Training-Free Single-Image Diffusion Model Achieves Fast, High-Quality Generation

Summary

This paper presents a training-free approach to single-image diffusion models. Instead of training a neural network on a single image (which is computationally expensive), the authors model the image using a dataset of its patches at different scales. They compute the score function for noisy patches using an optimal, closed-form denoiser, eliminating neural network training entirely. The method achieves state-of-the-art generation quality and diversity compared to trained single-image diffusion models, and enables applications like unconditional generation, text-guided stylization, image symmetrization, and retargeting. It also supports latent space diffusion and acceleration techniques for megapixel generation in one second and gigapixel generation in minutes.