app.py

import gradio as gr
import numpy as np
import random
import os
import spaces
import torch
from diffusers import DiffusionPipeline
from transformers import pipeline, AutoTokenizer
from huggingface_hub import login
from PIL import Image

hf_token = os.getenv("hf_token")
login(token=hf_token)

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 2048
PRELOAD_MODELS = False  # Easy switch for preloading

_text_gen_pipeline = None
_image_gen_pipeline = None

@spaces.GPU()
def get_image_gen_pipeline():
    global _image_gen_pipeline
    if _image_gen_pipeline is None:
        try:
            device = "cuda" if torch.cuda.is_available() else "cpu"
            dtype = torch.bfloat16
            _image_gen_pipeline = DiffusionPipeline.from_pretrained(
                "black-forest-labs/FLUX.1-schnell",
                torch_dtype=dtype,
            ).to(device)
            
            # Comment these out for now to match the working example
            # _image_gen_pipeline.enable_model_cpu_offload()
            # _image_gen_pipeline.enable_vae_slicing()
        except Exception as e:
            print(f"Error loading image generation model: {e}")
            return None
    return _image_gen_pipeline

@spaces.GPU()
def get_text_gen_pipeline():
    global _text_gen_pipeline
    if _text_gen_pipeline is None:
        try:
            device = "cuda" if torch.cuda.is_available() else "cpu"
            tokenizer = AutoTokenizer.from_pretrained(
                "mistralai/Mistral-7B-Instruct-v0.3",
                use_fast=True
            )
            tokenizer.pad_token = tokenizer.pad_token or tokenizer.eos_token
            
            _text_gen_pipeline = pipeline(
                "text-generation",
                model="mistralai/Mistral-7B-Instruct-v0.3",
                tokenizer=tokenizer,
                max_new_tokens=2048,
                device=device,
                pad_token_id=tokenizer.pad_token_id
            )
        except Exception as e:
            print(f"Error loading text generation model: {e}")
            return None
    return _text_gen_pipeline

@spaces.GPU()
def refine_prompt(prompt, progress=gr.Progress()):
    text_gen = get_text_gen_pipeline()
    if text_gen is None:
        return "", "Text generation model is unavailable."
    try:
        messages = [
            {"role": "system", "content": "Vous êtes un designer produit avec de solides connaissances dans la génération de texte en image. Vous recevrez une demande de produit sous forme de description succincte, et votre mission sera d'imaginer un nouveau design de produit répondant à ce besoin.\n\nLe livrable (réponse générée) sera exclusivement un texte de prompt pour l'IA de texte to image FLUX.1-schnell.\n\nCe prompt devra inclure une description visuelle de l'objet mentionnant explicitement les aspects indispensables de sa fonction.\nA coté de ça vous devez aussi explicitement mentionner dans ce prompt les caractéristiques esthétiques/photo du rendu image (ex : photoréaliste, haute qualité, focale, grain, etc.), sachant que l'image sera l'image principale de cet objet dans le catalogue produit. Le fond de l'image générée doit être entièrement blanc.\nLe prompt doit être sans narration, peut être long mais ne doit pas dépasser 77 jetons."},            {"role": "user", "content": prompt},
        ]
        
        # Indicate progress started
        progress(0, desc="Generating text")
        
        # Generate text
        refined_prompt = text_gen(messages)
        
        # Indicate progress complete
        progress(1)
        
        # Extract just the assistant's content from the response
        try:
            messages = refined_prompt[0]['generated_text']
            # Find the last message with role 'assistant'
            assistant_messages = [msg for msg in messages if msg['role'] == 'assistant']
            if not assistant_messages:
                return "", "Error: No assistant response found"
            assistant_content = assistant_messages[-1]['content']
            return assistant_content, "Prompt refined successfully!"
        except (KeyError, IndexError):
            return "", "Error: Unexpected response format from the model"
    except Exception as e:
        print(f"Error in refine_prompt: {str(e)}")  # Add debug print
        return "", f"Error refining prompt: {str(e)}"

def validate_dimensions(width, height):
    if width * height > MAX_IMAGE_SIZE * MAX_IMAGE_SIZE:
        return False, "Image dimensions too large"
    return True, None

@spaces.GPU()
def infer(prompt, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=4, progress=gr.Progress(track_tqdm=True)): 
    try:
        # Validate that prompt is not empty
        if not prompt or prompt.strip() == "":
            return None, "Please provide a valid prompt."

