app.py

# @spaces.GPU
import spaces
import warnings
warnings.filterwarnings("ignore", message="Can't initialize NVML")

import torch
import gradio as gr
from diffusers import StableDiffusionPipeline
from PIL import Image
import numpy as np
from sklearn.cluster import KMeans

# === Configuración base ===
BASE_MODEL = "runwayml/stable-diffusion-v1-5"
MODEL_CACHE = {}
DEFAULT_LORA = "marvin90/pixelartmaster-lite-faces"

# === Carga del modelo con ZeroGPU o CPU ===
@spaces.GPU
def load_model(repo_id):
    """Carga el modelo y LoRA solo una vez, usando GPU si está disponible."""
    device = "cuda" if torch.cuda.is_available() else "cpu"
    dtype = torch.float16 if device == "cuda" else torch.float32

    if repo_id not in MODEL_CACHE:
        print(f"🌐 Cargando modelo en {device}: {repo_id}")
        pipe = StableDiffusionPipeline.from_pretrained(
            BASE_MODEL,
            torch_dtype=dtype
        ).to(device)

        # Intentar cargar LoRA
        try:
            pipe.load_lora_weights(repo_id)
            print("✅ LoRA cargado correctamente")
        except Exception as e:
            print(f"⚠️ No se pudo cargar LoRA: {e}")

        MODEL_CACHE[repo_id] = pipe

    return MODEL_CACHE[repo_id]

# === Postprocesamiento con KMeans ===
def apply_kmeans_pil(image, k=12):
    """Reduce la cantidad de colores para estilo pixel art."""
    pixels = np.array(image)
    if pixels.ndim == 3 and pixels.shape[2] == 4:
        pixels = pixels[:, :, :3]
    shape = pixels.shape[:2]
    flat_pixels = pixels.reshape(-1, 3)

    kmeans = KMeans(n_clusters=k, n_init=10, random_state=42)
    labels = kmeans.fit_predict(flat_pixels)
    new_colors = np.clip(kmeans.cluster_centers_.astype("uint8"), 0, 255)
    clustered_pixels = new_colors[labels].reshape(shape[0], shape[1], 3)

    return Image.fromarray(clustered_pixels)

# === Generación principal ===
@spaces.GPU
def generar_pixelart(prompt):
    """Genera imagen pixel art con postprocesado."""
    pipe = load_model(DEFAULT_LORA)

    negative_prompt = (
        "low quality, bad anatomy, bad hands, text, error, "
        "missing fingers, extra digit, fewer digits, "
        "worst quality, blurry, smudged, grainy, noise, "
        "jpeg artifacts, signature, watermark, username"
    )

    # Generar imagen base
    image = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=30,
        guidance_scale=7.5
    ).images[0]

    # Pixelado
    pixelated = image.resize((128, 128), Image.NEAREST).resize((512, 512), Image.NEAREST)

    # Aplicar KMeans
    final_image = apply_kmeans_pil(pixelated, k=12)

    return final_image

# === Interfaz Gradio ===
with gr.Blocks() as demo:
    gr.Markdown("## 🎮 PixelArtMaster Lite Faces")
    gr.Markdown(
        "Generador ligero de retratos futuristas en pixel art desde texto, "
        "con reducción de colores usando K-means y escalado tipo pixelado. "
        "Modelo LoRA entrenado por [AmericaPixelGames](https://americapixelgames.com)."
    )

    prompt = gr.Textbox(label="Prompt de entrada")
    output = gr.Image(label="Resultado estilo Pixel Art")
    btn = gr.Button("🎨 Generar")

    btn.click(fn=generar_pixelart, inputs=prompt, outputs=output)

    gr.Markdown("---")
    gr.Markdown("Desarrollado por **[AmericaPixelGames](https://americapixelgames.com)** 🎮")

if __name__ == "__main__":
    demo.queue().launch(debug=True)