Synthetic Defect Generation Comprehensive Survey Report
kotai2003Executive Summary
Anomaly detection has the advantage that it can be trained using only good-product images; however, NG images are indispensable for threshold adjustment and performance evaluation, and when only one or two NG images can be obtained in the field, sufficient operation is difficult. This report comprehensively analyzes the latest research trends in Synthetic Defect Generation, which solves this issue, from three perspectives: web research, academic papers, and community practice, covering 2023 to 2026.
Most important finding: Since 2023, synthetic defect generation has diverged into four technical trends. (1) Diffusion-based methods (AnomalyDiffusion, SeaS, DefectFill, etc.) have become mainstream in terms of image quality, (2) feature-space synthesis methods (GLASS, CRAS) are the strongest in compatibility with PatchCore, (3) copy-paste methods (DRAEM, NSA) have become established as lightweight implementation baselines, and (4) few-shot-driven methods (TF-IDG, etc.) pursue practicality by generating from a single NG image. In particular, GLASS, presented at ECCV 2024, synthesizes anomalies in feature space using only normal images and achieved AUROC 99.9% on MVTec AD. It received the highest evaluation in two out of three sources as the method that can be most directly used for PatchCore threshold design.
Implications for practical operation: A workflow for designing thresholds using synthetic NG is promising, but there is a fundamental issue that the distribution gap between synthetic NG and real NG has not been systematically verified (explicitly pointed out only by academic sources). Thresholds determined using synthetic NG must always undergo final validation with real NG. As a recommended step-by-step introduction flow, this report proposes Phase 1: GLASS (no NG required) → Phase 2: TF-IDG/AnomalyDiffusion (expansion with one to two NG images) → Phase 3: SeaS + AnomalousPatchCore (advanced use with five or more NG images).
Risks to note: The domain gap problem, in which high accuracy on MVTec AD (97 to 99%) does not directly translate to actual factories, was pointed out by all three sources, and quality control of synthetic defects and review by domain experts are indispensable.
1. Overall Picture of Research Trends
1.1 Technical Transition from 2021 to 2026
Confidence: High (agreement across all three sources)
From 2021 to 2022, copy-paste-based methods such as CutPaste, NSA, and DRAEM appeared, establishing a framework that uses synthetic defects for learning from only good-product images. Since 2023, high-quality generation using Diffusion Models has become mainstream, and at the same time Feature-space synthesis (anomaly synthesis in feature space) has emerged as a method for direct integration with PatchCore.
2021 ── CutPaste (CVPR), DRAEM (ICCV): Establishment of copy-paste methods
2022 ── PatchCore (CVPR): Establishment of memory-bank-based AD SOTA, NSA (ECCV)
2023 ── Rise of diffusion-based methods, Optimizing PatchCore (few-shot)
2024 ── GLASS (ECCV), RealNet (CVPR), AnomalyDiffusion (AAAI),
DualAnoDiff, AnomalousPatchCore, COFT-AD (IEEE TIP),
AnoGen (ECCV)
2025 ── SeaS (ICCV), TF-IDG (ICCV), DefectFill (CVPR),
CRAS (TII), Sequential PatchCore
2026 ── Transition period toward zero-shot use of foundation models
(SAM, CLIP, DINOv2)1.2 Four Technical Trends
Confidence: High (agreement across all three sources)
| Trend | Representative methods | Maturity level | Compatibility with anomaly detection |
|---|---|---|---|
| Diffusion-based | AnomalyDiffusion, SeaS, DefectFill, TF-IDG | Rapidly developing | Indirect (image level) |
| Feature-space synthesis | GLASS, CRAS, COFT-AD | Promising at research level | Highest (direct operation in feature space) |
| Copy-paste-based | CutPaste, DRAEM, NSA, Sequential PatchCore | Mature | Medium (as preprocessing) |
| Few-shot-driven | AnomalousPatchCore, AnoGen | Developing | High (including direct PatchCore extension) |
1.3 Decline of GAN-based Methods
Confidence: High (agreement across all three sources)
GAN-based methods were mainstream before 2022, but due to the rise of diffusion-based methods, new research since 2023 has decreased significantly. Fundamental problems such as training instability, mode collapse, and difficulty in training with a small number of NG images have not been resolved, and the shift to diffusion-based methods is proceeding irreversibly. Their use is limited to cases such as CycleGAN-based domain transformation that continues to be used in semiconductor and PCB inspection.
