Research

Focal Log-Frequency Loss Cuts Flow-Matching Training Time by Up to 40%

3 days ago
arXiv logo

Image via arxiv.org

Researchers from multiple institutions have identified a fundamental spectral imbalance in how pixel-space flow-matching models are trained: because natural images follow a 1/f² spectral distribution, standard pixel-space reconstruction losses are dominated by low-frequency signals, causing models to learn fine-grained textures and edges only slowly. To fix this, the team proposes a Focal Log-Frequency Loss (f-loss) that equalizes the learning signal across the frequency spectrum, giving underrepresented high-frequency components more weight during optimization.

The proposed training strategy pairs f-loss with a two-phase schedule — prioritizing frequency-domain supervision early in training, then transitioning to a standard pixel-space v-loss for spatial refinement. The method requires no architectural changes and can serve as a drop-in replacement for existing flow-matching objectives. Experiments across multiple model scales show up to 40% faster convergence alongside consistent improvements in FID scores and perceptual fidelity.