Vision Banana turns image generators into all-purpose vision models

Computer vision has spent years building separate specialists for segmentation, depth, and other perception tasks. Google DeepMind's new paper suggests that stack may be collapsing into one broader pattern: train a good enough image generator, then teach it to answer vision problems through the same generative interface. The publication page frames the result as a possible paradigm shift, not a minor benchmark tweak.

The paper's central claim is bold. DeepMind says image generation pretraining can play the same foundational role for vision that language-model pretraining played for text. Its model, Vision Banana, starts from Nano Banana Pro and is instruction-tuned on a mix of original image-generation data plus a small amount of vision-task data. Instead of switching to separate heads or bespoke output formats, the team represents vision tasks as RGB image generation.

That design choice is what makes the paper interesting. If segmentation, depth estimation, and other perception tasks can all be expressed through the generator itself, model builders get a single interface instead of a growing pile of special-purpose systems. DeepMind says Vision Banana reaches state-of-the-art results across multiple 2D and 3D understanding tasks, beating or rivaling specialist families including Segment Anything on segmentation and Depth Anything on metric depth estimation.

Another part of the claim is economic, even though the paper does not present it as a pricing story. DeepMind says the gains come from lightweight instruction tuning rather than rebuilding the model from scratch, and that the system keeps its original image-generation ability after the adaptation. That points to a future where the same base model can create, segment, estimate depth, and possibly handle other perception jobs without spinning up a separate model family for each one.

This is still an arXiv result, so the next question is how broadly the approach holds up outside the paper's benchmark mix. But the directional signal is hard to miss. If image generators really are generalist vision learners, the industry may stop treating generation as a flashy side capability and start treating it as the training route to a universal visual foundation model.