2. Analysis by Method Category
2.1 Diffusion-based Methods (Mainstream from 2023 to 2025)
Confidence: High
A group of methods capable of generating high-quality and highly diverse defect images from a small number of NG images. They are based on Stable Diffusion / Latent Diffusion Models and specialize them for industrial defects by combining techniques such as spatial control, text control, and intensity control.
Strengths:
- Can generate visually natural defect images, making quality confirmation by humans easy [Community]
- Since they are generated with masks, they can also be used for segmentation evaluation [Academic]
- Many methods are explicitly designed to work in few-shot settings (one to five NG images) [Web, Academic]
Weaknesses:
- Fine-tuning requires high GPU cost (A100/V100 equivalent, several hours to one day) [Community]
- They do not directly operate the feature space of PatchCore, and their effect is indirect [Academic, Cross-validation]
- Generation of minute defects (one to five pixels) can become unstable during the denoising process [Academic]
| Method | Characteristics | Recommended scenario |
|---|---|---|
| AnomalyDiffusion (AAAI 2024) | Position control through spatial embedding. Highly reliable | When defect types are clear |
| DualAnoDiff (2024) | Strengthens background consistency with two branches | When high-quality generation is prioritized |
| SeaS (ICCV 2025) | Synthesizes unknown defects through attribute decomposition | When defect attributes are diverse |
| RealNet (CVPR 2024) | Synthesis with intensity control | Continuous generation from weak to obvious defects |
| DefectFill (CVPR 2025) | Precise local defects through inpainting | High-quality generation from a small number of reference defects |
| TF-IDG (ICCV 2025) | Training-free, one-shot | Lowest introduction barrier |
2.2 Feature-space Anomaly Synthesis Methods (Highest Compatibility with PatchCore)
Confidence: High
An approach that directly synthesizes anomalies in feature-vector space rather than image space. Since it operates in the same feature representation space as PatchCore’s memory bank, it can be used most directly for threshold design. It has been increasing rapidly from 2024 to 2025.
Strengths:
- Operates in the same space as PatchCore’s memory bank and directly connects to threshold calibration [Web, Academic, Community]
- Many methods work even without NG images (normal images only) [Academic]
- Can target and generate “weak defects” near the boundary of the normal distribution, making it strong for minute defects [Academic]
Weaknesses:
- Since the generated results are feature vectors, “visualization” is difficult, making intuitive quality confirmation by humans difficult [Academic]
- Hyperparameter adjustment for Gradient Ascent is necessary, and domain-specific tuning is required [Community]
| Method | Characteristics | Recommended scenario |
|---|---|---|
| GLASS (ECCV 2024) | Gradient Ascent + GAS/LAS. SOTA for minute defects | Best for PatchCore threshold design |
| CRAS (TII 2025) | Residual-based synthesis. Multi-class support | Unified management of multiple product categories |
| COFT-AD (IEEE TIP 2024) | Feature-space shaping through contrastive learning | Backbone enhancement without NG |
Relationship between GLASS and CRAS: Both methods are studies by the same author group (cqylunlun). GLASS is the representative feature-space synthesis method presented at ECCV 2024, and CRAS can be positioned as a version extended for multi-class support in TII 2025. When selecting an implementation, GLASS is suitable for single-class inspection, and CRAS is suitable for unified management of multiple classes.
2.3 Copy-paste Methods (Lightweight and Easy to Implement)
Confidence: High
The simplest approach for generating synthetic defects by cutting and pasting patches from good-product images or by overlaying external textures.
Strengths: Easy implementation, no additional data required, low computational cost [all three sources]
Weaknesses: Synthetic defects tend to diverge from actual industrial defects (scratches, dents, contamination, cracks) in terms of appearance and feature distribution [agreement across all three sources]
“Random texture pasting in DRAEM-style methods works on paper benchmarks, but actual metal dents and cracks have natural boundaries with the background, and this kind of synthesis produces a completely different feature distribution.” — Practitioner community [Community]
2.4 Direct PatchCore Extension Methods
Confidence: High
A group of methods that strengthen PatchCore’s architecture and pipeline with minimal changes.
