Reproduce “Balancing Frequencies and Pixels in Flow Matching” 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

Lucas Degeorge, Paul Couairon, Arijit Ghosh, Alexei A. Efros, David Picard, Vicky Kalogeiton

The authors introduce a scheduled frequency-to-pixel loss for JiT flow-matching image generators. It adaptively weights Fourier-domain prediction errors early in training, then shifts supervision to pixel-space v-loss. On class-conditional ImageNet at 256×256, guided JiT-L/16 shows 1.41× faster FID convergence by training epochs, where FID measures Fréchet distance between generated and real InceptionV3 features; the method adds 14% wall-clock time, with gains offsetting cost from 300 minutes onward.

Summary by alphaXiv.

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