Gather three or more near-identical items in one frame, photograph close enough to show the differences, write what distinguishes them. Obviously different objects, groups under three and unnamed features are refused; ten per contributor, from a global base strongest in Southeast Asia, the United States and Europe. The note is free text that tells you where to look; object classes are uncontrolled, spreading the set across household, retail and workshop items: general fine-grained discrimination, not one domain. Fine-grained data usually needs an expert labeler; here it is the owner, who knows which is which. A VLM that can name the distinguishing feature is doing something most still cannot; the human answer is the reference.