instantid

gradio_demo/model_util.py

+from typing import Literal, Union, Optional, Tuple, List
+
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer, CLIPTextModelWithProjection
+from diffusers import (
+    UNet2DConditionModel,
+    SchedulerMixin,
+    StableDiffusionPipeline,
+    StableDiffusionXLPipeline,
+    AutoencoderKL,
+)
+from diffusers.pipelines.stable_diffusion.convert_from_ckpt import (
+    convert_ldm_unet_checkpoint,
+)
+from safetensors.torch import load_file
+from diffusers.schedulers import (
+    DDIMScheduler,
+    DDPMScheduler,
+    LMSDiscreteScheduler,
+    EulerDiscreteScheduler,
+    EulerAncestralDiscreteScheduler,
+    UniPCMultistepScheduler,
+)
+
+from omegaconf import OmegaConf
+
+# DiffUsers版StableDiffusionのモデルパラメータ
+NUM_TRAIN_TIMESTEPS = 1000
+BETA_START = 0.00085
+BETA_END = 0.0120
+
+UNET_PARAMS_MODEL_CHANNELS = 320
+UNET_PARAMS_CHANNEL_MULT = [1, 2, 4, 4]
+UNET_PARAMS_ATTENTION_RESOLUTIONS = [4, 2, 1]
+UNET_PARAMS_IMAGE_SIZE = 64  # fixed from old invalid value `32`
+UNET_PARAMS_IN_CHANNELS = 4
+UNET_PARAMS_OUT_CHANNELS = 4
+UNET_PARAMS_NUM_RES_BLOCKS = 2
+UNET_PARAMS_CONTEXT_DIM = 768
+UNET_PARAMS_NUM_HEADS = 8
+# UNET_PARAMS_USE_LINEAR_PROJECTION = False
+
+VAE_PARAMS_Z_CHANNELS = 4
+VAE_PARAMS_RESOLUTION = 256
+VAE_PARAMS_IN_CHANNELS = 3
+VAE_PARAMS_OUT_CH = 3
+VAE_PARAMS_CH = 128
+VAE_PARAMS_CH_MULT = [1, 2, 4, 4]
+VAE_PARAMS_NUM_RES_BLOCKS = 2
+
+# V2
+V2_UNET_PARAMS_ATTENTION_HEAD_DIM = [5, 10, 20, 20]
+V2_UNET_PARAMS_CONTEXT_DIM = 1024
+# V2_UNET_PARAMS_USE_LINEAR_PROJECTION = True
+
+TOKENIZER_V1_MODEL_NAME = "CompVis/stable-diffusion-v1-4"
+TOKENIZER_V2_MODEL_NAME = "stabilityai/stable-diffusion-2-1"
+
+AVAILABLE_SCHEDULERS = Literal["ddim", "ddpm", "lms", "euler_a", "euler", "uniPC"]
+
+SDXL_TEXT_ENCODER_TYPE = Union[CLIPTextModel, CLIPTextModelWithProjection]
+
+DIFFUSERS_CACHE_DIR = None  # if you want to change the cache dir, change this
+
+
+def load_checkpoint_with_text_encoder_conversion(ckpt_path: str, device="cpu"):
+    # text encoderの格納形式が違うモデルに対応する ('text_model'がない)
+    TEXT_ENCODER_KEY_REPLACEMENTS = [
+        (
+            "cond_stage_model.transformer.embeddings.",
+            "cond_stage_model.transformer.text_model.embeddings.",
+        ),
+        (
+            "cond_stage_model.transformer.encoder.",
+            "cond_stage_model.transformer.text_model.encoder.",
+        ),
+        (
+            "cond_stage_model.transformer.final_layer_norm.",
+            "cond_stage_model.transformer.text_model.final_layer_norm.",
+        ),
+    ]
+
+    if ckpt_path.endswith(".safetensors"):
+        checkpoint = None
+        state_dict = load_file(ckpt_path)  # , device) # may causes error
+    else:
+        checkpoint = torch.load(ckpt_path, map_location=device)
+        if "state_dict" in checkpoint:
+            state_dict = checkpoint["state_dict"]
+        else:
+            state_dict = checkpoint
+            checkpoint = None
+
+    key_reps = []
+    for rep_from, rep_to in TEXT_ENCODER_KEY_REPLACEMENTS:
+        for key in state_dict.keys():
+            if key.startswith(rep_from):
+                new_key = rep_to + key[len(rep_from) :]
+                key_reps.append((key, new_key))
+
+    for key, new_key in key_reps:
+        state_dict[new_key] = state_dict[key]
+        del state_dict[key]
+
+    return checkpoint, state_dict
+
+
+def create_unet_diffusers_config(v2, use_linear_projection_in_v2=False):
+    """
+    Creates a config for the diffusers based on the config of the LDM model.
+    """
+    # unet_params = original_config.model.params.unet_config.params
+
+    block_out_channels = [
+        UNET_PARAMS_MODEL_CHANNELS * mult for mult in UNET_PARAMS_CHANNEL_MULT
