app.py

import os
import torch
import gc
import gradio as gr
import torch.nn.functional as F
import torchvision.transforms as T
from PIL import Image
from diffusers import StableDiffusionPipeline
from huggingface_hub import login

# Initialize model once and reuse
def init_pipeline():
    device = "cuda" if torch.cuda.is_available() else "cpu"
    torch_dtype = torch.float16 if device == "cuda" else torch.float32
    
    # Authenticate with HF token
    login(token=os.environ.get("HF_AUTH_TOKEN"))
    
    pipeline = StableDiffusionPipeline.from_pretrained(
        "CompVis/stable-diffusion-v1-4",
        torch_dtype=torch_dtype,
        use_auth_token=True
    ).to(device)
    
    # Optimize for performance
    if device == "cuda":
        pipeline.enable_attention_slicing()
        pipeline.enable_xformers_memory_efficient_attention()
    
    return pipeline, device

# Initialize pipeline at startup
pipe, device = init_pipeline()

# Define your original loss functions
def edge_loss(image_tensor):
    grayscale = image_tensor.mean(dim=0, keepdim=True)
    sobel_x = torch.tensor([[1, 0, -1], [2, 0, -2], [1, 0, -1]], 
                          device=device).float().unsqueeze(0).unsqueeze(0)
    sobel_y = sobel_x.transpose(2, 3)
    gx = F.conv2d(grayscale, sobel_x, padding=1)
    gy = F.conv2d(grayscale, sobel_y, padding=1)
    return -torch.mean(torch.sqrt(gx**2 + gy**2 + 1e-6))

def texture_loss(image_tensor):
    return F.mse_loss(image_tensor, torch.rand_like(image_tensor))

def entropy_loss(image_tensor):
    hist = torch.histc(image_tensor, bins=256, min=0, max=1)
    hist = hist / hist.sum()
    return -torch.sum(hist * torch.log(hist + 1e-7))

def symmetry_loss(image_tensor):
    width = image_tensor.shape[-1]
    left = image_tensor[..., :width//2]
    right = torch.flip(image_tensor[..., width//2:], dims=[-1])
    return F.mse_loss(left, right)

def contrast_loss(image_tensor):
    return -torch.std(image_tensor)

# Modified generation function
def generate_images(seed):
    # Clear memory before generation
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    gc.collect()

    # Generate base image
    generator = torch.Generator(device).manual_seed(int(seed))
    image = pipe(
        "A futuristic city skyline at sunset",
        generator=generator,
        num_inference_steps=50,
        guidance_scale=7.5
    ).images[0]

    # Process image
    transform = T.Compose([
        T.ToTensor(),
        T.Normalize([0.5], [0.5])
    ])
    image_tensor = transform(image).unsqueeze(0).to(device)

    # Calculate losses
    losses = {
        "Edge": edge_loss(image_tensor),
        "Texture": texture_loss(image_tensor),
        "Entropy": entropy_loss(image_tensor),
        "Symmetry": symmetry_loss(image_tensor),
        "Contrast": contrast_loss(image_tensor)
    }

    # Create thumbnail grid
    thumb = image.copy()
    thumb.thumbnail((256, 256))
    grid = Image.new('RGB', (256 * 2, 256))
    grid.paste(thumb, (0, 0))
    
    # Add visualization placeholder
    vis = Image.new('RGB', (256, 256), color='white')
    grid.paste(vis, (256, 0))

    return grid, "\n".join([f"{k}: {v.item():.4f}" for k,v in losses.items()])

# Gradio interface
def create_interface():
    return gr.Interface(
        fn=generate_images,
        inputs=gr.Number(label="Seed", value=42),
        outputs=[
            gr.Image(label="Generated Image & Visualizations"),
            gr.Textbox(label="Loss Values")
        ],
        title="Stable Diffusion Loss Analysis",
        description="Generate images and visualize different loss metrics"
    )

if __name__ == "__main__":
    import sys
    if sys.version_info < (3, 10):
        raise RuntimeError("Python 3.10 or later required")
    interface = create_interface()
    interface.launch(server_port=7860, share=False)