        pipe = get_image_gen_pipeline()
        if pipe is None:
            return None, "Image generation model is unavailable."
        
        is_valid, error_msg = validate_dimensions(width, height)
        if not is_valid:
            return None, error_msg

        if randomize_seed:
            seed = random.randint(0, MAX_SEED)
        
        # Use default torch generator instead of cuda-specific generator
        generator = torch.Generator().manual_seed(seed)
        
        # Match the working example's parameters
        output = pipe(
            prompt=prompt, 
            width=width,
            height=height,
            num_inference_steps=num_inference_steps, 
            generator=generator,
            guidance_scale=0.0,  # Changed from 7.5 to 0.0
        )
            
        image = output.images[0]
        return image, f"Image generated successfully with seed {seed}"
    except Exception as e:
        print(f"Error in infer: {str(e)}")  
        return None, f"Error generating image: {str(e)}"
 
examples = [
    "a backpack for kids, flower style",
    "medieval flip flops",
    "cat shaped cake mold",
]

css="""
#col-container {
    margin: 0 auto;
    max-width: 520px;
}
"""

def preload_models():
    global _text_gen_pipeline, _image_gen_pipeline
    
    print("Preloading models...")
    success = True
    
    try:
        _text_gen_pipeline = get_text_gen_pipeline()
        if _text_gen_pipeline is None:
            success = False
    except Exception as e:
        print(f"Error preloading text generation model: {str(e)}")
        success = False
    
    try:
        _image_gen_pipeline = get_image_gen_pipeline()
        if _image_gen_pipeline is None:
            success = False
    except Exception as e:
        print(f"Error preloading image generation model: {str(e)}")
        success = False
    
    status = "Models preloaded successfully!" if success else "Error preloading models"
    print(status)
    return success

def create_interface():
    # Preload models if needed
    if PRELOAD_MODELS:
        models_loaded = preload_models()
        model_status = "✅ Models loaded successfully!" if models_loaded else "⚠️ Error loading models"
    else:
        model_status = "ℹ️ Models will be loaded on demand"

    with gr.Blocks(css=css) as demo:
        gr.Info(model_status)

        with gr.Column(elem_id="col-container"):
            gr.Markdown("# Text to Product\nUsing Mistral-7B-Instruct-v0.3 + FLUX.1-dev + Trellis")
            
            # Basic inputs
            with gr.Row():
                prompt = gr.Text(
                    show_label=False,
                    max_lines=1,
                    placeholder="Enter basic object prompt",
                    container=False,
                )
                prompt_button = gr.Button("Refine prompt with Mistral")
            
            refined_prompt = gr.Text(
                show_label=False,
                max_lines=10, 
                placeholder="Detailed object prompt", 
                container=False,
                max_length=2048,
            )
            
            visual_button = gr.Button("Create visual with Flux")
            generated_image = gr.Image(show_label=False)
            error_box = gr.Textbox(
                label="Status Messages", 
                interactive=False,
                placeholder="Status messages will appear here",
            )
            
            # Accordion sections for advanced settings
            with gr.Accordion("Advanced Settings", open=False):
                with gr.Tab("Mistral"):
                    # Mistral settings
                    temperature = gr.Slider(
                        label="Temperature",
                        value=0.9,
                        minimum=0.0,
                        maximum=1.0,
                        step=0.05,
                        info="Higher values produce more diverse outputs",
                    )
                    
                with gr.Tab("Flux"):
                    # Flux settings
                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
                    
                    with gr.Row():
                        width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=512)
                        height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=512)
                    
                    num_inference_steps = gr.Slider(
                        label="Number of inference steps",
                        minimum=1,
                        maximum=50,
                        step=1,
                        value=10,
                    )

            # Examples section
            gr.Examples(
                examples=examples,
                fn=refine_prompt,
                inputs=[prompt],
                outputs=[refined_prompt],
                cache_examples=True,
            )

        # Event handlers
        gr.on(
            triggers=[prompt_button.click, prompt.submit],
            fn=refine_prompt,
            inputs=[prompt],
            outputs=[refined_prompt, error_box]
        )

        gr.on(
            triggers=[visual_button.click],
            fn=infer,
            inputs=[refined_prompt, seed, randomize_seed, width, height, num_inference_steps],
            outputs=[generated_image, error_box]
        )

    return demo

if __name__ == "__main__":
    demo = create_interface()
    demo.launch()