Google's Vision Banana turns visual perception into image generation

For years, computer vision has maintained a strict division of labor: image generators draw pretty pictures, while perception models like segmenters and depth estimators do the actual understanding.

Google DeepMind just decided to tear that wall down.

Google researchers published a paper titled "Image Generators are Generalist Vision Learners" and introduced Vision Banana. It is a unified model for both image understanding and generation, built by instruction-tuning Nano Banana Pro alongside a small mixture of vision task data.

The clever trick is deceptively simple: instead of designing custom prediction heads for each task, Vision Banana treats every perception problem as a standard RGB image generation task.

Perception reframed as rendering

Ask Vision Banana to segment a cat, an exit sign, or every single clove of garlic in a cluttered photo, and it simply generates a cleanly masked, color-coded image.

It works the same way for complex 3D geometry. Hand it a dense forest scene, and it generates an ultra-detailed metric depth map or a surface normal map showing how every leaf and branch is angled in 3D space. Combined with camera intrinsics, these generated depth maps can be directly unprojected into full 3D point clouds.

All of this was achieved through lightweight instruction-tuning without degrading the base model's original image generation performance.

Even more impressive is the benchmark performance. In zero-shot transfer evaluations across 2D and 3D tasks, Vision Banana achieved state-of-the-art results, beating or rivaling dedicated specialist models like Meta's Segment Anything Model 3 and the Depth Anything series.

Why it matters

We are likely watching computer vision experience its own GPT moment.

In natural language processing, broad text generation eventually replaced dozens of specialized narrow models. Vision Banana suggests that image generation pretraining does the exact same thing for computer vision: teaching the model a deep, general representation of the visual physical world.

For developers, this points toward a future where you do not need to juggle separate pipelines for segmentation, depth mapping, and creative rendering. A single unified generative interface can handle both perception and creation.

While outputting perception as RGB images raises questions about precision limits and inference latency, Google's proof of concept makes one thing clear: generative vision pretraining is no longer just for making art, it is becoming the foundation of machine vision.