feat: 初始提交

PhotoMaker/feature-extractor/preprocessor_config.json

+{
+    "crop_size": 224,
+    "do_center_crop": true,
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_resize": true,
+    "feature_extractor_type": "CLIPFeatureExtractor",
+    "image_mean": [
+      0.48145466,
+      0.4578275,
+      0.40821073
+    ],
+    "image_std": [
+      0.26862954,
+      0.26130258,
+      0.27577711
+    ],
+    "resample": 3,
+    "size": 224
+  }

PhotoMaker/gradio_demo/app.py

+import torch
+import numpy as np
+import random
+import os
+import sys
+
+from diffusers.utils import load_image
+from diffusers import EulerDiscreteScheduler
+
+from huggingface_hub import hf_hub_download
+import spaces
+import gradio as gr
+
+from photomaker import PhotoMakerStableDiffusionXLPipeline
+
+from style_template import styles
+from aspect_ratio_template import aspect_ratios
+
+# global variable
+base_model_path = 'SG161222/RealVisXL_V4.0'
+try:
+    if torch.cuda.is_available():
+        device = "cuda"
+    elif sys.platform == "darwin" and torch.backends.mps.is_available():
+        device = "mps"
+    else:
+        device = "cpu"
+except:
+    device = "cpu"
+
+MAX_SEED = np.iinfo(np.int32).max
+STYLE_NAMES = list(styles.keys())
+DEFAULT_STYLE_NAME = "Photographic (Default)"
+ASPECT_RATIO_LABELS = list(aspect_ratios)
+DEFAULT_ASPECT_RATIO = ASPECT_RATIO_LABELS[0]
+
+# download PhotoMaker checkpoint to cache
+photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker", filename="photomaker-v1.bin", repo_type="model")
+
+torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+if device == "mps":
+    torch_dtype = torch.float16
+
+pipe = PhotoMakerStableDiffusionXLPipeline.from_pretrained(
+    base_model_path, 
+    torch_dtype=torch_dtype,
+    use_safetensors=True, 
+    variant="fp16",
+    # local_files_only=True,
+).to(device)
+
+pipe.load_photomaker_adapter(
+    os.path.dirname(photomaker_ckpt),
+    subfolder="",
+    weight_name=os.path.basename(photomaker_ckpt),
+    trigger_word="img",
+    pm_version="v1",
+)
+pipe.id_encoder.to(device)
+
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+# pipe.set_adapters(["photomaker"], adapter_weights=[1.0])
+pipe.fuse_lora()
+pipe.to(device)
+
+@spaces.GPU
+def generate_image(upload_images, prompt, negative_prompt, aspect_ratio_name, style_name, num_steps, style_strength_ratio, num_outputs, guidance_scale, seed, progress=gr.Progress(track_tqdm=True)):
+    # check the trigger word
+    image_token_id = pipe.tokenizer.convert_tokens_to_ids(pipe.trigger_word)
+    input_ids = pipe.tokenizer.encode(prompt)
+    if image_token_id not in input_ids:
+        raise gr.Error(f"Cannot find the trigger word '{pipe.trigger_word}' in text prompt! Please refer to step 2️⃣")
+
+    if input_ids.count(image_token_id) > 1:
+        raise gr.Error(f"Cannot use multiple trigger words '{pipe.trigger_word}' in text prompt!")
+
+    # determine output dimensions by the aspect ratio
+    output_w, output_h = aspect_ratios[aspect_ratio_name]
+    print(f"[Debug] Generate image using aspect ratio [{aspect_ratio_name}] => {output_w} x {output_h}")
+
+    # apply the style template
+    prompt, negative_prompt = apply_style(style_name, prompt, negative_prompt)
+
+    if upload_images is None:
+        raise gr.Error(f"Cannot find any input face image! Please refer to step 1️⃣")
+
+    input_id_images = []
+    for img in upload_images:
+        input_id_images.append(load_image(img))
+    
+    generator = torch.Generator(device=device).manual_seed(seed)
+
+    print("Start inference...")
+    print(f"[Debug] Prompt: {prompt}, \n[Debug] Neg Prompt: {negative_prompt}")
+    start_merge_step = int(float(style_strength_ratio) / 100 * num_steps)
+    if start_merge_step > 30:
+        start_merge_step = 30
+    print(start_merge_step)
+    images = pipe(
+        prompt=prompt,
+        width=output_w,
+        height=output_h,
+        input_id_images=input_id_images,
+        negative_prompt=negative_prompt,
+        num_images_per_prompt=num_outputs,
+        num_inference_steps=num_steps,
+        start_merge_step=start_merge_step,
+        generator=generator,
+        guidance_scale=guidance_scale,
+    ).images
+    return images, gr.update(visible=True)
+
+def swap_to_gallery(images):
+    return gr.update(value=images, visible=True), gr.update(visible=True), gr.update(visible=False)
+
+def upload_example_to_gallery(images, prompt, style, negative_prompt):
+    return gr.update(value=images, visible=True), gr.update(visible=True), gr.update(visible=False)
+
+def remove_back_to_files():
+    return gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)
+    
+def remove_tips():
+    return gr.update(visible=False)
+
+def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    return seed
+
+def apply_style(style_name: str, positive: str, negative: str = ""):
+    p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME])
+    return p.replace("{prompt}", positive), n + ' ' + negative
+
+def get_image_path_list(folder_name):
+    image_basename_list = os.listdir(folder_name)
+    image_path_list = sorted([os.path.join(folder_name, basename) for basename in image_basename_list])
+    return image_path_list
+
+def get_example():
+    case = [
+        [
+            get_image_path_list('./examples/scarletthead_woman'),
+            "instagram photo, portrait photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain",
+            "(No style)",
+            "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth",
+        ],
+        [
+            get_image_path_list('./examples/newton_man'),
+            "sci-fi, closeup portrait photo of a man img wearing the sunglasses in Iron man suit, face, slim body, high quality, film grain",
+            "(No style)",
+            "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth",
+        ],
+    ]
+    return case
+
+### Description and style
+logo = r"""
+<center><img src='https://photo-maker.github.io/assets/logo.png' alt='PhotoMaker logo' style="width:80px; margin-bottom:10px"></center>
+"""
+title = r"""
+<h1 align="center">PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding</h1>
+"""
+
+description = r"""
+<b>Official 🤗 Gradio demo</b> for <a href='https://github.com/TencentARC/PhotoMaker' target='_blank'><b>PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding</b></a>.<br>
+<br>
+For stylization, you could use our other gradio demo [PhotoMaker-Style](https://huggingface.co/spaces/TencentARC/PhotoMaker-Style).
+<br>
+❗️❗️❗️[<b>Important</b>] Personalization steps:<br>
+1️⃣ Upload images of someone you want to customize. One image is ok, but more is better.  Although we do not perform face detection, the face in the uploaded image should <b>occupy the majority of the image</b>.<br>
+2️⃣ Enter a text prompt, making sure to <b>follow the class word</b> you want to customize with the <b>trigger word</b>: `img`, such as: `man img` or `woman img` or `girl img`.<br>
+3️⃣ Choose your preferred style template.<br>
+4️⃣ Click the <b>Submit</b> button to start customizing.
+"""
+
+article = r"""
+
+If PhotoMaker is helpful, please help to ⭐ the <a href='https://github.com/TencentARC/PhotoMaker' target='_blank'>Github Repo</a>. Thanks! 
+[![GitHub Stars](https://img.shields.io/github/stars/TencentARC/PhotoMaker?style=social)](https://github.com/TencentARC/PhotoMaker)
+---
+📝 **Citation**
+<br>
+If our work is useful for your research, please consider citing:
+
+```bibtex
+@article{li2023photomaker,
+  title={PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding},
+  author={Li, Zhen and Cao, Mingdeng and Wang, Xintao and Qi, Zhongang and Cheng, Ming-Ming and Shan, Ying},
+  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
+  year={2024}
+}
+```
+📋 **License**
+<br>
+Apache-2.0 LICENSE. Please refer to the [LICENSE file](https://huggingface.co/TencentARC/PhotoMaker/blob/main/LICENSE) for details.
+
+📧 **Contact**
+<br>
+If you have any questions, please feel free to reach me out at <b>zhenli1031@gmail.com</b>.
+"""
+
+tips = r"""
+### Usage tips of PhotoMaker
+1. Upload more photos of the person to be customized to **improve ID fidelty**. If the input is Asian face(s), maybe consider adding 'asian' before the class word, e.g., `asian woman img`
+2. When stylizing, does the generated face look too realistic? Try switching to our **other gradio demo** [PhotoMaker-Style](https://huggingface.co/spaces/TencentARC/PhotoMaker-Style). Adjust the **Style strength** to 30-50, the larger the number, the less ID fidelty, but the stylization ability will be better.
+3. For **faster** speed, reduce the number of generated images and sampling steps. However, please note that reducing the sampling steps may compromise the ID fidelity.
+"""
+# We have provided some generate examples and comparisons at: [this website]().
+# 3. Don't make the prompt too long, as we will trim it if it exceeds 77 tokens. 
+# 4. When generating realistic photos, if it's not real enough, try switching to our other gradio application [PhotoMaker-Realistic]().
+
+css = '''
+.gradio-container {width: 85% !important}
+'''
+with gr.Blocks(css=css) as demo:
+    gr.Markdown(logo)
+    gr.Markdown(title)
+    gr.Markdown(description)
+    # gr.DuplicateButton(
+    #     value="Duplicate Space for private use ",
+    #     elem_id="duplicate-button",
+    #     visible=os.getenv("SHOW_DUPLICATE_BUTTON") == "1",
+    # )
+    with gr.Row():
+        with gr.Column():
+            files = gr.Files(
+                        label="Drag (Select) 1 or more photos of your face",
+                        file_types=["image"]
+                    )
+            uploaded_files = gr.Gallery(label="Your images", visible=False, columns=5, rows=1, height=200)
+            with gr.Column(visible=False) as clear_button:
+                remove_and_reupload = gr.ClearButton(value="Remove and upload new ones", components=files, size="sm")
+            prompt = gr.Textbox(label="Prompt",
+                       info="Try something like 'a photo of a man/woman img', 'img' is the trigger word.",
+                       placeholder="A photo of a [man/woman img]...")
+            style = gr.Dropdown(label="Style template", choices=STYLE_NAMES, value=DEFAULT_STYLE_NAME)
+            aspect_ratio = gr.Dropdown(label="Output aspect ratio", choices=ASPECT_RATIO_LABELS, value=DEFAULT_ASPECT_RATIO)
+            submit = gr.Button("Submit")
+
+            with gr.Accordion(open=False, label="Advanced Options"):
+                negative_prompt = gr.Textbox(
+                    label="Negative Prompt", 
+                    placeholder="low quality",
+                    value="nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry",
+                )
+                num_steps = gr.Slider( 
+                    label="Number of sample steps",
+                    minimum=20,
+                    maximum=100,
+                    step=1,
+                    value=50,
+                )
+                style_strength_ratio = gr.Slider(
+                    label="Style strength (%)",
+                    minimum=15,
+                    maximum=50,
+                    step=1,
+                    value=20,
+                )
+                num_outputs = gr.Slider(
+                    label="Number of output images",
+                    minimum=1,
+                    maximum=4,
+                    step=1,
+                    value=2,
+                )
+                guidance_scale = gr.Slider(
+                    label="Guidance scale",
+                    minimum=0.1,
+                    maximum=10.0,
+                    step=0.1,
+                    value=5,
+                )
+                seed = gr.Slider(
+                    label="Seed",
+                    minimum=0,
+                    maximum=MAX_SEED,
+                    step=1,
+                    value=0,
+                )
+                randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+        with gr.Column():
+            gallery = gr.Gallery(label="Generated Images")
+            usage_tips = gr.Markdown(label="Usage tips of PhotoMaker", value=tips ,visible=False)
+
+        files.upload(fn=swap_to_gallery, inputs=files, outputs=[uploaded_files, clear_button, files])
+        remove_and_reupload.click(fn=remove_back_to_files, outputs=[uploaded_files, clear_button, files])
+
+        submit.click(
+            fn=remove_tips,
+            outputs=usage_tips,            
+        ).then(
+            fn=randomize_seed_fn,
+            inputs=[seed, randomize_seed],
+            outputs=seed,
+            queue=False,
+            api_name=False,
+        ).then(
+            fn=generate_image,
+            inputs=[files, prompt, negative_prompt, aspect_ratio, style, num_steps, style_strength_ratio, num_outputs, guidance_scale, seed],
+            outputs=[gallery, usage_tips]
+        )
+
+    gr.Examples(
+        examples=get_example(),
+        inputs=[files, prompt, style, negative_prompt],
+        run_on_click=True,
+        fn=upload_example_to_gallery,
+        outputs=[uploaded_files, clear_button, files],
+    )
+    
+    gr.Markdown(article)
+    
+demo.launch()

