app.py

"""
app.py — OOTDiffusion Hugging Face Space
Place this file in the ROOT of your Space repo alongside the
OOTDiffusion source folders: ootd/, run/, preprocess/, checkpoints/

README.md front-matter required:
---
title: OOTDiffusion Virtual Try-On
emoji: 👗
colorFrom: purple
colorTo: pink
sdk: gradio
sdk_version: 4.16.0
app_file: app.py
pinned: false
license: cc-by-nc-sa-4.0
---
"""

import sys
import os

# ── Path setup ────────────────────────────────────────────────────────────────
ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
RUN_DIR  = os.path.join(ROOT_DIR, "run")
sys.path.insert(0, ROOT_DIR)
sys.path.insert(0, RUN_DIR)

import torch
import numpy as np
import gradio as gr
from PIL import Image

# ── Device ────────────────────────────────────────────────────────────────────
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"[OOTDiffusion] Device: {DEVICE}")

# ── Lazy-load models (loaded once on first request) ───────────────────────────
_pipe_hd = None   # VITON-HD  — half-body
_pipe_dc = None   # Dress Code — full-body


def load_pipeline(model_type: str):
    """Import and cache the requested OOTDiffusion pipeline."""
    global _pipe_hd, _pipe_dc

    if model_type == "hd":
        if _pipe_hd is None:
            from ootd.inference_ootd_hd import OOTDiffusionHD
            print("[OOTDiffusion] Loading HD pipeline …")
            _pipe_hd = OOTDiffusionHD(ROOT_DIR)
        return _pipe_hd

    else:  # dc
        if _pipe_dc is None:
            from ootd.inference_ootd_dc import OOTDiffusionDC
            print("[OOTDiffusion] Loading DC pipeline …")
            _pipe_dc = OOTDiffusionDC(ROOT_DIR)
        return _pipe_dc


# ── Category mapping ──────────────────────────────────────────────────────────
CATEGORY_MAP = {
    "Upper-body":  0,
    "Lower-body":  1,
    "Dress":       2,
}


# ── Main inference function ───────────────────────────────────────────────────
def run_tryon(
    model_image,
    cloth_image,
    model_type,
    category_label,
    n_samples,
    n_steps,
    guidance_scale,
    seed,
):
    if model_image is None:
        raise gr.Error("Please upload a model (person) image.")
    if cloth_image is None:
        raise gr.Error("Please upload a garment image.")

    # Convert to PIL just in case Gradio passes numpy arrays
    if isinstance(model_image, np.ndarray):
        model_image = Image.fromarray(model_image)
    if isinstance(cloth_image, np.ndarray):
        cloth_image = Image.fromarray(cloth_image)

    model_image = model_image.convert("RGB")
    cloth_image = cloth_image.convert("RGB")

    category_idx = CATEGORY_MAP[category_label]

    try:
        pipe = load_pipeline(model_type)
    except Exception as e:
        raise gr.Error(
            f"Failed to load model: {e}\n"
            "Make sure checkpoints/ and ootd/ folders are present."
        )

    try:
        if model_type == "hd":
            result = pipe(
                model_type="hd",
                category=category_idx,
                image_garm=cloth_image,
                image_vton=model_image,
                mask=None,
                image_ori=model_image,
                num_samples=int(n_samples),
                num_steps=int(n_steps),
                guidance_scale=guidance_scale,
                seed=int(seed),
            )
        else:
            result = pipe(
                model_type="dc",
                category=category_idx,
                image_garm=cloth_image,
                image_vton=model_image,
                mask=None,
                image_ori=model_image,
                num_samples=int(n_samples),
                num_steps=int(n_steps),
                guidance_scale=guidance_scale,
                seed=int(seed),
            )
    except Exception as e:
        raise gr.Error(f"Inference failed: {e}")

    # result is expected to be a list of PIL Images
    if isinstance(result, (list, tuple)):
        return result
    return [result]


# ── Gradio UI ─────────────────────────────────────────────────────────────────
with gr.Blocks(title="OOTDiffusion Virtual Try-On", theme=gr.themes.Soft()) as demo:

    gr.Markdown(
        """
        # 👗 OOTDiffusion — Virtual Try-On
        **[AAAI 2025]** Upload a *model photo* and a *garment image*, choose settings, and click **Run Try-On**.
        > ⚠️ Non-commercial use only (CC-BY-NC-SA-4.0)
        """
    )

    with gr.Row():
        # ── Left column: inputs ───────────────────────────────────────────────
        with gr.Column(scale=1):
            model_img = gr.Image(
                label="Model Image (person)",
                type="pil",
                height=400,
            )
            cloth_img = gr.Image(
                label="Garment Image (clothing)",
                type="pil",
                height=400,
            )

        # ── Middle column: settings ───────────────────────────────────────────
        with gr.Column(scale=1):
            model_type = gr.Radio(
                choices=["hd", "dc"],
                value="hd",
                label="Model Type",
                info="hd = half-body (VITON-HD) | dc = full-body (Dress Code)",
            )
            category = gr.Dropdown(
                choices=list(CATEGORY_MAP.keys()),
                value="Upper-body",
                label="Garment Category",
                info="Only used when Model Type is 'dc'",
            )
            n_samples = gr.Slider(
                minimum=1, maximum=4, step=1, value=1,
                label="Number of Samples",
            )
            n_steps = gr.Slider(
                minimum=10, maximum=40, step=5, value=20,
                label="Denoising Steps",
                info="More steps = better quality but slower",
            )
            guidance_scale = gr.Slider(
                minimum=1.0, maximum=5.0, step=0.5, value=2.0,
                label="Guidance Scale",
            )
            seed = gr.Number(
                value=42,
                label="Seed (-1 = random)",
                precision=0,
            )

            run_btn = gr.Button("🚀 Run Try-On", variant="primary")

        # ── Right column: outputs ─────────────────────────────────────────────
        with gr.Column(scale=1):
            output_gallery = gr.Gallery(
                label="Try-On Results",
                columns=2,
                height=500,
                object_fit="contain",
            )

    gr.Markdown(
        """
        ### Tips
        - **HD model**: best for upper-body garments on half-body photos
        - **DC model**: supports upper-body / lower-body / dress on full-body photos
        - Increasing **steps** to 30–40 noticeably improves quality
        - Set **seed = -1** for random results each run
        """
    )

    # ── Wire up the button ────────────────────────────────────────────────────
    run_btn.click(
        fn=run_tryon,
        inputs=[
            model_img,
            cloth_img,
            model_type,
            category,
            n_samples,
            n_steps,
            guidance_scale,
            seed,
        ],
        outputs=output_gallery,
    )

# ── Launch ────────────────────────────────────────────────────────────────────
if __name__ == "__main__":
    demo.launch()