Origin of spectral bias in pixel-space flow-matching models

Determine the origin of spectral bias in modern pixel-space flow-matching models, including why a frequency objective plateaus when used alone during early training and why switching to pixel-space losses produces better performance at later training stages.

Background

The paper demonstrates that the pixel-space velocity loss used in JiT flow-matching models overemphasizes low spatial frequencies during early training and underestimates high-frequency details. The proposed frequency-domain loss mitigates this imbalance and improves early convergence, while a scheduled transition back to the pixel-space loss improves later-stage refinement.

Although the empirical benefit of combining frequency and pixel supervision is established, the mechanisms producing the observed training dynamics are not explained. In particular, the paper leaves unresolved both the underlying cause of spectral bias and the reason the frequency-only objective eventually plateaus whereas the scheduled objective performs better in later training.