3D-DefectStudio: High-Fidelity 3D Anomaly Synthesis

• 3D-DefectStudio is an interactive, knowledge-driven platform that generates anomalous 3D point clouds with exact point-wise ground-truth masks.
• It integrates a modular MPAS engine with 1D, 2D, and 3D anomaly operators to produce realistic, geometry-aware defect deformations.
• The system scales anomaly synthesis through automated mask generation and executable geometric transformations, enhancing industrial 3D anomaly detection.

3D-DefectStudio is an interactive software platform and programmable Python API for the controllable synthesis of anomalous 3D point clouds with point-wise ground-truth masks, introduced as the practical synthesis component of Synthesis4AD and built on the Multi-dimensional Primitive-guided Anomaly Synthesis (MPAS) engine. Its stated purpose is to transform normal 3D assets into synthetic anomalous samples at scale, while preserving enough geometric realism to support industrial 3D anomaly detection. The system was motivated by a specific deficiency in earlier synthesis pipelines: many existing methods rely on simple low-dimensional primitives such as spheres, scratches, or localized point perturbations, and therefore produce defects that are too regular, too local, too easy to detect, and insufficiently faithful to real industrial failures (Sun et al., 6 Apr 2026).

1. System identity and architectural role

The paper distinguishes three layers. MPAS is the algorithmic backend that performs anomaly injection through geometry-aware operators. 3D-DefectStudio is the user-facing platform that exposes MPAS through interactive workflows, scripted synthesis APIs, reproducible operator interfaces, and automatic mask generation. Synthesis4AD is the full end-to-end anomaly-detection framework that combines knowledge-driven synthesis, supervised detector training on synthetic anomaly-mask pairs, and prototype-based inference (Sun et al., 6 Apr 2026).

This separation is technically important. 3D-DefectStudio is not merely a visualization tool, and it is not itself the complete anomaly detector. Rather, it operationalizes MPAS so that engineering knowledge and product-design information can be converted into executable geometric transformations on point clouds. In that sense, it is the system component that closes the gap between product semantics and synthetic data generation.

The motivation for such a platform is the scarcity and long-tailed distribution of real abnormal samples in industrial inspection. The paper treats this as a fundamental bottleneck for 3D anomaly detection, because real defects are often spatially extended, irregular in boundary and shape, structurally coherent, and conditioned by product-specific geometry and design constraints. A plausible implication is that an overview system intended for this regime must model defect support regions and deformations in a geometry-aware way rather than as coordinate noise or ad hoc local perturbation.

2. Knowledge-driven synthesis workflow

The synthesis workflow is organized as a three-stage process whose first stage is knowledge-driven data synthesis. The system ingests product-side information

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

where $\mathcal{I}_{\text{mv}}$ denotes multi-view visual observations, $\mathcal{T}_{\text{spec}}$ textual product specifications, and $\mathcal{E}_{\text{prior}}$ expert prior knowledge about defect modes and sensitive regions. An MLLM then converts this heterogeneous input into an executable instruction tuple

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

with anomaly type $u$, suggested region $\mathcal{R}$, synthesis operator $\mathcal{A}$, and parameter set $\Theta$ (Sun et al., 6 Apr 2026).

The paper emphasizes that this is not a descriptive step but an executable one. The MLLM is therefore characterized as a semantic parser and, more specifically, as a “cognitive-to-geometric translator.” In the reported implementation, the MLLM is Gemini 3. Its role is to translate statements such as likely bending zones, crack-prone surfaces, or edge-sensitive regions into a concrete synthesis instruction that 3D-DefectStudio can invoke.

The instruction tuple is then passed through a validation module $\mathcal{V}(\mathbf{c}, X)$. If the proposed instruction is physically infeasible on the target point cloud $\mathcal{I}_{\text{mv}}$0, the system falls back to a rule-based alternative:

$\mathcal{I}_{\text{mv}}$1

The validated instruction drives deterministic anomaly injection through MPAS:

$\mathcal{I}_{\text{mv}}$2

yielding both the anomalous point cloud $\mathcal{I}_{\text{mv}}$3 and the point-wise anomaly mask $\mathcal{I}_{\text{mv}}$4.

This workflow makes 3D-DefectStudio an overview platform in a strict sense: it does not only apply geometric edits, but also produces exact supervision targets analytically during generation.

