PEEK：扩散采样中一步预看曝光偏差校正
原标题：PEEK: One-Step-Look-Ahead Exposure Bias Correction for Diffusion Sampling
AI 摘要
论文提出 PEEK，一种无需训练的确定性采样器插件，用于缓解扩散模型采样中的曝光偏差问题。PEEK 在每个去噪步骤先做一次临时单步更新，在预看状态上再评估一次去噪器，并将当前干净预测与一步后的预测插值形成下一噪声状态；通过贪心指标驱动的搜索优化插值调度，并借助前缀与噪声调度重基缓存提升效率。在 CIFAR-10、LSUN Bedroom、ImageNet 和 MS-COCO 上，覆盖像素空间与潜空间、无条件/标签引导/文本引导设置，PEEK 在相同 NFE 下持续优于基线，FID 最多降低 49.3%，每模型一次性校准耗时不超过 2.2 分钟。
正文节选
PEEK: One-Step-Look-Ahead Exposure Bias Correction for Diffusion Sampling Abstract Diffusion image generators owe much of their practical success to accurate denoising networks and strong numerical samplers. Yet the two components are not fully matched. During sampling, the denoiser predicts the next step from flawed inputs that differ substantially from real data. This train-test gap, or exposure bias, caps image generation quality (i.e., blurry images, missing details, or distorted shapes). To