Reproduce “Sequential Flow Straightening for Generative Modeling” 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
Jongmin Yoon, Juho Lee
Sequential Flow Straightening (SEQRF) introduces a method to reduce global truncation error in continuous-time generative models by segmenting and locally straightening probability flow trajectories. This approach enables significantly faster sampling while improving the quality of synthesized outputs, achieving an FID of 3.191 on CIFAR-10 with only 6 function evaluations.
Key insights
- Global truncation error (GTE) accumulates superlinearly, and its growth can be significantly suppressed by dividing the ODE's time domain into smaller segments and locally straightening the flow within each segment.
- Reducing the Lipschitz constant and the squared norm of the vector field gradient through sequential straightening leads to a smoother, less curved vector field, which directly contributes to lower GTE and improved ODE integration stability.
- Training from joint distributions generated from short trajectory segments acts as a variance reduction technique, contributing to improved training stability and efficiency in learning the vector field.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.