Spectral Prior for Reducing Exposure Bias in Diffusion Models
SPA targets diffusion sampling errors by matching intermediate frequency spectra to a learned training-data prior.
The paper says exposure bias shows up as frequency-dependent SNR mismatch between training and inference. That mismatch changes by model and timestep, so the authors argue fixed correction rules do not travel well. Their Spectral Alignment method fits a spectrum model offline, then applies FFT-based guidance during inference with reported overhead of 3-4%. They report gains across DDPM, ADM, SD2.0, SDXL, SD3.5, and FLUX, and say the method is complementary to CFG. HF Daily Papers' note
The paper says exposure bias shows up as frequency-dependent SNR mismatch between training and inference. That mismatch changes by model and timestep, so the authors argue fixed correction rules do not travel well. Their Spectral Alignment method fits a spectrum model offline, then applies FFT-based guidance during inference with reported overhead of 3-4%. They report gains across DDPM, ADM, SD2.0, SDXL, SD3.5, and FLUX, and say the method is complementary to CFG. HF Daily Papers' note
score 5