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.
More from Research
- Robot Sim2sim Transfer: Trained in <2 Minutes on One 4090 — yacineMTB · 2026-08-31
- Gumbel-Softmax Watermark Variant Can Hide LLM's True CoT Cryptographically — sytelus · 2026-08-31
- Researcher观点: Why Worst-of-N might be the one true metric for LLMs — willdepue · 2026-08-31
- Long CoT may not need single model call — CShorten30 · 2026-08-31
- Non-technical explanation of Gaussian Splatting technology — CSProfKGD · 2026-08-31
- Paper: Deliberate Practice for Robot Skills under a Budget — tomssilver · 2026-08-31