PhotoMaker/gradio_demo/app_v2.py

+import torch
+import torchvision.transforms.functional as TF
+import numpy as np
+import random
+import os
+import sys
+
+from diffusers.utils import load_image
+from diffusers import EulerDiscreteScheduler, T2IAdapter
+
+from huggingface_hub import hf_hub_download
+import spaces
+import gradio as gr
+
+from photomaker import PhotoMakerStableDiffusionXLAdapterPipeline
+from photomaker import FaceAnalysis2, analyze_faces
+
+from style_template import styles
+from aspect_ratio_template import aspect_ratios
+
+# global variable
+base_model_path = 'SG161222/RealVisXL_V4.0'
+face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
+face_detector.prepare(ctx_id=0, det_size=(640, 640))
+
+try:
+    if torch.cuda.is_available():
+        device = "cuda"
+    elif sys.platform == "darwin" and torch.backends.mps.is_available():
+        device = "mps"
+    else:
+        device = "cpu"
+except:
+    device = "cpu"
+
+MAX_SEED = np.iinfo(np.int32).max
+STYLE_NAMES = list(styles.keys())
+DEFAULT_STYLE_NAME = "Photographic (Default)"
+ASPECT_RATIO_LABELS = list(aspect_ratios)
+DEFAULT_ASPECT_RATIO = ASPECT_RATIO_LABELS[0]
+
+enable_doodle_arg = False
+photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker-V2", filename="photomaker-v2.bin", repo_type="model")
+
+torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+if device == "mps":
+    torch_dtype = torch.float16
+    
+# load adapter
+adapter = T2IAdapter.from_pretrained(
+    "TencentARC/t2i-adapter-sketch-sdxl-1.0", torch_dtype=torch_dtype, variant="fp16"
+).to(device)
+
+pipe = PhotoMakerStableDiffusionXLAdapterPipeline.from_pretrained(
+    base_model_path, 
+    adapter=adapter, 
+    torch_dtype=torch_dtype,
+    use_safetensors=True, 
+    variant="fp16",
+).to(device)
+
+pipe.load_photomaker_adapter(
+    os.path.dirname(photomaker_ckpt),
+    subfolder="",
+    weight_name=os.path.basename(photomaker_ckpt),
+    trigger_word="img",
+    pm_version="v2",
+)
+pipe.id_encoder.to(device)
+
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+# pipe.set_adapters(["photomaker"], adapter_weights=[1.0])
+pipe.fuse_lora()
+pipe.to(device)
+
+
+@spaces.GPU
+def generate_image(
+    upload_images, 
+    prompt, 
+    negative_prompt, 
+    aspect_ratio_name, 
+    style_name, 
+    num_steps, 
+    style_strength_ratio, 
+    num_outputs, 
+    guidance_scale, 
+    seed, 
+    use_doodle,
+    sketch_image,
+    adapter_conditioning_scale,
+    adapter_conditioning_factor,
+    progress=gr.Progress(track_tqdm=True)
+):
+    if use_doodle:
+        sketch_image = sketch_image["composite"]
+        r, g, b, a = sketch_image.split()
+        sketch_image = a.convert("RGB")
+        sketch_image = TF.to_tensor(sketch_image) > 0.5 # Inversion 
+        sketch_image = TF.to_pil_image(sketch_image.to(torch.float32))
+        adapter_conditioning_scale = adapter_conditioning_scale
+        adapter_conditioning_factor = adapter_conditioning_factor
+    else:
+        adapter_conditioning_scale = 0.
+        adapter_conditioning_factor = 0.
+        sketch_image = None
+
+    # check the trigger word
+    image_token_id = pipe.tokenizer.convert_tokens_to_ids(pipe.trigger_word)
+    input_ids = pipe.tokenizer.encode(prompt)
+    if image_token_id not in input_ids:
+        raise gr.Error(f"Cannot find the trigger word '{pipe.trigger_word}' in text prompt! Please refer to step 2️⃣")
+
+    if input_ids.count(image_token_id) > 1:
+        raise gr.Error(f"Cannot use multiple trigger words '{pipe.trigger_word}' in text prompt!")
+
+    # determine output dimensions by the aspect ratio
+    output_w, output_h = aspect_ratios[aspect_ratio_name]
+    print(f"[Debug] Generate image using aspect ratio [{aspect_ratio_name}] => {output_w} x {output_h}")
+
+    # apply the style template
+    prompt, negative_prompt = apply_style(style_name, prompt, negative_prompt)
+
+    if upload_images is None:
+        raise gr.Error(f"Cannot find any input face image! Please refer to step 1️⃣")
+
+    input_id_images = []
+    for img in upload_images:
+        input_id_images.append(load_image(img))
+    
+    id_embed_list = []
+
+    for img in input_id_images:
+        img = np.array(img)
+        img = img[:, :, ::-1]
+        faces = analyze_faces(face_detector, img)
+        if len(faces) > 0:
+            id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
+
+    if len(id_embed_list) == 0:
+        raise gr.Error(f"No face detected, please update the input face image(s)")
+    
+    id_embeds = torch.stack(id_embed_list)
+
+    generator = torch.Generator(device=device).manual_seed(seed)
+
+    print("Start inference...")
+    print(f"[Debug] Seed: {seed}")
+    print(f"[Debug] Prompt: {prompt}, \n[Debug] Neg Prompt: {negative_prompt}")
+    start_merge_step = int(float(style_strength_ratio) / 100 * num_steps)
+    if start_merge_step > 30:
+        start_merge_step = 30
+    print(start_merge_step)
+    images = pipe(
+        prompt=prompt,
+        width=output_w,
+        height=output_h,
+        input_id_images=input_id_images,
+        negative_prompt=negative_prompt,
+        num_images_per_prompt=num_outputs,
+        num_inference_steps=num_steps,
+        start_merge_step=start_merge_step,
+        generator=generator,
+        guidance_scale=guidance_scale,
+        id_embeds=id_embeds,
+        image=sketch_image,
+        adapter_conditioning_scale=adapter_conditioning_scale,
+        adapter_conditioning_factor=adapter_conditioning_factor,
+    ).images
+    return images, gr.update(visible=True)
+
+def swap_to_gallery(images):
+    return gr.update(value=images, visible=True), gr.update(visible=True), gr.update(visible=False)
+
+def upload_example_to_gallery(images, prompt, style, negative_prompt):
+    return gr.update(value=images, visible=True), gr.update(visible=True), gr.update(visible=False)
+
+def remove_back_to_files():
+    return gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)
+    
+def change_doodle_space(use_doodle):
+    if use_doodle:
+        return gr.update(visible=True)
+    else:
+        return gr.update(visible=False)
+
+def remove_tips():
+    return gr.update(visible=False)
+
+def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    return seed
+
+def apply_style(style_name: str, positive: str, negative: str = "") -> tuple[str, str]:
+    p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME])
+    return p.replace("{prompt}", positive), n + ' ' + negative
+
+def get_image_path_list(folder_name):
+    image_basename_list = os.listdir(folder_name)
+    image_path_list = sorted([os.path.join(folder_name, basename) for basename in image_basename_list])
+    return image_path_list
+
+def get_example():
+    case = [
+        [
+            get_image_path_list('./examples/scarletthead_woman'),
+            "instagram photo, portrait photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain",
+            "(No style)",
+            "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth",
+        ],
+        [
+            get_image_path_list('./examples/newton_man'),
+            "sci-fi, closeup portrait photo of a man img wearing the sunglasses in Iron man suit, face, slim body, high quality, film grain",
+            "(No style)",
+            "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth",
+        ],
+    ]
+    return case
+
+### Description and style
+logo = r"""
+<center><img src='https://photo-maker.github.io/assets/logo.png' alt='PhotoMaker logo' style="width:80px; margin-bottom:10px"></center>
+"""
+title = r"""
+<h1 align="center">PhotoMaker V2: Improved ID Fidelity and Better Controllability than PhotoMaker V1</h1>
+"""
+
+description = r"""
+<b>Official 🤗 Gradio demo</b> for <a href='https://github.com/TencentARC/PhotoMaker' target='_blank'><b>PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding</b></a>.<br>
+The details of PhotoMaker V2 can be found in 
+<br>
+<br>
+For previous version of PhotoMaker, you could use our original gradio demos [PhotoMaker](https://huggingface.co/spaces/TencentARC/PhotoMaker) and [PhotoMaker-Style](https://huggingface.co/spaces/TencentARC/PhotoMaker-Style).
+<br>
+❗️❗️❗️[<b>Important</b>] Personalization steps:<br>