+    ]
+
+    down_block_types = []
+    resolution = 1
+    for i in range(len(block_out_channels)):
+        block_type = (
+            "CrossAttnDownBlock2D"
+            if resolution in UNET_PARAMS_ATTENTION_RESOLUTIONS
+            else "DownBlock2D"
+        )
+        down_block_types.append(block_type)
+        if i != len(block_out_channels) - 1:
+            resolution *= 2
+
+    up_block_types = []
+    for i in range(len(block_out_channels)):
+        block_type = (
+            "CrossAttnUpBlock2D"
+            if resolution in UNET_PARAMS_ATTENTION_RESOLUTIONS
+            else "UpBlock2D"
+        )
+        up_block_types.append(block_type)
+        resolution //= 2
+
+    config = dict(
+        sample_size=UNET_PARAMS_IMAGE_SIZE,
+        in_channels=UNET_PARAMS_IN_CHANNELS,
+        out_channels=UNET_PARAMS_OUT_CHANNELS,
+        down_block_types=tuple(down_block_types),
+        up_block_types=tuple(up_block_types),
+        block_out_channels=tuple(block_out_channels),
+        layers_per_block=UNET_PARAMS_NUM_RES_BLOCKS,
+        cross_attention_dim=UNET_PARAMS_CONTEXT_DIM
+        if not v2
+        else V2_UNET_PARAMS_CONTEXT_DIM,
+        attention_head_dim=UNET_PARAMS_NUM_HEADS
+        if not v2
+        else V2_UNET_PARAMS_ATTENTION_HEAD_DIM,
+        # use_linear_projection=UNET_PARAMS_USE_LINEAR_PROJECTION if not v2 else V2_UNET_PARAMS_USE_LINEAR_PROJECTION,
+    )
+    if v2 and use_linear_projection_in_v2:
+        config["use_linear_projection"] = True
+
+    return config
+
+
+def load_diffusers_model(
+    pretrained_model_name_or_path: str,
+    v2: bool = False,
+    clip_skip: Optional[int] = None,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[CLIPTokenizer, CLIPTextModel, UNet2DConditionModel,]:
+    if v2:
+        tokenizer = CLIPTokenizer.from_pretrained(
+            TOKENIZER_V2_MODEL_NAME,
+            subfolder="tokenizer",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        )
+        text_encoder = CLIPTextModel.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="text_encoder",
+            # default is clip skip 2
+            num_hidden_layers=24 - (clip_skip - 1) if clip_skip is not None else 23,
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        )
+    else:
+        tokenizer = CLIPTokenizer.from_pretrained(
+            TOKENIZER_V1_MODEL_NAME,
+            subfolder="tokenizer",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        )
+        text_encoder = CLIPTextModel.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="text_encoder",
+            num_hidden_layers=12 - (clip_skip - 1) if clip_skip is not None else 12,
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        )
+
+    unet = UNet2DConditionModel.from_pretrained(
+        pretrained_model_name_or_path,
+        subfolder="unet",
+        torch_dtype=weight_dtype,
+        cache_dir=DIFFUSERS_CACHE_DIR,
+    )
+
+    vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder="vae")
+
+    return tokenizer, text_encoder, unet, vae
+
+
+def load_checkpoint_model(
+    checkpoint_path: str,
+    v2: bool = False,
+    clip_skip: Optional[int] = None,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[CLIPTokenizer, CLIPTextModel, UNet2DConditionModel,]:
+    pipe = StableDiffusionPipeline.from_single_file(
+        checkpoint_path,
+        upcast_attention=True if v2 else False,
+        torch_dtype=weight_dtype,
+        cache_dir=DIFFUSERS_CACHE_DIR,
+    )
+
+    _, state_dict = load_checkpoint_with_text_encoder_conversion(checkpoint_path)
+    unet_config = create_unet_diffusers_config(v2, use_linear_projection_in_v2=v2)
+    unet_config["class_embed_type"] = None
