Spectral Prior for Reducing Exposure Bias in Diffusion Models
Merged summary
TL;DR - Spectral Alignment (SPA) reduces diffusion-model exposure bias by guiding intermediate predictions toward a learned power-spectrum prior. It improves diverse diffusion and flow-matching architectures with only 3–4% computational overhead.
- Identifies model- and timestep-dependent frequency/SNR mismatches between training and inference.
- Fits a parametric spectrum prior offline from training data.
- Applies lightweight FFT-based gradient guidance during inference.
- Complements Classifier-Free Guidance and supports DDPM, ADM, SD2.0, SDXL, SD3.5, and FLUX.
Sources (1)
Spectral Prior for Reducing Exposure Bias in Diffusion Models
TL;DR - Spectral Alignment (SPA) reduces diffusion-model exposure bias by guiding intermediate predictions toward a learned power-spectrum prior. It improves diverse diffusion and flow-matching architectures with only 3–4% computational overhead.
- Identifies model- and timestep-dependent frequency/SNR mismatches between training and inference.
- Fits a parametric spectrum prior offline from training data.
- Applies lightweight FFT-based gradient guidance during inference.
- Complements Classifier-Free Guidance and supports DDPM, ADM, SD2.0, SDXL, SD3.5, and FLUX.