generate.py

import torch
from model import DiffusionModel, UNet
from torchvision.utils import save_image
import argparse
from PIL import Image

def generate(prompts, model_path="diffusion_model.pth", image_size=256, device="cuda"):
    # Load model
    model = UNet().to(device)
    model.load_state_dict(torch.load(model_path, map_location=device))
    model.eval()
    
    # Setup diffusion
    betas = torch.linspace(1e-4, 0.02, 1000).to(device)
    diffusion = DiffusionModel(model, betas, device)
    
    # Generate images
    with torch.no_grad():
        images = diffusion.sample(prompts, image_size=image_size, batch_size=len(prompts))
    
    # Save images
    os.makedirs("generated", exist_ok=True)
    for i, img in enumerate(images):
        img = Image.fromarray(img.permute(1, 2, 0).cpu().numpy())
        img.save(f"generated/sample_{i}.png")
    
    print(f"Generated {len(images)} images saved in 'generated' folder")

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--prompts", nargs="+", required=True, help="Text prompts for generation")
    parser.add_argument("--model", default="diffusion_model.pth", help="Path to trained model")
    parser.add_argument("--size", type=int, default=256, help="Image size")
    args = parser.parse_args()
    
    device = "cuda" if torch.cuda.is_available() else "cpu"
    generate(args.prompts, args.model, args.size, device)