+    unet_config["addition_embed_type"] = None
+    converted_unet_checkpoint = convert_ldm_unet_checkpoint(state_dict, unet_config)
+    unet = UNet2DConditionModel(**unet_config)
+    unet.load_state_dict(converted_unet_checkpoint)
+
+    tokenizer = pipe.tokenizer
+    text_encoder = pipe.text_encoder
+    vae = pipe.vae
+    if clip_skip is not None:
+        if v2:
+            text_encoder.config.num_hidden_layers = 24 - (clip_skip - 1)
+        else:
+            text_encoder.config.num_hidden_layers = 12 - (clip_skip - 1)
+
+    del pipe
+
+    return tokenizer, text_encoder, unet, vae
+
+
+def load_models(
+    pretrained_model_name_or_path: str,
+    scheduler_name: str,
+    v2: bool = False,
+    v_pred: bool = False,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[CLIPTokenizer, CLIPTextModel, UNet2DConditionModel, SchedulerMixin,]:
+    if pretrained_model_name_or_path.endswith(
+        ".ckpt"
+    ) or pretrained_model_name_or_path.endswith(".safetensors"):
+        tokenizer, text_encoder, unet, vae = load_checkpoint_model(
+            pretrained_model_name_or_path, v2=v2, weight_dtype=weight_dtype
+        )
+    else:  # diffusers
+        tokenizer, text_encoder, unet, vae = load_diffusers_model(
+            pretrained_model_name_or_path, v2=v2, weight_dtype=weight_dtype
+        )
+
+    if scheduler_name:
+        scheduler = create_noise_scheduler(
+            scheduler_name,
+            prediction_type="v_prediction" if v_pred else "epsilon",
+        )
+    else:
+        scheduler = None
+
+    return tokenizer, text_encoder, unet, scheduler, vae
+
+
+def load_diffusers_model_xl(
+    pretrained_model_name_or_path: str,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[List[CLIPTokenizer], List[SDXL_TEXT_ENCODER_TYPE], UNet2DConditionModel,]:
+    # returns tokenizer, tokenizer_2, text_encoder, text_encoder_2, unet
+
+    tokenizers = [
+        CLIPTokenizer.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="tokenizer",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        ),
+        CLIPTokenizer.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="tokenizer_2",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+            pad_token_id=0,  # same as open clip
+        ),
+    ]
+
+    text_encoders = [
+        CLIPTextModel.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="text_encoder",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        ),
+        CLIPTextModelWithProjection.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="text_encoder_2",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        ),
+    ]
+
+    unet = UNet2DConditionModel.from_pretrained(
+        pretrained_model_name_or_path,
+        subfolder="unet",
+        torch_dtype=weight_dtype,
+        cache_dir=DIFFUSERS_CACHE_DIR,
+    )
+    vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder="vae")
+    return tokenizers, text_encoders, unet, vae
+
+
+def load_checkpoint_model_xl(
+    checkpoint_path: str,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[List[CLIPTokenizer], List[SDXL_TEXT_ENCODER_TYPE], UNet2DConditionModel,]:
+    pipe = StableDiffusionXLPipeline.from_single_file(
+        checkpoint_path,
+        torch_dtype=weight_dtype,
+        cache_dir=DIFFUSERS_CACHE_DIR,
+    )
+
+    unet = pipe.unet
+    vae = pipe.vae
+    tokenizers = [pipe.tokenizer, pipe.tokenizer_2]
+    text_encoders = [pipe.text_encoder, pipe.text_encoder_2]
+    if len(text_encoders) == 2:
+        text_encoders[1].pad_token_id = 0
+
+    del pipe
+
+    return tokenizers, text_encoders, unet, vae
+
+
+def load_models_xl(
+    pretrained_model_name_or_path: str,
+    scheduler_name: str,
