MuSc-V2: Zero-Shot Industrial Anomaly Detection

• The paper introduces MuSc-V2, a training-free, zero-shot method for industrial anomaly classification and segmentation using mutual scoring among unlabeled samples.
• It employs specialized modules—IPG, SNAMD, MSM, CAE, and RsCon—to process 2D images and 3D point clouds, achieving up to +23.7% AP improvement and 5.6× faster performance.
• MuSc-V2 capitalizes on the cross-sample repetition of normal patches to effectively distinguish isolated, diverse anomalies without any labeled data or prompt guidance.

MuSc-V2 is a training-free, zero-shot framework for industrial anomaly classification (AC) and anomaly segmentation (AS) that operates on 2D images, 3D point clouds, or both jointly by using mutual scoring among unlabeled samples rather than labeled normals, prompts, or target-domain training (Li et al., 13 Nov 2025). Its central premise is that normal patches are highly repetitive across industrial products and therefore can find many similar patches among unlabeled test samples, whereas anomalies are rare, diverse, and isolated. On that basis, MuSc-V2 constructs patch representations, performs mutual scoring within each modality, enhances anomaly evidence across modalities when both are present, and refines sample-level classification through constrained neighborhood re-scoring. The framework is positioned as the first multimodal zero-shot industrial AC/AS method without prompts or training, and it extends the earlier MuSc paradigm from 2D test-set mutual scoring to 2D, 3D, and multimodal settings (Li et al., 13 Nov 2025).

1. Problem setting and conceptual basis

MuSc-V2 targets zero-shot industrial anomaly classification and segmentation. In this setting, AC determines whether an entire product or sample is anomalous, whereas AS localizes anomalous regions, pixels, or points. The framework is designed for 2D-only datasets such as MVTec AD and VisA, 3D-only settings, and multimodal 2D+3D datasets such as MVTec 3D-AD and Eyecandies (Li et al., 13 Nov 2025).

The method departs from two common zero-shot strategies. First, it does not compare test samples to labeled normal references. Second, it does not rely on text prompts or image-text alignment. Instead, it uses only unlabeled samples from the target industrial dataset. In the mutual scoring procedure, each sample is scored against the remaining unlabeled samples, denoted as $\{\mathcal{D} \setminus O_i\}$ for scoring $O_i$ (Li et al., 13 Nov 2025).

The empirical assumption underlying the framework is explicit: normal patches are highly self-consistent and can find many similar patches among the unlabeled test samples, while anomalous patches are diverse and therefore fail to find many strong matches. The paper grounds this in industrial data regularity and reports that most primitives are normal, including around $99.71\%$ normal primitives in MVTec 3D-AD, $99.77\%$ in Eyecandies 2D, and $98.57\%$–$99.31\%$ normal 3D points (Li et al., 13 Nov 2025). This distribution makes cross-sample recurrence a practical signal of normality.

A common misunderstanding is to treat MuSc-V2 as a conventional nearest-neighbor detector over test images. The framework is more structured: it introduces Iterative Point Grouping (IPG), Similarity Neighborhood Aggregation with Multi-Degrees (SNAMD), the Mutual Scoring Mechanism (MSM), Cross-modal Anomaly Enhancement (CAE), and Re-scoring with Constrained Neighborhood (RsCon), each of which addresses a distinct failure mode in zero-shot anomaly analysis (Li et al., 13 Nov 2025).

2. Relation to MuSc and the evolution of mutual scoring

MuSc-V2 inherits its basic paradigm from the earlier "MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images" (Li et al., 2024). The original MuSc established the core observation that, in industrial datasets, normal image patches tend to have many visually similar counterparts in other unlabeled images, while abnormal patches have few such matches. It operationalized that idea through Local Neighborhood Aggregation with Multiple Degrees (LNAMD), a Mutual Scoring Mechanism (MSM), and Re-scoring with Constrained Image-level Neighborhood (RsCIN), all in a test-set-only, prompt-free, zero-shot 2D framework (Li et al., 2024).

MuSc-V2 preserves the mutual-scoring logic while extending it substantially. It generalizes from 2D image-only anomaly analysis to single-modality 2D, single-modality 3D, and multimodal 2D+3D operation. It replaces LNAMD with SNAMD, introduces IPG to improve 3D point grouping, adds CAE to fuse cross-modal anomaly evidence, and uses RsCon for classification refinement. The paper also reports that MuSc-V2 is $5.6\times$ faster than the original MuSc (Li et al., 13 Nov 2025).

The conceptual continuity is strong. In both formulations, the anomaly score of a local region is derived from how well that region is supported by similar regions in other unlabeled samples. The main innovation of MuSc-V2 is not abandonment of mutual scoring but its extension to multimodal industrial data, together with additional mechanisms for geometric coherence, similarity-aware neighborhood aggregation, and cross-modal recovery of missed anomalies (Li et al., 2024).

