Open Vision Banana Beats Google on Cityscapes Benchmark

The University of Technology Nuremberg's Fundamental AI Lab released Open Vision Banana on November 13, 2026, a fully open reproduction of Google's proprietary Vision Banana unified perception model. The 9B parameter checkpoint achieves 73.9 mIoU on Cityscapes in zero-shot evaluation, surpassing Vision Banana's reported 69.9 score while using only public training data and releasing complete weights, code, and training pipelines.
Benchmark Performance and Architecture
Open Vision Banana builds on the open-weight FLUX.2-Klein image generator at 4B and 9B scales, fine-tuned on 2.1 to 3.8 million training pairs from public datasets including COCO-Stuff, ADE20K, BDD100K, and NYUv2. The headline Cityscapes result depends on a novel test-time method the authors call legend ensembling, which runs multiple forward passes under different color-to-class legends and fuses outputs via per-pixel majority vote. A single forward pass reaches only 65.4 mIoU; legend ensembling adds 8.7 points on Cityscapes and 15.0 points on COCO-Object without any retraining.
On depth estimation, the model rivals state-of-the-art affine-invariant estimators while retaining the base model's image-editing capabilities. The research collaboration includes teams from DLR, AIST, the University of Tübingen, TUM, and valeo.ai.
Performance Gaps on Other Tasks
While Open Vision Banana matches or exceeds Vision Banana on semantic segmentation, performance gaps remain on other perception tasks. The model comes within five points of Vision Banana on referring and reasoning segmentation benchmarks but trails significantly on instance segmentation, scoring 37.4 cgF1 versus Vision Banana's 47.5 on the SA-Co/Gold dataset. These results suggest the open reproduction captures core segmentation understanding but lacks some of the refinement present in Google's proprietary training pipeline or architecture choices.
Release Details and Training Costs
Officially released on November 13, 2026, the package includes the complete data pipeline, training and evaluation code, and both final and intermediate checkpoints. Training the 4B model required approximately 73 hours on 16 GH200 GPUs, while the 9B version took roughly 79 hours on 32 GH200s. The unusually complete release lowers the barrier for academic teams and independent researchers to build on unified perception models without proprietary dependencies.
- 9B model: 73.9 mIoU on Cityscapes (zero-shot)
- Legend ensembling adds 8.7 mIoU on Cityscapes, 15.0 on COCO-Object
- Instance segmentation: 37.4 cgF1 vs. Vision Banana's 47.5
- Training data: 2.1 to 3.8M pairs from public datasets
What This Means
Open Vision Banana demonstrates that competitive unified perception models can be built entirely from public components, breaking the perception that frontier vision capabilities require proprietary datasets or closed training infrastructure. The Cityscapes lead validates the legend ensembling approach as a practical inference-time improvement, though the instance-segmentation gap shows remaining work to match every dimension of proprietary baselines. By shipping complete artifacts, the release accelerates community development of open alternatives to closed vision systems and provides a benchmark for future open-source vision research.
on Emergent today