+1️⃣ Upload images of someone you want to customize. One image is ok, but more is better.  Although we do not perform face detection, the face in the uploaded image should <b>occupy the majority of the image</b>.<br>
+2️⃣ Enter a text prompt, making sure to <b>follow the class word</b> you want to customize with the <b>trigger word</b>: `img`, such as: `man img` or `woman img` or `girl img`.<br>
+3️⃣ Choose your preferred style template.<br>
+4️⃣ <b>(Optional: but new feature)</b> Select the ‘Enable Drawing Doodle...’ option and draw on the canvas<br>
+5️⃣ Click the <b>Submit</b> button to start customizing.
+"""
+
+article = r"""
+
+If PhotoMaker V2 is helpful, please help to ⭐ the <a href='https://github.com/TencentARC/PhotoMaker' target='_blank'>Github Repo</a>. Thanks! 
+[![GitHub Stars](https://img.shields.io/github/stars/TencentARC/PhotoMaker?style=social)](https://github.com/TencentARC/PhotoMaker)
+---
+📝 **Citation**
+<br>
+If our work is useful for your research, please consider citing:
+
+```bibtex
+@article{li2023photomaker,
+  title={PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding},
+  author={Li, Zhen and Cao, Mingdeng and Wang, Xintao and Qi, Zhongang and Cheng, Ming-Ming and Shan, Ying},
+  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
+  year={2024}
+}
+```
+📋 **License**
+<br>
+Apache-2.0 LICENSE. Please refer to the [LICENSE file](https://huggingface.co/TencentARC/PhotoMaker/blob/main/LICENSE) for details.
+
+📧 **Contact**
+<br>
+If you have any questions, please feel free to reach me out at <b>zhenli1031@gmail.com</b>.
+"""
+
+tips = r"""
+### Usage tips of PhotoMaker
+1. Upload **more photos**of the person to be customized to **improve ID fidelty**.
+2. If you find that the image quality is poor when using doodle for control, you can reduce the conditioning scale and factor of the adapter.
+Besides, you could enlarge the ratio for more consistency of your doodle. <br>
+If you have any issues, leave the issue in the discussion page of the space. For a more stable (queue-free) experience, you can duplicate the space.
+"""
+# We have provided some generate examples and comparisons at: [this website]().
+
+css = '''
+.gradio-container {width: 85% !important}
+'''
+with gr.Blocks(css=css) as demo:
+    gr.Markdown(logo)
+    gr.Markdown(title)
+    gr.Markdown(description)
+    # gr.DuplicateButton(
+    #     value="Duplicate Space for private use ",
+    #     elem_id="duplicate-button",
+    #     visible=os.getenv("SHOW_DUPLICATE_BUTTON") == "1",
+    # )
+    with gr.Row():
+        with gr.Column():
+            files = gr.Files(
+                        label="Drag (Select) 1 or more photos of your face",
+                        file_types=["image"]
+                    )
+            uploaded_files = gr.Gallery(label="Your images", visible=False, columns=5, rows=1, height=200)
+            with gr.Column(visible=False) as clear_button:
+                remove_and_reupload = gr.ClearButton(value="Remove and upload new ones", components=files, size="sm")
+            prompt = gr.Textbox(label="Prompt",
+                       info="Try something like 'a photo of a man/woman img', 'img' is the trigger word.",
+                       placeholder="A photo of a [man/woman img]...")
+            style = gr.Dropdown(label="Style template", choices=STYLE_NAMES, value=DEFAULT_STYLE_NAME)
+            aspect_ratio = gr.Dropdown(label="Output aspect ratio", choices=ASPECT_RATIO_LABELS, value=DEFAULT_ASPECT_RATIO)
+            submit = gr.Button("Submit")
+
+            enable_doodle = gr.Checkbox(
+                label="Enable Drawing Doodle for Control", value=enable_doodle_arg,
+                info="After enabling this option, PhotoMaker will generate content based on your doodle on the canvas, driven by the T2I-Adapter (Quality may be decreased)",
+            )
+            with gr.Accordion("T2I-Adapter-Doodle (Optional)", visible=False) as doodle_space:
+                with gr.Row():
+                    sketch_image = gr.Sketchpad(
+                        label="Canvas",
+                        type="pil",
+                        crop_size=[1024,1024],
+                        layers=False,
+                        canvas_size=(350, 350),
+                        brush=gr.Brush(default_size=5, colors=["#000000"], color_mode="fixed")
+                    )
+                    with gr.Group():
+                        adapter_conditioning_scale = gr.Slider(
+                            label="Adapter conditioning scale",
+                            minimum=0.5,
+                            maximum=1,
+                            step=0.1,
+                            value=0.7,
+                        )
+                        adapter_conditioning_factor = gr.Slider(
+                            label="Adapter conditioning factor",
+                            info="Fraction of timesteps for which adapter should be applied",
+                            minimum=0.5,
+                            maximum=1,
+                            step=0.1,
+                            value=0.8,
+                        )
+            with gr.Accordion(open=False, label="Advanced Options"):
+                negative_prompt = gr.Textbox(
+                    label="Negative Prompt", 
+                    placeholder="low quality",
+                    value="nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry",
+                )
+                num_steps = gr.Slider( 
+                    label="Number of sample steps",
+                    minimum=20,
+                    maximum=100,
+                    step=1,
+                    value=50,
+                )
+                style_strength_ratio = gr.Slider(
+                    label="Style strength (%)",
+                    minimum=15,
+                    maximum=50,
+                    step=1,
+                    value=20,
+                )
+                num_outputs = gr.Slider(
+                    label="Number of output images",
+                    minimum=1,
+                    maximum=4,
+                    step=1,
+                    value=2,
+                )
+                guidance_scale = gr.Slider(
+                    label="Guidance scale",
+                    minimum=0.1,
+                    maximum=10.0,
+                    step=0.1,
+                    value=5,
+                )
+                seed = gr.Slider(
+                    label="Seed",
+                    minimum=0,
+                    maximum=MAX_SEED,
+                    step=1,
+                    value=0,
+                )
+                randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+        with gr.Column():
+            gallery = gr.Gallery(label="Generated Images")
+            usage_tips = gr.Markdown(label="Usage tips of PhotoMaker", value=tips ,visible=False)
+
+        files.upload(fn=swap_to_gallery, inputs=files, outputs=[uploaded_files, clear_button, files])
+        remove_and_reupload.click(fn=remove_back_to_files, outputs=[uploaded_files, clear_button, files])
+        enable_doodle.select(fn=change_doodle_space, inputs=enable_doodle, outputs=doodle_space)
+
+        input_list = [
+            files, 
+            prompt, 
+            negative_prompt, 
+            aspect_ratio, 
+            style, 
+            num_steps, 
+            style_strength_ratio, 
+            num_outputs, 
+            guidance_scale, 
+            seed,
+            enable_doodle,
+            sketch_image,
+            adapter_conditioning_scale,
+            adapter_conditioning_factor
+        ]
+
+        submit.click(
+            fn=remove_tips,
+            outputs=usage_tips,            
+        ).then(
+            fn=randomize_seed_fn,
+            inputs=[seed, randomize_seed],
+            outputs=seed,
+            queue=False,
+            api_name=False,
+        ).then(
+            fn=generate_image,
+            inputs=input_list,
+            outputs=[gallery, usage_tips]
+        )
+
+    gr.Examples(
+        examples=get_example(),
+        inputs=[files, prompt, style, negative_prompt],
+        run_on_click=True,
+        fn=upload_example_to_gallery,
+        outputs=[uploaded_files, clear_button, files],
+    )
+    
+    gr.Markdown(article)
+    
+demo.launch()

