Organizations: the National Center of Technology Innovation for Intelligent Design and Numerical Control, Huazhong University of Science and Technology, Wuhan 430074, China · the Department of Intelligent Manufacturing, Contemporary Amperex Technology Ltd., Ningde 352000, China · the Institute for Infocomm Research, A*STAR, Singapore 138632 · the School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 639798 · the School of Artificial Intelligence and Robotics, Hunan University, Changsha 410082, China
Industrial 3D anomaly detection performance is fundamentally constrained by the scarcity and long-tailed distribution of abnormal samples. To address this challenge, we propose Synthesis4AD, an end-to-end paradigm that leverages large-scale, high-fidelity synthetic anomalies to learn more discriminative representations for 3D anomaly detection. At the core of Synthesis4AD is 3D-DefectStudio, a software platform built upon the controllable synthesis engine MPAS, which injects geometrically realistic defects guided by higher-dimensional support primitives while simultaneously generating accurate point-wise anomaly masks. Furthermore, Synthesis4AD incorporates a multimodal large language model (MLLM) to interpret product design information and automatically translate it into executable anomaly synthesis instructions, enabling scalable and knowledge-driven anomalous data generation. To improve the robustness and generalization of the downstream detector on unstructured point clouds, Synthesis4AD further introduces a training pipeline based on spatial-distribution normalization and geometry-faithful data augmentations, which alleviates the sensitivity of Point Transformer architectures to absolute coordinates and improves feature learning under realistic data variations. Extensive experiments demonstrate state-of-the-art performance on Real3D-AD, MulSen-AD, and a real-world industrial parts dataset. The proposed synthesis method MPAS and the interactive system 3D-DefectStudio will be publicly released at https://github.com/hustCYQ/Synthesis4AD.
Figures & tables
TABLE I: Comparison of supported anomaly categories between current 3D anomaly synthesis methods and our MPAS method. MPAS enables diverse and realistic defect generation across multiple dimensions of primitives.
Fig. 1: Comparison between previous methods and our method in 3D anomaly synthesis. In (b), from top to bottom, the rows illustrate anomalies synthesized based on 1D, 2D, and 3D primitives, respectively.
Fig. 2: MPAS framework leverages different dimensional primitives to automatically synthesize massive, realistic, and diverse 3D anomalous data.
Fig. 3: Overview of the proposed Synthesis4AD system. Stage I parses product-side knowledge, including expert priors, multi-view cues, and textual specifications, into executable synthesis instructions via an MLLM, and drives 3D-DefectStudio to inject controllable anomalies into large-scale normal 3D assets. Stage II trains the anomaly detector using the generated anomalous samples and their ground-truth masks. Stage III deploys the trained model for prototype-based online inference, producing both point-wise anomaly maps and object-wise anomaly scores from scanned test data.
Fig. 4: Visualization of anomalies. From top to bottom: real anomalies, Synthesized anomalies by MPAS with the same types, and two rows of more diverse compound anomalies synthesized by MPAS. Red insets highlight defect regions for detailed comparison.
Fig. 5: t-SNE visualization of feature distributions. Normal samples and real anomalies are compared with synthetic anomalies generated by MPAS, GLFM, and R3D-AD.
Method →
CPMF
R3D-AD
R3D-AD
Reg3D-AD
Group3AD
IMRNet
ISMP
PO3AD
MC3D-AD
GLFM
Synthesis4AD
Category ↓
PR’24
ECCV’24
+MPAS
NeurIPS’23
ACM MM’24
CVPR’24
AAAI’25
CVPR’25
IJCAI’25
TASE’25
Ours
Airplane
63.2/61.8
76.5/61.4
81.3/61.4
71.6/63.1
74.4/63.6
76.2/-
85.8 / 75.3
80.4/-
85.0 /62.8
54.6/74.3
62.1/ 83.6
Car
51.8/83.6
68.2/62.0
74.4/62.0
69.7/71.8
72.8/74.5
71.1/-
73.1/83.6
65.4/-
74.9/81.9
84.2 / 88.2
96.3 / 95.7
Candybar
71.8/73.4
66.4/61.2
72.0/61.2
82.7/72.4
84.7 /73.8
75.5/-
85.2 / 90.7
78.5/-
83.0/ 91.0
71.5/79.7
79.8/87.6
Chicken
64.0/55.9
69.7/53.9
69.7/53.9
85.2 /67.6
78.6 /75.9
78.0/-
71.4/ 79.8
68.6/-
71.5/64.0
68.8/62.4
74.2/ 78.5
Diamond
64.0/75.3
61.3/50.3
79.2/51.7
90.0/83.5
93.2/86.2
90.5/-
94.8 /92.6
80.1/-
95.5 / 94.2
71.2/76.8
76.3/ 95.3
TABLE II: Quantitative Results on Real3D-AD. The results are presented in O-ROC%/P-ROC%. The best performance is in bold , and the second best is underlined .
