It detects objects and gives you the bounding box. Then you draw a square on it and add a label.

No fancy LLM needed, just old fashioned machine learning models.