| Method | Approach | NG image requirement |
|---|---|---|
| AnomalousPatchCore (ECCV WS 2024) | Fine-tuning the feature extractor with NG images | One to several NG images |
| Sequential PatchCore (ECCV WS 2024) | Pre-training with synthetic impurities + fine-tuning with real data | Not required (replaced with synthesis) |
| Optimizing PatchCore (2023) | Hyperparameter tuning + leave-one-out threshold calibration | Not required |
3. Integrated Comparison Table of Major Papers
The following integrated evaluation reflects cross-validation results and corrects overestimation.
| # | Paper title | Year / Conference | Category | Paper link | GitHub link | Minute defects | Few NG | PatchCore compatibility | Evaluation dataset |
|---|---|---|---|---|---|---|---|---|---|
| 1 | GLASS | 2024 / ECCV | Feature-space | arXiv | GitHub | ★★★ | ★★★ | ★★★ | MVTec AD, VisA, MPDD |
| 2 | AnomalyDiffusion | 2024 / AAAI | Diffusion | arXiv | GitHub | ★★☆ | ★★★ | ★★☆ | MVTec AD |
| 3 | SeaS | 2025 / ICCV | Diffusion | arXiv | GitHub | ★★★ | ★★★ | ★★☆ | MVTec AD, MVTec 3D AD |
| 4 | RealNet | 2024 / CVPR | Diffusion+Feature | arXiv | GitHub | ★★★ | ★★☆ | ★★☆ | MVTec AD, VisA, MPDD, BTAD |
| 5 | DefectFill | 2025 / CVPR | Diffusion (Inpainting) | arXiv | Unreleased | ★★★ | ★★★ | ★★☆ | MVTec AD |
| 6 | TF-IDG | 2025 / ICCV | Diffusion (Training-free) | CVF | Unconfirmed | ★★★ | ★★★ | ★★☆ | MVTec AD, VisA |
| 7 | AnomalousPatchCore | 2024 / ECCV WS | PatchCore extension | arXiv | Unreleased | ★★☆ to ★★★ | ★★★ | ★★★ | MVTec AD |
| 8 | DualAnoDiff | 2024 / arXiv | Diffusion | arXiv | GitHub | ★★★ | ★★★ | ★★☆ | MVTec AD, VisA |
| 9 | CRAS | 2025 / TII | Feature-space | arXiv | GitHub | ★★★ | ★★☆ | ★★★ | MVTec AD, VisA, MPDD, ITDD |
| 10 | Sequential PatchCore | 2025 / ECCV WS | Copy-paste+PatchCore | arXiv | Unreleased | ★★☆ | ★★★ | ★★★ | Proprietary surface inspection data |
| 11 | COFT-AD | 2024 / IEEE TIP | Feature-space | arXiv | Unreleased | ★★☆ | ★★★ | ★★★ | MVTec AD |
| 12 | AnoGen | 2024 / ECCV | Diffusion+Few-shot | arXiv | GitHub | ★★☆ | ★★★ | ★★☆ | MVTec AD |
| 13 | Optimizing PatchCore | 2023 | PatchCore extension | arXiv | GitHub | ★★☆ | ★★★ | ★★★ | MVTec AD, VisA |
| 14 | AdaBLDM | 2024 | Diffusion | arXiv | GitHub | ★★☆ | ★★☆ | ★★☆ | MVTec AD |
Note on evaluation criteria: Through cross-validation, items overestimated in the Community report (RealNet’s few-NG support, SeaS/TF-IDG’s PatchCore compatibility, etc.) have been corrected. ★★★ means strongest class for the relevant perspective, ★★☆ means practical but with constraints, and ★☆☆ means baseline level.
4. Detailed Explanation of Important Papers
4.1 GLASS (ECCV 2024) — Strongest Feature-space Synthesis
Confidence: High (Academic/Community agreement, highest evaluation through cross-validation)
Official title: A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization
Synthesis method: Adopts a two-stage unified synthesis strategy.
- Global Anomaly Synthesis (GAS): Generates near-in-distribution anomaly vectors near the manifold boundary of normal features in feature space through Gradient Ascent + Gaussian noise. Truncated Projection prevents excessive distribution deviation.
- Local Anomaly Synthesis (LAS): Operates local textures/patterns at the image level and synthesizes realistic minute defects.
Support for minute defects: ★★★ Strongest class. Designed with “weak defect” detection as the core issue. MVTec AD AUROC 99.9% is a result of this.
Few-NG support: ★★★ Fully supported. Operates only with normal images. Can generate a large amount of synthetic NG for threshold calibration without requiring NG images.