+    weight_dtype: torch.dtype = torch.float32,
+    noise_scheduler_kwargs=None,
+) -> Tuple[
+    List[CLIPTokenizer],
+    List[SDXL_TEXT_ENCODER_TYPE],
+    UNet2DConditionModel,
+    SchedulerMixin,
+]:
+    if pretrained_model_name_or_path.endswith(
+        ".ckpt"
+    ) or pretrained_model_name_or_path.endswith(".safetensors"):
+        (tokenizers, text_encoders, unet, vae) = load_checkpoint_model_xl(
+            pretrained_model_name_or_path, weight_dtype
+        )
+    else:  # diffusers
+        (tokenizers, text_encoders, unet, vae) = load_diffusers_model_xl(
+            pretrained_model_name_or_path, weight_dtype
+        )
+    if scheduler_name:
+        scheduler = create_noise_scheduler(scheduler_name, noise_scheduler_kwargs)
+    else:
+        scheduler = None
+
+    return tokenizers, text_encoders, unet, scheduler, vae
+
+def create_noise_scheduler(
+    scheduler_name: AVAILABLE_SCHEDULERS = "ddpm",
+    noise_scheduler_kwargs=None,
+    prediction_type: Literal["epsilon", "v_prediction"] = "epsilon",
+) -> SchedulerMixin:
+    name = scheduler_name.lower().replace(" ", "_")
+    if name.lower() == "ddim":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/ddim
+        scheduler = DDIMScheduler(**OmegaConf.to_container(noise_scheduler_kwargs))
+    elif name.lower() == "ddpm":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/ddpm
+        scheduler = DDPMScheduler(**OmegaConf.to_container(noise_scheduler_kwargs))
+    elif name.lower() == "lms":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/lms_discrete
+        scheduler = LMSDiscreteScheduler(
+            **OmegaConf.to_container(noise_scheduler_kwargs)
+        )
+    elif name.lower() == "euler_a":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/euler_ancestral
+        scheduler = EulerAncestralDiscreteScheduler(
+            **OmegaConf.to_container(noise_scheduler_kwargs)
+        )
+    elif name.lower() == "euler":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/euler_ancestral
+        scheduler = EulerDiscreteScheduler(
+            **OmegaConf.to_container(noise_scheduler_kwargs)
+        )
+    elif name.lower() == "unipc":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/unipc
+        scheduler = UniPCMultistepScheduler(
+            **OmegaConf.to_container(noise_scheduler_kwargs)
+        )
+    else:
+        raise ValueError(f"Unknown scheduler name: {name}")
+
+    return scheduler
+
+
+def torch_gc():
+    import gc
+
+    gc.collect()
+    if torch.cuda.is_available():
+        with torch.cuda.device("cuda"):
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+
+
+from enum import Enum
+
+
+class CPUState(Enum):
+    GPU = 0
+    CPU = 1
+    MPS = 2
+
+
+cpu_state = CPUState.GPU
+xpu_available = False
+directml_enabled = False
+
+
+def is_intel_xpu():
+    global cpu_state
+    global xpu_available
+    if cpu_state == CPUState.GPU:
+        if xpu_available:
+            return True
+    return False
+
+
+try:
+    import intel_extension_for_pytorch as ipex
+
+    if torch.xpu.is_available():
+        xpu_available = True
+except:
+    pass
+
+try:
+    if torch.backends.mps.is_available():
+        cpu_state = CPUState.MPS
+        import torch.mps
+except:
+    pass
+
+
+def get_torch_device():
+    global directml_enabled
+    global cpu_state
+    if directml_enabled:
+        global directml_device
+        return directml_device
+    if cpu_state == CPUState.MPS:
+        return torch.device("mps")
+    if cpu_state == CPUState.CPU:
+        return torch.device("cpu")
+    else:
+        if is_intel_xpu():
+            return torch.device("xpu")
+        else:
+            return torch.device(torch.cuda.current_device())

