How few real defect images can I start with?

A few representative examples is the stated minimum. What matters more than the count is that they show the defect as it genuinely appears on your line.

How many good images do I need?

Up to about ten, chosen to cover normal variation in lighting, orientation and surface appearance.

Why duplicate the real defect images?

The synthetic set is much larger. Expanding the real images to roughly fifty keeps genuine defect appearances proportionally present during training.

Is the defect library reusable across projects?

Yes. A defect annotated once can be generated onto other parts and lines that show the same failure mode.

What can the trained model do?

Defect detection, localization and classification, ready to deploy into the inspection pipeline for automated quality control.

Does my data leave our network?

No. Generation and training run on-prem on your dedicated GPU server.

How to Deploy Inspection with Minimal Data

Most inspection projects stall because there are not enough defect images to train on — the defect is rare, which is exactly why you want to catch it. This workflow starts from a handful of real defect samples, turns them into a reusable defect library, generates realistic synthetic defects on your own good parts, and merges everything into one training set.

9 steps
A few real defect images to start
No code

Step 1 of 9

Collect a few real defect samples

Start by gathering a small number of images containing real defects — NOK images — that represent the defect you want to detect. This is the only part of the workflow that needs real defective parts, and it needs very few of them.

Typical input

A few representative defect examples.
Different appearances of the same defect, if you have them.

Real NOK sample images beside the ReliUI import dialog. — A handful of genuine defect images is enough to begin — the rest of the dataset is built from them.

Step 2 of 9

Annotate the defects

Annotate the defects in the selected NOK images. You are marking both what the defect looks like and where it sits on the part.

The annotation view with a defect marked, next to the Gallery listing the annotated NOK images. — Annotations feed both the model training and the synthetic generation later in the workflow.

Step 3 of 9

Create or extend the defect library

Add the new defect to the GenAI Defect Library by selecting the annotated defect dataset. If a library already exists for this product family, you are extending it rather than starting over.

The GenAI defect library with the add-defect dialog and the annotated dataset selected. — Each defect added to the library stays available for future projects on the same material.

Step 4 of 9

Select representative non-defective images

Choose a small set of OK images that represent normal production. These are the clean surfaces the synthetic defects will be rendered onto, so they set the realism ceiling for everything generated afterwards.

Typical selection

Up to 10 images.
Covering normal variations such as lighting, orientation and surface appearance.

OK part images being selected and imported into the Gallery. — Ten good images that cover the normal variation of the line are worth more than fifty near-identical ones.

Step 5 of 9

Define possible defect regions

Annotate the OK images to indicate where defects can realistically occur.

The annotation view marking allowed defect regions on an OK part, beside the cluster image gallery. — Marking where a defect could plausibly appear stops the GenAI engine placing one somewhere it never would.

Step 6 of 9

Generate synthetic defect images

Use the defect templates from the defect library to generate synthetic defects on the OK images. This is the step that turns a handful of real examples into a dataset large enough to train on.

The defect library with a template selected and the generation dialog targeting the OK image set. — Defect templates are applied to the good images inside the regions defined in the previous step.

Step 7 of 9

Expand the real defect dataset

The original real defect images are expanded slightly — for example to around 50 images — using duplication, so that genuine examples are not swamped by the far larger synthetic set.

The Gallery duplicating real defect images to expand the set. — Duplication keeps real examples proportionally present once thousands of synthetic images join the set.

Step 8 of 9

Merge synthetic and real data

Combine the three sets into the final training dataset.

Synthetic defect images.
Real defect images.
Non-defective images.

The Merge Image Sets dialog combining the synthetic, real and non-defective datasets. — The merge dialog is where the final training dataset is assembled from the three sources.

Step 9 of 9

Train the inspection model

Train the inspection model using the merged dataset. Once trained, the model can perform three things on the line.

Defect detection — is there a defect present?
Defect localization — where on the part is it?
Defect classification — which defect type is it?

The training-type selector and a training run showing loss and accuracy curves. — Training curves for the merged dataset. The model is then ready to deploy into the inspection pipeline.

Start from the defects you already have

Download ReliVision to run this workflow on your own parts, or explore workflows, training, deployment and synthetic data best practices in the Knowledge Hub.

Download ReliVision