Apple's research team has proposed a nested diffusion model called Matryoshka Diffusion Models (MDM) for generating high-resolution images. The central idea of the research is to train models that can generate 1024×1024 pixel single-pixel spatial models by using a low-resolution diffusion process as part of high-resolution generation and performing a joint diffusion process at multiple resolutions through a nested UNet architecture.
Paper address: https://arxiv.org/pdf/2310.15111.pdf
