For years, computer vision ran on two separate rails: generative models, which create images, and discriminative models, which understand them. The industry’s standing assumption — a model that paints well doesn’t necessarily read well — just got overturned. On April 22, 2026, the Google DeepMind team published a paper titled Image Generators are Generalist Vision Learners, introducing Vision Banana: a unified generalist vision model that matches or beats state-of-the-art specialist systems on semantic segmentation, instance segmentation, monocular metric depth estimation, and surface normal estimation — while keeping its original image-generation ability intact.
The Core Idea: The LLM Analogy, Transplanted to Vision
Anyone who knows large language models knows the two-stage recipe: generative pretraining on vast text, then instruction tuning to adapt to downstream tasks. Pretraining builds rich internal representations that transfer to almost anything. The Google team’s central claim: image-generation training plays exactly the role that LLM pretraining plays for language. Their base model, Nano Banana Pro (NBP), is Google’s most advanced image generator. By mixing a tiny proportion of computer-vision task data into NBP’s original training diet and running lightweight instruction tuning, they produced Vision Banana.
The key insight: generating photorealistic images implicitly requires understanding geometry, semantics, depth, and object relations. Vision Banana learns to express that latent knowledge in measurable, decodable formats. And crucially, the instruction-tuning data contains none of the evaluation benchmarks’ training data — every reported result reflects genuine generalization, not in-domain memorization.
How It Works: Perception as Image Generation
Vision Banana adds no specialized decoder heads or regression modules. Instead, every vision-task output is parameterized as an RGB image: the instruction-tuned model produces visualizations following precise, invertible color schemes, which means the generated images can be decoded back into quantitative outputs for benchmark scoring. The team identifies three advantages of this strategy. A unified model serves many tasks — after instruction tuning, switching tasks changes only the prompt, never the weights. It is data-efficient, because instruction tuning mostly teaches the model how to format CV outputs as RGB. And it preserves generation ability, since a task output is just another RGB image.
The Tasks, Concretely
Semantic segmentation: the model receives an instruction like “generate the segmentation visualization of this image using the color mapping: cat red, background yellow.” Every pixel is colored by predicted class, and because the color assignment lives in the prompt, no fixed label vocabulary is needed.
Instance segmentation: instance counts are unknown, so Vision Banana uses per-class inference — one pass per class, dynamically assigning each instance a unique color, then recovering masks by clustering similar-colored pixels.
Metric depth estimation: a bijective mapping folds unbounded depth values [0, ∞) into bounded RGB [0,1]³. A power transform (shape λ = -3, scale c = 10/3) first bends the depth values, then encodes them as a pseudo-color visualization traversing the edges of the RGB cube along a 3D Hilbert curve. The transform is strictly invertible, so generated depth images decode cleanly back to physical distances. Notably, no camera parameters — intrinsic or extrinsic — are needed at training or inference: the model infers absolute scale from visual cues and world knowledge embedded during pretraining, and the depth training data is entirely synthetic, rendered by simulation engines with zero real-world depth data.
Surface normal estimation: the mapping is almost free — surface normals are unit vectors (x, y, z) in [-1, 1], which maps naturally onto RGB channels. Left-facing normals encode pink; upward-facing normals light green; camera-facing normals light blue-purple.
Beating the Specialists, by the Numbers
All results are zero-shot transfer — the model never saw any training data from the evaluation sets:
2D semantic understanding: semantic segmentation on Cityscapes val — mIoU 0.699 versus SAM 3’s 0.652, a 4.7-point lead. Referring segmentation on RefCOCOg UMD val — cIoU 0.738, edging SAM 3 Agent’s 0.734. Reasoning segmentation on ReasonSeg val — gIoU 0.793, beating SAM 3 Agent’s 0.770 and even surpassing non-zero-shot methods trained with in-domain data. Instance segmentation on SA-Co/Gold — pmF1 0.540, on par with DINO-X (0.552) and clearly ahead of Gemini 2.5 (0.461), APE-D (0.369), and OWLv2 (0.420).
3D structural understanding: metric depth — average δ1 of 0.882 across six major benchmarks; on the four datasets evaluated by Depth Anything V3 (NYU, ETH3D, DIODE-Indoor, KITTI), Vision Banana scores 0.929 against 0.918 — with zero real-world training data and zero camera parameters. Surface normals — average angular error of 18.928° across four datasets versus Lotus-2’s 19.642°; on indoor datasets it posts the lowest mean (15.549°) and median (9.300°) angular error of any compared method.
Generation preserved: on generative benchmarks Vision Banana is statistically indistinguishable from its base — a 53.5 percent win rate over Nano Banana Pro on GenAI-Bench (text-to-image) and a competitive 47.8 percent on ImgEdit (image editing, against NBP’s 52.2 percent). Lightweight instruction tuning does not tax generation.
The Takeaways
- Image-generation pretraining produces generalist vision learners. As LLM pretraining unlocked emergent language understanding, training on image generation naturally develops internal visual representations that transfer to segmentation, depth, and normals.
- No expert architecture needed to beat expert models. Lightweight instruction tuning of Nano Banana Pro surpasses SAM 3 on three segmentation benchmarks, Depth Anything V3 on metric depth (δ1: 0.929 vs 0.918), and Lotus-2 on normals — all zero-shot.
- Every vision task becomes image generation. Parameterizing task outputs as decodable RGB means one weight set and prompt-only switching across segmentation, depth, and normals — no task-specific modules.
- Metric depth without cameras or real data. The bijective power transform maps depth into RGB color space; absolute scale comes from visual context alone, and training data is purely synthetic.
- Image generation as vision’s universal interface. As text generation unified language tasks, image generation may become the universal output interface for computer vision.
The Paradigm Signal
Vision Banana may mark a genuine paradigm shift in computer vision. Just as text generation became the unifying interface for language understanding and reasoning, image generation may become the universal interface for visual understanding — generative visual pretraining building one foundation model that both creates and comprehends. Practically, it suggests a future where nobody trains a specialist model per vision task; you instruction-tune one unified generator. Philosophically, it answers a long-running question — whether generation and understanding are fundamentally related — with a resounding yes: a model capable of creating visual content necessarily understands it too.
Paper: arXiv:2604.20329. Project page: vision-banana.github.io. Source: Google DeepMind.