PhotoMaker/gradio_demo/aspect_ratio_template.py

+# Note: Since output width & height need to be divisible by 8, the w & h -values do
+#       not exactly match the stated aspect ratios... but they are "close enough":)
+
+aspect_ratio_list = [
+    {
+        "name": "Instagram (1:1)",
+        "w": 1024,
+        "h": 1024,
+    },
+    {
+        "name": "35mm film / Landscape (3:2)",
+        "w": 1024,
+        "h": 680,
+    },
+    {
+        "name": "35mm film / Portrait (2:3)",
+        "w": 680,
+        "h": 1024,
+    },
+    {
+        "name": "CRT Monitor / Landscape (4:3)",
+        "w": 1024,
+        "h": 768,
+    },
+    {
+        "name": "CRT Monitor / Portrait (3:4)",
+        "w": 768,
+        "h": 1024,
+    },
+    {
+        "name": "Widescreen TV / Landscape (16:9)",
+        "w": 1024,
+        "h": 576,
+    },
+    {
+        "name": "Widescreen TV / Portrait (9:16)",
+        "w": 576,
+        "h": 1024,
+    },
+    {
+        "name": "Widescreen Monitor / Landscape (16:10)",
+        "w": 1024,
+        "h": 640,
+    },
+    {
+        "name": "Widescreen Monitor / Portrait (10:16)",
+        "w": 640,
+        "h": 1024,
+    },
+    {
+        "name": "Cinemascope (2.39:1)",
+        "w": 1024,
+        "h": 424,
+    },
+    {
+        "name": "Widescreen Movie (1.85:1)",
+        "w": 1024,
+        "h": 552,
+    },
+    {
+        "name": "Academy Movie (1.37:1)",
+        "w": 1024,
+        "h": 744,
+    },
+    {
+        "name": "Sheet-print (A-series) / Landscape (297:210)",
+        "w": 1024,
+        "h": 720,
+    },
+    {
+        "name": "Sheet-print (A-series) / Portrait (210:297)",
+        "w": 720,
+        "h": 1024,
+    },
+]
+
+aspect_ratios = {k["name"]: (k["w"], k["h"]) for k in aspect_ratio_list}

