app.py

import gradio as gr
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
from diffusers import StableDiffusionPipeline
import torch
from PIL import Image

# Configuración del dispositivo
device = "cpu"
if torch.cuda.is_available():
    device = "cuda"
elif torch.backends.mps.is_available():
    device = "mps"

# Configuración de modelos
model_id_image = "sd-legacy/stable-diffusion-v1-5"
model_id_image_description = "vikhyatk/moondream2"
revision = "2024-08-26"

torch_dtype = torch.float32
if torch.cuda.is_available():
    torch_dtype = torch.bfloat16  # Optimización en GPU

# Carga de modelos persistente
print("Cargando modelo de descripción de imágenes...")
model_description = AutoModelForCausalLM.from_pretrained(model_id_image_description, trust_remote_code=True, revision=revision)
tokenizer_description = AutoTokenizer.from_pretrained(model_id_image_description, revision=revision)

print("Cargando modelo de Stable Diffusion...")
pipe_sd = StableDiffusionPipeline.from_pretrained(model_id_image, torch_dtype=torch_dtype)
pipe_sd = pipe_sd.to(device)

# Opciones para optimizar memoria
pipe_sd.enable_attention_slicing()
if device == "cuda":
    pipe_sd.enable_sequential_cpu_offload()  # Liberar memoria gradualmente para GPUs pequeñas

def generate_description(image_path):
    image_test = Image.open(image_path)
    enc_image = model_description.encode_image(image_test)
    description = model_description.answer_question(enc_image, "Describe this image to create an avatar", tokenizer_description)
    return description

def generate_image_by_description(description, avatar_style=None):
    prompt = f"Create a pigeon profile avatar. Use the following description: {description}."
    if avatar_style:
        prompt += f" Use {avatar_style} style."

    result = pipe_sd(prompt)
    return result.images[0]

def process_and_generate(image, avatar_style):
    description = generate_description(image)
    return generate_image_by_description(description, avatar_style)

with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column(scale=2, min_width=300):
            selected_image = gr.Image(type="filepath", label="Upload an Image of the Pigeon", height=300)
            avatar_style = gr.Radio(
                ["Realistic", "Pixel Art", "Imaginative", "Cartoon"], 
                label="(optional) Select the avatar style:"
            )
            generate_button = gr.Button("Generate Avatar", variant="primary")
        with gr.Column(scale=2, min_width=300):
            generated_image = gr.Image(type="numpy", label="Generated Avatar", height=300)
            
        generate_button.click(process_and_generate, inputs=[selected_image, avatar_style], outputs=generated_image)

demo.launch()