gradio_demo/requirements.txt

+diffusers==0.25.1
+torch==2.0.0
+torchvision==0.15.1
+transformers==4.37.1
+accelerate
+safetensors
+einops
+onnxruntime-gpu
+spaces==0.19.4
+omegaconf
+peft
+huggingface-hub==0.20.2
+opencv-python
+insightface
+gradio
+controlnet_aux
+gdown
+peft
\ No newline at end of file

gradio_demo/style_template.py

+style_list = [
+    {
+        "name": "(No style)",
+        "prompt": "{prompt}",
+        "negative_prompt": "",
+    },
+    {
+        "name": "Watercolor",
+        "prompt": "watercolor painting, {prompt}. vibrant, beautiful, painterly, detailed, textural, artistic",
+        "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, anime, photorealistic, 35mm film, deformed, glitch, low contrast, noisy",
+    },
+    {
+        "name": "Film Noir",
+        "prompt": "film noir style, ink sketch|vector, {prompt} highly detailed, sharp focus, ultra sharpness, monochrome, high contrast, dramatic shadows, 1940s style, mysterious, cinematic",
+        "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
+    },
+    {
+        "name": "Neon",
+        "prompt": "masterpiece painting, buildings in the backdrop, kaleidoscope, lilac orange blue cream fuchsia bright vivid gradient colors, the scene is cinematic, {prompt}, emotional realism, double exposure, watercolor ink pencil, graded wash, color layering, magic realism, figurative painting, intricate motifs, organic tracery, polished",
+        "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
+    },
+    {
+        "name": "Jungle",
+        "prompt": 'waist-up "{prompt} in a Jungle" by Syd Mead, tangerine cold color palette, muted colors, detailed, 8k,photo r3al,dripping paint,3d toon style,3d style,Movie Still',
+        "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
+    },
+    {
+        "name": "Mars",
+        "prompt": "{prompt}, Post-apocalyptic. Mars Colony, Scavengers roam the wastelands searching for valuable resources, rovers, bright morning sunlight shining, (detailed) (intricate) (8k) (HDR) (cinematic lighting) (sharp focus)",
+        "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
+    },
+    {
+        "name": "Vibrant Color",
+        "prompt": "vibrant colorful, ink sketch|vector|2d colors, at nightfall, sharp focus, {prompt}, highly detailed, sharp focus, the clouds,colorful,ultra sharpness",
+        "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
+    },
+    {
+        "name": "Snow",
+        "prompt": "cinema 4d render, {prompt}, high contrast, vibrant and saturated, sico style, surrounded by magical glow,floating ice shards, snow crystals, cold, windy background, frozen natural landscape in background  cinematic atmosphere,highly detailed, sharp focus, intricate design, 3d, unreal engine, octane render, CG best quality, highres, photorealistic, dramatic lighting, artstation, concept art, cinematic, epic Steven Spielberg movie still, sharp focus, smoke, sparks, art by pascal blanche and greg rutkowski and repin, trending on artstation, hyperrealism painting, matte painting, 4k resolution",
+        "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
+    },
+    {
+        "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

