app.py

import patch_gradio
import os
import sys
import torch
import gradio as gr
from PIL import Image
import gc
import traceback
from huggingface_hub import snapshot_download

sys.path.insert(0, '/app/CatVTON')

from model.pipeline import CatVTONPipeline
from model.cloth_masker import AutoMasker
from utils import init_weight_dtype, resize_and_crop, resize_and_padding

pipeline = None
automasker = None

def load_models():
    global pipeline, automasker
    
    if pipeline is not None and automasker is not None:
        return
    
    print("🔄 Loading models...", file=sys.stderr)
    
    try:
        repo_path = snapshot_download(
            repo_id="zhengchong/CatVTON",
            cache_dir="/tmp/models"
        )
        
        nsfw_path = "/tmp/NSFW.jpg"
        if not os.path.exists(nsfw_path):
            Image.new('RGB', (512, 512), color='black').save(nsfw_path)
        
        pipeline = CatVTONPipeline(
            base_ckpt="booksforcharlie/stable-diffusion-inpainting",
            attn_ckpt=repo_path,
            attn_ckpt_version="mix",
            weight_dtype=torch.float16,
            use_tf32=True,
            device='cuda'
        )
        automasker = AutoMasker(
            densepose_ckpt=os.path.join(repo_path, "DensePose"),
            schp_ckpt=os.path.join(repo_path, "SCHP"),
            device='cpu'
        )
        
        print("✅ Models loaded!", file=sys.stderr)
        
    except Exception as e:
        print(f"❌ Error: {e}", file=sys.stderr)
        traceback.print_exc()
        raise

def generate_tryon(person_img, cloth_img, garment_category):  # Added parameter
    print("="*50, file=sys.stderr)
    print(f"Received - Person: {type(person_img)}, Cloth: {type(cloth_img)}, Category: {garment_category}", file=sys.stderr)
    
    if person_img is None or cloth_img is None:
        raise gr.Error("Both images required!")
    
    try:
        print("Images received as PIL", file=sys.stderr)
        
        load_models()
        
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        
        person_img = person_img.copy()
        cloth_img = cloth_img.copy()
        
        target_height = 1024
        target_width = 768
        person_img = resize_and_crop(person_img, (target_width, target_height))
        cloth_img = resize_and_padding(cloth_img, (target_width, target_height))
        
        # CHANGED: Use selected category instead of hardcoded "upper"
        mask = automasker(person_img, garment_category)['mask']
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        
        result = pipeline(
            image=person_img,
            condition_image=cloth_img,
            mask=mask,
            num_inference_steps=50,
            guidance_scale=2.5,
            seed=None,
            height=target_height,
            width=target_width
        )[0]
        
        print("✅ Success!", file=sys.stderr)
        return result
        
    except Exception as e:
        print(f"❌ Error: {e}", file=sys.stderr)
        traceback.print_exc()
        raise gr.Error(str(e))

# Updated Interface with category dropdown
demo = gr.Interface(
    fn=generate_tryon,
    inputs=[
        gr.Image(label="Person Image", type="pil"),
        gr.Image(label="Garment Image", type="pil"),
        gr.Dropdown(
            choices=["upper", "lower", "overall"],
            value="upper",
            label="Garment Category",
            info="Select: upper (tops), lower (pants/skirts), overall (dresses/full outfits)"
        )
    ],
    outputs=gr.Image(label="Result", type="pil"),
    title="Try-Space Virtual Try-On",
    description="Upload person and garment images. Select category. Processing takes 2-3 minutes on GPU T4.",
    api_name="generate_tryon",
    allow_flagging="never"
)

if __name__ == "__main__":
    print("🚀 Starting...", file=sys.stderr)
    try:
        load_models()
    except:
        pass
    demo.queue().launch(show_error=True)