probs[
+                :,
+                :,
+                self.M : -self.N,
+                self.textual_concept_idx[0] : self.textual_concept_idx[1],
+            ]
+            # Consider the ratio within context of text
+            textual_concept_mask_local = textual_concept_mask_local / torch.sum(
+                attention_probs[:, :, self.M : -self.N, : self.M], dim=-1, keepdim=True
+            )
+            # Average over words and head, Shape: [B, 1, N, 1]
+            textual_concept_mask_local = torch.mean(
+                textual_concept_mask_local, dim=(-1, 1), keepdim=True
+            )
+            # Normalize to average as 1
+            textual_concept_mask_local = textual_concept_mask_local / torch.mean(
+                textual_concept_mask_local, dim=-2, keepdim=True
+            )
+
+            self.textual_concept_mask = (
+                torch.zeros_like(attention_mask).unsqueeze(0).unsqueeze(0)
+            )
+            # log(A) in the paper
+            self.textual_concept_mask[:, :, self.M : -self.N, -self.N :] = torch.log(
+                textual_concept_mask_local
+            )
+
+        self.forward_count += 1
+
+        return attention_probs
+
+    def register(self, textual_concept_idx):
+        self.textual_concept_idx = textual_concept_idx
+
+    def visualize_attn_map(self, config_name: str, subject: str):
+        global global_concept_mask
+        global forward_count
+
+        save_dir = f"attn_maps/{config_name}/{subject}"
+        if not os.path.exists(save_dir):
+            os.makedirs(save_dir)
+        for attn_map_name, attn_maps in self.attn_maps.items():
+            if "token_to_token" in attn_map_name:
+                continue
+            plt.figure()
+
+            rows, cols = 8, 19
+            fig, axes = plt.subplots(
+                rows, cols, figsize=(64 * cols / 100, 64 * rows / 100)
+            )
+            fig.subplots_adjust(
+                wspace=0.1, hspace=0.1
+            )  # Adjust spacing between subplots
+
+            # Plot each array in the list on the grid
+            for i, ax in enumerate(axes.flatten()):
+                if i < len(attn_maps):  # Only plot existing arrays
+                    attn_map = attn_maps[i] / np.amax(attn_maps[i])
+                    ax.imshow(attn_map, cmap="viridis")
+                    ax.axis("off")  # Turn off axes for clarity
+                else:
+                    ax.axis("off")  # Turn off unused subplots
+
+            fig.set_size_inches(64 * cols / 100, 64 * rows / 100)
+            save_path = os.path.join(save_dir, f"{attn_map_name}.jpg")
+            plt.savefig(save_path)
+            plt.close()
+
+        for attn_map_name, attn_maps in self.attn_maps.items():
+            if "token_to_token" not in attn_map_name:
+                continue
+            plt.figure()
+
+            rows, cols = 8, 19
+            fig, axes = plt.subplots(
+                rows, cols, figsize=(2560 * cols / 100, 2560 * rows / 100)
+            )
+            fig.subplots_adjust(
+                wspace=0.1, hspace=0.1
+            )  # Adjust spacing between subplots
+
+            # Plot each array in the list on the grid
+            for i, ax in enumerate(axes.flatten()):
+                if i < len(attn_maps):  # Only plot existing arrays
+                    attn_map = attn_maps[i] / np.amax(attn_maps[i])
+                    ax.imshow(attn_map, cmap="viridis")
+                    ax.axis("off")  # Turn off axes for clarity
+                else:
+                    ax.axis("off")  # Turn off unused subplots
+
+            fig.set_size_inches(64 * cols / 100, 64 * rows / 100)
+            save_path = os.path.join(save_dir, f"{attn_map_name}.jpg")
+            plt.savefig(save_path)
+            plt.close()
+
+        for attn_map_name in self.attn_maps.keys():
+            self.attn_maps[attn_map_name] = []
+        global_concept_mask = None
+        forward_count = 0

omini_control/condition.py

@@ -0,0 +1,124 @@
+import torch
+from typing import Union, Tuple
+from diffusers.pipelines import FluxPipeline
+from PIL import Image, ImageFilter
+import numpy as np
+import cv2
+
+condition_dict = {
+    "depth": 0,
+    "canny": 1,
+    "subject": 4,
+    "coloring": 6,
+    "deblurring": 7,
+    "fill": 9,
+}
+
+
+class Condition(object):
+    def __init__(
+        self,
+        condition_type: str,
+        raw_img: Union[Image.Image, torch.Tensor] = None,
+        condition: Union[Image.Image, torch.Tensor] = None,
+        mask=None,
+    ) -> None:
+        self.condition_type = condition_type
+        assert raw_img is not None or condition is not None
+        if raw_img is not None:
+            self.condition = self.get_condition(condition_type, raw_img)
+        else:
+            self.condition = condition
+        # TODO: Add mask support
+        assert mask is None, "Mask not supported yet"
+
+    def get_condition(
+        self, condition_type: str, raw_img: Union[Image.Image, torch.Tensor]
+    ) -> Union[Image.Image, torch.Tensor]:
+        """
+        Returns the condition image.
