Field guide
Where an image-description model guesses
A description model produces the most probable wording for visible patterns. It does not mark every phrase with the evidence that supports it.
01
Small objects, occluded hands and distant text invite confident substitutions. Compare those nouns with the pixels before treating them as facts.
02
Relationships are harder than objects. The frame may show two people and a bicycle without proving who owns it, who is riding it or whether they arrived together.
03
Location and event labels often come from visual stereotypes. Architecture, clothing and weather can suggest a place without establishing one.
04
Use tags and captions as retrieval aids, then verify consequential details manually. Accessibility text should describe observable content and avoid unsupported motives, identities or diagnoses.