Reproduce “Few-Step Diffusion Sampling Through Instance-Aware Discretizations” 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
Liangyu Yuan, Ruoyu Wang, Tong Zhao, Dingwen Fu, Mingkun Lei, Beier Zhu, Chi Zhang
A framework called INDIS dynamically adapts timestep schedules for diffusion and flow matching models based on input-dependent priors, enhancing few-step sampling quality. This approach yields average FID improvements of 35.33%, 31.50%, and 15.62% for 3, 5, and 7 NFEs, respectively, across various diffusion probabilistic models.
Key insights
- Globally optimized timestep schedules are suboptimal; instance-specific adaptation of discretizations can significantly reduce endpoint errors and improve sample generation quality, particularly in few-step sampling.
- Conditioning timestep schedules on initial noise and auxiliary conditions enables dynamic, tailored adjustments that better capture the unique characteristics of each generative process.
- Instance-specific time and scale shift factors serve as an effective mechanism to alleviate exposure bias, enhancing the robustness and performance of the sampling process in diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.