Flownomaly

The choice of operating time depends on the size of the support set, and stability under the extreme 1-shot setting still needs to be verified on more classes.

Flownomaly does not fine-tune the DINOv2 backbone but instead performs conditional flow matching on frozen patch tokens.

Flownomaly proposes a Jacobian gain-noise criterion that relies only on normal samples, used to select the operating time (recovery step size) in few-shot settings where no anomalous validation samples are available.

In the 1-shot setting, Flownomaly achieves 97.8% image-level AUROC and 98.0% pixel-level AUROC on MVTec-AD.

In the 1-shot setting, Flownomaly achieves 95.3% image-level AUROC and 98.7% pixel-level AUROC on VisA, with image-level performance 2.1 percentage points higher than the strongest baseline.

The paper covers 1-, 2-, and 4-shot benchmarks on MVTec-AD and VisA, and conducts a sample-count variation evaluation on Real-IAD, achieving state-of-the-art results compared with recent class-specific methods in all cases.

Flownomaly performs conditional transport on frozen DINOv2 features, with the restoration target anchored to normal support samples, structurally reducing the risk of anomalous content being reproduced.

Flownomaly, in a frozen DINOv2 feature space, learns a velocity field through conditional flow matching to pull query tokens back toward their normal counterparts, and uses the squared residual before and after restoration as the anomaly signal.

Flownomaly's evaluation is concentrated on industrial inspection benchmarks, using class-specific modeling in which each class requires separate adaptation, and cross-class generalization capability has not yet been discussed.

Frozen DINOv2 features mean that Flownomaly's detection ceiling is constrained by the sensitivity of that feature space to fine-grained defects.

The Flownomaly code is open-sourced at github.com/b-rbmp/Flownomaly, and the few-shot benchmarks can be reproduced directly on MVTec-AD and VisA.