GOOD：基于引导扩散采样的训练-free out-of-distribution 检测
摘要
近期研究探索了将 text-to-image diffusion models 用于合成 out-of-distribution（OOD）样本，显著提升了 OOD 检测性能。然而，现有方法通常依赖扰动 text-conditioned embeddings，导致 semantic instability 和 insufficient shift diversity，限制了对真实 OOD 的 generalization。为此，我们提出 GOOD（Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection），一种 novel and flexible 框架，直接通过 off-the-shelf in-distribution（ID）classifier 指导 diffusion sampling 轨迹向 OOD region 移动。GOOD 包含 dual-level guidance：（1）基于 log partition 的 Image-level guidance，用于 reduce input likelihood，驱动 samples toward low-density regions in pixel space；（2）基于 classifier latent space中 k-NN distance 的 Feature-level guidance，promotes sampling in feature-sparse regions。因此，dual-guidance 设计 enables more controllable and diverse OOD sample generation。此外，我们引入 unified OOD score，adaptive combine image and feature discrepancies，enhance detection robustness。我们进行 thorough quantitative and qualitative analyses，以 evaluate GOOD 的 effectiveness，证明 training with samples generated by GOOD 可显著提升 OOD 检测性能。
引用
@article{arxiv.2510.17131,
title = {GOOD: Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection},
author = {Xin Gao and Jiyao Liu and Guanghao Li and Yueming Lyu and Jianxiong Gao and Weichen Yu and Ningsheng Xu and Liang Wang and Caifeng Shan and Ziwei Liu and Chenyang Si},
journal= {arXiv preprint arXiv:2510.17131},
year = {2025}
}
备注
28 pages, 16 figures, conference