Method →
PatchCore-FP
PatchCore-FP-R
PatchCore-PM
Reg3D-AD
R3D-AD
R3D-AD
IMRNet
GLFM
Synthesis4AD
Category ↓
CVPR’22
CVPR’22
CVPR’22
NeurIPS’23
ECCV’24
+MPAS
CVPR’24
TASE’25
Ours
Capsule
89.8/91.7
90.5/91.9
90.3/ 92.1
91.2 /87.7
78.1/59.8
78.1/59.8
60.1/42.3
96.7 / 93.0
91.2 /86.0
Cotton
25.3/55.4
26.3/54.6
19.7/52.8
43.0/52.1
79.0/38.3
87.5 /45.9
58.5/50.7
81.2/ 67.9
99.8 / 75.8
Cube
72.3/57.5
66.8/43.7
72.2/41.7
56.9/62.6
73.4/60.5
73.4/60.5
43.2/56.6
75.6 / 68.0
90.2 / 80.5
Spring pad
98.6/62.9
100 /60.1
96.5/62.1
95.1/ 80.2
44.4/49.5
82.6/59.7
65.1/40.1
100 /70.5
100 / 77.7
Screw
97.9/57.8
93.1/61.0
99.7 /59.7
97.2/54.0
56.2/49.7
80.2/ 61.3
74.2/45.6
63.6/60.9
99.3 / 66.3
TABLE III: Quantitative Results on MulSen-AD. The results are presented in O-ROC%/P-ROC%. The best performance is in bold , and the second best is underlined .
Fig. 6: Qualitative comparison of on representative categories. From top to bottom: input point clouds ( P ), ground-truth masks (GT), and predicted anomaly maps produced by R3D-AD, GLFM, and our Synthesis4AD. Warmer colors indicate higher anomaly scores.
Fig. 7: Feature distribution visualization of MC3D-AD, GLFM, and Synthesis4AD. All methods utilize the identical PointMAE backbone.
MPAS-1D
MPAS-2D
MPAS-3D
Real3D-AD
MulSen-AD
✓
✗
✗
75.6/79.4
81.3/68.3
✗
✓
✗
73.3/76.2
80.8/65.1
✗
✗
✓
77.6/81.8
84.6/69.9
✓
✓
✗
77.2/80.5
83.7/69.4
✓
✗
✓
78.4/83.2
86.2/70.8
✓
✓
✓
80.9/84.8
89.6/72.0
TABLE IV: Ablation results on the impact of different anomaly types used during training.
Fig. 8: Effect of synthetic anomaly scale on detection performance. O-ROC and P-ROC on Real3D-AD and MulSen-AD are reported as the number of synthesized anomalous training samples increases. The dashed lines indicate the previous SOTA results.
SDN
Rotations
Perturbations
Dropout
Real3D-AD
MulSen-AD
✗
✗
✗
✗
74.7/77.5
82.4/65.2
✓
✗
✗
✗
78.3/82.0
86.5/68.8
✓
✓
✗
✗
79.0/83.1
87.4/70.7
✓
✓
✓
✗
80.1/84.2
88.9/71.4
✓
✓
✗
✓
79.6/83.3
87.7/71.2
✓
✓
✓
✓
80.9/84.8
89.6/72.0
TABLE V: Ablation results on the impact of SDN and different data augmentation.
Fig. 9: Real-world industrial data collection and representative defect examples. (a) The 3D scanning setup is utilized to capture point clouds of industrial parts. (b) Abnormal samples from six object categories; red boxes highlight defective position.
Method →
R3D-AD
MC3D-AD
GLFM
Synthesis4AD
Category ↓
ECCV’24
IJCAI’25
TASE’25
Ours
Bevel Gear
63.7/42.5
80.7/62.1
83.7/64.1
100 / 72.1
Brake Caliper
64.0/47.7
83.0/65.6
83.0/66.6
93.7 / 70.1
Connecting Rod
56.8/54.2
80.0/63.7
68.0/66.3
97.3 / 72.9
Crankshaft
44.3/53.1
91.3/69.5
81.3/65.8
87.7 / 69.1
Piston
65.3/60.0
79.7/61.2
82.3/66.7
96.7 / 75.9
TABLE VI: Quantitative Results on Actual Industry Parts Dataset. The results are presented in O-ROC%/P-ROC%. The best performance is in bold .