infer.py

+import cv2
+import torch
+import numpy as np
+from PIL import Image
+
+from diffusers.utils import load_image
+from diffusers.models import ControlNetModel
+
+from insightface.app import FaceAnalysis
+from pipeline_stable_diffusion_xl_instantid import StableDiffusionXLInstantIDPipeline, draw_kps
+
+def resize_img(input_image, max_side=1280, min_side=1024, size=None, 
+               pad_to_max_side=False, mode=Image.BILINEAR, base_pixel_number=64):
+
+    w, h = input_image.size
+    if size is not None:
+        w_resize_new, h_resize_new = size
+    else:
+        ratio = min_side / min(h, w)
+        w, h = round(ratio*w), round(ratio*h)
+        ratio = max_side / max(h, w)
+        input_image = input_image.resize([round(ratio*w), round(ratio*h)], mode)
+        w_resize_new = (round(ratio * w) // base_pixel_number) * base_pixel_number
+        h_resize_new = (round(ratio * h) // base_pixel_number) * base_pixel_number
+    input_image = input_image.resize([w_resize_new, h_resize_new], mode)
+
+    if pad_to_max_side:
+        res = np.ones([max_side, max_side, 3], dtype=np.uint8) * 255
+        offset_x = (max_side - w_resize_new) // 2
+        offset_y = (max_side - h_resize_new) // 2
+        res[offset_y:offset_y+h_resize_new, offset_x:offset_x+w_resize_new] = np.array(input_image)
+        input_image = Image.fromarray(res)
+    return input_image
+
+
+if __name__ == "__main__":
+
+    app = FaceAnalysis(name='antelopev2', root='./', providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
+    app.prepare(ctx_id=0, det_size=(640, 640))
+
+    # Path to InstantID models
+    face_adapter = f'./checkpoints/ip-adapter.bin'
+    controlnet_path = f'./checkpoints/ControlNetModel'
+
+    # Load pipeline
+    controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16)
+
+    base_model_path = 'stabilityai/stable-diffusion-xl-base-1.0'
+
+    pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
+        base_model_path,
+        controlnet=controlnet,
+        torch_dtype=torch.float16,
+    )
+    pipe.cuda()
+    pipe.load_ip_adapter_instantid(face_adapter)
+
+    prompt = "analog film photo of a man. faded film, desaturated, 35mm photo, grainy, vignette, vintage, Kodachrome, Lomography, stained, highly detailed, found footage, masterpiece, best quality"
+    n_prompt = "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured (lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch,deformed, mutated, cross-eyed, ugly, disfigured"
+
+    face_image = load_image("./examples/yann-lecun_resize.jpg")
+    face_image = resize_img(face_image)
+
+    face_info = app.get(cv2.cvtColor(np.array(face_image), cv2.COLOR_RGB2BGR))
+    face_info = sorted(face_info, key=lambda x:(x['bbox'][2]-x['bbox'][0])*x['bbox'][3]-x['bbox'][1])[-1] # only use the maximum face
+    face_emb = face_info['embedding']
+    face_kps = draw_kps(face_image, face_info['kps'])
+
+    image = pipe(
+        prompt=prompt,
+        negative_prompt=n_prompt,
+        image_embeds=face_emb,
+        image=face_kps,
+        controlnet_conditioning_scale=0.8,
+        ip_adapter_scale=0.8,
+        num_inference_steps=30,
+        guidance_scale=5,
+    ).images[0]
+
+    image.save('result.jpg')
\ No newline at end of file

