What are the main paradigms and methods in industrial anomaly synthesis?

Why anomaly synthesis became its own research topic

Earlier surveys of industrial anomalies concentrated on anomaly detection, treating synthesis as a helper for building training data. This paper argues that synthesis deserves separate treatment because it now serves three distinct jobs: augmenting scarce defect data, training downstream detectors, and enabling controllable inspection in which the user specifies what kind of defect should appear and where.

That reframing matters for how results should be read. If synthesis is only a data-augmentation step, the only question is whether a detector improves. If synthesis is also a controllable inspection tool, then visual realism, spatial placement, and the ability to specify anomaly semantics become goals in their own right. The paper keeps these goals separate throughout, which is why it evaluates methods on three different axes rather than one.

Four paradigms differ by where anomalies are made

Hand-crafted synthesis applies manual rules and image operations to produce defects. The paper notes its limits directly: realism and diversity are limited, and manual rules may introduce synthetic artifacts.

Distribution hypothesis-based synthesis constructs abnormal samples or representations by introducing controlled deviations around the distribution of normal data. It splits into two subcategories. Prior-dependent methods assume an explicit geometry for normality — a manifold, hypersphere, or similar structured prior — and construct or interpret deviations relative to it. Data-driven methods drop that assumption and instead perturb, jitter, replace, or sample learned features directly; the paper's examples include adding Gaussian noise to adapted normal features, feature jittering, Gaussian perturbations inside binarized Perlin-noise masks, and replacing masked quantized features with codebook vectors sampled under similarity constraints.

Generative model-based synthesis learns image-space priors and outputs explicit synthetic images containing anomalies. Because it models image-level appearance directly, the paper states it generally offers a higher ceiling for visual realism and anomaly diversity than the first two paradigms, at the cost of higher training and inference expense and a stronger need to keep synthesized anomalies consistent with surrounding normal content.

Vision-language model-based synthesis exploits large-scale pre-trained vision–language, multimodal, or text-conditioned generative priors. The paper is explicit that the boundary with generative models is not the architecture: both may use diffusion models. The dividing line is dependence on large-scale pre-training, the knowledge-transfer and adaptation strategy, and the use of multimodal information; larger model scale is only a secondary characteristic.

Feature-space synthesis trades images for training signals

The distribution hypothesis paradigm is the one most easily misread, because its output is often not an image at all. Its abnormalities primarily remain in representation space. The paper draws the practical consequence: this paradigm suits detection and localization pipelines that can use feature-level perturbation as synthetic supervision or regularization, but it is less suitable when deployment requires explicit photorealistic abnormal images, visually inspectable synthetic samples, or reusable image-space anomaly datasets.

The two subcategories fail differently. Prior-dependent methods become unreliable when the assumed normal geometry is unstable, since highly heterogeneous normal representations may not be captured by a compact manifold or hypersphere. Data-driven methods depend strongly on feature quality and may produce abnormal signals that transfer weakly to real image-space defects. On cost, both require non-trivial training because they are built on learned encoders, autoencoders, feature extractors, score models, or auxiliary latent components, while inference overhead stays relatively low because synthesis is a matter of extraction followed by perturbation, projection, replacement, sampling, or restoration rather than iterative image generation.

Image-space synthesis buys realism with control risk

Generative model-based synthesis is where the realism ceiling rises, and where the failure modes become spatial. The paper attributes distinct image-space failure modes to its subcategories: full-image synthesis may suffer from mode collapse and domain overfitting, full-image translation often provides weak spatial controllability, and local anomaly synthesis may introduce boundary artifacts or depend strongly on mask quality.

This gives a concrete selection rule rather than a ranking. When the deployment objective is complete-image dataset expansion, the trade-off is different from when it is source-anchored translation, and different again from when it is localized spatial control. The paper also flags task-data dependence and possible target-domain overfitting as a general risk of this paradigm, and high computational cost with limited real-time applicability as a further constraint.

Pre-trained knowledge widens the control interface

Vision-language model-based synthesis matters less because it generates better pixels than because it changes what a user can specify. By leveraging text prompts, reference images, spatial conditions, and other auxiliary cues, it enables more explicit specification of anomaly semantics, appearance, and spatial properties. That is a different capability from the image-space control offered by mask-conditioned generative methods.

