ECCV Paper: Spectral Alignment Reduces Exposure Bias in Diffusion Models

mittu1204 · x · 2026-08-31

Addressing error accumulation (exposure bias) in diffusion models, this paper proposes Spectral Alignment (SPA). The study reveals frequency-dependent SNR discrepancies between training and inference that vary across models and timesteps. SPA calibrates the power spectrum of intermediate predictions using an offline-fitted parametric model and efficient FFT-based gradients at inference. With minimal overhead (3-4%) and compatibility with CFG, SPA consistently improves performance across architectures like DDPM, SDXL, and FLUX.

Original post →

More from Research

Research channel →