PatchCore compatibility: ★★★ Highest. Operates in the same feature space as PatchCore’s memory bank and can be directly used for threshold design.
Advantages: Can operate with zero NG images. Strong for minute defects. Official code available.
Weaknesses: Since it is feature-space synthesis, “visualizing” the generated results is difficult. Hyperparameter adjustment for Gradient Ascent is necessary [Community].
4.2 AnomalyDiffusion (AAAI 2024) — Representative Diffusion-based Method
Confidence: High (agreement across all three sources)
Official title: AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model
Synthesis method: Based on a Latent Diffusion Model, it introduces Spatial Anomaly Embedding (separately encoding the appearance and location of anomalies) and Adaptive Attention Re-weighting (dynamic enhancement of attention during generation).
Support for minute defects: ★★☆ Medium. Small defects can also be generated through mask-based spatial control, but minute details tend to be lost during the denoising process of diffusion. (Note: The Web report evaluated it as ◎, but ★★☆ from both Academic and Community sources is more technically accurate.)
Few-NG support: ★★★ Few-shot design is an explicit objective. It fine-tunes an LDM from several NG images and generates diverse defect patterns.
Advantages: Since NG data is obtained as images, human confirmation and selection are easy. Masked generation can also be used for segmentation evaluation.
Weaknesses: LDM fine-tuning requires GPU time. Generation diversity depends on the number of few-shot images.
4.3 SeaS (ICCV 2025) — Latest Few-shot Generation through Attribute Decomposition
Confidence: High (Academic/Community agreement)
Official title: SeaS: Few-shot Industrial Anomaly Image Generation with Separation and Sharing Fine-tuning
Synthesis method: Stable Diffusion fine-tuning based. Combines Separation (separate learning of anomaly tokens and normal tokens) and Sharing (generating both normal and anomalous images with the same U-Net). With Decoupled Anomaly Alignment Loss, anomaly attributes are decomposed and recombined into multiple tokens, enabling synthesis of unknown anomaly patterns.
Support for minute defects: ★★★ Synthesizes diverse minute defects such as small scratches, contamination, and unevenness through attribute decomposition. Generates high-resolution masks.
Few-NG support: ★★★ Strongest class. Can extract and recombine diverse defect attributes from several images.
PatchCore compatibility: ★★☆ Indirect. It is reported that generated normal/anomaly pairs improve unsupervised AD (including PatchCore) by +1.10% image-level AP and +8.66% pixel AP.
Advantages: Since normal images can also be generated at the same time, it can be used to augment the entire training dataset.
Weaknesses: Because it is based on Stable Diffusion, it requires a large amount of VRAM.
4.4 DefectFill (CVPR 2025) — Inpainting-type Precise Defect Generation
Confidence: Medium-High (Web/Community agreement, not mentioned in Academic)
Official title: DefectFill: Realistic Defect Generation with Inpainting Diffusion Model for Visual Inspection
- Paper: https://arxiv.org/abs/2503.13985
- GitHub: Unreleased (because it was presented in 2025)
Synthesis method: Fine-tunes an Inpainting Diffusion Model specifically for defects. Improves generation quality through a custom loss function combining defect loss, object loss, and attention loss, plus Low-Fidelity Selection.
Few-NG support: ★★★ Explicitly designed for generation from a small number of reference defect samples.
Weaknesses: Presented at CVPR 2025 and the official GitHub has not been released. Reproducibility has not yet been verified.
4.5 TF-IDG (ICCV 2025) — Training-free One-shot Generation
Confidence: Medium-High (Web/Community agreement)
Official title: Training-Free Industrial Defect Generation with Diffusion Models
- Paper: CVF PDF
- GitHub: https://github.com/rubymiaomiao/TF-IDG
Synthesis method: A training-free defect generation framework. Operates in a one-shot setting with a single NG image.
Greatest advantage: Since fine-tuning is unnecessary, GPU cost is low, and the barrier to field introduction is the lowest.
Weaknesses: GitHub unconfirmed, and community track record is limited.
4.6 AnomalousPatchCore (ECCV 2024 Workshop) — Direct PatchCore Extension
Confidence: High (Academic/Community agreement)
Official title: AnomalousPatchCore: Exploring the Use of Anomalous Samples in Industrial Anomaly Detection
- Paper: https://arxiv.org/abs/2408.15113
- GitHub: Unreleased
Synthesis method/approach: Fine-tunes PatchCore’s feature extractor through contrastive learning using normal + anomalous samples. Supports three auxiliary tasks: anomaly classification, localization, and class imbalance correction.
