app.py

import os
import gradio as gr
import torch
import spaces
from diffusers import Lumina2Pipeline
from transformers import AutoModelForCausalLM, AutoTokenizer

if torch.cuda.is_available():
    torch_dtype = torch.bfloat16
else:
    torch_dtype = torch.float32

# Load models
def load_models():
    model_name = "X-ART/LeX-Enhancer-full"
    
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype=torch.bfloat16,
        device_map="auto"
    )
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    
    pipe = Lumina2Pipeline.from_pretrained(
        "X-ART/LeX-Lumina", 
        torch_dtype=torch.bfloat16
    )
    device = "cuda" if torch.cuda.is_available() else "cpu"
    pipe.to("cuda")
    
    return model, tokenizer, pipe

model, tokenizer, pipe = load_models()

def truncate_caption_by_tokens(caption, max_tokens=256):
    """Truncate the caption to fit within the max token limit"""
    tokens = tokenizer.encode(caption)
    if len(tokens) > max_tokens:
        truncated_tokens = tokens[:max_tokens]
        caption = tokenizer.decode(truncated_tokens, skip_special_tokens=True)
        print(f"Caption was truncated from {len(tokens)} tokens to {max_tokens} tokens")
    return caption

@spaces.GPU(duration=70)
def generate_enhanced_caption(image_caption, text_caption):
    # model.to("cuda")
    """Generate enhanced caption using the LeX-Enhancer model"""
    combined_caption = f"{image_caption}, with the text on it: {text_caption}."
    instruction = """
Below is the simple caption of an image with text. Please deduce the detailed description of the image based on this simple caption. Note: 1. The description should only include visual elements and should not contain any extended meanings. 2. The visual elements should be as rich as possible, such as the main objects in the image, their respective attributes, the spatial relationships between the objects, lighting and shadows, color style, any text in the image and its style, etc. 3. The output description should be a single paragraph and should not be structured. 4. The description should avoid certain situations, such as pure white or black backgrounds, blurry text, excessive rendering of text, or harsh visual styles. 5. The detailed caption should be human readable and fluent. 6. Avoid using vague expressions such as "may be" or "might be"; the generated caption must be in a definitive, narrative tone. 7. Do not use negative sentence structures, such as "there is nothing in the image," etc. The entire caption should directly describe the content of the image. 8. The entire output should be limited to 200 words.
"""
    messages = [
        {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
        {"role": "user", "content": instruction + "\nSimple Caption:\n" + combined_caption}
    ]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

    generated_ids = model.generate(
        **model_inputs,
        max_new_tokens=1024
    )
    generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]

    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    enhanced_caption = response.split("</think>", -1)[-1].strip(" ").strip("\n")
    model.to("cpu")
    torch.cuda.empty_cache()
    
    return combined_caption, enhanced_caption

@spaces.GPU(duration=60)
def generate_image(enhanced_caption, seed, num_inference_steps, guidance_scale):
    # pipe.to("cuda")
    pipe.enable_model_cpu_offload()
    """Generate image using LeX-Lumina"""
    # Truncate the caption if it's too long
    enhanced_caption = truncate_caption_by_tokens(enhanced_caption, max_tokens=256)

    print(f"enhanced caption:\n{enhanced_caption}")
    
    generator = torch.Generator("cpu").manual_seed(seed) if seed != 0 else None
    
    image = pipe(
        enhanced_caption,
        height=1024,
        width=1024,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        cfg_trunc_ratio=1,
        cfg_normalization=True,
        max_sequence_length=256,
        generator=generator,
        system_prompt="You are an assistant designed to generate superior images with the superior degree of image-text alignment based on textual prompts or user prompts.",
    ).images[0]

    print(image)
    
    pipe.to("cpu")
    torch.cuda.empty_cache()
    
    return image

@spaces.GPU(duration=130)
def run_pipeline(image_caption, text_caption, seed, num_inference_steps, guidance_scale, enable_enhancer):
    """Run the complete pipeline from captions to final image"""
    combined_caption = f"{image_caption}, with the text on it: {text_caption}."
    
    if enable_enhancer:
        combined_caption, enhanced_caption = generate_enhanced_caption(image_caption, text_caption)
    else:
        enhanced_caption = combined_caption
    
    image = generate_image(enhanced_caption, seed, num_inference_steps, guidance_scale)
    
    return  image, combined_caption, enhanced_caption

# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# LeX-Enhancer & LeX-Lumina Demo")
    gr.Markdown("## Project Page: https://zhaoshitian.github.io/lexart/")
    gr.Markdown("Generate enhanced captions from simple image and text descriptions, then create images with LeX-Lumina")
    
    with gr.Row():
        with gr.Column():
            image_caption = gr.Textbox(
                lines=2,
                label="Image Caption",
                placeholder="Describe the visual content of the image",
                value="A picture of a group of people gathered in front of a world map"
            )
            text_caption = gr.Textbox(
                lines=2,
                label="Text Caption",
                placeholder="Describe any text that should appear in the image",
                value="\"Communicate\" in purple, \"Execute\" in yellow"
            )
            
            with gr.Accordion("Advanced Settings", open=False):
                enable_enhancer = gr.Checkbox(
                    label="Enable LeX-Enhancer",
                    value=True,
                    info="When enabled, the caption will be enhanced before image generation"
                )
                seed = gr.Slider(
                    minimum=0,
                    maximum=100000,
                    value=0,
                    step=1,
                    label="Seed (0 for random)"
                )
                num_inference_steps = gr.Slider(
                    minimum=20,
                    maximum=100,
                    value=40,
                    step=1,
                    label="Number of Inference Steps"
                )
                guidance_scale = gr.Slider(
                    minimum=1.0,
                    maximum=10.0,
                    value=7.5,
                    step=0.1,
                    label="Guidance Scale"
                )
            
            submit_btn = gr.Button("Generate", variant="primary")
        
        with gr.Column():
            output_image = gr.Image(label="Generated Image")
            combined_caption_box = gr.Textbox(
                label="Combined Caption",
                interactive=False
            )
            enhanced_caption_box = gr.Textbox(
                label="Enhanced Caption" if enable_enhancer.value else "Final Caption",
                interactive=False,
                lines=5
            )
    
    # Example prompts
    examples = [
        ["A modern office workspace", "\"Innovation\" in bold blue letters at the center"],
        ["A beach sunset scene", "\"Relax\" in cursive white text in the corner"],
        ["A futuristic city skyline", "\"The Future is Now\" in neon pink glowing letters"]
    ]
    gr.Examples(
        examples=examples,
        inputs=[image_caption, text_caption],
        label="Example Inputs"
    )
    
    # Update the label of enhanced_caption_box based on checkbox state
    def update_caption_label(enable_enhancer):
        return gr.Textbox(label="Enhanced Caption" if enable_enhancer else "Final Caption")
    
    enable_enhancer.change(
        fn=update_caption_label,
        inputs=enable_enhancer,
        outputs=enhanced_caption_box
    )
    
    submit_btn.click(
        fn=run_pipeline,
        inputs=[image_caption, text_caption, seed, num_inference_steps, guidance_scale, enable_enhancer],
        outputs=[output_image, combined_caption_box, enhanced_caption_box]
    )

if __name__ == "__main__":
    demo.launch(debug=True)