Spectral Prior for Reducing Exposure Bias in Diffusion Models

Researchers found a frequency-dependent mismatch between training and inference in diffusion sampling, which contributes to exposure bias and error accumulation.

They propose Spectral Alignment, a lightweight inference-time method that matches intermediate prediction spectra to a learned spectral prior.

The method fits the spectrum offline and uses FFT-based guidance during sampling with little added overhead.

It improves results across DDPM, ADM, SD2.0, SDXL, SD3.5, FLUX, and other diffusion or flow-matching models.