PatchCore compatibility: ★★★ Directly compatible. Can strengthen an existing PatchCore pipeline with minimal changes.
Weaknesses: Code has not been released, so reproduction implementation is required. Countermeasures against fine-tuning instability with few NG images are limited.
4.7 RealNet (CVPR 2024) — Intensity-controllable Diffusion Synthesis
Confidence: High (mentioned by all three sources, but evaluations conflict)
Official title: RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection
Synthesis method: Synthesizes anomalies by continuously controlling anomaly intensity using SDAS (Strength-controllable Diffusion Anomaly Synthesis). It has a three-component structure combining AFS (feature selection) and RRS (residual selection).
Important note: The Community report evaluated RealNet as “no NG required (★★★),” but as a result of cross-validation, the Web report’s “large-scale synthesis is assumed (★★☆)” is more accurate. RealNet is designed to use 360k pre-synthesized images, making it difficult to use as a direct generation flow from a small number of NG images.
5. Compatibility Analysis with AI Anomaly Detection and Threshold Design
5.1 Three Roles of NG Images in AI Anomaly Detection
Confidence: High (agreement across all three sources)
- Threshold setting: Check the distribution of anomaly scores and adjust the balance between FP/FN
- Performance evaluation: Calculate AUROC, F1 score, AU-PR, etc.
- Defect pattern analysis: Identify which kinds of defects are difficult to detect
5.2 Compatibility Mapping with AI Anomaly Detection by Method
Confidence: High (cross-validation corrected)
Tier 1: Directly Integrable with AI Anomaly Detection (★★★)
| Method | Integration method | NG data requirement |
|---|---|---|
| GLASS | Normal images only → synthetic NG generation in feature space → direct use for PatchCore threshold design | Not required |
| CRAS | Multi-class feature-space synthesis → memory bank expansion | Not required |
| AnomalousPatchCore | Fine-tune the feature extractor with NG images | One to several images |
| Sequential PatchCore | Pre-training with synthetic impurities → coreset update with real data | Not required (replaced with synthesis) |
| Optimizing PatchCore | Leave-one-out + statistical threshold calibration | Not required |
| COFT-AD | Directly insert the fine-tuned backbone into PatchCore | Not required |
Tier 2: Indirectly Effective as Data Augmentation (★★☆)
| Method | Usage | Notes |
|---|---|---|
| AnomalyDiffusion | Synthetic NG images → threshold adjustment data | High fine-tuning cost |
| SeaS | High-quality synthetic NG + masks → evaluation data augmentation | Requires a large amount of VRAM |
| DefectFill | Generate realistic defect images from a small number of reference defects | GitHub unreleased |
| TF-IDG | Mass-generate synthetic NG with one-shot, training-free → immediate construction of evaluation data | Not specialized for PatchCore |
| DualAnoDiff | Synthetic NG with high background consistency → evaluation data | High computational cost due to two branches |
| RealNet | Generate boundary samples with intensity control | Requires 360k pre-synthesized images |
5.3 Fundamental Caution Regarding Threshold Design
Confidence: High (unique academic-source point, extremely important)
The extent to which a threshold determined using synthetic NG is effective for the actual NG distribution has not been systematically verified. Threshold determination methods are not standardized across papers, making comparison difficult. Synthetic-NG-based thresholds must always undergo final verification with real NG. [Academic]
This issue is directly connected to the core issue of this survey. All three sources recommend “threshold setting with synthetic NG,” but only the Academic report explicitly points out the fragility of this premise itself, so sufficient caution is required in practical operation.
