app.py

from diffusers import AutoPipelineForText2Image
import torch
import gradio as gr
import threading
import time;
from queue import Queue
import spaces


@spaces.GPU(duration=120)
def GenerateImage(prompt,steps,progress,model):
    
    data = []
    
    queue = Queue();
    
    def StartThread():
        
        pipe_txt2img = AutoPipelineForText2Image.from_pretrained(
            model, torch_dtype=torch.float16, use_safetensors=True
        ).to("cuda")   
        
        vae = pipe_txt2img.vae
        
        def latents_callback(i, t, latents):
            latents = 1 / 0.18215 * latents
            image = vae.decode(latents).sample[0]
            image = (image / 2 + 0.5).clamp(0, 1)
            image = image.cpu().permute(1, 2, 0).numpy()
            FinalImage = pipe_txt2img.numpy_to_pil(image)
            queue.put({'type':'image', 'image':FinalImage[0], 'step': i})
        
        generator = torch.Generator(device="cpu").manual_seed(37)
        FinalImage = pipe_txt2img(prompt, generator=generator,  num_inference_steps=steps,callback=latents_callback, callback_steps=progress).images[0]
        queue.put({'type':'image', 'image':FinalImage, 'step': steps+1})
        queue.put({'type':'end'})
    
    t = threading.Thread(target=StartThread)
    t.start();
    
    while True:
        print("Waiting next item");
        nextItem = queue.get()
        
        if nextItem['type'] == 'end':
            break;
                 
        Image   = nextItem['image']
        Step    = nextItem['step']
        yield [Image,Step];

    print("Waiting thread finish...");
    t.join()
    
    print("Finished!");



with gr.Blocks() as demo:
    gr.Markdown("""
        This is a lab to demonstrate how we can implement a text-to-image generation using Gradio and Diffusers, showing the progress of each image produced at each step.  
        Type a prompt, choose the maximum number of steps and the frequency (in steps) at which progress is shown. You will see the diffusion process live!  
    """)
    
    with gr.Row():
        prompt          = gr.Text(label="prompt");
        TotalSteps      = gr.Slider(label="Steps", minimum=1,maximum=150,value=10);
        ProgressSteps   = gr.Number(label="Progress steps", value = 2);
        model           = gr.Text(label="Model", value="dreamlike-art/dreamlike-photoreal-2.0")
    
    with gr.Row():
        with gr.Column():
            btnRun      = gr.Button(value="Run!");
            btnStop   = gr.Button(value="Stop!");
        status  = gr.Text(label="Current Step");
   
        
    image   = gr.Image();
    
    
    GenerateEvent = btnRun.click( GenerateImage, [prompt,TotalSteps,ProgressSteps,model], [image,status] );
    btnStop.click( None,None,None, cancels=[GenerateEvent] )

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