This example is using bounding boxes, but it turns out Gemini 2.5 (both Pro and Flash) take that a step further and can return complex shaped segmentation masks identifying objects too: https://simonwillison.net/2025/Apr/18/gemini-image-segmentat...