Vision Banana unifies vision via generative learning

Vision Banana marks the transition from task-specific vision specialists to generalist "Visual GPT" models that understand the world by learning to "draw" it.

• –Reframes diverse perception outputs (masks, depth maps, surface normals) as RGB images, creating a single unified interface for all computer vision.
• –Demonstrates that generative image pretraining serves a foundational role equivalent to language modeling, unlocking emergent zero-shot capabilities.
• –Outperforms domain-specific models like SAM 3 and Depth Pro without requiring task-specific architectures or camera intrinsics.
• –Proves that training models to generate photorealistic images implicitly forces them to learn complex internal representations of geometry and semantics.
• –Signals a paradigm shift where specialized vision heads are replaced by a single instruction-following generative foundation model.