💻 computer science

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

The paper proposes FreqPrompt-AD, a training-free framework that enhances zero-shot industrial anomaly detection and segmentation by leveraging frequency-guided local semantic prompting to overcome the limitations of global image-text alignment in CLIP-style models, achieving state-of-the-art performance on multiple benchmarks.

Original authors: Siqi Qiao, Chang Liu

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Siqi Qiao, Chang Liu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a quality inspector at a factory. Your job is to spot tiny scratches, cracks, or stains on products like metal gears, circuit boards, or fabric. The problem? You've never seen this specific type of product before, and you don't have a manual or a list of what "bad" looks like for it. You just have to look at a picture and say, "That looks wrong."

This is the challenge of Zero-Shot Industrial Anomaly Detection. The paper introduces a new tool called FreqPrompt-AD to solve this.

Here is how it works, explained through simple analogies:

The Problem: The "Big Picture" vs. The "Tiny Scratch"

The researchers noticed that modern AI models (called Vision-Language Models, or VLMs) are like generalist art critics. They are great at looking at a whole painting and saying, "This is a landscape," or "This is a portrait." They understand the "big picture" very well.

However, industrial defects are often tiny, high-frequency details—like a hairline crack or a speck of dust. When you ask a generalist art critic to find a tiny scratch on a metal surface, they often miss it. They get too distracted by the overall shape of the object (e.g., "That's a gear") and ignore the tiny texture changes.

The paper shows that defective areas are "noisier" (they have more high-frequency energy) than normal areas, but standard AI models tend to smooth over this noise because they are trained to be calm and consistent.

The Solution: FreqPrompt-AD

The authors built a system that acts like a specialized detective who knows exactly where to look. Instead of just asking the AI "Is this object broken?", they give it a three-step strategy to find the trouble spots:

1. The "Static Filter" (Frequency-Aware Prompt Interaction)

Imagine you are listening to a song. Sometimes, the music is smooth, but a scratch on the record creates a sharp, high-pitched pop or hiss.

What the paper does: It takes the image and separates the "smooth music" (low-frequency background) from the "static" (high-frequency details).
The Analogy: It tells the AI, "Ignore the smooth parts of the image. Focus your attention only on the areas that look 'staticky' or 'crunchy.' Those are the places where a defect is likely hiding." This turns invisible texture changes into a spotlight for the AI.

2. The "Local Librarian" (Local Semantic Calibration)

Sometimes, the "static" isn't a defect; it's just the natural edge of the object (like the rim of a cup). If the AI only looks for noise, it might get confused.

What the paper does: It creates a "normality prototype" for this specific image. It looks at the clean, safe parts of the image and says, "Okay, this is what 'normal' looks like right here."
The Analogy: It's like a librarian who knows exactly what a "clean book" looks like on this specific shelf. If a book has a weird smudge, the librarian compares it to the other clean books on the shelf, not to a generic definition of a book. This stops the AI from getting confused by the object's natural shape.

3. The "Final Verdict" (Text-Anchored Decision)

Now the AI has two pieces of evidence:

"This area looks noisy/staticky" (from Step 1).
"This area looks different from the clean parts of this specific image" (from Step 2).

What the paper does: It combines these clues with the AI's original knowledge of what "broken" means (text prompts).
The Analogy: It's a judge weighing the evidence. "The static filter says 'suspicious,' the local librarian says 'unusual,' and the text prompt says 'this matches the description of a defect.' Therefore, we have a defect."

The Results

The researchers tested this on four different industrial datasets (MVTec AD, VisA, BTAD, and MPDD).

The "Training-Free" Version: The system works perfectly without needing to be taught on the specific factory line. It just uses the pre-trained AI and the three steps above. It beat all other similar "zero-shot" methods.
The "Adapter" Version: They also made a tiny, lightweight add-on (like a small plugin) that learns just a little bit from normal images. This version performed even better, achieving the highest scores overall, but it still didn't need to retrain the massive AI brain.

Why It Matters

The paper claims this is the best method currently available for finding defects in new products without needing to collect thousands of training photos first. It works because it stops the AI from just "guessing" based on the object's shape and forces it to look at the texture and frequency details where the actual damage hides.

In short: FreqPrompt-AD teaches a generalist AI to put on "noise-canceling headphones" for the background and "high-frequency glasses" for the defects, allowing it to spot tiny cracks and scratches that other models miss.