Fig. 10: Visualization of prediction results in the actual industry parts dataset using the proposed method and other methods. The first column is the original point clouds, while the second column is the ground truth. Subsequent columns depict various methods.
Point cloud anomaly detection identifies and localizes unusual 3D geometry or surface defects, often in industrial inspection with few labeled failures.
Detecting and localizing defects in 3D point clouds is challenging because abnormal samples are scarce and diverse, while training is often limited to normal data. We propose Anomaly Factory 3D (AF3AD), a modular framework that synthesizes diverse pseudo-anomalies from normal point clouds to expand the training data for unsupervised 3D anomaly detection methods that rely on pseudo-anomalies. AF3AD uses a center-conditioned parametric deformation model defined in local PCA frames, with kernel-controlled spatial falloff, anisotropy, directional gating, and normal/tangential displacement fields, enabling a broad set of geometric defect presets. We demonstrate its ease-of-use and effectiveness by integrating AF3AD with an offset-prediction detector and a reconstruction-based anomaly detection method, showing that AF3AD transfers across detection paradigms. Experiments on AnomalyShapeNet and Real3D-AD show consistent improvements in object- and point-level detection and localization, supported by ablations on preset groups and robustness under noise. AF3AD is designed as a standalone synthesis tool to facilitate adoption across different 3D anomaly detection paradigms. Code is available at github.com/vpc-ccg/AF3AD.
Ali Balapour, Faraz Hach
University of British Columbia, Vancouver, V6T 1Z4 BC, Canada · Vancouver Prostate Centre, M.H. Mohseni Institute of Urologic Sciences, Vancouver, V6H 3ZB BC, Canada · Department of Urologic Sciences, The University of British Columbia, Vancouver, V5Z 1M9 BC, Canada
3D point cloud anomaly detection plays a vital role in industrial manufacturing, yet it faces significant challenges due to the scarcity and high acquisition cost of real anomalous samples. The inherently anomaly-free training data further hinders detection methods from effectively learning discriminative features between normal and abnormal instances. To address these issues, we propose PA3AD, a novel framework that introduces a physics-inspired pseudo-anomaly generation strategy to create physically plausible anomalous samples from normal data. Additionally, we incorporate prototype features via a weight-sharing mechanism to guide the model in capturing the distribution shifts between normal and anomalous samples. Specifically, PA3AD introduces two key innovations to tackle the scarcity of real anomalies. First, a physics-inspired module generates diverse pseudo-anomalous point clouds from normal data via multi-physics modeling. Second, momentum-updated prototypes and a difference-aware fusion block capture stable normal representations and their discrepancies with pseudo-anomalies. This design effectively learns distribution shifts, achieving superior detection performance. Extensive experiments on the Anomaly-ShapeNet and Real3D-AD datasets demonstrate that our method consistently outperforms existing state-of-the-art approaches. Our code will be made publicly available at https://github.com/NingxiaoJian/PA3AD.
Jian Ning, Qin Zou, Linchun Wu +4
School of Computer Science, Wuhan University, Wuhan 430072, Hubei, China · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen 518060, Guangdong, China · School of Information and Software Engineering, East China Jiaotong University, Nanchang 330013, Jiangxi, China +1
Detecting and localizing 3D point-cloud defects is essential for industrial inspection. However, existing methods often suffer from imprecise localization due to the lack of anomaly supervision and reliance on single-granularity representations. To address these limitations, we propose Anomaly Type-Aware 3D Anomaly Detection (AT3D-AD), a unified framework for joint detection, localization, and classification. Specifically, we first design the Physics-Driven Parametric Anomaly Synthesis (PDPAS) module employing multiple parametric functions to generate synthetic anomalies, providing explicit anomaly supervision. Then, we propose the Hierarchical Global-Local Anomaly Alignment (HiGLA) module to align global and local representations within the normal and anomalous groups. Finally, we propose the Semantic-Geometric Anomaly Classification (SGAC) module to jointly learn localization and classification, yielding spatially precise and type-discriminative anomaly representations. Extensive experiments establish new state-of-the-art performance on all four benchmarks. AT3D-AD achieves Object/Point AUROC scores of 98.1%/98.9% on Anomaly-ShapeNet and 95.0%/95.2% on Real3D-AD, while reaching 74.2% Macro-F1 for anomaly-type recognition on Real3D-AD.
Jingyu Zeng, Haoquan Lu, Can Gao
College of Computer Science and Software Engineering and the Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen University, Shenzhen 518060, China