Synthesis4AD: 3D Anomaly Detection

• Synthesis4AD is an end-to-end paradigm for industrial 3D anomaly detection that uses synthetic supervision to mimic realistic defect patterns.
• It integrates 3D-DefectStudio, MPAS, and a Point Transformer detector to generate synthetic anomalies with explicit masks for robust segmentation training.
• The framework achieves state-of-the-art performance by combining geometry-aware synthesis with prototype-based inference and advanced normalization techniques.

Synthesis4AD is an end-to-end paradigm for industrial 3D anomaly detection that treats synthetic supervision as the primary mechanism for overcoming the scarcity and long-tailed distribution of real defective samples. It combines large-scale, high-fidelity anomaly synthesis, knowledge-driven instruction generation, and a detector-training pipeline tailored to unstructured point clouds. Its core components are 3D-DefectStudio, a software platform built upon the controllable synthesis engine MPAS, and a Point Transformer–based downstream detector trained on synthetic anomaly–mask pairs. The framework is reported to achieve state-of-the-art performance on Real3D-AD, MulSen-AD, and a real-world industrial parts dataset (Sun et al., 6 Apr 2026).

1. Problem formulation and conceptual basis

Synthesis4AD is motivated by a structural limitation of industrial 3D anomaly detection: real defective samples are scarce and follow a long-tailed distribution. In that setting, training purely on normal data yields representations that generalize poorly to unseen structural defects. The framework therefore departs from unsupervised normal-only paradigms and instead uses synthetic anomalous data as direct supervision for representation learning.

The paper also positions prior 3D anomaly-synthesis methods as inadequate because they perturb point clouds via low-dimensional primitives such as points or curves, producing “toy” anomalies including small protrusions or scratches. According to this formulation, such perturbations fail to model extended, realistic damage patterns. Synthesis4AD addresses that limitation by injecting large-scale, geometry-aware synthetic anomalies intended to close the train–test distribution gap and to yield more discriminative 3D representations.

At the system level, Synthesis4AD is organized into three tightly coupled stages. Stage I performs knowledge-driven synthesis from normal point-cloud assets and product-side knowledge. Stage II trains a segmentation-style anomaly detector on the synthesized anomalous samples and their point-wise masks. Stage III performs prototype-based inference, using features extracted from normal training samples to compute dense anomaly scores at test time. This decomposition is significant because it integrates data generation, detector optimization, and inference into a single pipeline rather than treating synthesis as a detached pre-processing step.

2. End-to-end architecture

The inputs to Stage I are normal 3D point-cloud assets $X$ and product-side knowledge

$$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$

A multimodal LLM $\Phi_{\psi}$ parses $\mathcal{Z}$ into an instruction tuple

$$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$

where $u$ denotes anomaly type $(1\text{D}, 2\text{D}, 3\text{D})$, $\mathcal{R}$ is a spatial region suggestion, $\mathcal{A}$ is the selected MPAS operator interface, and $\Theta$ contains operator parameters such as radius, depth, or rotation angle. A validator $$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$0 ensures the instruction is executable; invalid instructions are replaced with a rule-based fallback $$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$1.

3D-DefectStudio then invokes the MPAS engine $$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$2 to generate both the anomalous point cloud and the corresponding point-wise ground-truth mask:

$$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$3

This explicit mask generation is a defining property of the pipeline: synthetic data are not only visually plausible but also immediately usable for supervised localization training.

In Stage II, the detector uses a Point Transformer encoder and a lightweight MLP segmentation head to predict a binary anomaly mask $$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$4 for each synthesized $$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$5. Supervision is provided by a segmentation loss, exemplified in the paper by binary cross-entropy $$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$6. Stage III then discards synthetic supervision and returns to normal-data prototypes: point-wise features from normal training samples are clustered into a prototype set $$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$7, and test-time anomaly scores are computed as

$$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$8

A dense anomaly map is produced from these scores, while the object-wise score is defined as the mean of the top-$$\mathcal{Z} = \{\mathcal{I}_{\text{mv}}, \mathcal{T}_{\text{spec}}, \mathcal{E}_{\text{prior}}\}.$$9 point scores.

This architecture is technically notable because the training signal is synthetic and discriminative, whereas inference is prototype-based and anchored in normality. A plausible implication is that Synthesis4AD attempts to exploit the strengths of both supervised mask prediction and normal-feature reference modeling without requiring large real-defect corpora.

3. MPAS: multi-dimensional primitive-guided anomaly synthesis

MPAS, or Multi-dimensional Primitive-guided Anomaly Synthesis, is the synthesis engine underlying 3D-DefectStudio. Its central design choice is to formulate defects using 1D, 2D, and 3D support primitives, rather than relying only on point-level or curve-level perturbations. The result is a unified mechanism for generating simple protrusions, cracks and bends, and free-form surface distortions with accurate point-wise masks (Sun et al., 6 Apr 2026).

