A training-free diffusion model is proposed by leveraging patch distribution of a single reference image. It achieves state-of-the-art generation quality and diversity compared to trained single-image diffusion models, enabling megapixel generation in one second and gigapixel in minutes.
A training-free diffusion model is proposed by leveraging patch distribution of a single reference image. It uses a closed-form optimal denoiser for patches, eliminating the need for neural network training, which previously required hours of optimization.
Single-image generative modeling is useful when data is scarce, as it constrains the output to contain exactly the patches present in the input. This keeps style and structure closely aligned. Large models trained on internet-scale data may deviate from input patch statistics, whereas single-image models enforce strict adherence.
This method dramatically reduces computational cost, enabling megapixel generation in one second and gigapixel in minutes. It supports unconditional generation, text-guided stylization, symmetrization, and retargeting. However, the patch-based approach may struggle with complex global structures.