This paper introduces Patch-DM, a denoising diffusion model that synthesizes high-resolution images from small-sized patches. The model utilizes a novel “feature collage” strategy to avoid boundary artifacts when creating large images, thus ensuring seamless image generation. By forcing feature consistency across neighboring patches, this approach alleviates memory complexity issues associated with classic diffusion models while maintaining excellent image quality.
Publication date: 2 August 2023
Project Page: N/A
Paper: https://arxiv.org/pdf/2308.01316.pdf
