Reproduce “Perception Prioritized Training of Diffusion Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, Sungroh Yoon

Researchers at Seoul National University developed a Perception Prioritized (P2) weighting scheme for diffusion models, which rebalances the training objective to emphasize learning perceptually relevant image features. This method consistently reduced FID scores across various datasets and achieved state-of-the-art performance on Oxford Flowers and CelebA-HQ.

Key insights

  • Diffusion models implicitly learn distinct visual information across different noise levels during training, ranging from coarse global features to fine-grained details.
  • Optimizing the training objective by prioritizing noise levels where perceptually rich or coarse content is learned, while de-emphasizing imperceptible details, leads to better image generation.
  • A seemingly minor adjustment to the loss weighting can fundamentally improve the inductive bias of diffusion models, enhancing their ability to generate high-fidelity images.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.