Open Vision Banana
arXiv · 2026

Open Vision Banana

A Fully Open Recipe for
Generative Visual Perception

Arpit Jadon1,2,* Valentinos Pariza1,* Jona Ruthardt1 Ryousuke Yamada1,3 A. Sophia Koepke5,6 Andrei Bursuc4 Yuki M. Asano1
1University of Technology Nuremberg 2DLR 3AIST 4valeo.ai 5Technical University of Munich 6University of Tübingen

*Equal contribution

Paper arXiv Code Models
A banana character looking through a magnifying glass ✦ ✦

TL;DR

Turn an image generator into a
general-purpose perception model.

Open Vision Banana fine-tunes open-weight FLUX.2 Klein 4B and 9B Base image generators on public data. The same model generates an RGB image for segmentation or geometry, then a deterministic decoder reads out the prediction. No task-specific prediction heads or losses are added.

2open base model sizes
6perception tasks
163released checkpoints

The unified interface

Perception, expressed as image generation.

One model, multiple tasks. Explore inputs and predictions from the paper, including image editing retained after perception training.

Input image Cityscapes street scene with parked cars
Task instruction · summarized Paint the requested Cityscapes classes with their assigned colors. OVB image generator
Readout nearest prompted color → class label
Decoded semantic mask Hover to reveal input
Color-coded semantic mask of the street, buildings, cars, and sky
Semantic segmentation · Cityscapes
Example 1 of 3

Measured performance

A shared model for six perception tasks.

OVB takes on segmentation, relative depth, and surface normals with a shared model.

How does OVB compare?

OVB's non-zero-shot Cityscapes score exceeds the reported zero-shot scores of Vision Banana and SAM 3. Training and inference setups differ across methods.

Selected comparisons across six perception tasks from the main paper; OVB results are non-zero-shot and comparison setups differ
Task & benchmarkOpen Vision BananaVision BananaSelected specialists
Semantic segmentation Cityscapes val · mIoU ↑75.79B · non-zero-shot · 20-pass vote
69.7 mean single pass		69.9SAM 3 65.2
Referring segmentation RefCOCOg val (UMD) · cIoU ↑69.29B · non-zero-shot
Original expressions		73.8SAM 3 + Gemini 2.5 Pro 73.4HybridGL 51.3
Reasoning segmentation ReasonSeg val · gIoU ↑74.64B · non-zero-shot
+ Gemma4 31B		79.3+ Gemini 2.5 ProSAM 3 + Gemini 2.5 Pro 77.0SegZero-Qwen2.5-VL-7B 62.6
Instance segmentation SA-Co/Gold · cgF1 ↑40.19B · non-zero-shot · 11-pass consensus
+ Qwen3-VL-30B-A3B presence gate		47.5+ Gemini 3.1 Flash-LiteOWLv2 24.6DINO-X 21.3
Relative depth NYUv2 · δ₁ (%) ↑99.09B · non-zero-shot
Per-image disparity-affine alignment		94.8Metric depthMarigold V2 98.0Lotus-2 97.6
Surface normals NYUv2 · mean error (°) ↓13.214B · non-zero-shot
Camera-space normals		17.8Marigold V2 16.6Lotus-2 16.9DSINE 16.4

A central finding

Test-Time Scaling for Vision Models

The model often finds the right regions but assigns the wrong colors. Repeating predictions with different color legends and voting on the decoded labels improves accuracy without additional training.

  1. Change the color legend

    Request the same categories with different color assignments.

    Pass 1RoadCar
    Pass 2RoadCar

  2. Predict and decode

    The same OVB model generates each image. Decode colors into common class labels.

    PurpleRoadPass 1
    YellowRoadPass 2

  3. Vote on the labels

    For each pixel, keep the category selected by the most decoded predictions.

    RoadRoadCar
    MajorityRoad
Cityscapes · 9B · zero-shot65.4 → 74.06mIoU, first pass → peak vote (27, 28, or 30 passes)
COCO-Object · 9B · non-zero-shot44.9 → 59.3mIoU, first pass → 90-pass vote
Cityscapes mIoU climbs as predictions from more color legends are combined, while individual pass scores stay approximately flat
Cityscapes repeat sweep · zero-shot models

The gain comes from correcting color binding.

Voting improves label accuracy while region quality stays nearly unchanged. On COCO-Object, wrong-class foreground pixels fall from 19.0% to 10.3% over 90 passes.

WHAT MAKES THE RECIPE WORK

Designing Generative Perception Models

Three takeaways from studying target encoding, training data, and model adaptation.

01
RGB interface

How labels map to colors matters.

Training with a fixed color legend yields 54.3 mIoU on Cityscapes, but only 7.1 when the colors change. Randomized assignments teach the model to follow the prompt.

02
Multi-task recipe

Balance performance across tasks.

More data or longer training can help one task while hurting another. The final recipe balances segmentation, geometry, and retained generative abilities.

03
Model adaptation

Adapt more of the backbone.

Tuning only the last few transformer blocks performs poorly. Strong perception requires adapting a substantial fraction of the backbone.

Overview of the Open Vision Banana research questions, data, generator, RGB output and decoder
The complete recipe. Public data, target encoding, prompts, model adaptation, prediction readout, and test-time scaling.

Open by construction

A recipe others can inspect, reproduce, and extend.

Built with public base models and datasets. We release training and evaluation code alongside model weights, including intermediate and ablation checkpoints.

Model weights

View models ↗

Code repository

View code ↗

Paper

Read on arXiv ↗

Citation

Using Open Vision Banana?

Please cite OVB if you use it in your work.

@misc{jadon2026openvisionbanana,
  title   = {Open Vision Banana: A Fully Open Recipe for Generative Visual Perception},
  author  = {Jadon, Arpit and Pariza, Valentinos and Ruthardt, Jona and
             Yamada, Ryousuke and Koepke, A. Sophia and Bursuc, Andrei and Asano, Yuki M.},
  year    = {2026},
  eprint  = {2611.15844},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url     = {https://arxiv.org/abs/2611.15844}
}