Vision Banana General-Purpose Vision Model

Vision Banana is a general-purpose vision model created by instruction-tuning the image generation model Nano Banana Pro (NBP) with its original training data plus a small amount of vision task data. It reframes recognition problems as generation problems by parameterizing the output space of vision tasks as RGB images.

With this approach, the model achieved strong performance across both 2D and 3D understanding. In particular,

• in segmentation, it surpassed or matched Segment Anything Model 3,
• in metric depth estimation, it surpassed or matched the Depth Anything family.

The key point is that lightweight instruction-tuning alone added understanding capability while preserving the original image generation ability. The authors believe that image generation pretraining creates general-purpose visual representations, much like LLM pretraining, and that image generation can serve as a unified interface tying together vision tasks. This shows that generative vision pretraining could become a central pillar of future foundational vision models.