Cause of the mismatch between diffusion loss and sample quality

Determine the precise cause of why lower velocity-space diffusion regression loss for MFlow does not consistently translate into better FID relative to JiT and AsymFlow, including whether staged or decoupled optimization of the direction-wise gains can resolve this mismatch.

Background

MFlow jointly optimizes direction-wise parameterization gains and network weights under a velocity-space objective. Although the method achieves slightly lower training loss than JiT and AsymFlow in the reported ImageNet experiments, the lower regression loss does not consistently yield better distribution-level sample quality as measured by FID.

The paper attributes this discrepancy provisionally to the fact that improved regression conditioning may not directly improve generated distributions. The authors identify staged or decoupled optimization of the gains as an important direction for future work, but do not establish the mechanism responsible for the mismatch or determine whether such optimization improves sample quality.

References

Indeed, under the same velocity-space objective, MFlow attains a slightly lower training loss than JiT and AsymFlow on the ImageNet H-level experiment (Fig.~\ref{fig:imagenet_qual_uncurated2}); we observe similar trends for B-level models with and without LPIPS. However, the lower diffusion loss does not consistently translate into better FID. This mismatch suggests that improved regression conditioning need not directly improve distribution-level sample quality, and its precise cause remains unclear. Staged or decoupled optimization of the gains is an important direction for future work.

— Rethinking the design space of Vanilla Pixel Diffusion Transformers  (2610.23854 - Mukhopadhyay et al., 16 Oct 2026) in Appendix, subsection “Limitations”