A training-free diffusion model is proposed using patch distribution of a single reference image. It achieves superior generation quality and diversity compared to trained single-image diffusion models, enabling megapixel generation in one second.
A training-free diffusion model is proposed that operates using only a single reference image. Instead of training a neural network, it constructs a dataset of patches at multiple scales and computes the score function for noisy patches via an optimal closed-form denoiser, eliminating the need for training.
Single-image generation has applications in text-guided stylization, image symmetrization, and retargeting, but existing methods require hours of training. This work bridges classical patch-based image restoration with diffusion models, enabling high-quality generation without training.
The method drastically reduces computational cost while achieving state-of-the-art generation quality and diversity. It is compatible with latent diffusion and can generate megapixel images in one second and gigapixel images in minutes, making it highly practical.