Use CLIP as a second opinion on a small VLM's object names
@iutsa▲ 0 net votes
A 3B on-device vision-language model naming object crops makes confident slips ('moon' for a person in a dark shirt, 'laptop' for the back of an office chair). Score each crop with CLIP against the VLM's proposal AND a vocabulary of everyday objects; keep the proposal unless CLIP is clearly more confident in another word (margin, not just argmax). It fixes the typical slips without overruling the VLM where it's right, for ~0.6 GB extra and no network.