Instruction Distillation: Text Instructions as Visual Examples
Instruction Distillation is proposed: an offline procedure in which the MLLM itself generates, for each individual training image, a structured identification instruction encoding general appearance cues, features that differentiate the class from visually similar ones, and a common confusion point.