Diffusion Models Need Visual Priors for Image Generation

TL;DR

This paper introduces Diffusion on Diffusion (DoD), a multi-stage framework that enhances image generation by incorporating visual priors extracted from previous samples, significantly improving quality and reducing training costs.

Contribution

The paper proposes a novel multi-stage diffusion framework that leverages visual priors via a latent embedding module, leading to superior image quality with less training.

Findings

DoD reduces training cost by 7× compared to SiT and DiT.

DoD-XL achieves an FID-50K score of 1.83 with only 1 million steps.

DoD outperforms state-of-the-art methods in image quality without additional inference complexity.

Abstract

Conventional class-guided diffusion models generally succeed in generating images with correct semantic content, but often struggle with texture details. This limitation stems from the usage of class priors, which only provide coarse and limited conditional information. To address this issue, we propose Diffusion on Diffusion (DoD), an innovative multi-stage generation framework that first extracts visual priors from previously generated samples, then provides rich guidance for the diffusion model leveraging visual priors from the early stages of diffusion sampling. Specifically, we introduce a latent embedding module that employs a compression-reconstruction approach to discard redundant detail information from the conditional samples in each stage, retaining only the semantic information for guidance. We evaluate DoD on the popular ImageNet-$256 \times 256$ dataset, reducing 7$\times$ training cost compared to SiT and DiT with even better performance in terms of the FID-50K score. Our largest model DoD-XL achieves an FID-50K score of 1.83 with only 1 million training steps, which surpasses other state-of-the-art methods without bells and whistles during inference.