A plausible implication is that MuSc-V2 should be understood less as a single detector than as a modular family of test-time scoring procedures built around the same statistical asymmetry: normal structure is recurrent, anomalies are isolated.

3. Architecture and core modules

MuSc-V2 consists of four major stages: 2D/3D patch representation, SNAMD feature aggregation, MSM plus CAE, and RsCon for sample-level classification refinement (Li et al., 13 Nov 2025).

Module	Primary role	Main issue addressed
IPG	Surface-consistent 3D grouping	False positives from discontinuous surfaces
SNAMD	Multi-scale similarity-aware aggregation	Variable anomaly size and feature contamination
MSM	Cross-sample patch scoring	Zero-shot anomaly evidence from unlabeled data
CAE	Cross-modal score fusion	Modality-specific missed anomalies
RsCon	Sample-level score refinement	Fragility of max-based AC scores

Iterative Point Grouping

IPG addresses a specific defect of standard 3D grouping based on FPS plus KNN. The paper notes that KNN can merge points from discontinuous surfaces into one group, which can make a normal region appear anomalous. IPG begins with KNN grouping, computes curvature at the group center $p_c$, and, if curvature exceeds the threshold $\mathcal{C}_{thr}$, re-groups iteratively from a small seed group $\mathcal{L}^0$ of $O_i$0 nearest points. Candidate expansion uses

$O_i$1

followed by

$O_i$2

This produces more geometrically coherent 3D tokens and reduces false positives from discontinuous surfaces (Li et al., 13 Nov 2025).

Similarity Neighborhood Aggregation with Multi-Degrees

SNAMD is the feature aggregation stage. For each patch, neighborhoods are formed in 2D by reshaping tokens into a spatial grid and extracting an $O_i$3 neighborhood, and in 3D by Euclidean proximity in patch space. Rather than uniform averaging, SNAMD uses Similarity-Weighted Pooling (SWPooling):

$O_i$4

$O_i$5

For 2D, the degrees are $O_i$6. For 3D, high-curvature patches are restricted to $O_i$7 to preserve surface consistency. The purpose is to retain sensitivity to both small and large anomalies while reducing contamination from dissimilar neighbors (Li et al., 13 Nov 2025).

Mutual Scoring Mechanism

MSM is the core scoring engine. For a 2D patch $O_i$8 in image $O_i$9, the score assigned by another image $99.71\%$0 is

$99.71\%$1

The set of scores over all other unlabeled images is

$99.71\%$2

The same construction applies to 3D patches. To suppress outliers, MuSc-V2 applies Interval Average (IA), keeping only the lowest $99.71\%$3 of scores and averaging them:

$99.71\%$4

After IA, stage-wise scores are averaged:

$99.71\%$5

The resulting patch-wise anomaly maps encode low scores for recurrent, well-supported patches and high scores for isolated patches (Li et al., 13 Nov 2025).

Cross-modal Anomaly Enhancement

CAE is active in multimodal 2D+3D operation. It projects 3D patch scores into 2D patch space through camera parameters and point coordinates, aligns 3D anomaly evidence to each 2D patch, rescales the projected scores to the 2D score range, and fuses them by

$99.71\%$6

with

$99.71\%$7

The $99.71\%$8 term preserves the stronger anomaly signal from either modality, while $99.71\%$9 downweights unreliable cross-modal enhancement when the projected scores are inconsistent (Li et al., 13 Nov 2025).

Re-scoring with Constrained Neighborhood

For AC, MuSc-V2 begins with a max-based sample score, which is sensitive to local noise. RsCon refines this by constructing anomaly-salient sample features from the highest-scoring patch and building a graph with similarity weights

$99.77\%$0

A binary window mask

$99.77\%$1

restricts smoothing to nearby representative samples. The updated score vector is

$99.77\%$2

which the paper also expresses per sample as

$99.77\%$3

This blends a sample’s own score with a weighted neighborhood average and reduces false classification caused by noisy maxima (Li et al., 13 Nov 2025).

4. Operating modes, data flow, and zero-shot execution

MuSc-V2 is explicitly modular across modalities. In 2D-only mode, it uses 2D ViT features, applies SNAMD over 2D neighborhoods, runs MSM on 2D patches, and uses RsCon for classification refinement; CAE is absent because 3D is absent. In 3D-only mode, it uses point-cloud features from Point Transformer, preprocesses them with IPG, applies SNAMD in 3D patch space, runs MSM on 3D patches, and uses RsCon without CAE. In multimodal 2D+3D mode, it extracts both modalities, applies IPG to 3D, runs SNAMD and MSM independently in each modality, uses CAE for cross-modal anomaly recovery, fuses segmentation maps as

$99.77\%$4

and performs RsCon on concatenated sample features (Li et al., 13 Nov 2025).

