Reproduce “DriftXpress: Faster Drifting Models via Projected RKHS Fields” 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

Ali Falahati, Elliot Creager, Gautam Kamath, Shubhankar Mohapatra

DriftXpress reduces repeated training-time attraction calculations for one-step image generation. It precomputes landmark-based summaries of training data to estimate attraction, while computing repulsion exactly against generated samples. On CIFAR10, it reports 16,342 images per second versus 2,463 for standard drifting, with Best FID 5.52 ± 0.06 versus 5.64 ± 0.16. This comparison is specific to that setting; memory is higher.

Summary by alphaXiv.

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