Google's Vision Banana aims to unify vision tasks with a single pixel-generation interface

谷歌这把「香蕉」太狠了！何恺明等引爆视觉Transformer时刻

Google DeepMind and collaborators including Kaiming He introduced Vision Banana, claiming one pixel-generation interface can cover detection, segmentation, generation, and editing. The RSS snippet gives two head-to-head numbers versus Nano Banana Pro: 53.5% human win rate on GenAI-Bench and 47.8% on ImgEdit; it says only a small amount of reversible-format task data was mixed in, while data scale and full benchmark tables are not disclosed in the post.

Why it matters: HKR-H/K/R all pass: the story is a unified pixel-output interface spanning detection, segmentation, generation, and editing, with 53.5% and 47.8% benchmark figures. It stays in the 78-84 band because training scale and full benchmark coverage are not disclosed.