app.py

import spaces
import gradio as gr
import torch
import tempfile
import os
from PIL import Image
from gradio_client import Client, handle_file
from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler

HF_TOKEN = os.environ.get("HF_TOKEN", "")
_OOTD_HEADERS = {"Authorization": f"Bearer {HF_TOKEN}"} if HF_TOKEN else None

# ─── Carga de modelos ─────────────────────────────────────────────────────────

multiview_pipe = None


def load_models():
    global multiview_pipe

    multiview_pipe = DiffusionPipeline.from_pretrained(
        "sudo-ai/zero123plus-v1.2",
        custom_pipeline="sudo-ai/zero123plus-pipeline",
        torch_dtype=torch.float16,
        trust_remote_code=True,
    )
    multiview_pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(
        multiview_pipe.scheduler.config,
        timestep_spacing="trailing",
    )


load_models()


# ─── Helpers ──────────────────────────────────────────────────────────────────

def _preprocess_for_zero123(img: Image.Image, size: int = 320) -> Image.Image:
    """Centra y hace padding cuadrado con fondo blanco para Zero123++."""
    img = img.convert("RGBA")
    bbox = img.getbbox()
    if bbox:
        img = img.crop(bbox)
    max_dim = max(img.size)
    padded = Image.new("RGBA", (max_dim, max_dim), (255, 255, 255, 255))
    padded.paste(img, ((max_dim - img.width) // 2, (max_dim - img.height) // 2), img.split()[3])
    return padded.convert("RGB").resize((size, size), Image.LANCZOS)


# ─── Endpoint 1 — Try-on foto ─────────────────────────────────────────────────

def tryon(person_image: Image.Image, garment_image: Image.Image) -> Image.Image:
    """
    Recibe foto del usuario + imagen de prenda.
    Usa OOTDiffusion (open source) para aplicar la prenda sobre la persona.
    La foto del usuario debe ser de cuerpo completo, de frente.
    """
    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f_person:
        person_image.save(f_person.name, format="JPEG")
        person_path = f_person.name

    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f_garment:
        garment_image.save(f_garment.name, format="JPEG")
        garment_path = f_garment.name

    try:
        client = Client("eduardo4547/OOTDiffusion", headers=_OOTD_HEADERS, verbose=False)
        result = client.predict(
            vton_img=handle_file(person_path),
            garm_img=handle_file(garment_path),
            n_samples=1,
            n_steps=20,
            image_scale=2.0,
            seed=42,
            api_name="/process_hd",
        )
        # OOTDiffusion devuelve filepath directo (gr.Image output)
        output_path = result
        return Image.open(output_path).convert("RGB")
    finally:
        os.unlink(person_path)
        os.unlink(garment_path)


# ─── Endpoint 2 — Multi-vista ─────────────────────────────────────────────────

@spaces.GPU
def multiview(front_image: Image.Image) -> list[Image.Image]:
    """
    Genera 4 vistas (front, left, right, back) usando Zero123++.
    Recibe imagen del try-on con fondo limpio (rembg aplicado en el backend).
    """
    multiview_pipe.to("cuda")

    processed = _preprocess_for_zero123(front_image)
    result = multiview_pipe(processed, num_inference_steps=75).images[0]

    w, h = result.size
    tile_w, tile_h = w // 3, h // 2

    # Zero123++ v1.2: azimuths 30°, 90°, 150° (fila top) | 210°, 270°, 330° (fila bottom)
    views = {
        "front": result.crop((tile_w * 2, tile_h, w,          h)),      # 330° ≈ front-left
        "right": result.crop((tile_w,     0,      tile_w * 2, tile_h)), # 90°  = right profile
        "back":  result.crop((0,          tile_h, tile_w,     h)),      # 210° = back-left
        "left":  result.crop((tile_w,     tile_h, tile_w * 2, h)),      # 270° = left profile
    }

    multiview_pipe.to("cpu")
    return [views["front"], views["left"], views["right"], views["back"]]


# ─── Interfaz Gradio ──────────────────────────────────────────────────────────

with gr.Blocks() as demo:
    with gr.Tab("Try-On"):
        gr.Interface(
            fn=tryon,
            inputs=[
                gr.Image(type="pil", label="Foto del usuario (cuerpo completo, frontal)"),
                gr.Image(type="pil", label="Imagen de la prenda"),
            ],
            outputs=gr.Image(type="pil", label="Resultado"),
            api_name="tryon",
        )
    with gr.Tab("Multi-Vista"):
        gr.Interface(
            fn=multiview,
            inputs=gr.Image(type="pil", label="Imagen frontal del try-on"),
            outputs=[
                gr.Image(type="pil", label="Front"),
                gr.Image(type="pil", label="Left"),
                gr.Image(type="pil", label="Right"),
                gr.Image(type="pil", label="Back"),
            ],
            api_name="multiview",
        )

demo.launch(show_error=True)