FS-IAD: Few-Shot Industrial Anomaly Detection

• Few-Shot Industrial Anomaly Detection (FS-IAD) is a paradigm that detects anomalies in industrial imagery using only a handful of normal examples per category.
• It leverages advanced methods—including vision-language models, prototype refinement, and subspace analysis—to delineate the normal data manifold and identify subtle defects.
• The approach supports rapid adaptation in dynamic manufacturing environments, achieving high accuracy even when annotated data is extremely scarce.

Few-Shot Industrial Anomaly Detection (FS-IAD) is a research paradigm focused on detecting and localizing anomalous samples in industrial imagery (and, in some cases, audio or multimodal data), under the constraint that only a handful of normal exemplars per object category are available for model construction. Unlike classical industrial anomaly detection, which presumes abundant examples of defect-free products, FS-IAD addresses the realities faced in new product introduction, rare category changeover, or high-throughput manufacturing where annotated normal and anomalous data are extremely limited. The FS-IAD setting includes not only binary detection but, in some works, multi-class anomaly identification, and operates across both visual and multimodal industrial data.

1. Formal Problem Definition and Motivating Considerations

In FS-IAD, given a set of categories $\mathcal{C} = \{1, ..., M\}$, for each $c \in \mathcal{C}$ only $K$ defect-free (normal) support images $S_c = \{x_c^1,\dots,x_c^K\}$ are provided for training (or adaptation), typically with $K \in \{1,2,4,8\}$. At test time, the system must infer from a query image $x$ whether it is normal ($y=0$) or anomalous ($y=1$). No real anomalies are usually seen during training, making the problem a variant of unsupervised or “one-class” anomaly detection but in an N-way K-shot regime.

The main technical challenge is to reliably delineate the normal data manifold (potentially with multi-class or multi-modal structure) despite the fact that the available normal samples are very sparse and do not span the full range of non-defective instance variation (Deng et al., 27 Nov 2025). This data scarcity frequently leads to ambiguous boundary formation and poor generalization to subtle or rare defect patterns.

Industrial FS-IAD is motivated by practical scenarios: deployment to new product lines, rapid adaptation to process/tooling changes, or inspection of complex, multi-part assemblies in high-mix manufacturing—where collecting and curating abundant normal data is infeasible (Zhu et al., 1 Nov 2025).

2. Algorithmic Foundations and Key Methodologies

FS-IAD methods span a range of approaches, nearly all of which exploit advances in foundation vision (and vision-language) models to increase sample efficiency, typically by leveraging the strong feature invariances and prior knowledge induced by large-scale pretraining.

Prompt/Adapter-Based Vision-Language Methods: Methods like ABounD (Deng et al., 27 Nov 2025) and H2VLR (Huang et al., 16 Apr 2026) exploit frozen CLIP backbones, learning either prompt generators or semantic hypergraph heads. ABounD integrates a dynamic concept fusion (DCF) module that fuses general anomaly priors (e.g., “flawless,” “damaged”) with class-specific cues, and combines this with adversarial boundary forging (PGD-style optimization in feature space) to sharpen discrimination boundaries. H2VLR frames anomaly inference as a high-order hypergraph reasoning task, integrating patch-patch, patch-semantic, and semantic-semantic relations and performing message passing on this structure for improved global consistency.

Prototype and Memory-Based Methods: Many approaches build prototype representations from the limited support set. VisionAD (Wang et al., 16 Apr 2025) eschews prompt engineering, using nearest-neighbor search over augmented, multi-level DINOv2 features, and a class-indexed memory for both single- and multi-class detection. FastRef (Tian et al., 26 Jun 2025) refines support prototypes at test time via linear transformation and optimal transport-based anomaly suppression, systematically blending query statistics into memory and suppressing anomalous leakage.

Subspace and Manifold Modeling: SubspaceAD (Lendering et al., 26 Feb 2026) fits a low-dimensional PCA subspace to patch-level foundation model features from support images, computing residual norms as anomaly scores. FoundAD (Zhai et al., 2 Oct 2025) learns a lightweight, nonlinear projector to “snap” test features back onto the normal manifold, with reconstruction residuals used for scoring.

Hypergraph and Graph-Based Structural Methods: CIF (Lin et al., 8 Nov 2025) uses semantic-aware hypergraphs to extract intra-class structural commonality, storing these in a memory that aids structurally guided search at test time. GATE-AD (Psiris et al., 16 Mar 2026) leverages a graph attention network (GAT) over patch tokens, introducing masked, representation-aligned message passing and latent space residual anomaly scoring.