PhotoMaker/gradio_demo/style_template.py

+style_list = [
+    {
+        "name": "(No style)",
+        "prompt": "{prompt}",
+        "negative_prompt": "",
+    },
+    {
+        "name": "Cinematic",
+        "prompt": "cinematic still {prompt} . emotional, harmonious, vignette, highly detailed, high budget, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainy",
+        "negative_prompt": "anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured",
+    },
+    {
+        "name": "Disney Charactor",
+        "prompt": "A Pixar animation character of {prompt} . pixar-style, studio anime, Disney, high-quality",
+        "negative_prompt": "lowres, bad anatomy, bad hands, text, bad eyes, bad arms, bad legs, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, blurry, grayscale, noisy, sloppy, messy, grainy, highly detailed, ultra textured, photo",
+    },
+    {
+        "name": "Digital Art",
+        "prompt": "concept art {prompt} . digital artwork, illustrative, painterly, matte painting, highly detailed",
+        "negative_prompt": "photo, photorealistic, realism, ugly",
+    },
+    {
+        "name": "Photographic (Default)",
+        "prompt": "cinematic photo {prompt} . 35mm photograph, film, bokeh, professional, 4k, highly detailed",
+        "negative_prompt": "drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly",
+    },
+    {
+        "name": "Fantasy art",
+        "prompt": "ethereal fantasy concept art of  {prompt} . magnificent, celestial, ethereal, painterly, epic, majestic, magical, fantasy art, cover art, dreamy",
+        "negative_prompt": "photographic, realistic, realism, 35mm film, dslr, cropped, frame, text, deformed, glitch, noise, noisy, off-center, deformed, cross-eyed, closed eyes, bad anatomy, ugly, disfigured, sloppy, duplicate, mutated, black and white",
+    },
+    {
+        "name": "Neonpunk",
+        "prompt": "neonpunk style {prompt} . cyberpunk, vaporwave, neon, vibes, vibrant, stunningly beautiful, crisp, detailed, sleek, ultramodern, magenta highlights, dark purple shadows, high contrast, cinematic, ultra detailed, intricate, professional",
+        "negative_prompt": "painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured",
+    },
+    {
+        "name": "Enhance",
+        "prompt": "breathtaking {prompt} . award-winning, professional, highly detailed",
+        "negative_prompt": "ugly, deformed, noisy, blurry, distorted, grainy",
+    },
+    {
+        "name": "Comic book",
+        "prompt": "comic {prompt} . graphic illustration, comic art, graphic novel art, vibrant, highly detailed",
+        "negative_prompt": "photograph, deformed, glitch, noisy, realistic, stock photo",
+    },
+    {
+        "name": "Lowpoly",
+        "prompt": "low-poly style {prompt} . low-poly game art, polygon mesh, jagged, blocky, wireframe edges, centered composition",
+        "negative_prompt": "noisy, sloppy, messy, grainy, highly detailed, ultra textured, photo",
+    },
+    {
+        "name": "Line art",
+        "prompt": "line art drawing {prompt} . professional, sleek, modern, minimalist, graphic, line art, vector graphics",
+        "negative_prompt": "anime, photorealistic, 35mm film, deformed, glitch, blurry, noisy, off-center, deformed, cross-eyed, closed eyes, bad anatomy, ugly, disfigured, mutated, realism, realistic, impressionism, expressionism, oil, acrylic",
+    }
+]
+
+styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in style_list}
\ No newline at end of file

PhotoMaker/inference_scripts/inference_pmv2.py

+# !pip install opencv-python transformers accelerate
+import os
+import sys
+
+import numpy as np
+import torch
+from diffusers.utils import load_image
+from diffusers import EulerDiscreteScheduler
+from huggingface_hub import hf_hub_download
+from photomaker import PhotoMakerStableDiffusionXLPipeline
+from photomaker import FaceAnalysis2, analyze_faces
+
+face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
+face_detector.prepare(ctx_id=0, det_size=(640, 640))
+
+try:
+    if torch.cuda.is_available():
+        device = "cuda"
+    elif sys.platform == "darwin" and torch.backends.mps.is_available():
+        device = "mps"
+    else:
+        device = "cpu"
+except:
+    device = "cpu"
+
+torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+if device == "mps":
+    torch_dtype = torch.float16
+
+output_dir = "./outputs"
+os.makedirs(output_dir, exist_ok=True)
+
+photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker-V2", filename="photomaker-v2.bin", repo_type="model")
+
+prompt = "instagram photo, portrait photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain, best quality"
+negative_prompt = "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth"
+
+# initialize the models and pipeline
+### Load base model
+pipe = PhotoMakerStableDiffusionXLPipeline.from_pretrained(
+    "SG161222/RealVisXL_V4.0", torch_dtype=torch_dtype
+).to("cuda")
+
+### Load PhotoMaker checkpoint
+pipe.load_photomaker_adapter(
+    os.path.dirname(photomaker_path),
+    subfolder="",
+    weight_name=os.path.basename(photomaker_path),
+    trigger_word="img"  # define the trigger word
+)     
+### Also can cooperate with other LoRA modules
+# pipe.load_lora_weights(os.path.dirname(lora_path), weight_name=lora_model_name, adapter_name="lcm-lora")
+# pipe.set_adapters(["photomaker", "lcm-lora"], adapter_weights=[1.0, 0.5])
+
+pipe.fuse_lora()
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+
+pipe.enable_model_cpu_offload()
+
+### define the input ID images
+input_folder_name = './examples/scarletthead_woman'
+image_basename_list = os.listdir(input_folder_name)
+image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
+
+input_id_images = []
+for image_path in image_path_list:
+    input_id_images.append(load_image(image_path))
+
+id_embed_list = []
+
+for img in input_id_images:
+    img = np.array(img)
+    img = img[:, :, ::-1]
+    faces = analyze_faces(face_detector, img)
+    if len(faces) > 0:
+        id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
+
+if len(id_embed_list) == 0:
+    raise ValueError(f"No face detected in input image pool")
+
+id_embeds = torch.stack(id_embed_list)
+
+# generate image
+images = pipe(
+    prompt, 
+    negative_prompt=negative_prompt, 
+    input_id_images=input_id_images,
+    id_embeds=id_embeds,
+    num_images_per_prompt=2,
+    start_merge_step=10,
+).images
+
+for idx, img in enumerate(images): 
+    img.save(os.path.join(output_dir, f"output_pmv2_{idx}.jpg"))

PhotoMaker/inference_scripts/inference_pmv2_contronet.py

+# !pip install opencv-python transformers accelerate
+import os
+import sys
+
+import numpy as np
+import torch
+from diffusers.utils import load_image
+from diffusers import EulerDiscreteScheduler, ControlNetModel
+
+from huggingface_hub import hf_hub_download
+from controlnet_aux import OpenposeDetector
+from photomaker import PhotoMakerStableDiffusionXLControlNetPipeline
+from photomaker import FaceAnalysis2, analyze_faces
+
+face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
+face_detector.prepare(ctx_id=0, det_size=(640, 640))
+
+try:
+    if torch.cuda.is_available():
+        device = "cuda"
+    elif sys.platform == "darwin" and torch.backends.mps.is_available():
+        device = "mps"
+    else:
+        device = "cpu"
+except:
+    device = "cpu"
+
+torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+if device == "mps":
+    torch_dtype = torch.float16
+
+output_dir = "./outputs"
+os.makedirs(output_dir, exist_ok=True)
+
+photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker-V2", filename="photomaker-v2.bin", repo_type="model")
+
+openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
+controlnet_pose_model = "thibaud/controlnet-openpose-sdxl-1.0"
+controlnet_pose = ControlNetModel.from_pretrained(
+    controlnet_pose_model, torch_dtype=torch_dtype,
+).to("cuda")
+
+prompt = "instagram photo, a photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain, best quality"
+negative_prompt = "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth"
+
+# download an image
+pose_image = load_image(
+    "./examples/pos_ref.png"
+)
+pose_image = openpose(pose_image, detect_resolution=512, image_resolution=1024)
+
+# initialize the models and pipeline
+controlnet_conditioning_scale = 1.0  # recommended for good generalization
+
+### Load base model
+pipe = PhotoMakerStableDiffusionXLControlNetPipeline.from_pretrained(
+    "SG161222/RealVisXL_V4.0", 
+    controlnet=controlnet_pose,
+    torch_dtype=torch_dtype,
+).to("cuda")
+
+### Load PhotoMaker checkpoint
+pipe.load_photomaker_adapter(
+    os.path.dirname(photomaker_ckpt),
+    subfolder="",
+    weight_name=os.path.basename(photomaker_ckpt),
+    trigger_word="img"  # define the trigger word
+)     
+### Also can cooperate with other LoRA modules
+# pipe.load_lora_weights(os.path.dirname(lora_path), weight_name=lora_model_name, adapter_name="lcm-lora")
+# pipe.set_adapters(["photomaker", "lcm-lora"], adapter_weights=[1.0, 0.5])
+
+pipe.fuse_lora()
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+
+pipe.enable_model_cpu_offload()
+
+### define the input ID images
+input_folder_name = './examples/scarletthead_woman'
+image_basename_list = os.listdir(input_folder_name)
+image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
+
+input_id_images = []
+for image_path in image_path_list:
+    input_id_images.append(load_image(image_path))
+
+### extract insightface embedding
+input_id_images = []
+for image_path in image_path_list:
+    input_id_images.append(load_image(image_path))
+
+id_embed_list = []
+
+for img in input_id_images:
+    img = np.array(img)
+    img = img[:, :, ::-1]
+    faces = analyze_faces(face_detector, img)
+    if len(faces) > 0:
+        id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
+
+if len(id_embed_list) == 0:
+    raise ValueError(f"No face detected in input image pool")
+
+id_embeds = torch.stack(id_embed_list)
+
+# generate image
+images = pipe(
+    prompt, 
+    negative_prompt=negative_prompt, 
+    input_id_images=input_id_images,
+    id_embeds=id_embeds,
+    controlnet_conditioning_scale=controlnet_conditioning_scale, 
+    image=pose_image,
+    num_images_per_prompt=2,
+    start_merge_step=10,
+).images
+
+for idx, img in enumerate(images): 
+    img.save(os.path.join(output_dir, f"output_pmv2_cn_{idx}.jpg"))

