app.py

#!/usr/bin/env python

import os
import random

import gradio as gr
import numpy as np
import PIL.Image
import torch
import torchvision.transforms.functional as TF
from diffusers import (
    AutoencoderKL,
    EulerAncestralDiscreteScheduler,
    StableDiffusionXLAdapterPipeline,
    T2IAdapter,
)
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
from transformers import AutoTokenizer
import spaces
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

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

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

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

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

    OUTPUT_WIDTH, OUTPUT_HEIGHT = dict.size
    
    garm_img= garm_img.convert("RGB").resize((768,1024))
    
    human_img_orig = dict.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', area, 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")
     
    

    args = apply_net.create_argument_parser().parse_args(('show', './configs/densepose_rcnn_R_50_FPN_s1x.yaml', './ckpt/densepose/model_final_162be9.pkl', '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.resize((OUTPUT_WIDTH, OUTPUT_HEIGHT))
    else:
        return images[0].resize((OUTPUT_WIDTH, OUTPUT_HEIGHT))
    # 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]

image_blocks = gr.Blocks(css="style.css").queue()
with image_blocks as demo:
    gr.Markdown("## MyFit-AI")
    with gr.Row():
        with gr.Column():
            imgs = gr.Image(label='Human', type="pil")
            garm_img = gr.Image(label="Garment", type="pil")
            with gr.Row(elem_id="prompt-container"):
                with gr.Row():
                    prompt = gr.Textbox(placeholder="Description of garment ex) Neck T-shirts", show_label=False, elem_id="prompt")
            example = gr.Examples(
                inputs=garm_img,
                examples_per_page=8,
                examples=garm_list_path)
            
            try_button = gr.Button(value="Try-on", variant="primary")
            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)
                with gr.Row():
                    area = gr.Dropdown(["upper_body","lower_body"], value="upper_body", label="garment zone")
                with gr.Row():
                    is_checked = gr.Checkbox(label="Yes", info="Use auto-generated mask (Takes 5 seconds)",value=True,visible=False)
                with gr.Row():
                    is_checked_crop = gr.Checkbox(label="Yes", info="Use auto-crop & resizing",value=True,visible=False)
        with gr.Column():
            image_out = gr.Image(label="Output", elem_id="output-img",show_share_button=True)

    try_button.click(fn=start_tryon, inputs=[imgs, garm_img, prompt, is_checked,is_checked_crop, denoise_steps, seed, area], outputs=[image_out], api_name='tryon')

    # Doodly - T2I-Adapter-SDXL Sketch
    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": "3D Model",
            "prompt": "professional 3d model {prompt} . octane render, highly detailed, volumetric, dramatic lighting",
            "negative_prompt": "ugly, deformed, noisy, low poly, blurry, painting",
        },
        {
            "name": "Anime",
            "prompt": "anime artwork {prompt} . anime style, key visual, vibrant, studio anime,  highly detailed",
            "negative_prompt": "photo, deformed, black and white, realism, disfigured, low contrast",
        },
        {
            "name": "Digital Art",
            "prompt": "concept art {prompt} . digital artwork, illustrative, painterly, matte painting, highly detailed",
            "negative_prompt": "photo, photorealistic, realism, ugly",
        },
        {
            "name": "Photographic",
            "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": "Pixel art",
            "prompt": "pixel-art {prompt} . low-res, blocky, pixel art style, 8-bit graphics",
            "negative_prompt": "sloppy, messy, blurry, noisy, highly detailed, ultra textured, photo, realistic",
        },
        {
            "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": "Manga",
            "prompt": "manga style {prompt} . vibrant, high-energy, detailed, iconic, Japanese comic style",
            "negative_prompt": "ugly, deformed, noisy, blurry, low contrast, realism, photorealistic, Western comic style",
        },
    ]

    styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in style_list}
    STYLE_NAMES = list(styles.keys())
    DEFAULT_STYLE_NAME = "(No style)"