+        """
+        if condition_type == "depth":
+            from transformers import pipeline
+
+            depth_pipe = pipeline(
+                task="depth-estimation",
+                model="LiheYoung/depth-anything-small-hf",
+                device="cuda",
+            )
+            source_image = raw_img.convert("RGB")
+            condition_img = depth_pipe(source_image)["depth"].convert("RGB")
+            return condition_img
+        elif condition_type == "canny":
+            img = np.array(raw_img)
+            edges = cv2.Canny(img, 100, 200)
+            edges = Image.fromarray(edges).convert("RGB")
+            return edges
+        elif condition_type == "subject":
+            return raw_img
+        elif condition_type == "coloring":
+            return raw_img.convert("L").convert("RGB")
+        elif condition_type == "deblurring":
+            condition_image = (
+                raw_img.convert("RGB")
+                .filter(ImageFilter.GaussianBlur(10))
+                .convert("RGB")
+            )
+            return condition_image
+        elif condition_type == "fill":
+            return raw_img.convert("RGB")
+        return self.condition
+
+    @property
+    def type_id(self) -> int:
+        """
+        Returns the type id of the condition.
+        """
+        return condition_dict[self.condition_type]
+
+    @classmethod
+    def get_type_id(cls, condition_type: str) -> int:
+        """
+        Returns the type id of the condition.
+        """
+        return condition_dict[condition_type]
+
+    def _encode_image(self, pipe: FluxPipeline, cond_img: Image.Image) -> torch.Tensor:
+        """
+        Encodes an image condition into tokens using the pipeline.
+        """
+        cond_img = pipe.image_processor.preprocess(cond_img)
+        cond_img = cond_img.to(pipe.device).to(pipe.dtype)
+        cond_img = pipe.vae.encode(cond_img).latent_dist.sample()
+        cond_img = (
+            cond_img - pipe.vae.config.shift_factor
+        ) * pipe.vae.config.scaling_factor
+        cond_tokens = pipe._pack_latents(cond_img, *cond_img.shape)
+        cond_ids = pipe._prepare_latent_image_ids(
+            cond_img.shape[0],
+            cond_img.shape[2],
+            cond_img.shape[3],
+            pipe.device,
+            pipe.dtype,
+        )
+        return cond_tokens, cond_ids
+
+    def encode(self, pipe: FluxPipeline) -> Tuple[torch.Tensor, torch.Tensor, int]:
+        """
+        Encodes the condition into tokens, ids and type_id.
+        """
+        if self.condition_type in [
+            "depth",
+            "canny",
+            "subject",
+            "coloring",
+            "deblurring",
+            "fill",
+        ]:
+            tokens, ids = self._encode_image(pipe, self.condition)
+        else:
+            raise NotImplementedError(
+                f"Condition type {self.condition_type} not implemented"
+            )
+        type_id = torch.ones_like(ids[:, :1]) * self.type_id
+        return tokens, ids, type_id

omini_control/flux_conceptrol_pipeline.py

@@ -0,0 +1,368 @@
+import torch
+from diffusers.pipelines import FluxPipeline
+from typing import List, Union, Optional, Dict, Any, Callable
+from .transformer import tranformer_forward
+from .condition import Condition
+from .conceptrol import Conceptrol
+
+from diffusers.pipelines.flux.pipeline_flux import (
+    FluxPipelineOutput,
+    calculate_shift,
+    retrieve_timesteps,
+    np,
+)
+
+denoising_images = []
+
+
+def prepare_params(
+    prompt: Union[str, List[str]] = None,
+    prompt_2: Optional[Union[str, List[str]]] = None,
+    height: Optional[int] = 512,
+    width: Optional[int] = 512,
+    num_inference_steps: int = 28,
+    timesteps: List[int] = None,
+    guidance_scale: float = 3.5,
+    num_images_per_prompt: Optional[int] = 1,
+    generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+    latents: Optional[torch.FloatTensor] = None,
+    prompt_embeds: Optional[torch.FloatTensor] = None,
+    pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+    output_type: Optional[str] = "pil",
+    return_dict: bool = True,
+    joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+    callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+    max_sequence_length: int = 512,
+    **kwargs: dict,
+):
+    return (
+        prompt,
+        prompt_2,
+        height,
+        width,
+        num_inference_steps,
+        timesteps,
+        guidance_scale,
+        num_images_per_prompt,
+        generator,
+        latents,
+        prompt_embeds,
+        pooled_prompt_embeds,
+        output_type,
+        return_dict,
