Guide · Pillar: Prove
It depends — here are the variables and honest ranges.
For evaluation, you need fewer images than teams fear: 100–300 well-chosen examples expose most decisive differences between approaches. For training, ranges span orders of magnitude: zero for a prompted VLM, tens per class to start moving a fine-tuned model, hundreds or thousands per class for subtle distinctions. Any single number quoted without your task’s variability, class count, and accuracy bar attached is a guess — the table below states its uncertainty instead of hiding it.
Published August 31, 2026 · Pillar: Prove · By the VisionRoute team
Definitions
Two different questions
Representative
Who this is for
Why ranges, not numbers
The scenario table
Ranges by scenario, with the uncertainty stated
These are starting budgets drawn from practitioner experience, not measurements — the fourth column says how much to trust each row. Terms are in the glossary.
| Scenario | Training images | Evaluation images | Confidence in this range |
|---|---|---|---|
| Prompted VLM, no training | 0 — plus 5–10 exemplars if you use few-shot prompting. | 100–300, built backwards from failure. | High that zero works to start — that is the point of prompting. The evaluation need never goes away. |
| Specialist API, used as-is | 0. | 100–300, weighted toward your hard cases and negatives. | High — but "the API covers your task" is exactly the claim the evaluation set exists to check. |
| Fine-tune a pretrained classifier, visually distinct classes | Roughly 30–100 per class before results move. | 50+ per class you care about. | Medium — class similarity and image variability swing this by multiples in both directions. |
| Fine-tune for subtle distinctions (defects, grades, lookalikes) | Hundreds per class, often more. | 100+ per decisive class, borderline cases included. | Low–medium — subtlety is precisely what inflates data needs. Expect iteration, not a number. |
| Fine-tune a detector on custom classes | Tens to hundreds of labeled boxes per class. | 50+ images per class, including crowding and occlusion. | Medium — label quality (tight, consistent boxes) matters about as much as label count. |
| Train from scratch | Many thousands per class. | A large held-out set on top. | High that you should not start here — without a specialist team and evidence that pretrained models fail your task, scratch training is a research project. |
The stopping rule
Collect until the decisive metric stops moving
Label the evaluation set first — always
The first 100–300 labeled images become the evaluation set, untouched by training. Teams that train on everything and evaluate on leftovers end up grading the model on its own homework.Grow training data in increments
Start near the low end of your scenario’s range, then roughly double: 50, 100, 200, 400. Retrain and re-evaluate at each step, watching the decisive slice — not the average.Stop when the curve flattens
When the decisive-slice metric stops improving across two consecutive increments, more of the same data has stopped paying. Spend the remaining labeling budget on rare slices, hard negatives, and fresh failure cases instead.Let the costly class set the budget
With imbalanced classes, the average is a decoy. If the rare class is the expensive one — the defect, the violation — its count is your binding constraint, and "enough images overall" can still mean "not nearly enough of the ones that matter".
When this changes
More classes, more budget
A higher accuracy bar
High variability capture
Synthetic augmentation
Limitations
What this table cannot promise
- No universal number exists, here or anywhere. The ranges order scenarios relative to each other; your task can still land outside its row, which is what the stopping rule is for.
- The table assumes competent labels. Label noise quietly raises every count — disagreeing annotators are a definition problem to fix before a volume problem.
- It budgets data, not outcomes. Whether the resulting stack meets your bar is settled by evaluation — see how to test on your own data.
In short
Evaluation needs 100–300 well-chosen images for most narrow tasks; training needs anywhere from zero (prompted VLMs) to tens or hundreds per class (fine-tuning) to thousands (subtle distinctions or scratch training). Treat every count as a starting budget with a stopping rule — grow data in increments and stop when the decisive-slice metric flattens — and let the rare-but-costly class, not the average, set the budget.
Related: the feasibility checklist · test models on your data · feasibility testing explained