Backbone-Equated Diffusion OOD via Sparse Internal Snapshots

TL;DR

This paper introduces a new protocol and detectors for fair comparison of diffusion-based out-of-distribution detection, revealing that sparse internal states contain most of the OOD signal.

Contribution

It proposes the Mutualized Backbone-Equated protocol and Canonical Feature Snapshots, showing that sparse internal diffusion states are highly informative for OOD detection.

Findings

Sparse internal states capture most OOD signals in diffusion models.

A small number of internal activations can outperform full denoising trajectories.

The official implementation is available at the provided GitHub URL.

Abstract

Fair comparison between diffusion-based OOD detectors is challenging, as conclusions can vary with backbone choice, corruption parameterization, and test-time budget. We address this issue through a Mutualized Backbone-Equated (MBE) protocol that aligns canonical corruption levels and logical test-time cost across diffusion backbones. Within this setting, we introduce Canonical Feature Snapshots (CFS), a family of detectors that probes a frozen diffusion backbone using only a tiny number of native internal activations at canonical low-noise levels. On a controlled CIFAR-scale benchmark, the strongest one-forward CFS variant is CFS(1x2), while an even smaller decoder-only variant remains highly competitive. This shows that much of the relative-OOD signal exposed by frozen diffusion backbones is concentrated in a small number of sparse internal states, rather than requiring full denoising trajectories or high-capacity downstream heads. We further provide a local diagnostic theory explaining these observations through conditional encoder-decoder complementarity, diagonal-score separation, and low-noise corruption stability. The official implementation is available at https://github.com/RouzAY/cfs-diffusion-ood/.