app.py

import spaces
import gradio as gr
from PIL import Image
from src.tryon_pipeline import StableDiffusionXLInpaintPipeline as TryonPipeline
from src.unet_hacked_garmnet import UNet2DConditionModel as UNet2DConditionModel_ref
from src.unet_hacked_tryon import UNet2DConditionModel
from transformers import (
    CLIPImageProcessor,
    CLIPVisionModelWithProjection,
    CLIPTextModel,
    CLIPTextModelWithProjection,
)
from diffusers import DDPMScheduler,AutoencoderKL
from typing import List

import torch
import os
import io
import warnings
import requests
from transformers import AutoTokenizer
import numpy as np
from utils_mask import get_mask_location
from torchvision import transforms
import apply_net
from preprocess.humanparsing.run_parsing import Parsing
from preprocess.openpose.run_openpose import OpenPose
from detectron2.data.detection_utils import convert_PIL_to_numpy,_apply_exif_orientation
from torchvision.transforms.functional import to_pil_image
import pillow_heif  # HEIC 이미지 처리용 (아이폰 촬영 사진 포맷)
from urllib.parse import urlparse

# SSL 경고 억제
warnings.filterwarnings("ignore", message=".*OpenSSL.*")
warnings.filterwarnings("ignore", category=UserWarning, module="urllib3")

# requests 세션 설정
session = requests.Session()
session.verify = False  # SSL 검증 비활성화 (개발 환경용)

def pil_to_binary_mask(pil_image, threshold=0):
    np_image = np.array(pil_image)
    grayscale_image = Image.fromarray(np_image).convert("L")
    binary_mask = np.array(grayscale_image) > threshold
    mask = np.zeros(binary_mask.shape, dtype=np.uint8)
    for i in range(binary_mask.shape[0]):
        for j in range(binary_mask.shape[1]):
            if binary_mask[i,j] == True :
                mask[i,j] = 1
    mask = (mask*255).astype(np.uint8)
    output_mask = Image.fromarray(mask)
    return output_mask


base_path = 'yisol/IDM-VTON'
example_path = os.path.join(os.path.dirname(__file__), 'example')

unet = UNet2DConditionModel.from_pretrained(
    base_path,
    subfolder="unet",
    torch_dtype=torch.float16,
)
unet.requires_grad_(False)
tokenizer_one = AutoTokenizer.from_pretrained(
    base_path,
    subfolder="tokenizer",
    revision=None,
    use_fast=False,
)
tokenizer_two = AutoTokenizer.from_pretrained(
    base_path,
    subfolder="tokenizer_2",
    revision=None,
    use_fast=False,
)
noise_scheduler = DDPMScheduler.from_pretrained(base_path, subfolder="scheduler")

text_encoder_one = CLIPTextModel.from_pretrained(
    base_path,
    subfolder="text_encoder",
    torch_dtype=torch.float16,
)
text_encoder_two = CLIPTextModelWithProjection.from_pretrained(
    base_path,
    subfolder="text_encoder_2",
    torch_dtype=torch.float16,
)
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
    base_path,
    subfolder="image_encoder",
    torch_dtype=torch.float16,
    )
vae = AutoencoderKL.from_pretrained(base_path,
                                    subfolder="vae",
                                    torch_dtype=torch.float16,
)

# "stabilityai/stable-diffusion-xl-base-1.0",
UNet_Encoder = UNet2DConditionModel_ref.from_pretrained(
    base_path,
    subfolder="unet_encoder",
    torch_dtype=torch.float16,
)

parsing_model = Parsing(0)
openpose_model = OpenPose(0)

UNet_Encoder.requires_grad_(False)
image_encoder.requires_grad_(False)
vae.requires_grad_(False)
unet.requires_grad_(False)
text_encoder_one.requires_grad_(False)
text_encoder_two.requires_grad_(False)
tensor_transfrom = transforms.Compose(
            [
                transforms.ToTensor(),
                transforms.Normalize([0.5], [0.5]),
            ]
    )

pipe = TryonPipeline.from_pretrained(
        base_path,
        unet=unet,
        vae=vae,
        feature_extractor= CLIPImageProcessor(),
        text_encoder = text_encoder_one,
        text_encoder_2 = text_encoder_two,
        tokenizer = tokenizer_one,
        tokenizer_2 = tokenizer_two,
        scheduler = noise_scheduler,
        image_encoder=image_encoder,
        torch_dtype=torch.float16,
)
pipe.unet_encoder = UNet_Encoder


# 이미지 전처리 함수
def preprocess_image(image):
    # HEIC 이미지 처리
    if isinstance(image, np.ndarray):
        image = Image.fromarray(image)
    
