Spectral Prior for Reducing Exposure Bias in Diffusion Models

Researchers proposed Spectral Alignment, a lightweight inference-time method to reduce sampling instability in diffusion models.

The method targets frequency-dependent train-test mismatch by aligning intermediate spectra to a learned prior using FFT-based guidance.

It adds only small overhead and improves results across multiple diffusion and flow-matching families, including DDPM, ADM, SD2.0, SDXL, SD3.5, and FLUX.

The work addresses exposure bias and spectral mismatch, a common source of sampling error in generative models.