factory/handler.py

"""
THE FACTORY - Trabajo Iterativo Generativo
RunPod Serverless Handler

Tareas:
- image_generate: Genera imagen desde prompt
- image_variant: Genera variante de imagen existente
- image_upscale: Aumenta resolución
"""

import runpod
import base64
import os
from datetime import datetime
from io import BytesIO

# Force CUDA device
os.environ["CUDA_VISIBLE_DEVICES"] = "0"

# Modelos disponibles
MODELS = {
    "sdxl": "stabilityai/stable-diffusion-xl-base-1.0",
    "flux": "black-forest-labs/FLUX.1-schnell",
    "sdxl-turbo": "stabilityai/sdxl-turbo"
}

# Lazy loading de modelos
_loaded_models = {}


def get_model(model_name: str):
    """Carga modelo bajo demanda."""
    global _loaded_models

    if model_name not in _loaded_models:
        try:
            import torch
            from diffusers import AutoPipelineForText2Image

            # Force CUDA
            device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

            model_id = MODELS.get(model_name, MODELS["sdxl-turbo"])

            pipe = AutoPipelineForText2Image.from_pretrained(
                model_id,
                torch_dtype=torch.float16,
                variant="fp16",
                use_safetensors=True
            )
            pipe = pipe.to(device)

            _loaded_models[model_name] = pipe
        except Exception as e:
            return None, str(e)

    return _loaded_models[model_name], None


def generate_image(prompt: str, model: str = "sdxl-turbo",
                   width: int = 1024, height: int = 1024,
                   steps: int = 4, guidance: float = 0.0) -> dict:
    """Genera imagen desde prompt."""
    pipe, error = get_model(model)
    if error:
        return {"error": f"Model load failed: {error}"}

    try:
        image = pipe(
            prompt=prompt,
            width=width,
            height=height,
            num_inference_steps=steps,
            guidance_scale=guidance
        ).images[0]

        # Convertir a base64
        buffer = BytesIO()
        image.save(buffer, format="PNG")
        img_base64 = base64.b64encode(buffer.getvalue()).decode()

        return {
            "image_base64": img_base64,
            "width": width,
            "height": height,
            "model": model
        }
    except Exception as e:
        return {"error": str(e)}


def generate_variant(image_base64: str, prompt: str,
                    strength: float = 0.5, model: str = "sdxl-turbo") -> dict:
    """Genera variante de imagen existente."""
    try:
        import torch
        from diffusers import AutoPipelineForImage2Image
        from PIL import Image

        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

        # Decodificar imagen
        img_data = base64.b64decode(image_base64)
        init_image = Image.open(BytesIO(img_data)).convert("RGB")

        model_id = MODELS.get(model, MODELS["sdxl-turbo"])
        pipe = AutoPipelineForImage2Image.from_pretrained(
            model_id,
            torch_dtype=torch.float16,
            variant="fp16"
        ).to(device)

        image = pipe(
            prompt=prompt,
            image=init_image,
            strength=strength,
            num_inference_steps=4
        ).images[0]

        buffer = BytesIO()
        image.save(buffer, format="PNG")
        img_base64 = base64.b64encode(buffer.getvalue()).decode()

        return {"image_base64": img_base64}
    except Exception as e:
        return {"error": str(e)}


def upscale_image(image_base64: str, scale: int = 2) -> dict:
    """Upscale imagen usando PIL LANCZOS."""
    try:
        from PIL import Image

        img_data = base64.b64decode(image_base64)
        image = Image.open(BytesIO(img_data))

        new_size = (image.width * scale, image.height * scale)
        upscaled = image.resize(new_size, Image.LANCZOS)

        buffer = BytesIO()
        upscaled.save(buffer, format="PNG")
        img_base64 = base64.b64encode(buffer.getvalue()).decode()

        return {
            "image_base64": img_base64,
            "width": new_size[0],
            "height": new_size[1],
            "scale": scale
        }
    except Exception as e:
        return {"error": str(e)}


def handler(job):
    """
    Handler principal de THE FACTORY.

    Input esperado:
    {
        "task": "image_generate",        # Tarea a ejecutar
        "prompt": "...",                 # Prompt para generación
        "model": "sdxl-turbo",           # Modelo a usar
        "width": 1024,                   # Ancho (opcional)
        "height": 1024,                  # Alto (opcional)
        "image_base64": "...",           # Para variant/upscale
        "strength": 0.5,                 # Para variant
        "scale": 2                       # Para upscale
    }

    Tasks disponibles:
    - image_generate: Genera imagen desde prompt
    - image_variant: Genera variante
    - image_upscale: Aumenta resolución
    """
    job_input = job.get("input", {})
    trace_id = job_input.get("trace_id", str(datetime.utcnow().timestamp()))
    task = job_input.get("task", "image_generate")

    result = {"trace_id": trace_id, "task": task}

    if task == "image_generate":
        prompt = job_input.get("prompt")
        if not prompt:
            return {"error": "prompt es requerido para image_generate"}

        gen_result = generate_image(
            prompt=prompt,
            model=job_input.get("model", "sdxl-turbo"),
            width=job_input.get("width", 1024),
            height=job_input.get("height", 1024),
            steps=job_input.get("steps", 4),
            guidance=job_input.get("guidance", 0.0)
        )
        result.update(gen_result)

    elif task == "image_variant":
        image_base64 = job_input.get("image_base64")
        prompt = job_input.get("prompt", "")
        if not image_base64:
            return {"error": "image_base64 es requerido para image_variant"}

        var_result = generate_variant(
            image_base64=image_base64,
            prompt=prompt,
            strength=job_input.get("strength", 0.5),
            model=job_input.get("model", "sdxl-turbo")
        )
        result.update(var_result)

    elif task == "image_upscale":
        image_base64 = job_input.get("image_base64")
        if not image_base64:
            return {"error": "image_base64 es requerido para image_upscale"}

        up_result = upscale_image(
            image_base64=image_base64,
            scale=job_input.get("scale", 2)
        )
        result.update(up_result)

    else:
        return {"error": f"Task '{task}' no reconocida. Disponibles: image_generate, image_variant, image_upscale"}

    return result


# RunPod serverless
runpod.serverless.start({"handler": handler})