Vision banana!!!!

Google DeepMind has unveiled Vision Banana, a novel instruction-tuned image generator that achieves top-tier performance on both segmentation and metric depth estimation tasks. According to a recent paper, Vision Banana outperforms SAM 3 on segmentation benchmarks and Depth Anything V3 on metric depth estimation, two traditionally separate domains in computer vision. The model leverages instruction tuning, allowing it to interpret natural language prompts to generate precise object masks and depth maps from a single image input. This unified approach eliminates the need for multiple specialized models, streamlining workflows for applications like autonomous driving, robotics, and augmented reality.

Vision Banana's key innovation lies in its ability to handle diverse visual tasks through a single architecture, trained on a large-scale dataset combining segmentation and depth annotations. Early results show it achieves higher Intersection over Union (IoU) for segmentation and lower absolute relative error for depth estimation compared to prior state-of-the-art models. This breakthrough could simplify deployment in real-world systems, reducing computational overhead and latency. The model is open-source, with code and weights available on GitHub, enabling researchers and developers to integrate it into their pipelines. Vision Banana represents a step toward generalist vision models that can perform multiple tasks with human-like understanding, potentially transforming industries reliant on visual perception.