VLM verification lets an image model answer yes/no per picture and requires the attribute to be visible in at least 2 of 3. Slower and costs per search, but raises precision.

Fill in one or more fields — they are combined (AND). Region decides where appearance is compared (scene = the whole picture, content = what they say); the scene and content fields add further criteria.

Negative examples sharpen the search noticeably: the same kind of faces without the attribute. The attribute then becomes a one-click filter under Attributes.