DeepMind's Vision Banana beats SAM 3 and Depth Anything V3 — image-gen pretraining as the new vision paradigm

Google DeepMind introduced Vision Banana, an instruction-tuned image generator that outperforms SAM 3 on segmentation and Depth Anything V3 on metric depth estimation. The argument: image-generation pretraining is to vision what GPT-style next-token pretraining was to NLP.