+        joint_attention_kwargs,
+        callback_on_step_end,
+        callback_on_step_end_tensor_inputs,
+        max_sequence_length,
+    )
+
+
+def seed_everything(seed: int = 42):
+    torch.backends.cudnn.deterministic = True
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+
+
+def set_scale(pipe, condition_scale):
+    for name, module in pipe.transformer.named_modules():
+        if not name.endswith(".attn"):
+            continue
+        module.c_factor = torch.ones(1, 1) * condition_scale
+
+
+class FluxConceptrolPipeline(FluxPipeline):
+
+    def find_subsequence(self, text, sub):
+        sub_len = len(sub)
+        for i in range(len(text) - sub_len + 1):
+            if text[i : i + sub_len] == sub:
+                return i, i + sub_len  # Return start and end indices
+        return None
+
+    def locate_subject(self, prompt, subject, max_length=512):
+        text_inputs = self.tokenizer_2.tokenize(
+            prompt,
+            padding="max_length",
+            max_length=max_length,
+            truncation=True,
+            return_length=False,
+            return_overflowing_tokens=False,
+            return_tensors="pt",
+        )
+        subject_inputs = self.tokenizer_2.tokenize(
+            subject,
+            truncation=True,
+            return_length=False,
+            return_overflowing_tokens=False,
+            return_tensors="pt",
+        )
+        print("Text Inputs:", text_inputs)
+        print("Sbject Inputs:", subject_inputs)
+        print(self.find_subsequence(text_inputs, subject_inputs))
+        return self.find_subsequence(text_inputs, subject_inputs)
+
+        text_input_ids = text_inputs
+        return (
+            text_input_ids.index(subject_inputs[0]),
+            text_input_ids.index(subject_inputs[-1]) + 1,
+        )
+
+    def load_conceptrol(self, conceptrol):
+        self.conceptrol = conceptrol
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        image=None,
+        model_config: Optional[Dict[str, Any]] = {},
+        condition_scale: float = 1.0,
+        subject: Optional[str] = None,
+        control_guidance_start: float = 0.0,
+        control_guidance_end: float = 1.0,
+        conceptrol: Conceptrol = None,
+        seed: int = 42,
+        **params: dict,
+    ):
+        seed_everything(seed)
+
+        if conceptrol is None:
+            if not hasattr(self, "conceptrol"):
+                raise ValueError("Default conceptrol not loaded. Please call load_conceptrol() first.")
+            conceptrol = self.conceptrol
+
+        conditions = [Condition("subject", image.convert("RGB").resize((512, 512)))]
+        if condition_scale != 1:
+            for name, module in self.transformer.named_modules():
+                if not name.endswith(".attn"):
+                    continue
+                module.c_factor = torch.ones(1, 1) * condition_scale
+
+        (
+            prompt,
+            prompt_2,
+            height,
+            width,
+            num_inference_steps,
+            timesteps,
+            guidance_scale,
+            num_images_per_prompt,
+            generator,
+            latents,
+            prompt_embeds,
+            pooled_prompt_embeds,
+            output_type,
+            return_dict,
+            joint_attention_kwargs,
+            callback_on_step_end,
+            callback_on_step_end_tensor_inputs,
+            max_sequence_length,
+        ) = prepare_params(**params)
+
+        if subject is not None:
+            textual_concept_idx = self.locate_subject(params["prompt"], subject)
+        else:
+            raise ValueError("Subject has to be provided")
+
+        if textual_concept_idx is None:
+            raise ValueError("Textual concept idx has to be provided")
+
+        conceptrol.register(textual_concept_idx)
+
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+
+        # 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
+
+        lora_scale = (
+            self.joint_attention_kwargs.get("scale", None)
+            if self.joint_attention_kwargs is not None
+            else None
+        )
+        (
+            prompt_embeds,
+            pooled_prompt_embeds,
+            text_ids,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels // 4
+        latents, latent_image_ids = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 4.1. Prepare conditions
+        condition_latents, condition_ids, condition_type_ids = ([] for _ in range(3))
+        use_condition = conditions is not None or []
+        if use_condition:
+            assert len(conditions) <= 1, "Only one condition is supported for now."
