Proven technology that speaks
through results on the floor.

Vision models, anomaly detection, lightweight inference, operations feedback — connected end-to-end so the AI on your line keeps getting better.

Maximum performance from minimal data.

We achieved 100% detection of critical defects with 36× less data and manage 7,000+ production SKUs with a single model.

+27%
Overwhelming detection performance
+27.2%p detection over competing models A detection structure that lowers both false-positive and miss risk
36×
Maximum data efficiency
Compressed 155,282 training images down to 4,284 Cut initial-validation time and the burden of collecting defect data
100%
Never miss a critical defect
OOD techniques surface rare, unseen defects Reduces the risk of critical defects escaping the line
0.3s
Real-time high-speed processing
Industrial ultra-high-resolution images at 49.5MP inspected within 0.3 seconds per frame.

Key differentiators - AI Inspection

A sustainable operating structure with over-a-night model updates

AIVEX's AI operations aren't a one-shot deployment — performance compounds through a continual data loop on the shop floor.

Criteria

Scope of coverage

Response to environment shifts

Adapting to new SKUs

Ownership at rollout

Traditional rule-based inspection

Bound to initially defined rules

Re-author the rules from scratch

Rules must be edited for each SKU

Maintained by in-house engineers

AIVEX AI operations

Rapid response by learning from inspection results in real time

Periodic response via inspection-data feedback

New models created and rolled out with minimal resources

Simple feedback-driven AI ops

Generic AI solution

Limited to the initial training set

Spin up a new model

Repeat PoC per SKU

Build the ops backbone separately

Criteria

Scope of coverage

Response to environment shifts

Adapting to new SKUs

Ownership at rollout

Traditional rule-based inspection

Bound to initially defined rules

Re-author the rules from scratch

Rules must be edited for each SKU

Maintained by in-house engineers

AIVEX AI operations

Rapid response by learning from inspection results in real time

Periodic response via inspection-data feedback

New models created and rolled out with minimal resources

Simple feedback-driven AI ops

Generic AI solution

Limited to the initial training set

Spin up a new model

Repeat PoC per SKU

Build the ops backbone separately

Key differentiators - AI Robotics

Dexterous Manipulation in harsh environments, regardless of robot maker

AIVEX unifies robotics projects—traditionally run across robot makers, SIs, robot teaching, and wiring— into a single 'AI Robotics integrated control platform'.

No-CodeUsable by non-experts

AdaptiveAdapts to diverse environments

Hardware-AgnosticIndependent of robot makers

Connection flexibility
Hardware Agnostic
- A ROS2-based standard interface with no dependence on robot or camera makers, maximizing the long-term value of deployed equipment
AI-driven handling of irregular objects
AI Vision Driven Robotics
- 6-dimensional pose recognition of irregular objects with 3D-Vision-AI models, plus one-step Robot-Camera Calibration
Lego-block scalability
Modular Scalability
- No-Code Programming
- From Drag & Drop robot programming to AI plugins that extend robot capabilities
Imitation & reinforcement learning of expert behavior
Precision Physical AI
- Precision-assembly imitation learning, Sim2Real data augmentation, and reinforcement-learning fine-tuning for Physical-AI optimization

Validated research

Research that starts on the floor and reshapes it

AIVEX runs its own research program on problems that surface from real factory data and shares the results as papers and slides.

Memory-Distilled Selection for Noise-Robust Anomaly Detection
Industrial anomaly detection suffers a significant drop in performance when defect samples are mixed into the training data, and it typically has the limitation of requiring prior knowledge of the noise ratio or separate hyperparameter tuning for each ratio. This paper proposes MeDS, which performs well even without knowing the noise ratio. First, it randomly and sparsely samples memory and ensembles the results; this sparsity acts like a low-pass filter, providing a first-pass separation of normal and defective samples. The resulting scores are then distilled into a reconstruction network, and the model is fine-tuned with the clean data it has filtered out on its own, enabling precise localization down to the pixel level. It remains stable across various noise levels without ratio-specific tuning, achieving an image-level AUROC of 99.16% on MVTecAD with 40% noise and delivering state-of-the-art performance on VisA and Real-IAD as well. As an added benefit, it can also be used as a label correction tool to efficiently identify contaminated samples.
ICML 2026
- Learning with Noisy Label
- Anomaly Detection
View paper: Memory-Distilled Selection for Noise-Robust Anomaly Detection
TailedCore: Few-shot sampling for unsupervised long-tail noisy anomaly detection
In real industrial data, even normal samples often contain some defects mixed in as noise, and the amount of data can vary greatly by product type in long-tail distributions. This paper identifies a trade-off: filtering noise well can hurt rare-class performance, and vice versa, and handles the two separately. The key is TailSampler, which estimates class size by leveraging the symmetry of the embedding similarity distribution. Based on this, the authors propose TailedCore, a memory-based model that captures rare-class information while remaining robust to noise, achieving top performance in most settings on MVTecAD and VisA.
CVPR 2025
- Long-Tail
- Learning with Noisy Label
View paper: TailedCore: Few-shot sampling for unsupervised long-tail noisy anomaly detection
Few-Shot Segmentation Using Multi-Similarity and Attention Guidance
Few-shot segmentation aims to segment objects of novel classes using only a small number of examples, but a single prototype vector is not sufficient to capture all the information in the support images. This paper proposes MSANet, which includes a multi-similarity module that compares multiple feature maps to more accurately capture semantic relationships, as well as an attention module that enables the model to focus on important class information. Its performance is evaluated under 1-shot and 5-shot settings on PASCAL-5i and COCO-20i.
IEEE OJCS 2025
- Few-Shot Learning
- Image Segmentation
View paper: Few-Shot Segmentation Using Multi-Similarity and Attention Guidance
Background-Aware Defect Generation for Robust Industrial Anomaly Detection
Because industrial anomaly detection has very few defect samples, defects are often synthesized for training. However, conventional methods fail to account for the relationship between defects and their background, which can lead to unnatural anomalies. This paper proposes a background-aware generation framework that uses disentanglement loss and DDIM Inversion to generate only defects that fit the background context while leaving the background unchanged. It outperforms existing methods on both defect generation quality and anomaly detection performance on MVTec-AD and MVTec-LOCO.
arXiv Preprint 2024
- Generative Diffusion
- Anomaly Detection
View paper: Background-Aware Defect Generation for Robust Industrial Anomaly Detection

