Updated July 2026 · 7 min read · Adente Vision Engineering Team
What is synthetic data for defect detection, and how is it different from augmentation?
Synthetic data for defect detection is defect imagery a computer generates rather than a camera captures: rendered CAD models, simulated flaws, or images produced by a generative model. Data augmentation is different, it transforms real captured images by flipping, rotating, changing brightness or adding noise to squeeze more variety out of the examples you already have. Both aim at the same shortage, too few real defect images, but they start from opposite places: augmentation stretches real data, synthesis invents new data.
The distinction matters because the two carry different risks. Augmentation stays anchored to real defects, so its failure mode is limited variety. Synthesis is not anchored to anything real, so it can produce defects that look convincing to a person but never occur that way on the line. 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 it is designed so most inspection tasks need neither: it trains on good parts only.
When does synthetic data genuinely help?
Synthetic data genuinely helps when real defect examples are close to impossible to obtain. Some defects are dangerous to produce, some are destructive or costly to create deliberately, and some are so rare that waiting for them is not an option. In those cases a synthetic set can pad a sparse defect class enough for a supervised model to learn its general shape, and it can cover hazardous failure modes you would never manufacture on purpose.
Concrete examples make the boundary clear. A crack in a safety-critical weld, a rupture in a pressure part, a contamination you are not permitted to introduce into a food or pharmaceutical line: producing these on purpose is unsafe, unlawful or ruinously expensive, so a rendered or simulated version may be the only way to give a supervised model any exposure to them at all. That is synthetic data at its most defensible, filling a class you genuinely cannot capture from the line.
It also helps as a supplement rather than a substitute. A dataset that starts from a few real defects and adds carefully generated variations, validated against the real ones, can improve coverage of a known failure mode. The key qualifier is validation: synthetic data earns its place only when the model it trains is tested against real, held-out defects, not against more synthetic images.
When does synthetic data mislead?
Synthetic data misleads when the model learns the generator instead of the defect. If rendered or generated defects miss the real texture, optics and lighting of the line, the model can latch onto the artifacts of the renderer, the tell-tale signature of how the fake was made, and score well in testing while failing on real parts. This is the domain gap, and it is the central risk of synthetic defect data.
The failure mode is subtle because it hides behind good metrics. A model validated only on more synthetic data will look accurate, then miss the real defect it was supposed to catch, or over-reject a real good part whose normal texture the synthetic set never showed. Fake defects can also teach a flaw that never occurs that way physically, so the model chases a failure the line does not actually produce.
The domain gap widens over time as well. A synthetic set is frozen at the moment it was generated, while a real line drifts: material lots change, tooling wears, lighting ages. A model trained on last quarter's renders can fall behind this quarter's parts, and because the synthetic set cannot be recaptured from the line, closing that gap means regenerating and revalidating the whole set rather than photographing a fresh batch of real parts.
| Factor | Helps when | Misleads when |
|---|---|---|
| Defect availability | Real defects are dangerous or near impossible to capture | Real defects are actually collectable |
| Realism | Generator matches real optics, lighting and material | Renders miss real texture, so the model learns artifacts |
| Failure mode | Synthetic reproduces the true physical flaw | Synthetic teaches a flaw that never occurs that way |
| Validation | Tested against held-out real defects | Validated only on more synthetic images |
Why does good-parts-only anomaly detection reduce the need for synthetic defects?
Good-parts-only anomaly detection reduces the need for synthetic defects because it does not learn defects at all. It learns what a good part looks like from about 20 real good images and flags deviation, so there is no sparse defect class to pad and no failure mode to invent. The data you train on is real, abundant and already coming off the line, which sidesteps both the collection problem and the domain gap that synthetic data introduces.
There is an operational reason this matters on a real line. Good parts can be recaptured whenever normal shifts, so retraining is a matter of collecting about 20 fresh good images and training again under 48 hours, with the updated model deployed by USB even on an air-gapped line. A synthetic pipeline has no such shortcut: every change to the part means rebuilding the generator and revalidating it against real defects you still have to find.
That does not make synthetic data useless, it narrows where it is worth the trouble. If you have a specific, repeatable defect you want as a labeled class in the unit's Defect mode, a carefully validated synthetic set may help. For the rare, first-seen and unspecified defects, the Anomaly mode covers them from good parts only, and training finishes under 48 hours. Public benchmarks such as the MVTec Anomaly Detection dataset were built precisely because real defect data is scarce, and they measure anomaly methods on real, held-out defects, which is the standard synthetic data has to meet.
A trust checklist: should you use synthetic defect data?
Before you trust a synthetic defect set, run it through a short checklist. The goal is to keep the benefit, more coverage of a hard-to-capture defect, without importing the domain gap.
- Can you collect the real defect at all? If yes, prefer real images or good-parts-only anomaly detection.
- Does the generator reproduce the real optics, lighting and material, not just the shape?
- Is the model validated on real, held-out defects, never only on synthetic images?
- Does the synthetic defect match how the flaw physically occurs, or an imagined version of it?
- Could good-parts-only anomaly detection cover this without any defect generation at all?
If the honest answers point to real data or good-parts-only, that is usually the shorter and safer path. For the full few-shot method, see the pillar guide on AI visual inspection; for the practical 20-image walkthrough, see the sibling post on inspecting with 20 images; and to match the approach to a unit and mounting, see the system page.