Within this paradigm the paper separates single-stage from multi-stage pipelines by how the overall process is organized: single-stage methods produce the final abnormal output through one unified dominant synthesis process, while multi-stage methods explicitly introduce multiple stages or coordinated sub-pipelines such as mask synthesis, anomaly-part synthesis, or progressive refinement. The paper also identifies multimodal anomaly synthesis as underexplored in prior surveys, which is part of its stated contribution.

Three metric families answer three different questions

The paper's most useful warning is that realism, detector improvement, and spatial accuracy are not the same notion of synthesis quality, so evaluation is organized into three complementary perspectives.

Direct synthesis quality covers FID as a coarse feature-space realism indicator, and LPIPS, which must be read according to its protocol: reference-based LPIPS is lower-is-better for perceptual fidelity, whereas intra-cluster LPIPS (IC-LPIPS) is higher-is-better for anomaly diversity, computed as the mean pairwise LPIPS among generated anomalies grouped by real anomaly references or clusters. For local editing rather than full-image synthesis, PSNR and SSIM reflect background preservation and structural continuity outside the manipulated region, and Alignment IoU measures whether the anomaly lands in the intended mask region.

Downstream detection utility covers image-level AUROC, which is threshold-free and widely used, but the paper notes it may be insufficient under low anomaly prevalence and class imbalance, where image-level AP captures the precision–recall trade-off more directly and F1 reports balanced performance at a chosen operating point. Precision and recall clarify whether a gain comes from better anomaly coverage, fewer false alarms, or a compromise between the two.

Spatial localization accuracy covers pixel-level AUROC as a threshold-free starting point, but the paper argues pixel-wise ranking alone does not fully characterize localization quality in industrial settings. PRO and AU-PRO assess overlap at the connected-region level and are described as often better aligned with practical localization; AU-PRO is commonly reported with an FPR limit, often 0.3, while sPRO and AU-sPRO serve settings with logical or weakly pixel-defined anomalies, with AU-sPRO often used in MVTec LOCO-style evaluation at a limit of 0.05.

Benchmarks decide whether numbers are comparable

The paper states plainly that IAS results are not directly comparable unless benchmark choice, supervision setting, and evaluation context are considered together. Its benchmark table spans real, synthetic, and mixed data sources that differ not only in data type but in scale, inspection setting, and annotation characteristics.

Among real-world benchmarks, MVTec-AD and VisA are described as widely adopted standard references for 2D industrial anomaly detection. BTAD, MPDD, KSDD, and KSDD2 are tied to more specific inspection scenarios with more constrained object categories or acquisition conditions. MVTec LOCO-AD extends the setting by including logical anomalies alongside structural anomalies, and Real-IAD introduces large-scale, high-resolution, multi-view data collected from real production lines. Datasets also differ in sensing modality and construction strategy.

The practical implication is that metric selection is not a neutral reporting detail. Some studies test only whether synthesized anomalies improve downstream detectors; others additionally examine visual realism, structural preservation, editing faithfulness, or image–mask alignment. A reported gain therefore means different things depending on which of these the study actually measured.

Choose the paradigm by the output you need

The paper's deployment-oriented comparison turns the taxonomy into a decision structure. If the downstream task can consume distribution- or representation-level abnormal signals, distribution hypothesis-based synthesis is attractive and cheap at inference, but its limitations become apparent as requirements shift toward visually explicit, photorealistic, and directly reusable synthetic anomalies in image space. If explicit image-space samples with learned anomaly appearance are required, generative model-based synthesis is the relevant paradigm, and the choice within it should follow whether the objective prioritizes complete-image dataset expansion, source-anchored translation, or localized spatial control. If the requirement is to specify anomaly semantics, appearance, and spatial properties through prompts, references, and spatial cues, vision-language model-based synthesis is the paradigm that provides that interface.

The paper's own framing of the field's three central challenges — limited anomaly distribution coverage, the difficulty of synthesizing realistic anomaly samples, and the underexplored role of multimodal information — maps onto the same split. The most consequential judgment for a reader is therefore not which paradigm is strongest overall, but which output form the deployment actually needs: feature-level signals, photorealistic image-space samples, or prompt- and reference-controllable synthesis. The paper's stated limitations for each paradigm, from unstable normal-geometry assumptions to mask-quality dependence and high computational cost, are the conditions under which that choice should be revisited.

Sources used in this answer