Does "trains on good parts only" mean it can never learn specific defects?

No. It starts from good parts only, so you can inspect without collecting defects, but you can still add labeled defects later where a specific, repeatable flaw needs its own class. The Defect mode handles known defect types; the Anomaly mode handles the first-seen and rare ones by scoring deviation from good. Most lines start good-parts-only and add specific defect classes only where they pay off.

How many good images do you need to start?

About 20 representative good reference images. They should cover the natural variation a good part actually shows, different lots, positions and lighting, not 20 identical perfect examples. Training finishes under 48 hours, and the unit mounts in about 30 minutes, so a new part becomes a two-day task rather than a multi-month data-collection project.

If it never sees a defect, how does it decide a part is bad?

It scores how far each part deviates from the learned normal and compares that score to a threshold. A part inside the normal range passes, a part far enough outside it is flagged. You set the threshold to your cost of a false reject versus an escape, which is the single knob that balances catching every deviation against over-rejecting acceptable variation.

Will a good-parts-only model over-reject normal variation?

It can, if the good set was too clean. If the 20 images capture only pristine parts, ordinary in-spec variation reads as anomalous. The fix is to build the acceptable spread into the good set and tune the threshold. A representative good set is what keeps false rejects low, and it is a quality decision, not a software setting.

Does good-parts-only work at production line speed?

Yes. On a live cap-inspection line an Adente Vision unit trained on good parts only reaches 99.65% F1 with a 0.69% false-negative rate at about 30 ms per part. The catalog bound is 0.5 s per part at 100+ parts per minute. The figure you can commit to for your own cycle needs an application-specific measurement.

Train on Good Parts Only: What It Means

Updated July 2026 · 7 min read · Adente Vision Engineering Team

Training on good parts only means an inspection model learns what a normal, in-spec part looks like from good examples, then flags anything that deviates. You never collect, stage or label defect images to start. About 20 good reference images are enough, and training finishes under 48 hours.

What does "trains on good parts only" actually mean?

Trains on good parts only means an inspection model learns the appearance of a good, in-spec part from a small set of good examples, builds a picture of what normal looks like, and then scores every new part by how far it departs from that normal. No defect image is shown during training. The method is also called normal-only or good-samples-only training, and it is the mechanism behind few-shot anomaly detection.

The contrast is with supervised classification, which learns from both good and defective parts and needs each defect type labeled in advance. Good-parts-only inverts that: it models only the good class and treats anything sufficiently far from it as a candidate defect. You are not teaching the model a list of flaws to look for, you are teaching it what right looks like.

Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators, and this good-parts-only approach is how it starts inspecting a new part from about 20 reference images.

Why does skipping defect collection collapse the project timeline?

Skipping defect collection removes the single slowest step in a traditional inspection project. When a model must be trained on labeled defects, someone has to find, produce or wait for enough examples of every failure mode, photograph them consistently, and label each one. On a line that runs well, those defects are scarce, so the collection phase alone can run for weeks or months. That range is industry-typical, not an Adente figure.

Good-parts-only skips that phase entirely. Good parts are the one thing a production line has in abundance, so the data you need is already coming off the line. About 20 representative good images are enough to begin, model training finishes under 48 hours, and the unit itself mounts in about 30 minutes. A new inspection task becomes a two-day job instead of a multi-month data-collection exercise.

It also changes what a line changeover costs. Adding a new part or variant no longer means restarting defect collection, it means capturing about 20 good images of the new part and retraining.

How does a good-parts-only project compare to a defect-labeled one?

The two approaches diverge from the first day of the project, not just on accuracy but on what work you have to do and when you can go live.

Dimension	Defect-labeled (supervised)	Good-parts-only (anomaly)
Data to start	Hundreds to thousands of labeled defect images (industry range)	About 20 good reference images
Defect collection	Collect, stage and label every failure mode first	None needed to start
First-seen defects	Misses what was never labeled	Flags deviation from normal
Time to first model	Weeks to months (industry range)	Under 48 hours
Line changeover	Re-collect defects for the new part	Capture ~20 good images, retrain

How can a model catch a defect it has never seen?

A good-parts-only model catches an unseen defect because it does not look for defects at all, it looks for deviation from normal. Once it has modeled what a good part looks like, any part that strays far enough from that learned normal scores as anomalous, whether or not that particular flaw ever appeared in training. A first-seen scratch, a short-shot, a contamination mark nobody specified: each reads as a departure from good.

That is the structural advantage over a rule or a supervised classifier, which can only catch a defect that was described or labeled in advance. On the four inspection modes the unit runs, Anomaly, Defect, Counting and Quality, the Anomaly mode is the one built on this reject-by-deviation logic, and it is combined with classical computer vision for measurement so the result stays auditable rather than a black box.

The trade is that deviation scoring returns a distance, not a verdict, so you set a threshold that turns the score into a pass or fail signal to the PLC. Where you set it decides the balance between catching every deviation and over-rejecting normal spread.

What must the "good" set actually contain?

The good set has to contain the full natural variation of a good part, not a handful of perfect ones. Real good parts differ in colour, texture, position, surface finish and lighting response within the range that quality accepts. If the 20 images capture only pristine, identical examples, the model learns an unrealistically narrow normal and rejects ordinary in-spec variation as if it were a defect. That is the most common cause of over-rejection in a good-parts-only deployment.

So the goal when choosing the good images is representativeness, not perfection. Include the acceptable spread: different lots, the range of positions the part actually presents, the lighting the line actually uses. The unit's configurable lighting and 12 MP global-shutter capture help keep those images consistent, but the judgment of what counts as acceptable variation is a quality decision, and it belongs in the good set on purpose.

A clean, representative good set and a threshold tuned to your cost of a false reject are what separate a good-parts-only model that runs quietly from one that nags the operator.

Does trains-on-good-parts-only hold up in production?

Good-parts-only holds up in production where it counts, on accuracy and speed. On a live cap-inspection line, an Adente Vision unit trained this way reaches a 99.65% F1-score with a 0.69% false-negative rate at about 30 ms per part, rejecting broken, unclosed and hinge-damaged caps the moment they pass. That is a good-parts-only model doing production inspection at line speed, not a lab result.

The catalog throughput bound is 0.5 s per part with 100+ parts per minute, and the measured 30 ms is the field figure from that installation. The number you can commit to on your own line still needs an application-specific measurement on your parts and lighting, but the mechanism is the same one behind the cap-line result. For the full method behind few-shot and anomaly detection, see the pillar guide on AI visual inspection; for the practical "20 images is enough" walkthrough, see the sibling post on inspecting with 20 images; and to match the approach to a unit and mounting, see the system page.

What "trains on good parts only" actually means.