Google DeepMind presents a generalist model that pushes the boundaries of computer vision

The Google DeepMind research team has demonstrated with the Vision Banana model that image generation precursors serve as a strong foundation for general understanding of the visual world, much like large language models (LLMs) develop language understanding through next-word prediction. The system is based on Nano Banana Pro, Google's most advanced image generator, which has been transformed into Vision Banana through lightweight instruction-based learning. The key innovation is that various computer vision tasks, such as segmentation, depth determination, and surface normal estimation, have been transformed into RGB image generation tasks.

Vision Banana achieved superior results in so-called “zero-shot” environments, where the model has no prior experience with specific datasets. It outperformed the SAM 3 model in image segmentation, while achieving a depth metric score of 0.929 (δ1 parameter), beating the previous record holder Depth Anything V3 (0.918). What is particularly impressive is that the model does not require any information about camera parameters to determine depth, which has been a major obstacle for such systems until now.

This approach provides three key advantages. A single model where a single neural network can perform a wide range of tasks, with only the text prompt changing. Only a small amount of specific visual data was required to adapt the model. Furthermore, despite the new analytical capabilities, Vision Banana still fully retains its original function of generating superb photorealistic images.

The researchers believe that we are witnessing a paradigm shift where generative pre-learning will become the standard for building general visual models of the future. Vision Banana is not just a new tool, but evidence that the ability to create visual content implicitly requires a deep understanding of geometry, semantics, and spatial relationships in the real world.