MPAS-1D

MPAS-1D produces point- or line-guided deformations. It selects ordered anchors

$\Phi_{\psi}$0

and constructs a geodesic skeleton on a $\Phi_{\psi}$1-NN graph $\Phi_{\psi}$2. For each consecutive anchor pair, it computes the shortest path

$\Phi_{\psi}$3

The mask region is obtained by radius expansion around the path union $\Phi_{\psi}$4:

$\Phi_{\psi}$5

After averaging normals inside $\Phi_{\psi}$6, each point is displaced according to

$\Phi_{\psi}$7

with $\Phi_{\psi}$8 and peak magnitude $\Phi_{\psi}$9. This yields a controlled local deformation and an explicit anomaly mask $\mathcal{Z}$0.

MPAS-2D

MPAS-2D operates on plane-guided bends and cracks. It defines a primitive plane

$\mathcal{Z}$1

with signed distance $\mathcal{Z}$2. An intersection band

$\mathcal{Z}$3

localizes the deformation. For bending, the hinge axis is estimated via PCA on $\mathcal{Z}$4, and each point is rotated with a position-dependent weight

$\mathcal{Z}$5

leading to

$\mathcal{Z}$6

with mask $\mathcal{Z}$7.

For cracking, the signed distance is perturbed as $\mathcal{Z}$8 with $\mathcal{Z}$9, points satisfying $$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$0 are removed, and the boundary mask is

$$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$1

This mechanism introduces discontinuities rather than merely deforming local neighborhoods.

MPAS-3D

MPAS-3D generates free-form surface distortions. It samples random anchors $$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$2, builds a convex hull $$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$3, and defines the support mask

$$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$4

A local frame is estimated by PCA on $$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$5, yielding tangent axes $$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$6 and centroid $$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$7. A height field over projected coordinates is then specified as

$$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$8

and deformation is applied along the local normal:

$$\mathbf{c} = \Phi_{\psi}(\mathcal{Z}) = \{u,\;\mathcal{R},\;\mathcal{A},\;\Theta\},$$9

A subsequent $u$0-NN smoothing step produces the final geometry:

$u$1

Taken together, these three operators formalize a progression from low-dimensional supports to extended surface- and volume-like deformations. This suggests that the realism claim of Synthesis4AD is rooted not only in appearance but also in the dimensionality and geometry of the defect support.

4. Knowledge-driven synthesis with 3D-DefectStudio and an MLLM

3D-DefectStudio is the software layer that operationalizes MPAS. It uses a multimodal LLM to interpret product design information and translate it into executable anomaly-synthesis instructions. The paper specifies the off-the-shelf Gemini 3 MLLM, which is prompted with multimodal cues $u$2 and produces structured instructions $u$3 without additional task-specific training or loss terms.

The validator $u$4 performs schema-checking and grounding to guarantee feasible instructions. If an instruction is invalid, it is replaced by rule-based templates $u$5. The role of this component is not cosmetic: it is what allows the “cognitive→geometric” translation described in the paper, where engineering guidance such as a desired bend location or deformation angle is converted into precise MPAS calls.

This MLLM-mediated interface matters because Synthesis4AD is presented as scalable across many product categories. The paper argues that the combination of the 3D-DefectStudio API and MLLM prompting can automatically generate tens of thousands of paired anomaly–mask samples. A plausible implication is that the bottleneck shifts from manual defect engineering toward specification design and validation, which is more compatible with heterogeneous industrial inventories.

5. Detector training, normalization, and inference

The detector-training stage is built around a Point Transformer encoder and a lightweight MLP segmentation head. The paper emphasizes that Point Transformer architectures are sensitive to absolute coordinates, and the training pipeline therefore introduces spatial-distribution normalization (SDN) and geometry-faithful augmentations to improve robustness on unstructured point clouds.

For each category $u$6, SDN computes a bounding sphere center $u$7 and radius $u$8, then normalizes points as

$u$9

The normalized point cloud is then voxel-downsampled with fixed voxel size $(1\text{D}, 2\text{D}, 3\text{D})$0, producing uniform point density across categories. The reported effect is removal of scale bias and reduced sensitivity of positional embeddings to global coordinates.

