FreqPrompt-AD: Frequency-guided local semantic prompting for zero-shot industrial anomaly detection and segmentation

Abstract Zero-shot industrial anomaly detection aims to identify and segment defects in unseen categories without target-specific model fitting. CLIP-style vision-language models (VLMs) provide useful semantic transferability, but their global image-text alignment often emphasizes object-level semantics and overlooks local high-frequency defects. We propose \ours, a training-free framework that constructs local-frequency semantic evidence with a frozen VLM. Frequency-aware prompt interaction (FPI) converts high-frequency residuals into spatial guidance for patch-text anomaly evidence; local semantic calibration (LSC) calibrates patch tokens with image-adaptive normality prototypes; and text-anchored decision (TAD) fuses semantic and prototype-deviation evidence into image-level anomaly scores and dense anomaly maps. We further evaluate \oursft, an adapter-enhanced variant in which only a lightweight adapter is optimized and the VLM backbone and text encoder remain frozen. Full-scale experiments on MVTec AD, VisA, BTAD, and MPDD show that \ours{} achieves the best performance among zero-shot VLM-based detectors, improving average image-level and pixel-level AUROC over DLVP-CLIP by 2.05 and 2.15 percentage points, respectively. \oursft{} obtains the best overall average results. Ablations, sensitivity analysis, efficiency comparison, and qualitative results verify the effectiveness of the proposed local-frequency semantic evidence formulation.