    # HEIC 이미지를 JPEG로 변환
    try:
        output = io.BytesIO()
        image.convert("RGB").save(output, format="JPEG", quality=95)
        output.seek(0)
        image = Image.open(output)
    except Exception as e:
        print(f"Error converting image: {e}")
        # 변환 실패 시 원본 이미지 사용
        image = image.convert("RGB")

    # 이미지 크기 가져오기
    width, height = image.size
    
    # 3:4 비율로 중앙 자르기
    target_width = int(min(width, height * (3 / 4)))
    target_height = int(min(height, width * (4 / 3)))
    left = (width - target_width) / 2
    top = (height - target_height) / 2
    right = (width + target_width) / 2
    bottom = (height + target_height) / 2
    
    # 이미지 자르기
    cropped_img = image.crop((left, top, right, bottom))
    
    # 768x1024로 리사이징
    resized_img = cropped_img.resize((768, 1024), resample=Image.Resampling.LANCZOS)
    
    return resized_img


# URL에서 이미지 가져오기 함수
def load_image_from_url(url):
    try:
        response = session.get(url, stream=True, timeout=10)
        response.raise_for_status()  # HTTP 오류 확인
        
        # 이미지 다운로드
        img = Image.open(response.raw).convert("RGB")
        
        # JPEG로 변환
        output = io.BytesIO()
        img.save(output, format="JPEG", quality=95)
        output.seek(0)
        
        # 변환된 JPEG 이미지 반환
        jpeg_img = Image.open(output)
        return jpeg_img
        
    except requests.exceptions.RequestException as e:
        print(f"Error downloading image from URL: {e}")
        return None
    except Exception as e:
        print(f"Error processing image from URL: {e}")
        return None
    

#URL 입력 처리 함수
def process_url_image(url):
    if not url or not url.strip():
        return None
    
    # URL 유효성 검사
    try:
        result = urlparse(url)
        if not all([result.scheme, result.netloc]):
            print("Invalid URL format")
            return None
    except Exception as e:
        print(f"Error parsing URL: {e}")
        return None
    
    img = load_image_from_url(url)
    if img is None:
        print("Failed to load image from URL")
        return None
        
    return preprocess_image(img)


# Hugging Face LFS OID를 사용하여 densepose 모델 다운로드
def download_densepose_model():
    """
    Hugging Face LFS OID를 사용하여 densepose 모델을 다운로드합니다.
    OID: b8a7382001b16e453bad95ca9dbc68ae8f2b839b304cf90eaf5c27fbdb4dae91
    """
    model_path = './ckpt/densepose/model_final_162be9.pkl'
    
    # 모델 파일이 이미 존재하는지 확인
    if os.path.exists(model_path):
        print(f"DensePose model already exists at {model_path}")
        return model_path
    
    # 디렉토리 생성
    os.makedirs('./ckpt/densepose', exist_ok=True)
    
    # Hugging Face LFS OID를 사용한 다운로드 URL
    oid = "b8a7382001b16e453bad95ca9dbc68ae8f2b839b304cf90eaf5c27fbdb4dae91"
    download_url = f"https://huggingface.co/datasets/hf-internal-testing/fixtures/resolve/{oid}/model_final_162be9.pkl"
    
    try:
        print(f"Downloading DensePose model from {download_url}")
        response = session.get(download_url, stream=True, timeout=300)
        response.raise_for_status()
        
        with open(model_path, 'wb') as f:
            for chunk in response.iter_content(chunk_size=8192):
                f.write(chunk)
        
        print(f"DensePose model downloaded successfully to {model_path}")
        return model_path
        
    except Exception as e:
        print(f"Error downloading DensePose model: {e}")
        # 대체 URL 시도
        fallback_url = "https://dl.fbaipublicfiles.com/densepose/densepose_rcnn_R_50_FPN_s1x/165712039/model_final_162be9.pkl"
        try:
            print(f"Trying fallback URL: {fallback_url}")
            response = session.get(fallback_url, stream=True, timeout=300)
            response.raise_for_status()
            
            with open(model_path, 'wb') as f:
                for chunk in response.iter_content(chunk_size=8192):
                    f.write(chunk)
            
            print(f"DensePose model downloaded successfully from fallback URL to {model_path}")
            return model_path
            
        except Exception as e2:
            print(f"Error downloading from fallback URL: {e2}")
            return None