Meta-Learning and Bayesian Approaches: BayPrAnoMeta (Sarkar et al., 27 Jan 2026) extends Proto-MAML, replacing point prototypes by a Normal-Inverse-Wishart (NIW) Bayesian prior, yielding Student-$t$ predictive likelihoods for uncertainty-aware scoring and enabling robust adaptation even when $K < d$.

Data Augmentation Strategies: The role of data augmentation is nuanced. Systematic experiments show that while augmentations (rotation, flip, scale) can expand the patch-level feature manifold and thus improve memory-bank or teacher-student methods, no single strategy is universally optimal and multi-augmentation mixtures benefit only select algorithms like PatchCore (Zhang et al., 2023).

Vision-Language Foundation Models and In-Context/Large VLMs: FADE (Li et al., 2024) and IADGPT (Zhao et al., 14 Aug 2025) adapt large-scale VLMs (CLIP, Qwen2.5-VL), leveraging multi-scale adaptation, prompt ensembles, and in-context learning pipelines to perform detection, segmentation, and even reasoning, with in-context learning delivering strong zero- and few-shot transfer.

3. Architecture Variants and Technical Mechanisms

Distinct technical innovations address the central FS-IAD challenge—discriminative normal/anomaly separation with few support samples:

• Dynamic Prompting and Hypergraph Reasoning: ABounD's DCF module fuses mixture-of-experts priors and class-specific text prompts, while ABF sculpts the margin using adversarial samples. H2VLR constructs a hypergraph over patches and prompts, generating anomaly heatmaps via L-layer hypergraph convolutions with context-weighted hyperedges.
• Prototype Refinement: FastRef's two-stage pipeline employs linear reconstruction from query to support space, followed by Sinkhorn-optimized OT anomaly suppression, and is compatible with PatchCore, FastRecon, and WinCLIP variants. GraphCore (Xie et al., 2023) enforces rotation invariance through visual isometric invariant features computed via learned graph neural nets, compressing memory via coreset sampling.
• Subspace Modeling: SubspaceAD fits a PCA to aggregated normal patch features (with rotations for data augmentation), using orthogonal projection residuals as statistically grounded scores, exploiting the near-linear structure of DINOv2 patch features.
• Large Vision-LLMs: FADE uses ensembles of CLIP prompts and multi-scale embeddings (with GEM blocks) for both language- and vision-guided scoring, while IADGPT employs a progressive training curriculum on industrial data, in-context few-shot adaptation, and interpretable reasoning outputs.
• Bayesian Meta-Learning: BayPrAnoMeta’s Bayesian formalism allows for well-posed adaptation and heavy-tailed scoring in extreme few-shot, federated, and heterogeneous settings, leveraging NIW conjugacy for analytic posterior updates.
• Adversarial Feature Alignment: Integrating adversarial loss in Siamese or reconstruction-based architectures yields feature robustness in the few-shot regime by explicitly aligning source and target distributions (Lee et al., 2023).

4. Empirical Performance and Comparative Results

On core industrial benchmarks—MVTec-AD and VisA—state-of-the-art FS-IAD approaches achieve competitive image- and pixel-level AUROC and PRO scores even in the extreme 1-shot regime. Representative numbers:

Method	MVTec-AD I-AUROC	MVTec P-AUROC/PRO	VisA I-AUROC	VisA P-AUROC/PRO
VisionAD (Wang et al., 16 Apr 2025)	97.4%	99.0 / 92.5	94.8%	97.6
ABounD (Deng et al., 27 Nov 2025)	94.8%	-- / 91.1	87.3%	-- / 97.4
FoundAD (Zhai et al., 2 Oct 2025)	96.1%	96.8	92.6%	99.7
SubspaceAD (Lendering et al., 26 Feb 2026)	98.0%	97.6	93.3%	98.3
GATE-AD (Psiris et al., 16 Mar 2026)	97.7%	--	93.5%	--
H2VLR (Huang et al., 16 Apr 2026)	84.9%	--	--	--

Several methods report saturation of pixel-level metrics (pixel-AUC, PRO) at $c \in \mathcal{C}$096% under 1- or 2-shot settings for standard textures/objects; architectural improvements shift the SOTA by 1–3 AUC points per generation.

Scaling to larger, more challenging benchmarks (e.g. Real-IAD Variety, 160 classes), vision-language foundation models maintain robust performance as class diversity increases, with only minor drops in I-AUROC. One- or two-shot adaptation (e.g., via AdaptCLIP) achieves nearly the same performance as full-shot baselines (Zhu et al., 1 Nov 2025).

