How do you know a synthetic defect label is correct?
Visynex labels every defect image by what a separate detection pass actually finds, and discards the images where it can’t find the defect ordered. Most synthetic defect data doesn’t check. The label is whatever the generator was asked to draw.
What Visynex isVisynex makes synthetic defect images for training machine vision inspection models. Each image is made from one photograph of a good part, and every defect image ships only if a separate detection pass finds the defect ordered.
Labelling by construction, and why it fails
A generator is asked to put a scratch at a certain place. The pipeline records “scratch, at these coordinates” as the label. Nothing looks at the result.
If the generator drew the scratch faintly, or a few centimetres away, or not at all, the label doesn’t change. It is still a confident, precise box, around nothing, or around the wrong place. You can’t tell from the label. A model trained on it learns that a defect sometimes looks like nothing.
Labelling by independent detection
We treat the generator’s placement as a request, not a fact.
Every defect image goes to a separate detection pass. It receives:
- the generated image;
- your clean reference photograph, for comparison;
- the name of the defect to look for.
It is not told where the defect was placed. It has to find it.
- If it finds the defect with confidence, and the defect is on the part, its box becomes the label, even if the defect landed somewhere other than where it was placed. The label describes what is in the picture.
- If it doesn’t find it, isn’t confident (below 0.5), or what it finds isn’t on the part, the image is discarded. It never reaches you.
The claim is exact, and narrow: every defect image is checked for the defect you ordered. Apart from the pose and drawn-mark checks, which look only for a turned part and a drawn mark, nothing checks the rest of the image for other changes. Defect-free images get no detection check, because there’s no defect to look for.

What discards tell you
Discards aren’t only waste. When one class is discarded far more than the others, the class usually doesn’t work from a photograph. On our own test part, both images of kerf deviation, a dimensional error measured against a drawing, were discarded. We’d rather learn that from the check than sell labels for it.
The limits of the check
- It’s the same family of model as the generator, run separately. It isn’t fooled by the generator’s intent, because it never sees it. But it isn’t a different technology either. It’s a strong filter, not a proof.
- It looks only for the defect you ordered. It doesn’t check the rest of the image for other changes.
- A wrong box can occasionally get through. We’ve seen it. So every defect image ships with its box visible, and you can review any of them.
- It can only judge what the photograph shows. Dimensional and hidden defects are outside what any image-based check can confirm.

Questions
Are synthetic defect labels accurate?
Only if something checks them. Labels taken from the generator’s instructions are as accurate as the generator was obedient. Visynex labels each defect image from a separate detection pass and discards images where it can’t find the defect ordered.
What does “independent” mean here?
The detection pass is a separate model call that is never told where the defect was placed. It’s the same family of model as the generator, so it’s independent of the generator’s intent, not of the technology.
What happens to images that fail the check?
They’re discarded and never delivered. Replacements are made at our cost. If a class comes up short, we refill it at no extra charge. If anything is still missing after that, the difference is credited to your next order.
Can I see the labels before training?
Yes. Every defect image comes with its bounding box, in YOLO and COCO formats.
Try it on your own part
Sign in with your work email, upload one photograph of a good part, and get up to 10 images of it back, labelled in YOLO and COCO. Each takes about a minute. Free, once per email address.
Try it freeTalk to an engineer
Thirty minutes with someone who built the pipeline. No upload, no account.
Book a call