Memory-Distilled Selection for Noise-Robust Anomaly Detection
Industrial anomaly detection suffers a significant drop in performance when defect samples are mixed into the training data, and it typically has the limitation of requiring prior knowledge of the noise ratio or separate hyperparameter tuning for each ratio. This paper proposes MeDS, which performs well even without knowing the noise ratio. First, it randomly and sparsely samples memory and ensembles the results; this sparsity acts like a low-pass filter, providing a first-pass separation of normal and defective samples. The resulting scores are then distilled into a reconstruction network, and the model is fine-tuned with the clean data it has filtered out on its own, enabling precise localization down to the pixel level. It remains stable across various noise levels without ratio-specific tuning, achieving an image-level AUROC of 99.16% on MVTecAD with 40% noise and delivering state-of-the-art performance on VisA and Real-IAD as well. As an added benefit, it can also be used as a label correction tool to efficiently identify contaminated samples.
- Learning with Noisy Label
- Anomaly Detection
ICML 2026
View paper
TailedCore: Few-shot sampling for unsupervised long-tail noisy anomaly detection
In real industrial data, even normal samples often contain some defects mixed in as noise, and the amount of data can vary greatly by product type in long-tail distributions. This paper identifies a trade-off: filtering noise well can hurt rare-class performance, and vice versa, and handles the two separately. The key is TailSampler, which estimates class size by leveraging the symmetry of the embedding similarity distribution. Based on this, the authors propose TailedCore, a memory-based model that captures rare-class information while remaining robust to noise, achieving top performance in most settings on MVTecAD and VisA.
- Long-Tail
- Learning with Noisy Label
- Anomaly Detection
CVPR 2025
View paper
Few-Shot Segmentation Using Multi-Similarity and Attention Guidance
Few-shot segmentation aims to segment objects of novel classes using only a small number of examples, but a single prototype vector is not sufficient to capture all the information in the support images. This paper proposes MSANet, which includes a multi-similarity module that compares multiple feature maps to more accurately capture semantic relationships, as well as an attention module that enables the model to focus on important class information. Its performance is evaluated under 1-shot and 5-shot settings on PASCAL-5i and COCO-20i.
- Few-Shot Learning
- Image Segmentation
IEEE OJCS 2025
View paper
Background-Aware Defect Generation for Robust Industrial Anomaly Detection
Because industrial anomaly detection has very few defect samples, defects are often synthesized for training. However, conventional methods fail to account for the relationship between defects and their background, which can lead to unnatural anomalies. This paper proposes a background-aware generation framework that uses disentanglement loss and DDIM Inversion to generate only defects that fit the background context while leaving the background unchanged. It outperforms existing methods on both defect generation quality and anomaly detection performance on MVTec-AD and MVTec-LOCO.
- Generative Diffusion
- Anomaly Detection
arXiv Preprint 2024
View paper

View use cases

See for yourself the innovation proven on the floor.

See what AIVEX's AI operations can change for your line.

Proven technology that speaksthrough results on the floor.

Maximum performance from minimal data.

A sustainable operating structure with over-a-night model updates

Dexterous Manipulation in harsh environments, regardless of robot maker

Hardware Agnostic

AI Vision Driven Robotics

Modular Scalability

Precision Physical AI

Research that starts on the floor and reshapes it

Memory-Distilled Selection for Noise-Robust Anomaly Detection

TailedCore: Few-shot sampling for unsupervised long-tail noisy anomaly detection

Few-Shot Segmentation Using Multi-Similarity and Attention Guidance