Spectral Prior for Reducing Exposure Bias in Diffusion Models

The paper identifies systematic, frequency-dependent discrepancies between diffusion-model training and iterative inference, interpreting them as frequency-dependent SNR errors. It proposes Spectral Alignment (SPA), a lightweight guidance method that calibrates the power spectrum of intermediate predictions against a precomputed parametric prior. The prior is fitted offline from training data, while inference uses efficient FFT-based gradients. The authors report 3–4% computational overhead, compatibility with Classifier-Free Guidance, and improvements across pixel-space models such as DDPM and ADM, latent models including SD2.0 and SDXL, and flow-matching models including SD3.5 and FLUX.