@spaces.GPU
def start_tryon(dict,garm_img,garment_des,is_checked,denoise_steps,seed, is_checked_crop):
    device = "cuda"
    
    openpose_model.preprocessor.body_estimation.model.to(device)
    pipe.to(device)
    pipe.unet_encoder.to(device)

    garm_img= garm_img.convert("RGB").resize((768,1024))
    human_img_orig = dict["background"].convert("RGB")    
    
    if is_checked_crop:
        width, height = human_img_orig.size
        target_width = int(min(width, height * (3 / 4)))
        target_height = int(min(height, width * (4 / 3)))
        left = (width - target_width) / 2
        top = (height - target_height) / 2
        right = (width + target_width) / 2
        bottom = (height + target_height) / 2
        cropped_img = human_img_orig.crop((left, top, right, bottom))
        crop_size = cropped_img.size
        human_img = cropped_img.resize((768,1024))
    else:
        human_img = human_img_orig.resize((768,1024))


    if is_checked:
        keypoints = openpose_model(human_img.resize((384,512)))
        model_parse, _ = parsing_model(human_img.resize((384,512)))
        mask, mask_gray = get_mask_location('hd', "upper_body", model_parse, keypoints)
        mask = mask.resize((768,1024))
    else:
        mask = pil_to_binary_mask(dict['layers'][0].convert("RGB").resize((768, 1024)))
        # mask = transforms.ToTensor()(mask)
        # mask = mask.unsqueeze(0)
    mask_gray = (1-transforms.ToTensor()(mask)) * tensor_transfrom(human_img)
    mask_gray = to_pil_image((mask_gray+1.0)/2.0)


    human_img_arg = _apply_exif_orientation(human_img.resize((384,512)))
    human_img_arg = convert_PIL_to_numpy(human_img_arg, format="BGR")
     
    

    # DensePose 모델 다운로드 및 경로 설정
    densepose_model_path = download_densepose_model()
    if densepose_model_path is None:
        print("Failed to download DensePose model, using default path")
        densepose_model_path = './ckpt/densepose/model_final_162be9.pkl'
    
    args = apply_net.create_argument_parser().parse_args(('show', './configs/densepose_rcnn_R_50_FPN_s1x.yaml', densepose_model_path, 'dp_segm', '-v', '--opts', 'MODEL.DEVICE', 'cuda'))
    # verbosity = getattr(args, "verbosity", None)
    pose_img = args.func(args,human_img_arg)    
    pose_img = pose_img[:,:,::-1]    
    pose_img = Image.fromarray(pose_img).resize((768,1024))
    
    with torch.no_grad():
        # Extract the images
        with torch.cuda.amp.autocast():
            with torch.no_grad():
                prompt = "model is wearing " + garment_des
                negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
                with torch.inference_mode():
                    (
                        prompt_embeds,
                        negative_prompt_embeds,
                        pooled_prompt_embeds,
                        negative_pooled_prompt_embeds,
                    ) = pipe.encode_prompt(
                        prompt,
                        num_images_per_prompt=1,
                        do_classifier_free_guidance=True,
                        negative_prompt=negative_prompt,
                    )
                                    
                    prompt = "a photo of " + garment_des
                    negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
                    if not isinstance(prompt, List):
                        prompt = [prompt] * 1
                    if not isinstance(negative_prompt, List):
                        negative_prompt = [negative_prompt] * 1
                    with torch.inference_mode():
                        (
                            prompt_embeds_c,
                            _,
                            _,
                            _,
                        ) = pipe.encode_prompt(
                            prompt,
                            num_images_per_prompt=1,
                            do_classifier_free_guidance=False,
                            negative_prompt=negative_prompt,
                        )



                    pose_img =  tensor_transfrom(pose_img).unsqueeze(0).to(device,torch.float16)
                    garm_tensor =  tensor_transfrom(garm_img).unsqueeze(0).to(device,torch.float16)
                    generator = torch.Generator(device).manual_seed(seed) if seed is not None else None
                    images = pipe(
                        prompt_embeds=prompt_embeds.to(device,torch.float16),
                        negative_prompt_embeds=negative_prompt_embeds.to(device,torch.float16),
                        pooled_prompt_embeds=pooled_prompt_embeds.to(device,torch.float16),
                        negative_pooled_prompt_embeds=negative_pooled_prompt_embeds.to(device,torch.float16),
                        num_inference_steps=denoise_steps,
                        generator=generator,
                        strength = 1.0,
                        pose_img = pose_img.to(device,torch.float16),
                        text_embeds_cloth=prompt_embeds_c.to(device,torch.float16),
                        cloth = garm_tensor.to(device,torch.float16),
                        mask_image=mask,
                        image=human_img, 
                        height=1024,
                        width=768,
                        ip_adapter_image = garm_img.resize((768,1024)),
                        guidance_scale=2.0,
                    )[0]

