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) # Global constants and default values MAX_SEED = np.iinfo(np.int32).max MAX_IMAGE_SIZE = 2048 PRELOAD_MODELS = False # Default system prompt for text generation DEFAULT_SYSTEM_PROMPT = """You are a product designer with strong knowledge in text-to-image generation. You will receive a product request in the form of a brief description, and your mission will be to imagine a new product design that meets this need. The deliverable (generated response) will be exclusively a text prompt for the FLUX.1-schnell text-to-image AI. This prompt should include a visual description of the object explicitly mentioning the essential aspects of its function. Additionally, you should explicitly mention in this prompt the aesthetic/photo characteristics of the image rendering (e.g., photorealistic, high quality, focal length, grain, etc.), knowing that the image will be the main image of this object in the product catalog. The background of the generated image must be entirely white. The prompt should be without narration, can be long but must not exceed 77 tokens.""" # Default Flux parameters DEFAULT_SEED = 42 DEFAULT_RANDOMIZE_SEED = True DEFAULT_WIDTH = 512 DEFAULT_HEIGHT = 512 DEFAULT_NUM_INFERENCE_STEPS = 6 DEFAULT_GUIDANCE_SCALE = 0.0 DEFAULT_TEMPERATURE = 0.9 _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, system_prompt=DEFAULT_SYSTEM_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": system_prompt}, {"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'] # Remove quotation marks at the beginning and end if assistant_content.startswith('"') and assistant_content.endswith('"'): assistant_content = assistant_content[1:-1] 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=DEFAULT_SEED, randomize_seed=DEFAULT_RANDOMIZE_SEED, width=DEFAULT_WIDTH, height=DEFAULT_HEIGHT, num_inference_steps=DEFAULT_NUM_INFERENCE_STEPS, 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." progress(0.1, desc="Loading model") 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) progress(0.3, desc="Running inference") # Match the working example's parameters output = pipe( prompt=prompt, width=width, height=height, num_inference_steps=num_inference_steps, generator=generator, guidance_scale=DEFAULT_GUIDANCE_SCALE, ) progress(0.8, desc="Processing output") image = output.images[0] progress(1.0, desc="Complete") 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)}" # Format: [prompt, system_prompt] 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(): print("Preloading models...") text_success = get_text_gen_pipeline() is not None image_success = get_image_gen_pipeline() is not None success = text_success and image_success status = "Models preloaded successfully!" if success else "Error preloading models" print(status) return success # Create a combined function that handles the whole pipeline from example to image # This version gets the parameters from the UI components @spaces.GPU() def process_example_pipeline(example_prompt, system_prompt=DEFAULT_SYSTEM_PROMPT, progress=gr.Progress()): # Step 1: Update status progress(0, desc="Starting example processing") # Step 2: Refine the prompt progress(0.1, desc="Refining prompt with Mistral") refined, status = refine_prompt(example_prompt, system_prompt, progress) if not refined: return "", "Failed to refine prompt: " + status # Return only the refined prompt and status - don't generate image return refined, "Prompt refined successfully!" 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") 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) gen3d_button = gr.Button("Create 3D visual with Trellis") message_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=DEFAULT_TEMPERATURE, minimum=0.0, maximum=1.0, step=0.05, info="Higher values produce more diverse outputs", ) system_prompt = gr.Textbox( label="System Prompt", value=DEFAULT_SYSTEM_PROMPT, lines=10, info="Instructions for the Mistral model" ) with gr.Tab("Flux"): # Flux settings seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=DEFAULT_SEED) randomize_seed = gr.Checkbox(label="Randomize seed", value=DEFAULT_RANDOMIZE_SEED) with gr.Row(): width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=DEFAULT_WIDTH) height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=DEFAULT_HEIGHT) num_inference_steps = gr.Slider( label="Number of inference steps", minimum=1, maximum=50, step=1, value=DEFAULT_NUM_INFERENCE_STEPS, ) # Examples section - simplified version that only updates the prompt fields gr.Examples( examples=examples, # Now just a list of prompts fn=process_example_pipeline, inputs=[prompt], # Add system_prompt as input outputs=[refined_prompt, message_box], # Don't output image cache_examples=True, ) # Event handlers gr.on( triggers=[prompt_button.click, prompt.submit], fn=refine_prompt, inputs=[prompt, system_prompt], # Add system_prompt as input outputs=[refined_prompt, message_box] ) gr.on( triggers=[visual_button.click], fn=infer, inputs=[refined_prompt, seed, randomize_seed, width, height, num_inference_steps], outputs=[generated_image, message_box] ) return demo if __name__ == "__main__": demo = create_interface() demo.launch()