Model Training
Your defect. Your images. Your model.
Some defects are specific to your product and need a custom inspection model. Train your own model offline, directly on the machine.
| 1,000,000+ | Real production images It starts from our visual foundation model, trained on over a million real production images. |
It starts from our visual foundation model, trained on over a million real production images.
| You train the model yourself. | Built-in model training Collect images, label examples and train your model—all within the inspection software, with no cloud upload required. |
Collect images, label examples and train your model—all within the inspection software, with no cloud upload required.
Train a model in five steps.
- 01 Add to the dataset One click on any inspected region, or a whole product at once, puts the images into the dataset. Feedback from the review screen lands there too.
- 02 Label OK or NG Label each image as OK or NG. The software tracks unlabelled images so you can pause and resume later.
- 03 Train Name your model, start training and monitor its performance. Stop at any time and keep the best result so far.
- 04 Check the result Evaluate the model on images not used for training. Review correct predictions, false calls and missed defects in a results table.
- 05 Apply the model Apply the model to selected inspection regions and set it as the default for the group. It runs alongside your existing inspection checks.
Watch: training a model, start to finish
Teach the model what matters on your boards.
Build on our visual foundation model to inspect niche defects specific to your products. Train with labelled images from your own production to extend its inspection capabilities.



Improvement in practice
This demo shows 40, 19 and 6 alerts across three inspections, with feedback and adjustments between runs. Counts shown are alerts, not confirmed defects.
Your data stays in your facilities.
If an existing inspection produces false calls, you may not need a separate custom model. Submit feedback and run AI Update to improve the existing model.