🤯 Vision Banana: AI Learns Like a Human! 🍌

VISION BANANA: A UNIFIED MODEL REVOLUTIONIZING COMPUTER VISION
The computer vision community has historically operated with two distinct approaches: generative models focused on image creation and discriminative models dedicated to understanding images. The prevailing assumption was that models proficient in generating images wouldn’t necessarily excel at visual understanding tasks. However, a recent groundbreaking paper from Google, “Image Generators are Generalist Vision Learners,” published in April 2026, challenges this established paradigm.

INTRODUCING VISION BANANA: A SINGLE MODEL ACROSS MULTIPLE TASKS
A team of Google DeepMind researchers introduced Vision Banana, a unified model that surpasses or matches the performance of specialized systems across a broad spectrum of visual understanding tasks. These tasks include semantic segmentation, instance segmentation, monocular metric depth estimation, and surface normal estimation – all while retaining the image generation capabilities of its underlying base model, Nano Banana Pro (NBP). This innovative approach mirrors the established playbook for large language models: initial pretraining followed by instruction tuning.

NANO BANANA PRO (NBP) AS THE FOUNDATION
At the core of Vision Banana is NBP, Google’s state-of-the-art image generator. The team achieved this breakthrough through a lightweight instruction-tuning process, subtly incorporating computer vision task data into NBP’s original training mixture. This key insight—that generating photorealistic images inherently requires a model to understand geometry, semantics, depth, and object relationships—led to Vision Banana’s ability to express this latent knowledge in measurable, decodable formats.

INSTRUCTION TUNING AND RGB FORMATS
Crucially, no training data from the evaluation benchmarks was included in the instruction-tuning mixture. This ensured that all results reflected true performance. The model was instruction-tuned to produce visualizations that adhere to precise, invertible color schemes – meaning generated images can be decoded back into quantitative outputs for benchmark evaluation. This strategy yields three significant advantages: first, it supports a wide variety of tasks with a single unified model – only the prompt changes, not the model weights; second, it requires relatively little new training data, as instruction-tuning focuses on teaching the model how to format computer vision outputs as RGB; and third, it preserves the model’s original image generation capabilities, as the outputs are simply new RGB images.

SPECIFIC TASK IMPLEMENTATIONS
For semantic segmentation, the model is prompted with instructions such as: “Generate a segmentation visualization of this image, using the color mapping: {‘cat’: ‘red’, ‘background’: ‘yellow’}.” Each pixel is colored according to its predicted class, and the color assignments are specified in the prompt, eliminating the need for a fixed label vocabulary. Instance segmentation utilizes a per-class inference strategy, running a separate pass for each class and dynamically assigning unique colors to each instance, with masks recovered by clustering pixels with similar colors using a threshold. Metric depth estimation employs a bijective mapping between unbounded metric depth values and bounded RGB values, using a power transform to “curve” the depth values and encode them as a false-color visualization following a 3D Hilbert curve. This transform is strictly invertible, allowing the generated depth image to decode cleanly back to physical distances. Notably, no camera parameters—intrinsics or extrinsics—are required during training or inference.

SURFACE NORMAL ESTIMATION AND COLOR CODING
Surface normal estimation employs a direct mapping: surface normals, represented as unit vectors ranging from -1.0 to 1.0, map naturally to RGB channels. Facing-left normals encode as pinkish-red; facing-up normals encode as light green; normals pointing toward the camera encode as light blue/purple.

KEY RESULTS AND ZERO-SHOT PERFORMANCE
The research team’s results across benchmarks, conducted in zero-shot transfer settings (where the model has never seen training data from the evaluated datasets), demonstrate the significant impact of this approach. On generative benchmarks, Vision Banana achieves a 53.5% win rate against Nano Banana Pro on GenAI-Bench (text-to-image) and a 47.8% win rate on ImgEdit (image editing), while Nano Banana Pro scores 52.2% and 52.2% respectively. These findings confirm that lightweight instruction-tuning does not degrade the model’s generative capabilities, solidifying Vision Banana as a revolutionary advancement in computer vision.