The geometry-faithful augmentations are applied on the normalized, downsampled cloud $(1\text{D}, 2\text{D}, 3\text{D})$1. They consist of random rotation to mitigate pose sensitivity, noise perturbation to simulate sensor uncertainty, and point dropout to mimic occlusion and missing data. The training objective is binary cross-entropy between predicted mask $(1\text{D}, 2\text{D}, 3\text{D})$2 and synthetic mask $(1\text{D}, 2\text{D}, 3\text{D})$3. Optimization uses AdamW with learning rate $(1\text{D}, 2\text{D}, 3\text{D})$4, batch size $(1\text{D}, 2\text{D}, 3\text{D})$5, and $(1\text{D}, 2\text{D}, 3\text{D})$6 training iterations. The paper further lists representative hyperparameters including voxel size $(1\text{D}, 2\text{D}, 3\text{D})$7 uniformly across all categories, for example $(1\text{D}, 2\text{D}, 3\text{D})$8 units; prototype count $(1\text{D}, 2\text{D}, 3\text{D})$9 for inference, for example $\mathcal{R}$0; and augmentation strengths including noise $\mathcal{R}$1, dropout ratio $\mathcal{R}$2, and rotation $\mathcal{R}$3.

Inference is prototype-based. Features extracted from normal training samples are clustered into prototypes, and each test-point anomaly score is the minimum distance to that prototype set. This design indicates that the model is not used purely as a discriminative segmenter at test time; instead, learned features are repurposed for nearest-prototype scoring. The paper’s qualitative description of the resulting maps is that they are compact and mask-aligned, with fewer false positives on smooth surfaces.

6. Empirical performance and ablation structure

The experimental evaluation covers Real3D-AD, MulSen-AD, and an additional real-world industrial parts dataset with six categories. The reported metrics are object-wise AUROC (O-ROC) for sample detection and point-wise AUROC (P-ROC) for localization quality (Sun et al., 6 Apr 2026).

Benchmark	Baselines	Synthesis4AD
Real3D-AD (mean O-ROC / P-ROC)	PatchCore-FP: 62.5 / 75.9; GLFM: 75.0 / 76.7	80.9 / 84.8
MulSen-AD (mean O-ROC / P-ROC)	PatchCore-FP: 86.0 / 64.0; GLFM: 78.5 / 66.5	89.6 / 72.0
Industry parts (mean O-ROC / P-ROC)	R3D-AD: 60.9 / 51.5; GLFM: 82.8 / 65.8	95.9 / 73.8

The ablation studies are organized around three questions: anomaly type, synthesis scale, and training-pipeline robustness. On Real3D-AD, the contribution of anomaly dimensionality is reported as follows: 1D only yields $\mathcal{R}$4, 2D only yields $\mathcal{R}$5, 3D only yields $\mathcal{R}$6, and the combined 1D+2D+3D configuration yields $\mathcal{R}$7. This result supports the claim that multi-dimensional synthesis is more effective than any single primitive family in isolation.

The scale study reports monotonic performance increases when raising the number of synthetic anomalies from $\mathcal{R}$8 on Real3D-AD and from $\mathcal{R}$9 on MulSen-AD. The robustness study on Real3D-AD reports the following sequence: base (no SDN/aug) $\mathcal{A}$0, +SDN only $\mathcal{A}$1, +SDN + rotation $\mathcal{A}$2, +SDN + rotation + noise $\mathcal{A}$3, +SDN + rotation + dropout $\mathcal{A}$4, and +all $\mathcal{A}$5. These numbers directly support the paper’s claim that SDN and geometry-faithful augmentations stabilize Point Transformer training and improve cross-category generalization.

Qualitatively, the paper states that MPAS reproduces irregular boundaries, heterogeneous bulges, and compound defect modes, and that t-SNE feature alignment places MPAS-generated anomalies closer to real-anomaly clusters than GLFM or R3D-AD. Within the article’s argument, these observations are intended to explain why synthetic supervision improves both localization and object-level discrimination rather than merely increasing training set size.

7. Significance, limitations, and scope

The principal significance of Synthesis4AD lies in its claim that synthetic anomalies are sufficient to achieve state-of-the-art 3D anomaly detection and localization, provided that synthesis is geometry-aware, mask-generating, and integrated with robust detector training. The paper summarizes this position through several “key insights and best practices”: higher-dimensional support primitives produce rich, realistic anomalies that more closely match real defect statistics; MLLM-driven instruction parsing automates knowledge-guided data generation across product categories; and SDN plus geometry-faithful augmentations reduce scale and pose sensitivity in Point Transformer training.

The framework is also presented as a practical deployment path. The proposed synthesis method MPAS and the interactive system 3D-DefectStudio are slated for public release at https://github.com/hustCYQ/Synthesis4AD. In the paper’s framing, this release is relevant because it exposes not only a trained detector but also the anomaly-generation machinery needed to scale supervision in industrial settings.

The main stated limitation is that the current workflow is open-loop: synthesis is not adaptively refined based on detector feedback. The paper identifies “closing the loop,” described as reinforcement of synthesis by detection performance, as a promising future direction. This suggests that Synthesis4AD should be understood less as a fixed generator–detector stack than as a foundation for iterative synthetic-supervision systems in which anomaly generation and anomaly detection may eventually co-evolve (Sun et al., 6 Apr 2026).