New Diffusion OOD Detector Uses Sparse Internal Snapshots

A new research article presents Canonical Feature Snapshots (CFS), a set of detectors aimed at out-of-distribution (OOD) detection utilizing frozen diffusion backbones. The authors introduce a Mutualized Backbone-Equated (MBE) protocol designed to ensure standardized comparisons among backbones by synchronizing corruption levels and test-time expenses. CFS examines only a limited number of native internal activations at canonical low-noise levels, thus bypassing the need for complete denoising trajectories. In a CIFAR-scale benchmark, the most effective single-forward variant is identified as CFS(1x2), with a smaller decoder-only variant also proving competitive. Results suggest that the relative OOD signal is primarily found in sparse internal states. The paper can be accessed on arXiv with ID 2605.11014.