+            self.set_adapters(conditions[0].condition_type)
+            for condition in conditions:
+                tokens, ids, type_id = condition.encode(self)
+                condition_latents.append(tokens)  # [batch_size, token_n, token_dim]
+                condition_ids.append(ids)  # [token_n, id_dim(3)]
+                condition_type_ids.append(type_id)  # [token_n, 1]
+            condition_latents = torch.cat(condition_latents, dim=1)
+            condition_ids = torch.cat(condition_ids, dim=0)
+            if condition.condition_type == "subject":
+                condition_ids[:, 2] += width // 16
+            condition_type_ids = torch.cat(condition_type_ids, dim=0)
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        image_seq_len = latents.shape[1]
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.base_image_seq_len,
+            self.scheduler.config.max_image_seq_len,
+            self.scheduler.config.base_shift,
+            self.scheduler.config.max_shift,
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(
+            len(timesteps) - num_inference_steps * self.scheduler.order, 0
+        )
+        self._num_timesteps = len(timesteps)
+
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if (i / len(timesteps) < control_guidance_start) or (
+                    (i + 1) / len(timesteps) > control_guidance_end
+                ):
+                    set_scale(self, 0.5)  # Warmup required for the first few steps
+                else:
+                    set_scale(self, condition_scale)
+                if self.interrupt:
+                    continue
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                # handle guidance
+                if self.transformer.config.guidance_embeds:
+                    guidance = torch.tensor([guidance_scale], device=device)
+                    guidance = guidance.expand(latents.shape[0])
+                else:
+                    guidance = None
+                noise_pred = tranformer_forward(
+                    self.transformer,
+                    model_config=model_config,
+                    conceptrol=conceptrol,
+                    # Inputs of the condition (new feature)
+                    condition_latents=condition_latents if use_condition else None,
+                    condition_ids=condition_ids if use_condition else None,
+                    condition_type_ids=condition_type_ids if use_condition else None,
+                    # Inputs to the original transformer
+                    hidden_states=latents,
+                    timestep=timestep / 1000,
+                    guidance=guidance,
+                    pooled_projections=pooled_prompt_embeds,
+                    encoder_hidden_states=prompt_embeds,
+                    txt_ids=text_ids,
+                    img_ids=latent_image_ids,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, return_dict=False
+                )[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(
+                        self, latents, callback_kwargs
+                    )
+
+                    global denoising_images
+                    denoising_images.append(callback_outputs)
+
+                # 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 output_type == "latent":
+            image = latents
+
+        else:
+            latents = self._unpack_latents(
+                latents, height, width, self.vae_scale_factor
+            )
+            latents = (
+                latents / self.vae.config.scaling_factor
+            ) + self.vae.config.shift_factor
+            image = self.vae.decode(latents, return_dict=False)[0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if condition_scale != 1:
+            for name, module in self.transformer.named_modules():
+                if not name.endswith(".attn"):
+                    continue
+                del module.c_factor
+
+        if not return_dict:
+            return (image,)
+
+        return FluxPipelineOutput(images=image)

omini_control/lora_controller.py

@@ -0,0 +1,75 @@
+from peft.tuners.tuners_utils import BaseTunerLayer
+from typing import List, Any, Optional, Type
+
+
+class enable_lora:
+    def __init__(self, lora_modules: List[BaseTunerLayer], activated: bool) -> None:
+        self.activated: bool = activated
+        if activated:
+            return
+        self.lora_modules: List[BaseTunerLayer] = [
+            each for each in lora_modules if isinstance(each, BaseTunerLayer)
+        ]
+        self.scales = [
+            {
+                active_adapter: lora_module.scaling[active_adapter]
+                for active_adapter in lora_module.active_adapters
+            }
+            for lora_module in self.lora_modules
+        ]
+
+    def __enter__(self) -> None:
+        if self.activated:
+            return
+
+        for lora_module in self.lora_modules:
+            if not isinstance(lora_module, BaseTunerLayer):
+                continue
+            lora_module.scale_layer(0)
+
+    def __exit__(
+        self,
+        exc_type: Optional[Type[BaseException]],
+        exc_val: Optional[BaseException],
+        exc_tb: Optional[Any],
+    ) -> None:
+        if self.activated:
+            return
+        for i, lora_module in enumerate(self.lora_modules):
+            if not isinstance(lora_module, BaseTunerLayer):
+                continue
+            for active_adapter in lora_module.active_adapters:
+                lora_module.scaling[active_adapter] = self.scales[i][active_adapter]
+
+
+class set_lora_scale:
+    def __init__(self, lora_modules: List[BaseTunerLayer], scale: float) -> None:
+        self.lora_modules: List[BaseTunerLayer] = [
+            each for each in lora_modules if isinstance(each, BaseTunerLayer)
+        ]
+        self.scales = [
+            {
+                active_adapter: lora_module.scaling[active_adapter]
+                for active_adapter in lora_module.active_adapters
+            }
+            for lora_module in self.lora_modules
+        ]
+        self.scale = scale
+
+    def __enter__(self) -> None:
+        for lora_module in self.lora_modules:
+            if not isinstance(lora_module, BaseTunerLayer):
+                continue
+            lora_module.scale_layer(self.scale)
+
+    def __exit__(
+        self,
+        exc_type: Optional[Type[BaseException]],
+        exc_val: Optional[BaseException],
+        exc_tb: Optional[Any],
+    ) -> None:
+        for i, lora_module in enumerate(self.lora_modules):
+            if not isinstance(lora_module, BaseTunerLayer):
+                continue
+            for active_adapter in lora_module.active_adapters:
+                lora_module.scaling[active_adapter] = self.scales[i][active_adapter]

omini_control/transformer.py

@@ -0,0 +1,273 @@
+import torch
+from typing import Optional, Dict, Any
+from .block import block_forward, single_block_forward
+from .lora_controller import enable_lora
+from .conceptrol import Conceptrol
+from diffusers.models.transformers.transformer_flux import (
+    FluxTransformer2DModel,
+    Transformer2DModelOutput,
+    USE_PEFT_BACKEND,
+    is_torch_version,
+    scale_lora_layers,
+    unscale_lora_layers,
+    logger,
+)
+import numpy as np
+
+
+def prepare_params(
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor = None,
+    pooled_projections: torch.Tensor = None,
+    timestep: torch.LongTensor = None,
+    img_ids: torch.Tensor = None,
+    txt_ids: torch.Tensor = None,
+    guidance: torch.Tensor = None,
+    joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    controlnet_block_samples=None,
+    controlnet_single_block_samples=None,
+    return_dict: bool = True,
+    **kwargs: dict,
+):
+    return (
+        hidden_states,
+        encoder_hidden_states,
+        pooled_projections,
+        timestep,
+        img_ids,
+        txt_ids,
+        guidance,
+        joint_attention_kwargs,
+        controlnet_block_samples,
+        controlnet_single_block_samples,
+        return_dict,
+    )
+
+
+def tranformer_forward(
+    transformer: FluxTransformer2DModel,
+    condition_latents: torch.Tensor,
+    condition_ids: torch.Tensor,
+    condition_type_ids: torch.Tensor,
+    model_config: Optional[Dict[str, Any]] = {},
+    return_conditional_latents: bool = False,
+    c_t=0,
+    conceptrol: Conceptrol = None,
+    **params: dict,
+):
+    self = transformer
+    use_condition = condition_latents is not None
+    use_condition_in_single_blocks = model_config.get(
+        "use_condition_in_single_blocks", True
+    )
+    # if return_conditional_latents is True, use_condition and use_condition_in_single_blocks must be True
+    assert not return_conditional_latents or (
+        use_condition and use_condition_in_single_blocks
+    ), "`return_conditional_latents` is True, `use_condition` and `use_condition_in_single_blocks` must be True"
+
+    (
+        hidden_states,
+        encoder_hidden_states,
+        pooled_projections,
+        timestep,
+        img_ids,
+        txt_ids,
+        guidance,
+        joint_attention_kwargs,
+        controlnet_block_samples,
+        controlnet_single_block_samples,
+        return_dict,
+    ) = prepare_params(**params)
+
+    if joint_attention_kwargs is not None:
+        joint_attention_kwargs = joint_attention_kwargs.copy()
+        lora_scale = joint_attention_kwargs.pop("scale", 1.0)
+    else:
+        lora_scale = 1.0
+
+    if USE_PEFT_BACKEND:
+        # weight the lora layers by setting `lora_scale` for each PEFT layer
+        scale_lora_layers(self, lora_scale)
+    else:
+        if (
+            joint_attention_kwargs is not None
+            and joint_attention_kwargs.get("scale", None) is not None
+        ):
+            logger.warning(
+                "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+            )
+    with enable_lora((self.x_embedder,), model_config.get("latent_lora", False)):
+        hidden_states = self.x_embedder(hidden_states)
+    condition_latents = self.x_embedder(condition_latents) if use_condition else None
+
+    timestep = timestep.to(hidden_states.dtype) * 1000
+    if guidance is not None:
+        guidance = guidance.to(hidden_states.dtype) * 1000
+    else:
+        guidance = None
+    temb = (
+        self.time_text_embed(timestep, pooled_projections)
+        if guidance is None
+        else self.time_text_embed(timestep, guidance, pooled_projections)
+    )
+    cond_temb = (
+        self.time_text_embed(torch.ones_like(timestep) * c_t * 1000, pooled_projections)
+        if guidance is None
+        else self.time_text_embed(
+            torch.ones_like(timestep) * c_t * 1000, guidance, pooled_projections
+        )
+    )
+    if hasattr(self, "cond_type_embed") and condition_type_ids is not None:
+        cond_type_proj = self.time_text_embed.time_proj(condition_type_ids[0])
+        cond_type_emb = self.cond_type_embed(cond_type_proj.to(dtype=cond_temb.dtype))
+        cond_temb = cond_temb + cond_type_emb
+    encoder_hidden_states = self.context_embedder(encoder_hidden_states)
+
+    if txt_ids.ndim == 3:
+        logger.warning(
+            "Passing `txt_ids` 3d torch.Tensor is deprecated."