5.4 Best Practices for Threshold Design
Confidence: Medium-High (integration of insights from multiple sources)
- Leave-one-out estimation: Estimate a beta distribution using only normal images and determine the initial threshold [Web, Community]
- Percentile-based: Set the 99.9th percentile of the anomaly score distribution of normal images as a provisional threshold [Community]
- ROC construction with synthetic NG: Generate 20 to 50 synthetic NG images with GLASS/TF-IDG and draw an ROC curve [Web]
- Final verification with real NG: Confirm and fine-tune the threshold derived from synthetic NG using one to two real NG images on hand [Academic]
- Review by domain experts: Inspectors perform final confirmation of the quality of synthetic NG and the validity of the threshold [Community]
6. Quantitative Benchmark Results
6.1 MVTec AD (Image-level AUROC)
Confidence: Medium-High (reported values from each paper. Note differences in experimental settings)
| Method | MVTec AD AUROC | Notes |
|---|---|---|
| PatchCore (baseline) | 99.1% | CVPR 2022 |
| GLASS | 99.9% | ECCV 2024, specialized for weak defects |
| RealNet | SOTA class | CVPR 2024 |
| SeaS (contribution to unsupervised AD) | +1.10% AP improvement | ICCV 2025 |
| AnoGen (contribution to DRAEM) | +5.8% AU-PR | ECCV 2024 |
| CRAS | 98.3% (Image), 98.0% (Pixel) | TII 2025 |
Note: MVTec AD is almost saturated with AUROC in the 97 to 99% range. The field is moving toward more difficult benchmarks (MVTec AD 2, VisA, MPDD, etc.). [Web, Academic]
6.2 List of Major Datasets
| Dataset | Characteristics | Number of categories | Difficulty |
|---|---|---|---|
| MVTec AD | Industrial parts and textures. Industry standard | 15 | Somewhat saturated |
| MVTec AD 2 | Eight new scenarios, 8000+ images | 8+ | High (released in 2025) |
| VisA | High-resolution industrial products | 12 | Higher than MVTec |
| MPDD | Dents and scratches on metal parts | 6 | Many minute defects |
| MVTec LOCO | Logical anomalies (misplacement, etc.) | 5 | Special |
| BTAD | Industrial products | 3 | Medium |
7. Practitioner Insights and Best Practices
7.1 Voices from the Field
Confidence: Medium-High (Community report only, but based on examples)
SEA Vision (IEEE ETFA 2025):
“By using Anti-Aliased WRN50 as the backbone, setting the input resolution to 448×448, and combining dedicated data augmentation, practical anomaly detection became possible even with only one normal image.”
DataRoots (PatchCore implementation blog 2024):
“PatchCore’s greatest strength is that it can be trained using only normal images. However, NG images are still required for threshold setting, which becomes a headache in the field from the beginning.”
Issue trends in the official PatchCore repository (amazon-science):
- “Is there a way to automatically determine the threshold?” — most common question category
- “When there are few NG images, the ROC curve cannot be drawn accurately” — common concern
7.2 Implementation Tuning Knowledge
Confidence: Medium-High [Community]
- Backbone: WideResNet50 is the most stable. Switching to DINOv2/ViT-based models is a recent trend
- Coreset ratio: 1 to 10% is sufficient (using all features is excessive)
- Input resolution: Adjust in the range of 224 to 448 px according to product size
- Integrated framework: anomalib is the most practical because it supports multiple methods including PatchCore
7.3 Typical Failure Patterns When Introducing Synthetic Defects
Confidence: Medium [Community]
- “Implemented the paper method as-is, but it did not work” → Domain adaptation to the company’s own product category is required
- “Using synthetic defects for training increased false positives” → Quality control and filtering of synthetic defects are essential
- “Tried diffusion fine-tuning but gave up due to insufficient GPU” → Training-free methods (TF-IDG, etc.) are realistic alternatives
8. Recommended Classification by Three Perspectives
8.1 Methods Strong in Minute Defect Generation (Scratches, Dents, Cracks, Foreign Matter, Unevenness)
Confidence: High
| Priority | Method | Category | Reason |
|---|---|---|---|
| ★★★ First recommendation | GLASS | Feature-space | Targets anomalies at the normal boundary through Gradient Ascent. Countermeasures for weak defects are the core of the design. Official code available [Academic, Community] |
| ★★★ Second recommendation | SeaS | Diffusion | Generates fine scratches, contamination, and unevenness with detailed control through attribute decomposition. High-resolution masks included [Academic] |
| ★★★ Third recommendation | DefectFill | Diffusion (Inpainting) | Fine defect generation through inpainting, quality assurance through Low-Fidelity Selection [Web] |
| ★★☆ Complementary | DualAnoDiff | Diffusion | Generates minute defects that do not “float” due to background consistency. High mask accuracy [Academic] |