PhotoMaker/inference_scripts/inference_pmv2_ip_adapter.py

+# !pip install opencv-python transformers accelerate
+import os
+import sys
+
+import numpy as np
+import torch
+from diffusers.utils import load_image
+from diffusers import EulerDiscreteScheduler
+from huggingface_hub import hf_hub_download
+from photomaker import PhotoMakerStableDiffusionXLPipeline
+from photomaker import FaceAnalysis2, analyze_faces
+
+face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
+face_detector.prepare(ctx_id=0, det_size=(640, 640))
+
+try:
+    if torch.cuda.is_available():
+        device = "cuda"
+    elif sys.platform == "darwin" and torch.backends.mps.is_available():
+        device = "mps"
+    else:
+        device = "cpu"
+except:
+    device = "cpu"
+
+torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+if device == "mps":
+    torch_dtype = torch.float16
+
+output_dir = "./outputs"
+os.makedirs(output_dir, exist_ok=True)
+
+photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker-V2", filename="photomaker-v2.bin", repo_type="model")
+
+prompt = "portrait photo of a woman img, colorful, perfect face, best quality"
+negative_prompt = "(asymmetry, worst quality, low quality, illustration), open mouth"
+
+# # initialize the models and pipeline
+### Load base model
+pipe = PhotoMakerStableDiffusionXLPipeline.from_pretrained(
+    "SG161222/RealVisXL_V4.0", torch_dtype=torch_dtype,
+).to("cuda")
+
+pipe.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name="ip-adapter_sdxl.bin")
+pipe.set_ip_adapter_scale(0.7)
+
+print("Loading images...")
+style_images = [load_image(f"./examples/statue.png")]
+
+### Load PhotoMaker checkpoint
+pipe.load_photomaker_adapter(
+    os.path.dirname(photomaker_ckpt),
+    subfolder="",
+    weight_name=os.path.basename(photomaker_ckpt),
+    trigger_word="img"  # define the trigger word
+)     
+### Also can cooperate with other LoRA modules
+# pipe.load_lora_weights(os.path.dirname(lora_path), weight_name=lora_model_name, adapter_name="lcm-lora")
+# pipe.set_adapters(["photomaker", "lcm-lora"], adapter_weights=[1.0, 0.5])
+
+pipe.fuse_lora()
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+
+pipe.enable_model_cpu_offload()
+
+### define the input ID images
+input_folder_name = './examples/scarletthead_woman'
+image_basename_list = os.listdir(input_folder_name)
+image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
+
+input_id_images = []
+for image_path in image_path_list:
+    input_id_images.append(load_image(image_path))
+
+id_embed_list = []
+
+for img in input_id_images:
+    img = np.array(img)
+    img = img[:, :, ::-1]
+    faces = analyze_faces(face_detector, img)
+    if len(faces) > 0:
+        id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
+
+if len(id_embed_list) == 0:
+    raise ValueError(f"No face detected in input image pool")
+
+id_embeds = torch.stack(id_embed_list)
+
+# generate image
+images = pipe(
+    prompt, 
+    negative_prompt=negative_prompt, 
+    input_id_images=input_id_images,
+    id_embeds=id_embeds,
+    ip_adapter_image=[style_images],
+    num_images_per_prompt=2,
+    start_merge_step=10,
+).images
+
+for idx, img in enumerate(images): 
+    img.save(os.path.join(output_dir, f"output_pmv2_ipa_{idx}.jpg"))

PhotoMaker/inference_scripts/inference_pmv2_t2i_adapter.py

+# !pip install opencv-python transformers accelerate
+import os
+import sys
+
+import numpy as np
+import torch
+from diffusers.utils import load_image
+from diffusers import EulerDiscreteScheduler, T2IAdapter
+
+from huggingface_hub import hf_hub_download
+from controlnet_aux import OpenposeDetector
+from photomaker import PhotoMakerStableDiffusionXLAdapterPipeline
+from photomaker import FaceAnalysis2, analyze_faces
+
+face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
+face_detector.prepare(ctx_id=0, det_size=(640, 640))
+
+try:
+    if torch.cuda.is_available():
+        device = "cuda"
+    elif sys.platform == "darwin" and torch.backends.mps.is_available():
+        device = "mps"
+    else:
+        device = "cpu"
+except:
+    device = "cpu"
+
+torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+if device == "mps":
+    torch_dtype = torch.float16
+
+output_dir = "./outputs"
+os.makedirs(output_dir, exist_ok=True)
+
+photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker-V2", filename="photomaker-v2.bin", repo_type="model")
+
+openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
+# load adapter
+adapter = T2IAdapter.from_pretrained(
+  "TencentARC/t2i-adapter-openpose-sdxl-1.0", torch_dtype=torch_dtype,
+).to("cuda")
+
+prompt = "instagram photo, a photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain, best quality"
+negative_prompt = "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth"
+
+# download an image
+pose_image = load_image(
+    "./examples/pos_ref.png"
+)
+pose_image = openpose(pose_image, detect_resolution=512, image_resolution=1024)
+
+# initialize the models and pipeline
+adapter_conditioning_scale = 0.8  # recommended for good generalization
+adapter_conditioning_factor = 0.8
+
+### Load base model
+pipe = PhotoMakerStableDiffusionXLAdapterPipeline.from_pretrained(
+    "SG161222/RealVisXL_V4.0", 
+    adapter=adapter,
+    torch_dtype=torch_dtype,
+).to("cuda")
+
+### Load PhotoMaker checkpoint
+pipe.load_photomaker_adapter(
+    os.path.dirname(photomaker_ckpt),
+    subfolder="",
+    weight_name=os.path.basename(photomaker_ckpt),
+    trigger_word="img"  # define the trigger word
+)     
+### Also can cooperate with other LoRA modules
+# pipe.load_lora_weights(os.path.dirname(lora_path), weight_name=lora_model_name, adapter_name="lcm-lora")
+# pipe.set_adapters(["photomaker", "lcm-lora"], adapter_weights=[1.0, 0.5])
+
+pipe.fuse_lora()
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+
+pipe.enable_model_cpu_offload()
+
+### define the input ID images
+input_folder_name = './examples/scarletthead_woman'
+image_basename_list = os.listdir(input_folder_name)
+image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
+
+input_id_images = []
+for image_path in image_path_list:
+    input_id_images.append(load_image(image_path))
+
+id_embed_list = []
+
+for img in input_id_images:
+    img = np.array(img)
+    img = img[:, :, ::-1]
+    faces = analyze_faces(face_detector, img)
+    if len(faces) > 0:
+        id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
+
+if len(id_embed_list) == 0:
+    raise ValueError(f"No face detected in input image pool")
+
+id_embeds = torch.stack(id_embed_list)
+
+# generate image
+images = pipe(
+    prompt, 
+    negative_prompt=negative_prompt, 
+    input_id_images=input_id_images,
+    id_embeds=id_embeds,
+    adapter_conditioning_scale=adapter_conditioning_scale,
+    image=pose_image,
+    num_images_per_prompt=2,
+    start_merge_step=10,
+).images
+
+for idx, img in enumerate(images): 
+    img.save(os.path.join(output_dir, f"output_pmv2_t2ia_{idx}.jpg"))