+            "Please remove the batch dimension and pass it as a 2d torch Tensor"
+        )
+        txt_ids = txt_ids[0]
+    if img_ids.ndim == 3:
+        logger.warning(
+            "Passing `img_ids` 3d torch.Tensor is deprecated."
+            "Please remove the batch dimension and pass it as a 2d torch Tensor"
+        )
+        img_ids = img_ids[0]
+
+    ids = torch.cat((txt_ids, img_ids), dim=0)
+    image_rotary_emb = self.pos_embed(ids)
+    if use_condition:
+        cond_ids = condition_ids
+        cond_rotary_emb = self.pos_embed(cond_ids)
+
+    # hidden_states = torch.cat([hidden_states, condition_latents], dim=1)
+
+    for index_block, block in enumerate(self.transformer_blocks):
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module, return_dict=None):
+                def custom_forward(*inputs):
+                    if return_dict is not None:
+                        return module(*inputs, return_dict=return_dict)
+                    else:
+                        return module(*inputs)
+
+                return custom_forward
+
+            ckpt_kwargs: Dict[str, Any] = (
+                {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+            )
+            encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(block),
+                hidden_states,
+                encoder_hidden_states,
+                temb,
+                image_rotary_emb,
+                **ckpt_kwargs,
+            )
+
+        else:
+            encoder_hidden_states, hidden_states, condition_latents = block_forward(
+                block,
+                model_config=model_config,
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                condition_latents=condition_latents if use_condition else None,
+                temb=temb,
+                cond_temb=cond_temb if use_condition else None,
+                cond_rotary_emb=cond_rotary_emb if use_condition else None,
+                image_rotary_emb=image_rotary_emb,
+                conceptrol=conceptrol,
+            )
+
+        # controlnet residual
+        if controlnet_block_samples is not None:
+            interval_control = len(self.transformer_blocks) / len(
+                controlnet_block_samples
+            )
+            interval_control = int(np.ceil(interval_control))
+            hidden_states = (
+                hidden_states
+                + controlnet_block_samples[index_block // interval_control]
+            )
+    hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+    for index_block, block in enumerate(self.single_transformer_blocks):
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module, return_dict=None):
+                def custom_forward(*inputs):
+                    if return_dict is not None:
+                        return module(*inputs, return_dict=return_dict)
+                    else:
+                        return module(*inputs)
+
+                return custom_forward
+
+            ckpt_kwargs: Dict[str, Any] = (
+                {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+            )
+            hidden_states = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(block),
+                hidden_states,
+                temb,
+                image_rotary_emb,
+                **ckpt_kwargs,
+            )
+
+        else:
+            result = single_block_forward(
+                block,
+                model_config=model_config,
+                hidden_states=hidden_states,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+                **(
+                    {
+                        "condition_latents": condition_latents,
+                        "cond_temb": cond_temb,
+                        "cond_rotary_emb": cond_rotary_emb,
+                        "conceptrol": conceptrol,
+                    }
+                    if use_condition_in_single_blocks and use_condition
+                    else {}
+                ),
+            )
+            if use_condition_in_single_blocks and use_condition:
+                hidden_states, condition_latents = result
+            else:
+                hidden_states = result
+
+        # controlnet residual
+        if controlnet_single_block_samples is not None:
+            interval_control = len(self.single_transformer_blocks) / len(
+                controlnet_single_block_samples
+            )
+            interval_control = int(np.ceil(interval_control))
+            hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
+                hidden_states[:, encoder_hidden_states.shape[1] :, ...]