3. Primitive-guided anomaly operators

At the core of 3D-DefectStudio is MPAS, which organizes defect generation around multi-dimensional support primitives. The paper presents this as the geometry-aware, “physics-isomorphic” basis of the synthesis process.

MPAS module	Support primitive	Defect forms
MPAS-1D	points or geodesic paths	protrusions, depressions, holes, scratches, grooves
MPAS-2D	planes and plane-intersection bands	bending, cracking
MPAS-3D	convex-hull-guided support regions and free-form surface deformation	free-form structural damage

MPAS-1D

MPAS-1D is designed for point-like or line-like defects. It begins from an ordered anchor set

$\mathcal{I}_{\text{mv}}$5

then constructs shortest surface paths on a kNN graph via Dijkstra’s algorithm:

$\mathcal{I}_{\text{mv}}$6

The resulting path union defines a support skeleton, which is expanded into a deformation neighborhood

$\mathcal{I}_{\text{mv}}$7

The deformation is then applied along the average surface normal of this region with a distance-decayed displacement:

$\mathcal{I}_{\text{mv}}$8

where $\mathcal{I}_{\text{mv}}$9 switches protrusion versus depression (Sun et al., 6 Apr 2026).

The significance of this formulation is that support geometry is surface-aware rather than Euclidean and volumetric. Defects such as grooves or scratches therefore follow the object topology through geodesic structure instead of being painted as straight-line artifacts in ambient space.

MPAS-2D

MPAS-2D uses a plane primitive

$\mathcal{T}_{\text{spec}}$0

with signed point-to-plane distance

$\mathcal{T}_{\text{spec}}$1

A narrow intersection band

$\mathcal{T}_{\text{spec}}$2

acts as the support region for two distinct operators.

For bending, the cloud is gradually rotated across the band using a hinge axis estimated by centroid and PCA. The interpolation coefficient is

$\mathcal{T}_{\text{spec}}$3

and the transformed point is

$\mathcal{T}_{\text{spec}}$4

For cracking, the signed distance is perturbed with Gaussian noise,

$\mathcal{T}_{\text{spec}}$5

and the removed region is

$\mathcal{T}_{\text{spec}}$6

This yields a cracked cloud by deletion, together with a labeled boundary band on the retained points near the fracture.

MPAS-3D

MPAS-3D is described as the most expressive module. It starts from anchors

$\mathcal{T}_{\text{spec}}$7

and forms the convex hull

$\mathcal{T}_{\text{spec}}$8

Rather than using the full volume, MPAS-3D uses the hull boundary $\mathcal{T}_{\text{spec}}$9 as a support surface and selects a narrow neighborhood around it as the deformation mask.

A local tangent frame is estimated with PCA, producing projected coordinates

$\mathcal{E}_{\text{prior}}$0

The local deformation is then modeled through a Gaussian-mixture height field,

$\mathcal{E}_{\text{prior}}$1

followed by displacement along the point normal and a kNN smoothing step to remove artifacts (Sun et al., 6 Apr 2026).

The result is a class of smooth, free-form structural damage that is intended to be difficult to obtain with earlier low-complexity synthesis schemes.

4. Integrated mask generation and supervision

A defining feature of 3D-DefectStudio is that anomaly masks are generated analytically inside each synthesis operator rather than annotated after the fact. In MPAS-1D, the mask is the geodesic support expanded by radius $\mathcal{E}_{\text{prior}}$2. In MPAS-2D, the bending mask is the set of points with $\mathcal{E}_{\text{prior}}$3, while the crack mask is a boundary band near the removed region. In MPAS-3D, the mask consists of points close to the convex-hull boundary. The paper treats this as a major scaling advantage, because large volumes of synthetic anomaly-mask pairs can be produced automatically (Sun et al., 6 Apr 2026).

These outputs supervise a Point Transformer-based detector in the second stage of Synthesis4AD. The reported detector configuration uses a Point Transformer / PointMAE-style point backbone as encoder and a lightweight MLP segmentation head, trained under a GLFM-style dense supervision protocol. The conceptual role of 3D-DefectStudio here is straightforward: it supplies geometrically plausible anomalous point clouds together with exact point-wise labels, thereby compensating for the scarcity of real defect annotations.

