How do you add a new part variant to a running inspection line?

Capture about 20 good images of the new variant on the line, train a good-parts-only model in under 48 hours, validate it on the first parts, and deploy it by USB. No defect samples are collected, and per-part confidence recognises the variant at run time, so a mixed run needs no manual sorting.

How long does onboarding a new variant take?

The capture step is minutes and the model trains in under 48 hours, unattended. The line keeps running its current work meanwhile, and the new model is deployed when it is ready, so adding a variant is a two-day task at most rather than a multi-month project.

Do you need a vision engineer to add a variant?

No. Capturing about 20 good images and training on-device is an operator or integrator task, aided by the on-device preview for framing and lighting. No labelling of defects and no specialist tuning are required, which is what keeps a new variant off the critical path.

How does the line handle a mix of variants?

By per-part recognition. The unit identifies the variant of each part as it arrives, at per-part confidence above 0.9 in a bumper deployment, and applies the right inspection or signals the right robot program. A safe default handles an ambiguous part rather than guessing.

What if the new variant needs a different fixture or lighting?

Lighting is configurable in the unit and set during the aim step with the on-device preview, no laptop needed. If the geometry changes the framing, you re-aim and re-capture the roughly 20 good images. The previous model is kept as a fallback until the new one is validated.

Add a New Part Variant to Inspection

Inspection camera aimed at a new part variant on a running conveyor

Updated July 2026 · 6 min read · Adente Vision Engineering Team

Adding a new variant takes about 20 good images, no defect samples. The model trains good-parts-only in under 48 hours while the line keeps running, deploys by USB, and recognises each part at run time with per-part confidence. It is an operator or integrator task, not a multi-month project or a vision-engineer hire.

The question integrators get asked most often is not about accuracy, it is about change: a customer adds a part, and how fast can the line inspect it, and who does the work? On a traditional vision system the answer is a new project. On a good-parts-only unit it is a two-day task.

What does adding a variant actually require?

Four things, none of which is a defect library. You capture about 20 good images of the new variant on the line, train a model on those good parts only, validate it on the first real parts, and deploy it. Because the model learns the good part rather than a catalogue of defects, there is nothing to stage, seed or label. For the mechanism behind good-parts-only training, see how to inspect a part with only 20 images.

Onboarding a variant, step by step

Step	What it involves	Owner and time
Capture about 20 good images	On the line, no defect samples	Operator or integrator, minutes
Train the model	Good-parts-only, on-device	Under 48 hours, unattended
Validate on first parts	Check false rejects and escapes	Quality or integrator
Deploy the model	USB update, keep the old one	Integrator, minutes
Run with per-part confidence	Variant recognised each part	Automatic

The line does not stop for any of this. It keeps inspecting current work while the new model trains in the background, and the deployment is a USB update that keeps the previous model as a fallback.

How does the running line tell the variants apart?

01The unit recognises the variant of each part as it arrives, not in batches.
02A per-part confidence score drives the decision, above 0.9 in a bumper deployment.
03An ambiguous part takes a safe default rather than being guessed.
04The right inspection, or the right robot program, follows from the recognised variant.

This is the same recognition that lets a unit signal left versus right headlight to a robot at a leading automotive OEM. 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, so the integrator who runs the line owns the changeover and keeps the customer. See the integrator model, and how a 30-minute field install sets it up, plus the pillar on AI visual inspection.

Adding a new part variant to a running inspection line.

What does adding a variant actually require?

Onboarding a variant, step by step

How does the running line tell the variants apart?

Frequently asked questions

Onboarding a new SKU or variant?