Ablation studies uniformly confirm that boundary-shaping (e.g., adversarial forging, prototype refinement, OT) and feature manifold regularization (e.g., subspace projection, rotation invariance) are critical for few-shot robustness, with each core regularizer removing 0.5–2% from AUROC when ablated (Deng et al., 27 Nov 2025, Tian et al., 26 Jun 2025, Lendering et al., 26 Feb 2026).

5. Challenges, Pitfalls, and Open Research Directions

While state-of-the-art FS-IAD approaches achieve strong AUROC/PRO metrics under academic conditions, challenges remain:

• Boundary Ambiguity and Out-of-Distribution Sensitivity: Very subtle surface defects, low-contrast anomalies, or logical/structural defects may remain within the few-shot normal manifold, leading to missed detections (Deng et al., 27 Nov 2025).
• Viewpoint and Background Sensitivity: Severe viewpoint changes degrade nearest-neighbor models; spatial misalignment can cause even DINO-type features to confound (Zhai et al., 2 Oct 2025, Wang et al., 16 Apr 2025).
• Memory and Computational Overhead: Some methods (especially memory-bank or VLM architectures) remain expensive in GPU memory or runtime; scalable memory management, efficient inference, and low latency are active areas of optimization (Psiris et al., 16 Mar 2026).
• Augmentation Strategy Tuning: No universal data augmentation exists; algorithm-specific choices must be made, with PatchCore and similar methods benefiting the most from mixed augmentation (Zhang et al., 2023).
• Multi-class and Multimodal Scenarios: Extracting structural commonality in the presence of multimodal support (RGB, 3D) is still an active research challenge, with methods like CIF (Lin et al., 8 Nov 2025) and H2VLR (Huang et al., 16 Apr 2026) providing principled, yet open-ended, frameworks.
• Reasoning and Human-Interpretable Explanations: Integrating reasoning and attribute-level diagnosis is nascent; progressive VLM curricula (IADGPT (Zhao et al., 14 Aug 2025)) offer promising avenues for direct accommodation of industrial logic.

6. Practical Guidelines and Deployment Considerations

Pragmatic recommendations for FS-IAD deployment are summarized from empirical studies:

• Training-free pipelines (VisionAD, SubspaceAD) are performant and highly attractive for practical adaptation, as they obviate retraining.
• Prototype refinement and memory compression (FastRef, GraphCore) substantially reduce inference costs and memory while improving discriminative power.
• Hyperparameter Sensitivity: Optimal configuration (e.g., number of augmentation rotations, memory size, number of hypergraph layers) is method- and dataset-dependent.
• Class Scaling: Foundation VLM/adaptor-based FS-IAD systems generalize across hundreds of classes with only 1–2 normals per class (Zhu et al., 1 Nov 2025), making them suitable for large, dynamic manufacturing lines.
• Latency: Runtime-optimized GNN or subspace models (GATE-AD, SubspaceAD) achieve $c \in \mathcal{C}$1 ms per image on modern hardware, supporting real-time industrial QC (Psiris et al., 16 Mar 2026).
• Failure Detection and Further Learning: Detection of extreme out-of-distribution or logical anomalies may require continual learning, active adaptation, or richer data (including synthetic/simulated anomalies).

7. Directions for Future FS-IAD Research

Emerging directions, as outlined across foundational works, include:

• Multimodal and 3D Data: Generalizing hypergraph and message-passing approaches to point cloud, infrared, or X-ray modalities (Lin et al., 8 Nov 2025, Huang et al., 16 Apr 2026).
• Meta-Learning and Federated Adaptation: Combining federated Bayesian meta-learners for heterogeneous, cross-plant industrial conditions (Sarkar et al., 27 Jan 2026).
• Automatic Augmentation and Dynamic Manifold Expansion: Data-driven augmentation policy search for robust patch manifold coverage (Zhang et al., 2023).
• Integrating Reasoning and Human Knowledge: Unified VLMs capable of simultaneous detection and attribute-level explanation or identification of defect cause (Zhao et al., 14 Aug 2025).
• Continual and Incremental Few-Shot Adaptation: Engineering pipelines that allow online updating of the normal manifold as new few-shot data streams arrive.

A plausible implication is that as foundation models continue to improve in semantic robustness and coverage, the distinction between zero-shot and few-shot IAD may blur further, with simple adaptation techniques sufficing for robust industrial anomaly localization and classification at unprecedented scales.