+                + controlnet_single_block_samples[index_block // interval_control]
+            )
+
+    hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
+
+    hidden_states = self.norm_out(hidden_states, temb)
+    output = self.proj_out(hidden_states)
+    if return_conditional_latents:
+        condition_latents = (
+            self.norm_out(condition_latents, cond_temb) if use_condition else None
+        )
+        condition_output = self.proj_out(condition_latents) if use_condition else None
+
+    if USE_PEFT_BACKEND:
+        # remove `lora_scale` from each PEFT layer
+        unscale_lora_layers(self, lora_scale)
+
+    if not return_dict:
+        return (
+            (output,) if not return_conditional_latents else (output, condition_output)
+        )
+
+    return Transformer2DModelOutput(sample=output)

requirements.txt

@@ -1,6 +1,9 @@
-accelerate
-diffusers
-invisible_watermark
-torch
 transformers
-xformers
+diffusers
+peft
+opencv-python
+protobuf
+sentencepiece
+gradio
+jupyter
+torchao

style.css

@@ -0,0 +1,95 @@
+h1 {
+    text-align: center;
+    justify-content: center;
+}
+
+[role="tabpanel"] {
+    border: 0
+}
+
+#duplicate-button {
+    margin: auto;
+    color: #fff;
+    background: #1565c0;
+    border-radius: 100vh;
+}
+
+.gradio-container {
+    max-width: 690px ! important;
+}
+
+.equal-height {
+    display: flex;
+    flex: 1;
+}
+
+.grid-container {
+    display: grid;
+    grid-template-columns: 1fr 1fr;  /* 两列宽度相等 */
+    gap: 20px;
+    height: 100%;  /* 确保容器高度为100% */
+}
+
+.grid-item {
+    display: flex;
+    flex-direction: column;
+    height: 100%;
+}
+
+.flex-grow {
+    flex-grow: 1;  /* 使该元素占据剩余的高度 */
+    display: flex;
+    flex-direction: column;
+}
+
+#share-btn-container {
+    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;
+    margin-left: auto;
+    margin-top: 0.35em;
+}
+
+div#share-btn-container>div {
+    flex-direction: row;
+    background: black;
+    align-items: center
+}
+
+#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;
+    font-size: 15px;
+}
+
+#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
+}

utils.py

@@ -0,0 +1,212 @@
+import os
+from typing import Optional
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+from lpips import LPIPS
+from PIL import Image
+from torchvision.transforms import Normalize
+
+
+def show_images_horizontally(
+    list_of_files: np.array, output_file: Optional[str] = None, interact: bool = False
+) -> None:
+    """
+    Visualize the list of images horizontally and save the figure as PNG.
+
+    Args:
+        list_of_files: The list of images as numpy array with shape (N, H, W, C).
+        output_file: The output file path to save the figure as PNG.
+        interact: Whether to show the figure interactively in Jupyter Notebook or not in Python.
+    """
+    number_of_files = len(list_of_files)
+
+    heights = [a[0].shape[0] for a in list_of_files]
+    widths = [a.shape[1] for a in list_of_files[0]]
+
+    fig_width = 8.0  # inches
+    fig_height = fig_width * sum(heights) / sum(widths)
+
+    # Create a figure with subplots
+    _, axs = plt.subplots(
+        1, number_of_files, figsize=(fig_width * number_of_files, fig_height)
+    )
+    plt.tight_layout()
+    for i in range(number_of_files):
+        _image = list_of_files[i]
+        axs[i].imshow(_image)
+        axs[i].axis("off")
+
+    # Save the figure as PNG
+    if interact:
+        plt.show()
+    else:
+        plt.savefig(output_file, bbox_inches="tight", pad_inches=0.25)
+
+
+def image_grids(images, rows=None, cols=None):
+    if not images:
+        raise ValueError("The image list is empty.")
+
+    n_images = len(images)
+    if cols is None:
+        cols = int(n_images**0.5)
+    if rows is None:
+        rows = (n_images + cols - 1) // cols
+
+    width, height = images[0].size
+    grid_width = cols * width
+    grid_height = rows * height
+
+    grid_image = Image.new("RGB", (grid_width, grid_height))
+
+    for i, image in enumerate(images):
+        row, col = divmod(i, cols)
+        grid_image.paste(image, (col * width, row * height))
+
+    return grid_image
+
+
+def save_image(image: np.array, file_name: str) -> None:
+    """
+    Save the image as JPG.