    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

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    if torch.cuda.is_available():
        model_id = "stabilityai/stable-diffusion-xl-base-1.0"
        adapter = T2IAdapter.from_pretrained(
            "TencentARC/t2i-adapter-sketch-sdxl-1.0", torch_dtype=torch.float16, variant="fp16"
        )
        scheduler = EulerAncestralDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler")
        pipe = StableDiffusionXLAdapterPipeline.from_pretrained(
            model_id,
            vae=AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16),
            adapter=adapter,
            scheduler=scheduler,
            torch_dtype=torch.float16,
            variant="fp16",
        )
        pipe.to(device)
    else:
        pipe = None

    MAX_SEED = np.iinfo(np.int32).max

    def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
        if randomize_seed:
            seed = random.randint(0, MAX_SEED)
        return seed

    def run(
        image: PIL.Image.Image,
        prompt: str,
        negative_prompt: str,
        style_name: str = DEFAULT_STYLE_NAME,
        num_steps: int = 25,
        guidance_scale: float = 5,
        adapter_conditioning_scale: float = 0.8,
        adapter_conditioning_factor: float = 0.8,
        seed: int = 0,
        progress=gr.Progress(track_tqdm=True),
    ) -> PIL.Image.Image:
        image = image.convert("RGB")
        image = TF.to_tensor(image) > 0.5
        image = TF.to_pil_image(image.to(torch.float32))

        prompt, negative_prompt = apply_style(style_name, prompt, negative_prompt)

        generator = torch.Generator(device=device).manual_seed(seed)
        out = pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            image=image,
            num_inference_steps=num_steps,
            generator=generator,
            guidance_scale=guidance_scale,
            adapter_conditioning_scale=adapter_conditioning_scale,
            adapter_conditioning_factor=adapter_conditioning_factor,
        ).images[0]
        return out

    gr.Markdown("# Doodly - T2I-Adapter-SDXL **Sketch**")
    with gr.Row():
        with gr.Column():
            with gr.Group():
                image = gr.Image(
                    type="pil",
                    image_mode="L",
                    shape=(1024, 1024),
                    brush_radius=4,
                    height=440
                )
                prompt = gr.Textbox(label="Prompt")
                style = gr.Dropdown(label="Style", choices=STYLE_NAMES, value=DEFAULT_STYLE_NAME)
                run_button = gr.Button("Run")
            with gr.Accordion("Advanced options", open=False):
                negative_prompt = gr.Textbox(
                    label="Negative prompt",
                    value=" extra digit, fewer digits, cropped, worst quality, low quality, glitch, deformed, mutated, ugly, disfigured",
                )
                num_steps = gr.Slider(
                    label="Number of steps",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=25,
                )
                guidance_scale = gr.Slider(
                    label="Guidance scale",
                    minimum=0.1,
                    maximum=10.0,
                    step=0.1,
                    value=5,
                )
                adapter_conditioning_scale = gr.Slider(
                    label="Adapter conditioning scale",
                    minimum=0.5,
                    maximum=1,
                    step=0.1,
                    value=0.8,
                )
                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,
                )
                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():
            result = gr.Image(label="Result", height=400)

    inputs = [
        image,
        prompt,
        negative_prompt,
        style,
        num_steps,
        guidance_scale,
        adapter_conditioning_scale,
        adapter_conditioning_factor,
        seed,
    ]
    prompt.submit(
        fn=randomize_seed_fn,
        inputs=[seed, randomize_seed],
        outputs=seed,
        queue=False,
        api_name=False,
    ).then(
        fn=run,
        inputs=inputs,
        outputs=result,
        api_name=False,
    )
    negative_prompt.submit(
        fn=randomize_seed_fn,
        inputs=[seed, randomize_seed],
        outputs=seed,
        queue=False,
        api_name=False,
    ).then(
        fn=run,
        inputs=inputs,
        outputs=result,
        api_name=False,
    )
    run_button.click(
        fn=randomize_seed_fn,
        inputs=[seed, randomize_seed],
        outputs=seed,
        queue=False,
        api_name=False,
    ).then(
        fn=run,
        inputs=inputs,
        outputs=result,
        api_name=False,
    )

if __name__ == "__main__":
    demo.queue(max_size=20).launch(auth=("gini", "pick"))