PhotoMaker/photomaker/__init__.py

+from .model import PhotoMakerIDEncoder
+from .model_v2 import PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken
+from .resampler import FacePerceiverResampler
+from .pipeline import PhotoMakerStableDiffusionXLPipeline
+from .pipeline import PhotoMakerStableDiffusionXLPipeline
+from .pipeline_controlnet import PhotoMakerStableDiffusionXLControlNetPipeline
+from .pipeline_t2i_adapter import PhotoMakerStableDiffusionXLAdapterPipeline
+
+# InsightFace Package
+from .insightface_package import FaceAnalysis2, analyze_faces
+
+__all__ = [
+    "FaceAnalysis2",
+    "analyze_faces",
+    "FacePerceiverResampler",
+    "PhotoMakerIDEncoder",
+    "PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken",
+    "PhotoMakerStableDiffusionXLPipeline",
+    "PhotoMakerStableDiffusionXLControlNetPipeline",
+    "PhotoMakerStableDiffusionXLAdapterPipeline",
+]
\ No newline at end of file

PhotoMaker/photomaker/insightface_package.py

+import numpy as np
+# pip install insightface==0.7.3
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+
+### 
+# https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/165#issue-2055829543
+###
+class FaceAnalysis2(FaceAnalysis):
+    # NOTE: allows setting det_size for each detection call.
+    # the model allows it but the wrapping code from insightface
+    # doesn't show it, and people end up loading duplicate models
+    # for different sizes where there is absolutely no need to
+    def get(self, img, max_num=0, det_size=(640, 640)):
+        if det_size is not None:
+            self.det_model.input_size = det_size
+
+        return super().get(img, max_num)
+
+def analyze_faces(face_analysis: FaceAnalysis, img_data: np.ndarray, det_size=(640, 640)):
+    # NOTE: try detect faces, if no faces detected, lower det_size until it does
+    detection_sizes = [None] + [(size, size) for size in range(640, 256, -64)] + [(256, 256)]
+
+    for size in detection_sizes:
+        faces = face_analysis.get(img_data, det_size=size)
+        if len(faces) > 0:
+            return faces
+
+    return []

PhotoMaker/photomaker/model.py

+# Merge image encoder and fuse module to create an ID Encoder
+# send multiple ID images, we can directly obtain the updated text encoder containing a stacked ID embedding
+
+import torch
+import torch.nn as nn
+from transformers.models.clip.modeling_clip import CLIPVisionModelWithProjection
+from transformers.models.clip.configuration_clip import CLIPVisionConfig
+from transformers import PretrainedConfig
+
+VISION_CONFIG_DICT = {
+    "hidden_size": 1024,
+    "intermediate_size": 4096,
+    "num_attention_heads": 16,
+    "num_hidden_layers": 24,
+    "patch_size": 14,
+    "projection_dim": 768
+}
+
+class MLP(nn.Module):
+    def __init__(self, in_dim, out_dim, hidden_dim, use_residual=True):
+        super().__init__()
+        if use_residual:
+            assert in_dim == out_dim
+        self.layernorm = nn.LayerNorm(in_dim)
+        self.fc1 = nn.Linear(in_dim, hidden_dim)
+        self.fc2 = nn.Linear(hidden_dim, out_dim)
+        self.use_residual = use_residual
+        self.act_fn = nn.GELU()
+
+    def forward(self, x):
+        residual = x
+        x = self.layernorm(x)
+        x = self.fc1(x)
+        x = self.act_fn(x)
+        x = self.fc2(x)
+        if self.use_residual:
+            x = x + residual
+        return x
+
+
+class FuseModule(nn.Module):
+    def __init__(self, embed_dim):
+        super().__init__()
+        self.mlp1 = MLP(embed_dim * 2, embed_dim, embed_dim, use_residual=False)
+        self.mlp2 = MLP(embed_dim, embed_dim, embed_dim, use_residual=True)
+        self.layer_norm = nn.LayerNorm(embed_dim)
+
+    def fuse_fn(self, prompt_embeds, id_embeds):
+        stacked_id_embeds = torch.cat([prompt_embeds, id_embeds], dim=-1)
+        stacked_id_embeds = self.mlp1(stacked_id_embeds) + prompt_embeds
+        stacked_id_embeds = self.mlp2(stacked_id_embeds)
+        stacked_id_embeds = self.layer_norm(stacked_id_embeds)
+        return stacked_id_embeds
+
+    def forward(
+        self,
+        prompt_embeds,
+        id_embeds,
+        class_tokens_mask,
+    ) -> torch.Tensor:
+        # id_embeds shape: [b, max_num_inputs, 1, 2048]
+        id_embeds = id_embeds.to(prompt_embeds.dtype)
+        num_inputs = class_tokens_mask.sum().unsqueeze(0) # TODO: check for training case
+        batch_size, max_num_inputs = id_embeds.shape[:2]
+        # seq_length: 77
+        seq_length = prompt_embeds.shape[1]
+        # flat_id_embeds shape: [b*max_num_inputs, 1, 2048]
+        flat_id_embeds = id_embeds.view(
+            -1, id_embeds.shape[-2], id_embeds.shape[-1]
+        )
+        # valid_id_mask [b*max_num_inputs]
+        valid_id_mask = (
+            torch.arange(max_num_inputs, device=flat_id_embeds.device)[None, :]
+            < num_inputs[:, None]
+        )
+        valid_id_embeds = flat_id_embeds[valid_id_mask.flatten()]
+
+        prompt_embeds = prompt_embeds.view(-1, prompt_embeds.shape[-1])
+        class_tokens_mask = class_tokens_mask.view(-1)
+        valid_id_embeds = valid_id_embeds.view(-1, valid_id_embeds.shape[-1])
+        # slice out the image token embeddings
+        image_token_embeds = prompt_embeds[class_tokens_mask]
+        stacked_id_embeds = self.fuse_fn(image_token_embeds, valid_id_embeds)
+        assert class_tokens_mask.sum() == stacked_id_embeds.shape[0], f"{class_tokens_mask.sum()} != {stacked_id_embeds.shape[0]}"
+        prompt_embeds.masked_scatter_(class_tokens_mask[:, None], stacked_id_embeds.to(prompt_embeds.dtype))
+        updated_prompt_embeds = prompt_embeds.view(batch_size, seq_length, -1)
+        return updated_prompt_embeds
+
+class PhotoMakerIDEncoder(CLIPVisionModelWithProjection):
+    def __init__(self):
+        super().__init__(CLIPVisionConfig(**VISION_CONFIG_DICT))
+        self.visual_projection_2 = nn.Linear(1024, 1280, bias=False)
+        self.fuse_module = FuseModule(2048)
+
+    def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask):
+        b, num_inputs, c, h, w = id_pixel_values.shape
+        id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
+
+        shared_id_embeds = self.vision_model(id_pixel_values)[1]
+        id_embeds = self.visual_projection(shared_id_embeds)
+        id_embeds_2 = self.visual_projection_2(shared_id_embeds)
+
+        id_embeds = id_embeds.view(b, num_inputs, 1, -1)
+        id_embeds_2 = id_embeds_2.view(b, num_inputs, 1, -1)    
+
+        id_embeds = torch.cat((id_embeds, id_embeds_2), dim=-1)
+        updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask)
+
+        return updated_prompt_embeds
+
+
+if __name__ == "__main__":
+    PhotoMakerIDEncoder()
\ No newline at end of file