infer_full.py

+import cv2
+import torch
+import numpy as np
+from PIL import Image
+
+from diffusers.utils import load_image
+from diffusers.models import ControlNetModel
+from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
+
+from insightface.app import FaceAnalysis
+from pipeline_stable_diffusion_xl_instantid_full import StableDiffusionXLInstantIDPipeline, draw_kps
+
+from controlnet_aux import MidasDetector
+
+def convert_from_image_to_cv2(img: Image) -> np.ndarray:
+    return cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
+
+def resize_img(input_image, max_side=1280, min_side=1024, size=None, 
+               pad_to_max_side=False, mode=Image.BILINEAR, base_pixel_number=64):
+
+    w, h = input_image.size
+    if size is not None:
+        w_resize_new, h_resize_new = size
+    else:
+        ratio = min_side / min(h, w)
+        w, h = round(ratio*w), round(ratio*h)
+        ratio = max_side / max(h, w)
+        input_image = input_image.resize([round(ratio*w), round(ratio*h)], mode)
+        w_resize_new = (round(ratio * w) // base_pixel_number) * base_pixel_number
+        h_resize_new = (round(ratio * h) // base_pixel_number) * base_pixel_number
+    input_image = input_image.resize([w_resize_new, h_resize_new], mode)
+
+    if pad_to_max_side:
+        res = np.ones([max_side, max_side, 3], dtype=np.uint8) * 255
+        offset_x = (max_side - w_resize_new) // 2
+        offset_y = (max_side - h_resize_new) // 2
+        res[offset_y:offset_y+h_resize_new, offset_x:offset_x+w_resize_new] = np.array(input_image)
+        input_image = Image.fromarray(res)
+    return input_image
+
+
+if __name__ == "__main__":
+
+    # Load face encoder
+    app = FaceAnalysis(name='antelopev2', root='./', providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
+    app.prepare(ctx_id=0, det_size=(640, 640))
+
+    # Path to InstantID models
+    face_adapter = f'./checkpoints/ip-adapter.bin'
+    controlnet_path = f'./checkpoints/ControlNetModel'
+    controlnet_depth_path = f'diffusers/controlnet-depth-sdxl-1.0-small'
+    
+    # Load depth detector
+    midas = MidasDetector.from_pretrained("lllyasviel/Annotators")
+
+    # Load pipeline
+    controlnet_list = [controlnet_path, controlnet_depth_path]
+    controlnet_model_list = []
+    for controlnet_path in controlnet_list:
+        controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16)
+        controlnet_model_list.append(controlnet)
+    controlnet = MultiControlNetModel(controlnet_model_list)
+    
+    base_model_path = 'stabilityai/stable-diffusion-xl-base-1.0'
+
+    pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
+        base_model_path,
+        controlnet=controlnet,
+        torch_dtype=torch.float16,
+    )
+    pipe.cuda()
+    pipe.load_ip_adapter_instantid(face_adapter)
+
+    # Infer setting
+    prompt = "analog film photo of a man. faded film, desaturated, 35mm photo, grainy, vignette, vintage, Kodachrome, Lomography, stained, highly detailed, found footage, masterpiece, best quality"
+    n_prompt = "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured (lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch,deformed, mutated, cross-eyed, ugly, disfigured"
+
+    face_image = load_image("./examples/yann-lecun_resize.jpg")
+    face_image = resize_img(face_image)
+
+    face_info = app.get(cv2.cvtColor(np.array(face_image), cv2.COLOR_RGB2BGR))
+    face_info = sorted(face_info, key=lambda x:(x['bbox'][2]-x['bbox'][0])*x['bbox'][3]-x['bbox'][1])[-1] # only use the maximum face
+    face_emb = face_info['embedding']
+
+    # use another reference image
+    pose_image = load_image("./examples/poses/pose.jpg")
+    pose_image = resize_img(pose_image)
+
+    face_info = app.get(cv2.cvtColor(np.array(pose_image), cv2.COLOR_RGB2BGR))
+    pose_image_cv2 = convert_from_image_to_cv2(pose_image)
+    face_info = sorted(face_info, key=lambda x:(x['bbox'][2]-x['bbox'][0])*x['bbox'][3]-x['bbox'][1])[-1] # only use the maximum face
+    face_kps = draw_kps(pose_image, face_info['kps'])
+
+    width, height = face_kps.size
+
+    # use depth control
+    processed_image_midas = midas(pose_image)
+    processed_image_midas = processed_image_midas.resize(pose_image.size)
+    
+    # enhance face region
+    control_mask = np.zeros([height, width, 3])
+    x1, y1, x2, y2 = face_info["bbox"]
+    x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
+    control_mask[y1:y2, x1:x2] = 255
+    control_mask = Image.fromarray(control_mask.astype(np.uint8))
+
+    image = pipe(
+        prompt=prompt,
+        negative_prompt=n_prompt,
+        image_embeds=face_emb,
+        control_mask=control_mask,
+        image=[face_kps, processed_image_midas],
+        controlnet_conditioning_scale=[0.8,0.8],
+        ip_adapter_scale=0.8,
+        num_inference_steps=30,
+        guidance_scale=5,
+    ).images[0]
+
+    image.save('result.jpg')
\ No newline at end of file

