Reproduce “Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps” 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

Mingxiao Li, Tingyu Qu, Ruicong Yao, Wei Sun, Marie-Francine Moens

This paper addresses exposure bias in Diffusion Probabilistic Models by introducing a training-free "Time-Shift Sampler." The sampler dynamically adjusts denoising time steps based on the predicted sample's variance, consistently improving FID scores and enhancing generation quality, especially when using significantly fewer sampling steps for faster inference.

Key insights

  • Exposure bias in DPMs manifests as accumulating prediction errors that disproportionately affect sample quality in later denoising stages, partly due to the varying sensitivity of the network to input variations across different time steps.
  • For a given network-predicted sample, an 'alternate time step' can exist that provides a superior statistical coupling (alignment) compared to the conventionally dictated next time step.
  • This optimal shifted time step can be theoretically derived by minimizing KL divergence, suggesting that the shifted step's variance should closely match the variance of the predicted sample.

Summary by alphaXiv.

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