The training pipeline further incorporates Spatial-Distribution Normalization (SDN). For a point cloud $\mathcal{E}_{\text{prior}}$4, category center $\mathcal{E}_{\text{prior}}$5, and radius $\mathcal{E}_{\text{prior}}$6, SDN maps the cloud into a canonical unit-ball space:

$\mathcal{E}_{\text{prior}}$7

A fixed voxel size $\mathcal{E}_{\text{prior}}$8 is then used for voxel downsampling. The stated purpose is to reduce category-scale bias, stabilize feature statistics, and improve cross-category generalization.

After SDN, the detector is trained with geometry-faithful augmentations: random rotation, noise perturbation, and point dropout. The paper motivates these augmentations as a way to reduce sensitivity to global pose, simulate measurement noise, and simulate occlusion or partial scans, thereby forcing the model to rely more strongly on local geometric cues.

5. Inference mechanism and empirical results

At deployment, the trained encoder builds a prototype set of normal features

$\mathcal{E}_{\text{prior}}$9

For a test point cloud, point features $$\mathbf{c}=\Phi_{\psi}(\mathcal{Z}) = \left\{u,\mathcal{R},\mathcal{A},\Theta\right\},$$0 are extracted and scored by distance to the nearest prototype:

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

An object-level anomaly score is then computed as

$$\mathbf{c}=\Phi_{\psi}(\mathcal{Z}) = \left\{u,\mathcal{R},\mathcal{A},\Theta\right\},$$2

where $$\mathbf{c}=\Phi_{\psi}(\mathcal{Z}) = \left\{u,\mathcal{R},\mathcal{A},\Theta\right\},$$3 is the mean of the top-$$\mathbf{c}=\Phi_{\psi}(\mathcal{Z}) = \left\{u,\mathcal{R},\mathcal{A},\Theta\right\},$$4 scores (Sun et al., 6 Apr 2026).

The system is evaluated on Real3D-AD, MulSen-AD, and a real-world industrial parts dataset whose categories include Bevel Gear, Brake Caliper, Connecting Rod, Crankshaft, Piston, and Turbine Blade. The reported state-of-the-art results are:

Dataset	Object-level	Point-level
Real3D-AD	80.9% O-ROC	84.8% P-ROC
MulSen-AD	89.6% O-ROC	72.0% P-ROC
Real industrial dataset	95.9% O-ROC	73.8% P-ROC

The ablation results are especially informative for understanding 3D-DefectStudio’s contribution. MPAS-3D alone outperforms MPAS-1D or MPAS-2D alone, while combining all three anomaly types yields the best overall results. Increasing the number of synthetic samples improves performance monotonically. SDN and the augmentation pipeline each contribute measurable gains. The paper also reports that replacing older synthesis with MPAS improves the baseline from 66.6/52.0 to 75.2/56.5 on Real3D-AD and from 68.2/53.5 to 81.0/58.7 on MulSen-AD, which directly attributes a substantial share of the performance gain to higher-fidelity synthetic defect generation.

6. Significance, scope, and interpretation

3D-DefectStudio is significant because it makes scalable, knowledge-driven, geometry-faithful anomaly synthesis operational for industrial 3D inspection. Its contribution is not limited to generating more samples. It also formalizes the connection between semantic prior knowledge, executable geometric operations, and analytically derived point-wise supervision. In the paper’s framing, this is the operational bridge that turns MPAS from an overview method into a complete industrial data-generation system (Sun et al., 6 Apr 2026).

Three aspects define its scope. First, it targets realism: defects are synthesized through support primitives and geometry-consistent deformation rather than low-dimensional perturbation. Second, it targets scalability: automatic mask generation eliminates manual annotation as the limiting factor. Third, it targets knowledge integration: multi-view observations, textual specifications, and expert priors are explicitly brought into the generation loop through the MLLM.

At the same time, the paper’s decomposition clarifies what 3D-DefectStudio does not do. It is not the anomaly-detection backbone, and it is not a substitute for the downstream representation-learning pipeline. Its role is to generate the synthetic anomalous data and labels that make such learning feasible under real-world data scarcity. This suggests that its broader importance lies in data-centric industrial vision: improving the geometric fidelity and semantic controllability of synthetic anomalies can materially improve detector quality even when the downstream architecture remains within the Point Transformer family.