Google DeepMind Unveils Vision Banana: A Unified Image Generator Excelling in Segmentation and Depth Estimation

The field of computer vision has traditionally divided into generative and discriminative models, with the belief that proficiency in image creation does not equate to image comprehension. However, a recent study from Google DeepMind challenges this notion with the introduction of Vision Banana, a versatile model that excels in multiple visual understanding tasks. This includes semantic and instance segmentation, as well as metric depth estimation, while maintaining its image generation capabilities. The underlying model, Nano Banana Pro, serves as a foundation, where a unique instruction-tuning process integrates minimal computer vision data into its training. This approach allows Vision Banana to leverage its generative training to grasp essential visual concepts such as depth and object relationships. The model's design eliminates the need for specialized components for different tasks, instead producing outputs as RGB images that can be easily interpreted. In benchmark tests, Vision Banana has shown impressive results, surpassing specialist models like SAM 3 and Depth Anything V3 in various tasks, all without requiring real-world training data or camera parameters. This advancement signifies a major leap in how visual tasks can be approached, emphasizing the potential of generative models in understanding visual data.