+
+    Args:
+        image: The input image as numpy array with shape (H, W, C).
+        file_name: The file name to save the image.
+    """
+    image = Image.fromarray(image)
+    image.save(file_name)
+
+
+def load_and_process_images(load_dir: str) -> np.array:
+    """
+    Load and process the images into numpy array from the directory.
+
+    Args:
+        load_dir: The directory to load the images.
+
+    Returns:
+        images: The images as numpy array with shape (N, H, W, C).
+    """
+    images = []
+    print(load_dir)
+    filenames = sorted(
+        os.listdir(load_dir), key=lambda x: int(x.split(".")[0])
+    )  # Ensure the files are sorted numerically
+    for filename in filenames:
+        if filename.endswith(".jpg"):
+            img = Image.open(os.path.join(load_dir, filename))
+            img_array = (
+                np.asarray(img) / 255.0
+            )  # Convert to numpy array and scale pixel values to [0, 1]
+            images.append(img_array)
+    return images
+
+
+def compute_lpips(images: np.array, lpips_model: LPIPS) -> np.array:
+    """
+    Compute the LPIPS of the input images.
+
+    Args:
+        images: The input images as numpy array with shape (N, H, W, C).
+        lpips_model: The LPIPS model used to compute perceptual distances.
+
+    Returns:
+        distances: The LPIPS of the input images.
+    """
+    # Get device of lpips_model
+    device = next(lpips_model.parameters()).device
+    device = str(device)
+
+    # Change the input images into tensor
+    images = torch.tensor(images).to(device).float()
+    images = torch.permute(images, (0, 3, 1, 2))
+    normalize = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+    images = normalize(images)
+
+    # Compute the LPIPS between each adjacent input images
+    distances = []
+    for i in range(images.shape[0]):
+        if i == images.shape[0] - 1:
+            break
+        img1 = images[i].unsqueeze(0)
+        img2 = images[i + 1].unsqueeze(0)
+        loss = lpips_model(img1, img2)
+        distances.append(loss.item())
+    distances = np.array(distances)
+    return distances
+
+
+def compute_gini(distances: np.array) -> float:
+    """
+    Compute the Gini index of the input distances.
+
+    Args:
+        distances: The input distances as numpy array.
+
+    Returns:
+        gini: The Gini index of the input distances.
+    """
+    if len(distances) < 2:
+        return 0.0  # Gini index is 0 for less than two elements
+
+    # Sort the list of distances
+    sorted_distances = sorted(distances)
+    n = len(sorted_distances)
+    mean_distance = sum(sorted_distances) / n
+
+    # Compute the sum of absolute differences
+    sum_of_differences = 0
+    for di in sorted_distances:
+        for dj in sorted_distances:
+            sum_of_differences += abs(di - dj)
+
+    # Normalize the sum of differences by the mean and the number of elements
+    gini = sum_of_differences / (2 * n * n * mean_distance)
+    return gini
+
+
+def compute_smoothness_and_consistency(images: np.array, lpips_model: LPIPS) -> tuple:
+    """
+    Compute the smoothness and efficiency of the input images.
+
+    Args:
+        images: The input images as numpy array with shape (N, H, W, C).
+        lpips_model: The LPIPS model used to compute perceptual distances.
+
+    Returns:
+        smoothness: One minus gini index of LPIPS of consecutive images.
+        consistency: The mean LPIPS of consecutive images.
+        max_inception_distance: The maximum LPIPS of consecutive images.
+    """
+    distances = compute_lpips(images, lpips_model)
+    smoothness = 1 - compute_gini(distances)
+    consistency = np.mean(distances)
+    max_inception_distance = np.max(distances)
+    return smoothness, consistency, max_inception_distance
+
+
+def separate_source_and_interpolated_images(images: np.array) -> tuple:
+    """
+    Separate the input images into source and interpolated images.
+    The input source is the start and end of the images, while the interpolated images are the rest.
+
+    Args:
+        images: The input images as numpy array with shape (N, H, W, C).
+
+    Returns:
+        source: The source images as numpy array with shape (2, H, W, C).
+        interpolation: The interpolated images as numpy array with shape (N-2, H, W, C).
+    """
+    # Check if the array has at least two elements
+    if len(images) < 2:
+        raise ValueError("The input array should have at least two elements.")
+
+    # Separate the array into two parts
+    # First part takes the first and last element
+    source = np.array([images[0], images[-1]])
+    # Second part takes the rest of the elements
+    interpolation = images[1:-1]
+    return source, interpolation