    if is_checked_crop:
        out_img = images[0].resize(crop_size)        
        human_img_orig.paste(out_img, (int(left), int(top)))    
        return human_img_orig, mask_gray
    else:
        return images[0], mask_gray
    # return images[0], mask_gray

garm_list = os.listdir(os.path.join(example_path,"cloth"))
garm_list_path = [os.path.join(example_path,"cloth",garm) for garm in garm_list]

human_list = os.listdir(os.path.join(example_path,"human"))
human_list_path = [os.path.join(example_path,"human",human) for human in human_list]

human_ex_list = []
for ex_human in human_list_path:
    ex_dict= {}
    ex_dict['background'] = ex_human
    ex_dict['layers'] = None
    ex_dict['composite'] = None
    human_ex_list.append(ex_dict)

##default human


image_blocks = gr.Blocks().queue()
with image_blocks as demo:
    gr.Markdown("## DXCO : GENAI-VTON")
    gr.Markdown("임성남, 윤지영, 김준기 based on IDM-VTON")
    gr.Markdown("이미지는 3:4비율(384x512 또는 768x1024)로 올려주세요")
    with gr.Row():
        with gr.Column():
            imgs = gr.ImageEditor(sources='upload', type="pil", label='대상 이미지', interactive=True)
            img_url_input = gr.Textbox(label="대상 이미지 URL", placeholder="예) https://example.com/human_image.jpg")
            with gr.Row():
                with gr.Row():
                    is_checked = gr.Checkbox(label="Yes", info="자동 마스킹",value=True)
            example = gr.Examples(
                inputs=imgs,
                examples_per_page=8,
                examples=human_ex_list
            )

        with gr.Column():
            garm_img = gr.Image(label="의상 이미지", sources='upload', type="pil")
            garm_url_input = gr.Textbox(label="의상 이미지 URL", placeholder="예) https://example.com/garment.jpg")
            with gr.Row(elem_id="prompt-container"):
                with gr.Row():
                    prompt = gr.Textbox(placeholder="Description of garment ex) Short Sleeve Round Neck T-shirts", show_label=False, elem_id="prompt")
            example = gr.Examples(
                inputs=garm_img,
                examples_per_page=8,
                examples=garm_list_path)
    with gr.Row():
        try_button = gr.Button(value="Try-on")
    with gr.Row():
        with gr.Column():
            # image_out = gr.Image(label="Output", elem_id="output-img", height=400)
            masked_img = gr.Image(label="Masked image output", elem_id="masked-img",show_share_button=False)
        with gr.Column():
            # image_out = gr.Image(label="Output", elem_id="output-img", height=400)
            image_out = gr.Image(label="Output", elem_id="output-img",show_share_button=False)
        
    
        # with gr.Accordion(label="Advanced Settings", open=False):
        #     with gr.Row():
        #         denoise_steps = gr.Number(label="Denoising Steps", minimum=20, maximum=40, value=30, step=1)
        #         seed = gr.Number(label="Seed", minimum=-1, maximum=2147483647, step=1, value=42)

    # is_checked = gr.Number(value=True)

    # 이미지 업로드 시 전처리
    imgs.change(
        fn=preprocess_image,
        inputs=imgs,
        outputs=imgs,  # 전처리된 이미지를 ImageEditor에 다시 표시
    )


    # 대상 이미지: URL 입력 처리
    img_url_input.change(
        fn=lambda url: process_url_image(url),
        inputs=img_url_input,
        outputs=imgs,
    )

    # 의상 이미지: URL 입력 처리
    garm_url_input.change(
        fn=lambda url: process_url_image(url),
        inputs=garm_url_input,
        outputs=garm_img,
    )

    is_checked_crop = True
    denoise_steps = 30
    seed = 42
    try_button.click(
        fn=lambda *args: start_tryon(*args, is_checked_crop=is_checked_crop, denoise_steps=denoise_steps, seed=seed),
        inputs=[imgs, garm_img, prompt, is_checked],
        outputs=[image_out, masked_img],
        api_name='tryon'
    )

# 앱 시작 시 DensePose 모델 미리 다운로드
print("Initializing DensePose model...")
try:
    download_densepose_model()
    print("DensePose model initialization completed.")
except Exception as e:
    print(f"Warning: Could not download DensePose model: {e}")
    print("The model will be downloaded when needed during inference.")

# 앱 실행
if __name__ == "__main__":
    try:
        print("Starting GENAI-VTON application...")
        image_blocks.launch(server_name="0.0.0.0", server_port=7860, share=False)
    except Exception as e:
        print(f"Error starting the application: {e}")
        print("Please check if all required dependencies are installed.")