Few-shot industrial anomaly detection has to model what normal looks like from a handful of defect-free images, and most methods tie inference to stored exemplars, impose rigid linear structure, or risk reproducing the anomalous content they are trying to detect. Flownomaly recasts the task as normal-counterpart recovery in frozen DINOv2 feature space: a velocity field is learned by conditional flow matching over frozen patch tokens to transport Gaussian noise toward normal features.
The payoff is the step count. Along straight paths, a single recovery step using one evaluation of the optimal field yields the posterior mean of the clean normal token, with no trajectory integration. The operating time is selected from the supports alone by a normal-only Jacobian gain-to-noise criterion, so no anomalous validation data is needed. On MVTec-AD and VisA at 1, 2 and 4 shots, plus sample-shot evaluations on Real-IAD, Flownomaly reaches state-of-the-art performance against recent category-specific methods: image-level and pixel-level AUROC of 97.8 percent and 98.0 percent on MVTec-AD, and 95.3 percent and 98.7 percent on VisA - a 2.1 point image-level AUROC gain over the strongest baseline on VisA.
The caveat is that this is a vision-features result on standard industrial benchmarks, and the anomaly signal comes from residuals between a query token and its recovered counterpart rather than from a direct defect classifier.
https://arxiv.org/abs/2610.32964
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.