Google DeepMind's Vision Banana Beats SAM 3 and Depth Anything V3 — With Zero Specialist Training
A new Google DeepMind paper, “Image Generators are Generalist Vision Learners” (arXiv:2604.20329, published April 22, 2026), collapses the long-standing divide between generative and discriminative computer vision. The result is Vision Banana: a single model that generates photorealistic images and outperforms specialist systems across segmentation, depth estimation, and surface normals — without ever being trained on the evaluation benchmarks.
The paper draws a direct parallel to how large language models are built. LLM pretraining on text creates rich representations that transfer to downstream tasks via fine-tuning. The Google team argues image generation training does the same for vision: to generate photorealistic images, a model must internally represent geometry, semantics, depth, and object boundaries. Vision Banana simply teaches its base model, Nano Banana Pro, to express that latent knowledge in decodable formats.
The instruction-tuning pass is minimal. A small proportion of computer vision task data is mixed into the original training distribution at a low ratio. All task outputs are represented as RGB images using invertible color mappings, so the model produces a new image as its answer — one that can be decoded back into segmentation masks, metric depths, or surface normals. No decoder heads, no task-specific architectures, no real-world training data for depth estimation.
Benchmark Results
All results are in zero-shot transfer settings — no evaluation dataset training data included:
| Task | Vision Banana | Specialist | Delta |
|---|---|---|---|
| Semantic seg. (Cityscapes val, mIoU) | 0.699 | SAM 3: 0.652 | +4.7 pts |
| Referring expression seg. (RefCOCOg UMD val, cIoU) | 0.738 | SAM 3 Agent: 0.734 | +0.4 pts |
| Reasoning segmentation (ReasonSeg val, gIoU) | 0.793 | SAM 3 Agent: 0.770 | +2.3 pts |
| Metric depth estimation (avg δ1, 4 datasets) | 0.929 | Depth Anything V3: 0.918 | +1.1 pts |
| Surface normal estimation (mean angle error, 4 datasets) | 18.928° | Lotus-2: 19.642° | −0.7° |
On instance segmentation (SA-Co/Gold), Vision Banana posts pmF1 of 0.540 — on par with DINO-X at 0.552, and ahead of Gemini 2.5 (0.461) and OWLv2 (0.420) in zero-shot transfer. The depth model uses no camera intrinsics or extrinsics; absolute scale is inferred from visual cues embedded during pretraining, and the training data is entirely synthetic.
Critically, Vision Banana does not degrade as an image generator. On GenAI-Bench it achieves a 53.5% win rate against Nano Banana Pro — effectively parity — and a 47.8% win rate on ImgEdit, where the base model scores 52.2%.
Why This Matters
The implication is structural. If generation pretraining creates generalist visual representations in the same way language pretraining creates generalist semantic representations, the field may not need separate architectures for generative and discriminative vision tasks. A single model trained to generate images could, with minimal instruction-tuning, handle most computer vision applications.
The paper is authored by a team at Google DeepMind with He Kaiming and Xie Saining among the contributors — two researchers whose past work on ResNets and masked autoencoders has defined modern vision architectures. Their involvement signals that this is not a side project.
The practical near-term implication: Nano Banana Pro, already Google’s flagship image generator, appears to have learned a generalist visual understanding that no one explicitly trained it for. The next version of Vision Banana, or a future Nano Banana variant with integrated vision tasks, could become the single model for generation, segmentation, and depth in production pipelines. Labs building on separate CLIP, SAM, and depth model stacks should be watching this closely.