PhotoMaker/photomaker/model_v2.py

+# Merge image encoder and fuse module to create an ID Encoder
+# send multiple ID images, we can directly obtain the updated text encoder containing a stacked ID embedding
+
+import torch
+import torch.nn as nn
+from transformers.models.clip.modeling_clip import CLIPVisionModelWithProjection
+from transformers.models.clip.configuration_clip import CLIPVisionConfig
+
+from .resampler import FacePerceiverResampler
+
+VISION_CONFIG_DICT = {
+    "hidden_size": 1024,
+    "intermediate_size": 4096,
+    "num_attention_heads": 16,
+    "num_hidden_layers": 24,
+    "patch_size": 14,
+    "projection_dim": 768
+}
+
+
+class MLP(nn.Module):
+    def __init__(self, in_dim, out_dim, hidden_dim, use_residual=True):
+        super().__init__()
+        if use_residual:
+            assert in_dim == out_dim
+        self.layernorm = nn.LayerNorm(in_dim)
+        self.fc1 = nn.Linear(in_dim, hidden_dim)
+        self.fc2 = nn.Linear(hidden_dim, out_dim)
+        self.use_residual = use_residual
+        self.act_fn = nn.GELU()
+
+    def forward(self, x):
+        residual = x
+        x = self.layernorm(x)
+        x = self.fc1(x)
+        x = self.act_fn(x)
+        x = self.fc2(x)
+        if self.use_residual:
+            x = x + residual
+        return x
+
+
+class QFormerPerceiver(nn.Module):
+    def __init__(self, id_embeddings_dim, cross_attention_dim, num_tokens, embedding_dim=1024, use_residual=True, ratio=4):
+        super().__init__()
+
+        self.num_tokens = num_tokens
+        self.cross_attention_dim = cross_attention_dim
+        self.use_residual = use_residual
+        print(cross_attention_dim*num_tokens)
+        self.token_proj = nn.Sequential(
+            nn.Linear(id_embeddings_dim, id_embeddings_dim*ratio),
+            nn.GELU(),
+            nn.Linear(id_embeddings_dim*ratio, cross_attention_dim*num_tokens),
+        )
+        self.token_norm = nn.LayerNorm(cross_attention_dim)
+        self.perceiver_resampler = FacePerceiverResampler(
+            dim=cross_attention_dim,
+            depth=4,
+            dim_head=128,
+            heads=cross_attention_dim // 128,
+            embedding_dim=embedding_dim,
+            output_dim=cross_attention_dim,
+            ff_mult=4,
+        )
+
+    def forward(self, x, last_hidden_state):
+        x = self.token_proj(x)
+        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
+        x = self.token_norm(x) # cls token
+        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
+        if self.use_residual: # TODO: if use_residual is not true
+            out = x + 1.0 * out 
+        return out
+
+
+class FuseModule(nn.Module):
+    def __init__(self, embed_dim):
+        super().__init__()
+        self.mlp1 = MLP(embed_dim * 2, embed_dim, embed_dim, use_residual=False)
+        self.mlp2 = MLP(embed_dim, embed_dim, embed_dim, use_residual=True)
+        self.layer_norm = nn.LayerNorm(embed_dim)
+
+    def fuse_fn(self, prompt_embeds, id_embeds):
+        stacked_id_embeds = torch.cat([prompt_embeds, id_embeds], dim=-1)
+        stacked_id_embeds = self.mlp1(stacked_id_embeds) + prompt_embeds
+        stacked_id_embeds = self.mlp2(stacked_id_embeds)
+        stacked_id_embeds = self.layer_norm(stacked_id_embeds)
+        return stacked_id_embeds
+
+    def forward(
+        self,
+        prompt_embeds,
+        id_embeds,
+        class_tokens_mask,
+    ) -> torch.Tensor:
+        # id_embeds shape: [b, max_num_inputs, 1, 2048]
+        id_embeds = id_embeds.to(prompt_embeds.dtype)
+        num_inputs = class_tokens_mask.sum().unsqueeze(0) # TODO: check for training case
+        batch_size, max_num_inputs = id_embeds.shape[:2]
+        # seq_length: 77
+        seq_length = prompt_embeds.shape[1]
+        # flat_id_embeds shape: [b*max_num_inputs, 1, 2048]
+        flat_id_embeds = id_embeds.view(
+            -1, id_embeds.shape[-2], id_embeds.shape[-1]
+        )
+        # valid_id_mask [b*max_num_inputs]
+        valid_id_mask = (
+            torch.arange(max_num_inputs, device=flat_id_embeds.device)[None, :]
+            < num_inputs[:, None]
+        )
+        valid_id_embeds = flat_id_embeds[valid_id_mask.flatten()]
+
+        prompt_embeds = prompt_embeds.view(-1, prompt_embeds.shape[-1])
+        class_tokens_mask = class_tokens_mask.view(-1)
+        valid_id_embeds = valid_id_embeds.view(-1, valid_id_embeds.shape[-1])
+        # slice out the image token embeddings
+        image_token_embeds = prompt_embeds[class_tokens_mask]
+        stacked_id_embeds = self.fuse_fn(image_token_embeds, valid_id_embeds)
+        assert class_tokens_mask.sum() == stacked_id_embeds.shape[0], f"{class_tokens_mask.sum()} != {stacked_id_embeds.shape[0]}"
+        prompt_embeds.masked_scatter_(class_tokens_mask[:, None], stacked_id_embeds.to(prompt_embeds.dtype))
+        updated_prompt_embeds = prompt_embeds.view(batch_size, seq_length, -1)
+        return updated_prompt_embeds
+
+
+class PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken(CLIPVisionModelWithProjection):
+    def __init__(self, id_embeddings_dim=512):
+        super().__init__(CLIPVisionConfig(**VISION_CONFIG_DICT))
+        self.fuse_module = FuseModule(2048)
+        self.visual_projection_2 = nn.Linear(1024, 1280, bias=False)
+
+        cross_attention_dim = 2048
+        # projection
+        self.num_tokens = 2
+        self.cross_attention_dim = cross_attention_dim
+        self.qformer_perceiver = QFormerPerceiver(
+                                    id_embeddings_dim, 
+                                    cross_attention_dim, 
+                                    self.num_tokens,
+                                )
+
+    def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds):
+        b, num_inputs, c, h, w = id_pixel_values.shape
+        id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
+
+        last_hidden_state = self.vision_model(id_pixel_values)[0]
+        id_embeds = id_embeds.view(b * num_inputs, -1)
+
+        id_embeds = self.qformer_perceiver(id_embeds, last_hidden_state)
+        id_embeds = id_embeds.view(b, num_inputs, self.num_tokens, -1)
+        updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask)
+
+        return updated_prompt_embeds
+
+if __name__ == "__main__":
+    PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken()

PhotoMaker/photomaker/resampler.py

+#### Borrowed from https://github.com/tencent-ailab/IP-Adapter/blob/main/ip_adapter/resampler.py 
+# 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
+
+
+class FacePerceiverResampler(torch.nn.Module):
+    def __init__(
+        self,
+        *,
+        dim=768,
+        depth=4,
+        dim_head=64,
+        heads=16,
+        embedding_dim=1280,
+        output_dim=768,
+        ff_mult=4,
+    ):
+        super().__init__()
+        
+        self.proj_in = torch.nn.Linear(embedding_dim, dim)
+        self.proj_out = torch.nn.Linear(dim, output_dim)
+        self.norm_out = torch.nn.LayerNorm(output_dim)
+        self.layers = torch.nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                torch.nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, latents, x):
+        x = self.proj_in(x)
+        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)
+
+# 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)
+
+
+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)