Reproduce “Unbiased Image Synthesis via Manifold Guidance in 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

Xingzhe Su, Daixi Jia, Fengge Wu, Junsuo Zhao, Changwen Zheng, Wenwen Qiang

Manifold Guidance Sampling (MGS), developed by researchers at the Chinese Academy of Sciences, provides an unsupervised, plug-and-play method to reduce bias amplification in Diffusion Models, bringing gender representation in generated CelebA images from 148% overrepresentation down to near the training data's 57.9%, concurrently enhancing image quality and diversity.

Key insights

  • High-dimensional data intrinsically lies on lower-dimensional submanifolds, and an unbiased generative model should ideally distribute samples uniformly across this true data manifold.
  • An unsupervised network architecture can be effectively trained to learn this intrinsic data manifold by optimizing for maximal variance between different submanifolds and minimal variance within them.
  • Integrating adherence to a learned data manifold as a guidance mechanism during the diffusion model's inference process can mitigate bias amplification and improve diversity without requiring costly model retraining or prior bias labels.

Summary by alphaXiv.

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