ip_adapter/attention_processor.py

+# 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
+
+try:
+    import xformers
+    import xformers.ops
+    xformers_available = True
+except Exception as e:
+    xformers_available = False
+
+class RegionControler(object):
+    def __init__(self) -> None:
+        self.prompt_image_conditioning = []
+region_control = RegionControler()
+
+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 forward(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        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 forward(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        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)
+
+        if xformers_available:
+            hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
+        else:
+            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)
+        
+        if xformers_available:
+            ip_hidden_states = self._memory_efficient_attention_xformers(query, ip_key, ip_value, None)
+        else:
+            ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+            ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+        ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+
+        # region control
+        if len(region_control.prompt_image_conditioning) == 1:
+            region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None)
+            if region_mask is not None:
+                h, w = region_mask.shape[:2]
+                ratio = (h * w / query.shape[1]) ** 0.5
+                mask = F.interpolate(region_mask[None, None], scale_factor=1/ratio, mode='nearest').reshape([1, -1, 1])
+            else:
+                mask = torch.ones_like(ip_hidden_states)
+            ip_hidden_states = ip_hidden_states * mask     
+
+        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
+
+
+    def _memory_efficient_attention_xformers(self, query, key, value, attention_mask):
+        # TODO attention_mask
+        query = query.contiguous()
+        key = key.contiguous()
+        value = value.contiguous()
+        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
+        # hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+
+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 forward(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        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 forward(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        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)
+
+        # region control
+        if len(region_control.prompt_image_conditioning) == 1:
+            region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None)
+            if region_mask is not None:
+                query = query.reshape([-1, query.shape[-2], query.shape[-1]])
+                h, w = region_mask.shape[:2]
+                ratio = (h * w / query.shape[1]) ** 0.5
+                mask = F.interpolate(region_mask[None, None], scale_factor=1/ratio, mode='nearest').reshape([1, -1, 1])
+            else:
+                mask = torch.ones_like(ip_hidden_states)
+            ip_hidden_states = ip_hidden_states * mask
+
+        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

ip_adapter/resampler.py

+# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+import math
+
+import torch
+import torch.nn as nn
+
+
+# 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,
+    ):
+        super().__init__()
+        
+        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.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):
+        
+        latents = self.latents.repeat(x.size(0), 1, 1)
+        
+        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)
\ No newline at end of file

ip_adapter/utils.py

+import torch.nn.functional as F
+
+
+def is_torch2_available():
+    return hasattr(F, "scaled_dot_product_attention")

model.properties

+# 模型唯一标识
+modelCode = 531
+# 模型名称
+modelName = instantid_pytorch
+# 模型描述
+modelDescription = instantid可以生成包含输入人物特征及给定动作的图像。
+# 应用场景
+appScenario = 推理,AIGC,零售,广媒,设计
+# 框架类型
+frameType = pytorch

requirements.txt

+torch==1.13.1
+torchvision==0.14.1
+diffusers==0.25.1
+opencv-python 
+transformers 
+accelerate 
+insightface
+omegaconf
+gradio
+peft
+controlnet_aux
\ No newline at end of file