Start with the smallest feature
State the detail the model must use: a printed character, a product shape, a surface defect, or a small object in context. Total pixel dimensions alone do not show whether that feature is legible. Ask for representative full-resolution samples and inspect the target at the intended processing size rather than judging only attractive previews.
Separate original and exported size
Record captured dimensions where known and delivered dimensions for every asset. An enlarged derivative should not be treated as newly captured detail. Specify crop, resize, and orientation rules and preserve links to original identifiers. Mixed-resolution archives may still be useful, but their limitations should be explicit before they are accepted into a shared collection.
Test the whole preparation path
Review the result after the transformations your pipeline will actually apply. A crop can remove context; a resize can make a label unreadable; a change in aspect ratio can distort shapes. Check that annotations still align and that minimum target size survives. This is more informative than imposing a large minimum image size without checking task relevance.
Choose sensible acceptance bands
Define acceptable ranges and exception handling rather than one ambiguous high-quality requirement. Separate files that fail to decode from those that decode but lack useful detail. A pilot can reveal whether stricter requirements improve task fit enough to justify narrower sourcing. Fiund should review those tradeoffs against specific proposed material, with no implied guarantee of stock coverage.
Review checklist
- Identify the smallest task-relevant visual feature.
- Request original and delivered dimensions separately.
- Inspect samples after intended transformations.
- Define acceptable ranges and exception handling.