app.py

import gradio as gr
import torch
from diffusers import FluxPipeline
import huggingface_hub
from huggingface_hub import InferenceClient
import os

huggingface_hub.login(token=os.getenv("HUGGINGFACE_API_TOKEN"))
# Initialize the Flux pipeline
def initialize_flux_pipeline():
    pipe = FluxPipeline.from_pretrained(
        "black-forest-labs/FLUX.1-dev",
        torch_dtype=torch.bfloat16
    )
    pipe.enable_model_cpu_offload()
    pipe.load_lora_weights("EvanZhouDev/open-genmoji")
    return pipe

flux_pipeline = initialize_flux_pipeline()

# Initialize the language model client
llm_client = InferenceClient("Qwen/Qwen2.5-72B-Instruct", token=os.getenv("HUGGINGFACE_API_TOKEN"))

# Function to refine the prompt
def refine_prompt(original_prompt):
    messages = [
        {
            "role": "system",
            "content": (
                "You are helping create a prompt for a Emoji generation image model. An emoji must be easily "
                "interpreted when small so details must be exaggerated to be clear. Your goal is to use descriptions "
                "to achieve this.\n\nYou will receive a user description, and you must rephrase it to consist of "
                "short phrases separated by periods, adding detail to everything the user provides.\n\nAdd describe "
                "the color of all parts or components of the emoji. Unless otherwise specified by the user, do not "
                "describe people. Do not describe the background of the image. Your output should be in the format:\n\n"
                "```emoji of {description}. {addon phrases}. 3D lighting. no cast shadows.```\n\nThe description "
                "should be a 1 sentence of your interpretation of the emoji. Then, you may choose to add addon phrases."
                " You must use the following in the given scenarios:\n\n- \"cute.\": If generating anything that's not "
                "an object, and also not a human\n- \"enlarged head in cartoon style.\": ONLY animals\n- \"head is "
                "turned towards viewer.\": ONLY humans or animals\n- \"detailed texture.\": ONLY objects\n\nFurther "
                "addon phrases may be added to ensure the clarity of the emoji."
            ),
        },
        {"role": "user", "content": original_prompt},
    ]
    completion = llm_client.chat_completion(messages, max_tokens=100)
    refined = completion["choices"][0]["message"]["content"].strip()
    return refined

# Define the process function
def process(prompt, guidance_scale, num_inference_steps, height, width, seed):
    print(f"Original Prompt: {prompt}")

    # Refine the prompt
    try:
        refined_prompt = refine_prompt(prompt)
        print(f"Refined Prompt: {refined_prompt}")
    except Exception as e:
        return f"Error refining prompt: {str(e)}"

    # Set the random generator seed
    generator = torch.Generator(device="cuda").manual_seed(seed)

    try:
        # Generate the image
        output = flux_pipeline(
            prompt=refined_prompt,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            height=height,
            width=width,
            generator=generator,
        )
        image = output.images[0]
        return image
    except Exception as e:
        return f"Error generating image: {str(e)}"

# Create the Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# Flux Text-to-Image Generator with Prompt Refinement")

    # User inputs
    with gr.Row():
        prompt_input = gr.Textbox(label="Enter a Prompt", placeholder="Describe your image")
        guidance_scale_input = gr.Slider(
            label="Guidance Scale", minimum=1.0, maximum=10.0, value=3.5, step=0.1
        )
    with gr.Row():
        num_inference_steps_input = gr.Slider(
            label="Inference Steps", minimum=1, maximum=100, value=50, step=1
        )
        seed_input = gr.Number(label="Seed", value=42, precision=0)
    with gr.Row():
        height_input = gr.Slider(label="Height", minimum=256, maximum=2048, value=768, step=64)
        width_input = gr.Slider(label="Width", minimum=256, maximum=2048, value=1360, step=64)

    # Output components
    refined_prompt_output = gr.Textbox(label="Refined Prompt", interactive=False)
    image_output = gr.Image(label="Generated Image")

    # Button to generate the image
    generate_button = gr.Button("Generate Image")

    # Define button click behavior
    generate_button.click(
        fn=lambda prompt, *args: (refine_prompt(prompt), process(prompt, *args)),
        inputs=[
            prompt_input,
            guidance_scale_input,
            num_inference_steps_input,
            height_input,
            width_input,
            seed_input,
        ],
        outputs=[refined_prompt_output, image_output],
    )

# Launch the app
if __name__ == "__main__":
    demo.launch(show_error=True)