| ★★☆ Complementary | RealNet | Diffusion+Feature | Continuous generation from weak to obvious defects through intensity-variable Diffusion [Academic, Web] |
Methods unsuitable for minute defects: Copy-paste methods (CutPaste, DRAEM) are often unsuitable for reproducing minute defects [agreement across all three sources]
8.3 Methods Suitable for AI Anomaly Detection Evaluation and Threshold Adjustment
Confidence: High
| Priority | Method | Specific usage |
|---|---|---|
| ★★★ | GLASS | Normal images only → synthetic NG generation in feature space → directly design PatchCore threshold |
| ★★★ | Optimizing PatchCore | Leave-one-out + statistical fitting for threshold calibration with almost no NG data required |
| ★★★ | AnomalousPatchCore | Fine-tune the feature extractor with one to two NG images → immediately strengthen PatchCore |
| ★★☆ | CRAS | Multi-class support. Unified threshold management for multiple product categories |
| ★★☆ | TF-IDG | One-shot, training-free mass generation of synthetic NG → immediate construction of evaluation data |
| ★★☆ | AnomalyDiffusion | Several NG images → generation of diverse synthetic NG → use as threshold adjustment data |
9 Reference Surveys and Awesome Lists
| Resource | URL | Content |
|---|---|---|
| A Survey on Industrial Anomalies Synthesis (2025) | arXiv | IAS taxonomy covering about 40 methods |
| Anomaly Detection with Diffusion Models Survey (2025) | arXiv / GitHub | Survey specialized in diffusion-based methods |
| awesome-industrial-anomaly-detection | GitHub | Comprehensive paper list for industrial AD |
| awesome-anomaly-synthesis | GitHub | List specialized in synthesis methods |
| Awesome-Few-Shot-Defect-Image-Generation | GitHub | Few-shot defect generation list |
| Awesome-Anomaly-Generation | GitHub | Hierarchical taxonomy of anomaly generation methods |
10. Unresolved Issues and Risks
10.1 Real-environment Effectiveness of Synthetic NG Thresholds (Most Important)
Confidence: High (unique academic-source point but a fundamental issue)
All sources recommend the flow of “adjusting PatchCore thresholds with synthetic NG,” but there is no systematic validation data on how effective thresholds determined using synthetic NG are for the actual NG distribution. This is the most important knowledge gap identified in this survey and concerns the reliability of the entire field of synthetic defect research.
Countermeasure: Treat synthetic NG thresholds as “initial setting values” and incorporate verification with real NG as a mandatory process.
10.2 Domain Gap Problem
Confidence: High (Academic/Community agreement)
There are many reported cases in which methods that achieved high accuracy on MVTec AD/VisA suffer a significant performance drop when applied to actual factory data.
“Even if it reaches 99% on MVTec, it can become only 60% on our own products.” — Practitioner [Community]
Beyond Academic Benchmarks (arXiv 2503.23451) analyzes this problem in detail.
10.3 Generation Quality Limitations for Ultra-minute Defects
Confidence: Medium (Academic only)
Scratches and dents on the scale of one to five pixels are unstable to generate even with diffusion models. Generation resolution and denoising randomness become inhibiting factors. Feature-space synthesis (GLASS, etc.) may be more suitable for this type of minute defect.
10.4 Lack of Quantitative Data on Computational Cost
Confidence: Medium
All sources lack quantitative comparisons of GPU time, memory usage, and inference speed for each method. Diffusion-based fine-tuning requires a GPU equivalent to A100/V100, and depending on the method, training time can take several hours to one day, but concrete benchmarks are not yet established.
10.5 Absence of a Standard Definition of Minute Defects
Confidence: Medium
The definition of “minute defect” is not standardized. There are no quantitative criteria such as definitions by pixel size or signal-to-noise ratio (SNR) relative to normal regions, making fair comparison between methods difficult.
10.6 Lack of Introduction Cases in Actual Factories
Confidence: Medium
There are case reports from SEA Vision, Edge Impulse, and others, but concrete accuracy values and actual threshold-adjustment performance data have not been made public.
Limitations:
- The investigation of each source is based on information as of April 17, 2026
- Self-reported benchmark values in papers are difficult to compare directly because of differences in experimental settings (backbone, resolution, etc.)
- The Community report tended to overestimate some methods, and corrections were made through cross-validation
- For the latest methods presented in 2025 (DefectFill, TF-IDG, SeaS), community track records are insufficient
- Quantitative success-case data in actual factories is limited
This report is based on multifaceted literature and practical research as of 2026-04-17. Because the technology is advancing rapidly, please check arXiv, GitHub, and Papers with Code as needed for the latest information.
