Vision Banana Shows That Image Generation Equals Understanding – DeepMind’s GPT‑like Leap

Background

For years, researchers have speculated that a model capable of creating high‑fidelity visual content should also be able to comprehend that content. While language models have proven this intuition—generative pre‑training yields strong understanding—vision has lagged behind, relying on supervised, contrastive, or auto‑encoding methods that do not match generative approaches.

Core Method: Vision Banana

DeepMind builds on the Nano Banana Pro (NBP) image‑generation model, creating Vision Banana without adding any dedicated perception modules. All visual tasks are reformulated as image‑generation problems: the output space of each task is parameterized as an RGB image. During training, a small fraction of task‑specific data is mixed into the generative dataset, and lightweight instruction‑fine‑tuning teaches the model to follow prompts and “draw” the desired result.

For semantic segmentation, a prompt such as “paint the skateboard pure yellow <255,255,0>” leads the model to generate an RGB mask, which can be decoded by extracting the specified color. For monocular depth, a bijective mapping bends depth values via a power‑law transform onto the edges of the RGB cube, producing a pseudo‑color image that can be precisely decoded back to metric depth. Surface normals are directly encoded as RGB channels using a right‑handed coordinate system.

2D Understanding Results

On Cityscapes, Vision Banana attains mIoU 0.699, surpassing SAM 3 (0.652) and narrowing the gap to closed‑set specialist models. For instance segmentation on SA‑Co/Gold, it achieves pmF1 0.540, comparable to DINO‑X (0.552) and well above Gemini 2.5 (0.461). In referring‑expression segmentation, Vision Banana reaches cIoU 0.738 on RefCOCOg and gIoU 0.793 on ReasonSeg, both exceeding SAM 3 Agent.

3D Understanding Results

Vision Banana estimates monocular depth without any camera intrinsics, relying solely on geometric priors learned during generative pre‑training. Across four benchmarks (NYU, ETH3D, DIODE‑indoor, KITTI) its average δ₁ score is 0.929, beating Depth Anything V3 (0.918) and leading UniK3D by ~6 percentage points. Absolute relative error is ~20 % lower than MoGe‑2. A field test measured a 13.71 m depth estimate versus a ground‑truth 12.87 m, yielding an error of only 0.065.

For surface‑normal estimation, Vision Banana achieves the lowest mean and median angular errors on indoor datasets and matches Lotus‑2 on outdoor scenes, while its rendered normal maps show superior visual fidelity.

Generation Capability Verification

Human‑preference evaluations on GenAI‑Bench (text‑to‑image) and ImgEdit (image editing) show Vision Banana retains the generative strength of Nano Banana Pro, with win rates of 53.5 % and 47.8 % respectively, confirming “generation without forgetting understanding.”

Paradigm Shift and Limitations

The study validates two key claims: (1) image generators act as general‑purpose visual learners, and (2) image generation can serve as a universal interface for diverse vision tasks. However, the current evaluation focuses on single‑image inputs; multi‑view or video extensions remain open, and inference cost is still higher than lightweight expert models.

Conclusion

Vision Banana provides concrete evidence that generative pre‑training can unify visual generation and understanding, marking a potential “GPT moment” for computer vision and paving the way toward foundational visual models and vision‑centric AGI.