New f-loss Cures Spectral Bias in Pixel-Space Flow Matching, Speeding Convergence

serrjoa · x · 2026-09-04

An arXiv paper identifies spectral imbalance as a key inefficiency in pixel-space flow matching: pixel losses treat all spatial errors equally, so low frequencies dominate and fine details are learned late. The authors propose a Focal Log-Frequency Loss (f-loss) that equalizes learning signal across frequencies, plus a curriculum that starts with frequency-domain supervision and transitions to pixel-space v-loss. Requiring no architectural changes, the drop-in loss speeds up convergence and improves FID across multiple model scales.

Original post →

More from Multimodal

Multimodal channel →