app.py

import os
import gradio as gr
import numpy as np
import random
from huggingface_hub import AsyncInferenceClient, InferenceClient
from PIL import Image
from gradio_client import Client, handle_file
from gradio_imageslider import ImageSlider

MAX_SEED = np.iinfo(np.int32).max
HF_TOKEN = os.environ.get("HF_TOKEN")
HF_TOKEN_UPSCALER = os.environ.get("HF_TOKEN_UPSCALER")

client = AsyncInferenceClient()
llm_client = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")

def enable_lora(lora_add, basemodel):
    return basemodel if not lora_add else lora_add

async def generate_image(prompt, model, lora_word, width, height, scales, steps, seed):
    try:
        if seed == -1:
            seed = random.randint(0, MAX_SEED)
        seed = int(seed)
        text = prompt + "," + lora_word
        image = await client.text_to_image(prompt=text, height=height, width=width, guidance_scale=scales, num_inference_steps=steps, model=model)
        return image, seed
    except Exception as e:
        return f"Error al generar imagen: {e}", None

def get_upscale_finegrain(prompt, img_path, upscale_factor):
    try:
        client = Client("finegrain/finegrain-image-enhancer", hf_token=HF_TOKEN_UPSCALER)
        result = client.predict(input_image=handle_file(img_path), prompt=prompt, negative_prompt="", seed=42, upscale_factor=upscale_factor, controlnet_scale=0.6, controlnet_decay=1, condition_scale=6, tile_width=112, tile_height=144, denoise_strength=0.35, num_inference_steps=18, solver="DDIM", api_name="/process")
        return result[1]
    except Exception as e:
        return None

async def gen(prompt, basemodel, width, height, scales, steps, seed, upscale_factor, process_upscale, lora_model, process_lora):
    model = enable_lora(lora_model, basemodel) if process_lora else basemodel
    
    improved_prompt = await improve_prompt(prompt)
    combined_prompt = f"{prompt} {improved_prompt}"

    image, seed = await generate_image(combined_prompt, model, "", width, height, scales, steps, seed)
    
    if isinstance(image, str) and image.startswith("Error"):
        return [image, None]
    
    image_path = "temp_image.jpg"
    image.save(image_path, format="JPEG")
    
    if process_upscale:
        upscale_image_path = get_upscale_finegrain(combined_prompt, image_path, upscale_factor)
        if upscale_image_path is not None:
            upscale_image = Image.open(upscale_image_path)
            upscale_image.save("upscale_image.jpg", format="JPEG")
            return [image_path, "upscale_image.jpg"]
        else:
            return [image_path, image_path]
    else:
        return [image_path, image_path]

async def improve_prompt(prompt):
    try:
        instruction = "Improve and translate this prompt into English, adding detailed descriptions of style, cinematography, cameras, atmosphere, and lighting for the best quality, up to 200 words."
        formatted_prompt = f"{instruction}: {prompt}"
        response = llm_client.text_generation(formatted_prompt, max_new_tokens=300)  # Allowing more tokens for detailed description
        improved_text = response['generated_text'].strip() if 'generated_text' in response else response.strip()

        return improved_text
    except Exception as e:
        return f"Error mejorando el prompt: {e}"

css = """
#col-container{ margin: 0 auto; max-width: 1024px;}
"""

with gr.Blocks(css=css, theme="Nymbo/Nymbo_Theme") as demo:
    with gr.Column(elem_id="col-container"):
        with gr.Row():
            with gr.Column(scale=3):
                output_res = ImageSlider(label="Flux / Upscaled")
            with gr.Column(scale=2):
                prompt = gr.Textbox(label="Descripción de imágen")
                basemodel_choice = gr.Dropdown(label="Modelo", choices=["black-forest-labs/FLUX.1-schnell", "black-forest-labs/FLUX.1-DEV"], value="black-forest-labs/FLUX.1-schnell")
                lora_model_choice = gr.Dropdown(label="LORA Realismo", choices=["Shakker-Labs/FLUX.1-dev-LoRA-add-details", "XLabs-AI/flux-RealismLora"], value="XLabs-AI/flux-RealismLora")
                
                with gr.Row():
                    process_lora = gr.Checkbox(label="Procesar LORA")
                    process_upscale = gr.Checkbox(label="Procesar Escalador")

                upscale_factor = gr.Radio(label="Factor de Escala", choices=[2, 4, 8], value=2)

                improved_prompt = gr.Textbox(label="Prompt Mejorado", interactive=False)

                improve_btn = gr.Button("Mejora mi prompt")
                improve_btn.click(fn=improve_prompt, inputs=[prompt], outputs=improved_prompt)
                
                with gr.Accordion(label="Opciones Avanzadas", open=False):
                    width = gr.Slider(label="Ancho", minimum=512, maximum=1280, step=8, value=1280)
                    height = gr.Slider(label="Alto", minimum=512, maximum=1280, step=8, value=768)
                    scales = gr.Slider(label="Escalado", minimum=1, maximum=20, step=1, value=10)
                    steps = gr.Slider(label="Pasos", minimum=1, maximum=100, step=1, value=20)
                    seed = gr.Number(label="Semilla", value=-1)
    
                btn = gr.Button("Generar")
                btn.click(fn=gen, inputs=[prompt, basemodel_choice, width, height, scales, steps, seed, upscale_factor, process_upscale, lora_model_choice, process_lora], outputs=output_res)
    demo.launch()