The framework is training-free in a strict sense. It uses no anomaly labels, no training on the target dataset, no fine-tuning on normal samples, and no prompt engineering. The components that are used are pretrained backbones: DINO ViT-B-8 for most multimodal experiments, CLIP ViT-L-14-336 for comparison or fair 2D settings, and Point Transformer pretrained with Point-MAE for 3D (Li et al., 13 Nov 2025).

The implementation details reported in the paper fix several important hyperparameters: both ViT and Point Transformer are divided into $99.77\%$5 stages ($99.77\%$6), images are resized to $99.77\%$7, point clouds are clustered into $99.77\%$8 groups of $99.77\%$9 points each, the IPG iterative increment is $98.57\%$0, the curvature threshold is $98.57\%$1, SNAMD uses $98.57\%$2, MSM uses a minimum $98.57\%$3 IA interval, and RsCon uses window size $98.57\%$4 (Li et al., 13 Nov 2025).

The paper also evaluates robustness under reduced data pools by splitting the full dataset into $98.57\%$5 or $98.57\%$6 subsets and performing mutual scoring within each subset. Performance drops only slightly. This suggests that the framework does not require the full unlabeled set to retain most of its effectiveness, although its scoring principle remains dependent on cross-sample redundancy (Li et al., 13 Nov 2025).

5. Empirical performance and ablation evidence

MuSc-V2 reports its headline gains in multimodal segmentation. On MVTec 3D-AD, it achieves a $98.57\%$7 AP gain over the second-best zero-shot baseline; on Eyecandies, it reports a $98.57\%$8 AP gain. The paper also reports AC AUROC gains of $98.57\%$9 on MVTec 3D-AD and $99.31\%$0 on Eyecandies, and states that the method often exceeds some few-shot approaches despite using no training (Li et al., 13 Nov 2025).

On 2D datasets such as MVTec AD and VisA, MuSc-V2 improves over zero-shot 2D CLIP-based methods and in some cases slightly outperforms the prior MuSc version, especially on segmentation. The paper therefore presents the multimodal extension not merely as broader in scope but also as beneficial even when evaluated against strong 2D zero-shot baselines (Li et al., 13 Nov 2025).

The ablation studies identify the contribution of each major component. IPG improves classification more than segmentation, which is consistent with its stated role of removing geometrically implausible but actually normal 3D patches. SNAMD benefits from multi-degree aggregation, and removing SWPooling causes substantial performance drops, especially on Eyecandies where anomalies are often small. IA reduces overlap between normal and abnormal score distributions by suppressing outlier scores. CAE recovers anomalies missed by one modality, while the variance-based $99.31\%$1 improves robustness by limiting unreliable cross-modal boosts. RsCon consistently improves AC metrics, and window masking is reported as critical; performance is fairly stable for reasonable $99.31\%$2 values around $99.31\%$3 (Li et al., 13 Nov 2025).

These results support the framework’s two principal claims. First, unlabeled target-domain data can function as an internal reference set for anomaly scoring. Second, multimodal anomaly analysis benefits from explicit mechanisms that preserve 3D surface consistency and exchange anomaly evidence across modalities rather than merely concatenating features.

6. Positioning, limitations, and interpretive significance

MuSc-V2 is positioned against both prior zero-shot and few-shot industrial anomaly methods. Prior zero-shot methods are described as relying on CLIP-like image-text alignment, prompt tuning or fine-tuning, text descriptions, pretrained normal prototypes, or, in 3D, multi-view rendering instead of native point-cloud processing. Few-shot methods such as BTF, M3DM, and CFM require labeled normal samples or some training process. MuSc-V2 claims novelty in combining zero-shot, prompt-free, training-free operation with support for 2D, 3D, and multimodal anomaly analysis in a single framework (Li et al., 13 Nov 2025).

Two clarifications are important. First, MuSc-V2 is not a supervised multimodal detector adapted to a zero-shot label regime; it is a test-time framework that exploits recurrence statistics within the unlabeled target set. Second, it is not simply the 3D or multimodal analogue of CLIP-based prompting. The backbone features are pretrained, but anomaly evidence is derived from mutual agreement and disagreement among unlabeled samples, not from text-conditioned semantics (Li et al., 13 Nov 2025).

The paper’s reported limitations are primarily structural rather than conceptual. Because the method relies on mutual scoring among unlabeled samples, its effectiveness is tied to the presence of many normal patches and to cross-sample consistency. The dataset-splitting experiment indicates robustness to smaller pools, but not independence from sample recurrence. A plausible implication is that the framework is best matched to production settings where normal structures are strongly repeated across items.

In historical terms, MuSc-V2 formalizes a broader version of the argument introduced by MuSc: unlabeled test data are not merely evaluation inputs but can serve as the reference structure for zero-shot anomaly inference (Li et al., 2024). Its contribution is to extend that argument to multimodal industrial inspection through IPG, SNAMD, CAE, and RsCon, while retaining a fully training-free and prompt-free operating regime (Li et al., 13 Nov 2025).