app.py

import spaces
from diffusers import StableDiffusionXLPipeline
from diffusers import DiffusionPipeline
from pydantic import BaseModel
from PIL import Image
import gradio as gr
import torch
import uuid
import io
import os

# Load the base & refiner pipelines
base = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", 
    torch_dtype=torch.float16, 
    variant="fp16", 
    use_safetensors=True
)
base.to("cuda:0")

# Load your model
pipe = StableDiffusionXLPipeline.from_pretrained(
    "segmind/SSD-1B", 
    torch_dtype=torch.float16, 
    use_safetensors=True, 
    variant="fp16"
)
pipe.to("cuda:0")

refiner = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-refiner-1.0",
    text_encoder_2=base.text_encoder_2,
    vae=base.vae,
    torch_dtype=torch.float16,
    use_safetensors=True,
    variant="fp16",
)
refiner.to("cuda:0")
refiner.unet = torch.compile(refiner.unet, mode="reduce-overhead", fullgraph=True)

@spaces.GPU  # Apply the GPU decorator
def generate_and_save_image(prompt, negative_prompt=''):
    # Generate image using the provided prompts
    image = pipe(prompt=prompt, negative_prompt=negative_prompt).images[0]

    # Generate a unique UUID for the filename
    unique_id = str(uuid.uuid4())
    image_path = f"generated_images/{unique_id}.jpeg"

    # Save generated image locally
    os.makedirs('generated_images', exist_ok=True)
    image.save(image_path, format='JPEG')

    # Return the path of the saved image to display in Gradio interface
    return image_path

def generate_image_with_refinement(prompt):
    n_steps = 40
    high_noise_frac = 0.8

    # run both experts
    image = base(
        prompt=prompt,
        num_inference_steps=n_steps,
        denoising_end=high_noise_frac,
        output_type="latent",
    ).images
    image = refiner(
        prompt=prompt,
        num_inference_steps=n_steps,
        denoising_start=high_noise_frac,
        image=image,
    ).images[0]

    # Save the image as before
    unique_id = str(uuid.uuid4())
    image_path = f"generated_images_refined/{unique_id}.jpeg"
    os.makedirs('generated_images_refined', exist_ok=True)
    image.save(image_path, format='JPEG')

    return image_path

# Start of the Gradio Blocks interface
with gr.Blocks() as demo:
    with gr.Column():
        gr.Markdown("# Image Generation with SSD-1B")
        gr.Markdown("Enter a prompt and (optionally) a negative prompt to generate an image.")
        
        # Input fields for positive and negative prompts
        with gr.Row():
            prompt1 = gr.Textbox(label="Enter prompt")
            negative_prompt = gr.Textbox(label="Enter negative prompt (optional)")
        
        # Button for generating the image
        generate_button1 = gr.Button("Generate Image")

        # Output image display, set to a larger default size
        output_image1 = gr.Image(label="Generated Image")

        # Click event for the generate button
        generate_button1.click(
            generate_and_save_image, 
            inputs=[prompt1, negative_prompt], 
            outputs=output_image1
        )

    with gr.Column():
        gr.Markdown("## Refined Image Generation")
        gr.Markdown("Enter a prompt to generate a refined image.")
        
        # Input field for the prompt
        prompt2 = gr.Textbox(label="Enter prompt for refined generation")
        
        # Button for generating the refined image
        generate_button2 = gr.Button("Generate Refined Image")
        
        # Output refined image display, set to a larger default size
        output_image2 = gr.Image(label="Generated Refined Image")

        # Click event for the generate button
        generate_button2.click(
            generate_image_with_refinement, 
            inputs=[prompt2], 
            outputs=output_image2
        )

    # Set the image display to be the largest element for both SSD-1B and refined generation
    demo.update(
        output_image1.style(width='100%', height='auto', min_height='400px'),
        output_image2.style(width='100%', height='auto', min_height='400px')
    )


# Launch the combined Gradio app
demo.launch()