The most common question in a first vision conversation, and the one with the least satisfying answer: it depends, and anyone quoting a number before seeing your parts is guessing.
What actually drives the number
Visual distinctiveness matters more than volume. A defect that changes the shape of a silhouette may need a few hundred examples. A subtle finish variation under uneven light can need an order of magnitude more. Variation matters too — every product variant, fixture and lighting condition multiplies what you need to cover.
The scarcity problem
A well-run process produces few defects, which is exactly what makes the dataset hard to build. Waiting for natural occurrence can take a year for classes you most need to catch.
What to do about it
Three things, in order. Pull historical footage and scrap records, which often exist and are usually forgotten. Capture deliberately during changeovers and known-difficult runs, when defects are most likely. And where examples remain genuinely rare, reframe the problem: train on known-good and flag deviation, rather than trying to classify a defect you have twelve pictures of.
What not to do
Do not synthesise defects and treat them as real validation data. Generated examples can help a model generalise during training; they cannot tell you whether it works. The held-out set has to be real parts from your line.
Getting to an answer
A feasibility test on a sample of your own material during discovery will tell you the range for your specific case in about a week — which is cheaper than committing to a program built on a guess.