Reproduce “Guidance Without Labels: Flow Matching from Coupled Gaussian Mixtures” 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
Stanislas Ducotterd, Bassam El Rawas, Michael Unser
The method adapts pretrained flow-matching image generators by fitting a Gaussian mixture to training data and coupling its components probabilistically to data points. It reuses frozen velocity networks and adds a correction. In tests on two published image checkpoints, it reaches CelebA-HQ FID 11.24 at 64 Euler steps versus 11.45 at 128, reducing base-network evaluations, not establishing a general speedup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.