Technical Summary: FreqPrompt-AD

Problem Statement
Industrial anomaly detection (IAD) traditionally relies on one-class or unsupervised learning using normal samples, which limits applicability in zero-shot scenarios where target-category data is unavailable due to confidentiality or cold-start constraints. While Vision-Language Models (VLMs) like CLIP offer strong zero-shot transferability through semantic alignment, they suffer from a "granularity mismatch." VLMs are trained for object-level semantic invariance, often overlooking local, high-frequency defects such as thin scratches, cracks, or boundary irregularities. Furthermore, standard VLM embeddings tend to suppress high-frequency responses, making them suboptimal for dense industrial defect localization where defects manifest as subtle texture discontinuities rather than category changes.

Methodology: FreqPrompt-AD
The authors propose FreqPrompt-AD, a training-free framework that reformulates zero-shot anomaly detection as a "local-frequency semantic evidence construction" problem. The framework utilizes a frozen VLM backbone and consists of three core modules:

Frequency-aware Prompt Interaction (FPI):
- The input image is decomposed into a low-frequency component and a high-frequency residual using a Gaussian low-pass operator.
- High-frequency residuals are averaged over image patches to generate a frequency prior ( $w_f$ ).
- This prior modulates the standard patch-text anomaly probability. Specifically, the raw semantic anomaly score is amplified for patches located in high-frequency regions, converting spectral residuals into spatial guidance for patch-text alignment.
Local Semantic Calibration (LSC):
- To mitigate object-level semantic bias (where normal and abnormal images share the same category identity), LSC removes the global semantic component from patch tokens via projection.
- An image-adaptive "normality prototype" is constructed by aggregating the least suspicious patch tokens (selected based on low anomaly scores and low frequency weights).
- Local anomaly evidence is derived by measuring the deviation of calibrated patch tokens from this image-specific normality prototype, complementing the text-based evidence.
Text-Anchored Decision (TAD):
- TAD fuses the frequency-conditioned semantic evidence (from FPI) and the prototype-deviation evidence (from LSC) using a weighted sum.
- The resulting patch scores are upsampled and smoothed to generate dense anomaly maps.
- Image-level anomaly scores are computed via top- $p$ pooling over the anomaly map to avoid dilution of small defects by global averaging.

Adapter-Enhanced Variant (FreqPrompt-AD-FT)
The authors also introduce FreqPrompt-AD-FT, a variant where a lightweight bottleneck adapter is inserted into the visual encoder. Only the adapter parameters are optimized using normal training images to minimize normal patch scores and suppress false positives, while the VLM backbone, text encoder, and prompts remain frozen.

Key Contributions

Framework Design: Proposes a unified, training-free framework that explicitly injects frequency-domain priors into the patch-text interaction process of a frozen VLM, addressing the granularity mismatch in industrial defect detection.
Evidence Chain: Designs a cohesive evidence chain (FPI, LSC, TAD) where frequency cues expose texture discontinuities, normality prototypes suppress object-level bias, and language anchors provide decision direction.
Performance: Demonstrates that this formulation achieves state-of-the-art results among zero-shot VLM-based detectors without requiring target-specific memory banks or full model fine-tuning.

Experimental Results
The method was evaluated on four industrial benchmarks: MVTec AD, VisA, BTAD, and MPDD.

Zero-Shot Performance: FreqPrompt-AD outperformed all existing zero-shot VLM-based detectors. Compared to the strongest baseline, DLVP-CLIP, it improved average image-level AUROC by 2.05% and pixel-level AUROC by 2.15%.
Adapter Variant: FreqPrompt-AD-FT achieved the best overall average results across all datasets and metrics, validating that the evidence formulation benefits from lightweight adaptation.
Efficiency: The strict zero-shot version requires no trainable parameters and no memory bank construction, offering lower inference latency than methods like AF-CLIP and VisualAD. The FT variant adds only 0.68M parameters while significantly boosting accuracy.
Qualitative Analysis: Visualizations show that FreqPrompt-AD produces sharper, more localized heatmaps for thin cracks and irregular contaminations, reducing background activation on normal object structures compared to global-prompt methods.

Significance and Claims
The paper claims that FreqPrompt-AD addresses a critical limitation in current zero-shot anomaly detection: the inability of global VLM semantics to capture local, high-frequency defects. By treating anomaly detection as a construction of "local-frequency semantic evidence," the method bridges the gap between semantic transferability and the physical nature of industrial defects. The authors assert that their approach is particularly suitable for rapid industrial deployment in single-GPU environments where target data is unavailable. They emphasize that frequency priors must be calibrated by semantic information (via LSC and TAD) to avoid false positives on normal high-frequency structures like edges, a balance achieved by their unified evidence chain. Future work is suggested to explore adaptive frequency-band selection and extension to multimodal 2D-3D reasoning.