app.py

import gradio as gr
import spaces
from diffusers import AutoPipelineForText2Image
import torch
import time
# import logging
from threading import Timer
from nsfw_detector import NSFWDetector

# logging.basicConfig(level=logging.INFO)
# logger = logging.getLogger(__name__)

# Global variables
pipe = None
last_use_time = None
unload_timer = None
TIMEOUT_SECONDS = 120  # 2 minutes
BATCH_SIZE = 4

def chunk_generations(num_images):
    """Split number of images into batches of BATCH_SIZE"""
    return [min(BATCH_SIZE, num_images - i) for i in range(0, num_images, BATCH_SIZE)]

@spaces.GPU(duration=25)
def generate_image(
    prompt,
    num_inference_steps=1,
    num_images=1,
    height=512,
    width=512,
):    
    global pipe
    start_time = time.time()
    
    # Load model if needed
    if pipe is None:
        yield None, "Loading model..."
        pipe = AutoPipelineForText2Image.from_pretrained(
            "stabilityai/sdxl-turbo", 
            torch_dtype=torch.float16, 
            variant="fp16"
        ).to("cuda")
        yield None, "Model loaded, starting generation..."
    
    reset_timer()
    
    # Process in batches if more than BATCH_SIZE images
    if num_images > BATCH_SIZE:
        yield None, f"Generating {num_images} images in batches..."
        all_images = []
        batches = chunk_generations(num_images)
        
        for i, batch_size in enumerate(batches):
            yield None, f"Generating batch {i+1}/{len(batches)} ({batch_size} images)..."
            batch_images = pipe(
                prompt=prompt,
                num_inference_steps=num_inference_steps,
                height=height,
                width=width,
                guidance_scale=0.0,
                num_images_per_prompt=batch_size
            ).images
            all_images.extend(batch_images)
        images = all_images
    else:
        yield None, f"Generating {num_images} image(s) with {num_inference_steps} steps..."
        images = pipe(
            prompt=prompt,
            num_inference_steps=num_inference_steps,
            height=height,
            width=width,
            guidance_scale=0.0,
            num_images_per_prompt=num_images
        ).images
    
    total_time = time.time() - start_time
    avg_time = total_time / num_images
    status_msg = f"Generated {num_images} image(s) in {total_time:.2f} seconds (avg {avg_time:.2f}s per image)"
    # logger.info(status_msg)
    
    # Check for NSFW content
    detector = NSFWDetector()
    is_nsfw, category, confidence = detector.check_image(images[0])
        
    if category == "SAFE":
        yield images, status_msg
    else:
        return

def unload_model():
    global pipe, last_use_time
    current_time = time.time()
    if last_use_time and (current_time - last_use_time) >= TIMEOUT_SECONDS:
        # logger.info("Unloading model due to inactivity...")
        pipe = None
        torch.cuda.empty_cache()
        return "Model unloaded due to inactivity"

def reset_timer():
    global unload_timer, last_use_time
    if unload_timer:
        unload_timer.cancel()
    last_use_time = time.time()
    unload_timer = Timer(TIMEOUT_SECONDS, unload_model)
    unload_timer.start()

# Create the Gradio interface
with gr.Blocks() as demo:
    with gr.Column():
        prompt = gr.Textbox(label="Prompt", placeholder="Enter your prompt here...")
        with gr.Row():
            steps = gr.Slider(minimum=1, maximum=10, value=1, step=1, label="Number of inference steps")
            num_images = gr.Slider(minimum=1, maximum=64, value=1, step=1, label="Number of images to generate")
        with gr.Row():
            height = gr.Slider(minimum=512, maximum=1024, value=512, step=64, label="Height")
            width = gr.Slider(minimum=512, maximum=1024, value=512, step=64, label="Width")
        
        generate_btn = gr.Button("Generate")
        gallery = gr.Gallery()
        status = gr.Textbox(
            label="Status", 
            value="Model not loaded - will load on first generation", 
            interactive=False
        )
        
        generate_btn.click(
            fn=generate_image,
            inputs=[prompt, steps, num_images, height, width],
            outputs=[gallery, status]
        )

    gr.Markdown("""
        This model works best with 512x512 resolution and 1-4 inference steps. 
        Values above 4 steps may not improve quality significantly.
        The model will automatically unload after 2 minutes of inactivity.
    """)

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