An AI method designed to find industrial defects from only a few normal examples reported average image-level AUROC scores of 95.2% to 96.0% and pixel-level AUROC scores of 97.1% to 97.6% as the support budget rose from one to four shots.
Its reported average AUPR scores rose from 96.8% to 97.3% over the same settings, while PRO, the paper's per-region overlap measure, rose from 89.5% to 90.5%.
These results came from benchmark experiments on MVTec-AD and VisA using a few-shot protocol built around normal support samples.
A text descriptor that changes with the image
DriftAD's central question is whether frozen CLIP text embeddings can be conditioned on local visual context, so anomaly descriptors adapt to different parts of an image instead of remaining static.
The framework combines ASA for spatial and frequency amplification, VGTD for adaptive text drift and DGSG for probe-based spatial gating that helps decide where the anomaly signal should be emphasized.
Training adds gate-supervision and drift-separation losses. They are intended to encourage spatially discriminative gating and keep normal and abnormal drifted descriptors from collapsing together.
The tests used two industrial benchmarks
Tests used MVTec-AD, which has 5,354 images across 15 categories—five textures and 10 objects—with 3,629 normal training images and 1,725 test images. VisA has 10,821 images across 12 subsets, including 9,621 normal images and 1,200 anomalous images.
The protocol trains on normal samples, tests on mixed normal and anomalous samples, and uses NSA-synthesized pseudo-anomalies for pixel supervision. It trains on one dataset and tests on the other to avoid leakage.
Inputs were resized to 224 by 224 pixels, and the evaluation used frozen OpenCLIP ViT-H/14 image and text encoders from ImageBind-H, with vision features taken from layers 8, 16, 24 and 36.
The benchmark comparisons come with a caveat
With four-shot support, the paper reports DriftAD as competitive or superior to the listed 8-shot and full-shot baselines. Against full-shot DiAD on MVTec-AD, the reported differences favored DriftAD by 0.8 percentage points in AUROC and 0.4 points in pAUROC; on VisA, it remained competitive with full-shot methods.
That comparison needs care: four-shot, eight-shot and full-shot results use different support regimes, so they are not equivalent sample-budget comparisons.
Which pieces mattered
In the architecture ablation, the version that added VGTD and DGSG to the ASA-only baseline was associated with gains of 1.1 percentage points in AUROC and 0.8 points in pAUROC on MVTec-AD, and 0.6 and 0.2 points on VisA.
The loss ablation reported an incremental benefit from adding Lgate to the base architecture and further gains from Ldrift; the full loss configuration performed best on both datasets.
A layer-resolution test favored the adaptive grid {4, 8, 8, 16}, which reached 96.8% pAUROC. Uniform-16 performed worst, and the authors report that deeper layers benefit from finer-grained drift fields.
In a visualization-based patch-level analysis on MVTec-AD, the drifted-text condition showed a 0.063 cosine-similarity gap between normal and anomalous examples, versus 0.024 with clean text. The drifted gap was 2.6 times larger, but this was not an independent statistical test.
What the benchmark cannot settle
The evidence is limited to offline benchmark experiments on MVTec-AD and VisA. Because pixel-level supervision used NSA-synthesized anomalies, the study does not isolate performance on naturally occurring defects.
The supplied analysis reports no confidence intervals, significance tests or repeated-run variability, so the size and stability of the reported differences remain uncertain.
The mean scores also do not establish that DriftAD was best on every category or metric; the supplied analysis notes exceptions in the subset results.
The paper is supplied as an arXiv preprint dated August 24, 2026; its front matter also presents an ACM MM '26 proceedings citation for November 10-14, 2026. The abstract says code is available through a GitHub repository.
Paper data and sources
Original title: DriftAD: Visually-Guided Text Drift for Few-Shot Industrial Anomaly Detection
Authors: Wenyang Liu, Tianyi Liu, Dongshuo Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text