Tying loss bias and guidance intervals to spectral properties

Ascertain whether the sigmoid‑weighted ELBO loss bias parameter and the classifier‑free guidance interval bounds in pixel diffusion models can be determined as functions of an image’s radially averaged power spectral density, thereby removing manual tuning across resolutions.

Background

Although the proposed spectrally guided schedules adapt across resolutions without hyperparameter changes, the training loss bias and sampling guidance interval still require manual tuning. These hyperparameters materially affect optimization and sampling behavior.

The authors explicitly state uncertainty about whether these remaining components can also be parameterized using spectral information, analogous to how the schedule is derived from the RAPSD.

References

While our noise schedules successfully adapt to different resolutions with no hyperparameter changes, other aspects of the model still need tuning; namely, the loss bias and guidance intervals. It remains to be seen whether these could also be tied to spectral properties.

— Spectrally-Guided Diffusion Noise Schedules  (2603.19222 - Esteves et al., 19 Mar 2026) in Conclusion and limitations