Open Vision Banana Beats Proprietary Baseline on Cityscapes

A new paper from the University of Technology Nuremberg's Fundamental AI Lab, with collaborators at DLR, AIST, Tübingen, TUM and valeo.ai, claims to reproduce the proprietary Vision Banana model using only public models and public data, and to beat it on at least one headline benchmark. Open Vision Banana fine-tunes the open-weight FLUX.2-Klein generator at 4B and 9B scale on 2.1 to 3.8 million training pairs drawn from public datasets including COCO-Stuff, ADE20K, BDD100K and NYUv2.

The headline figure is zero-shot semantic segmentation on Cityscapes, where the 9B model scores 73.9 mIoU against Vision Banana's reported 69.9. On depth estimation the paper reports results that "rivals state-of-the-art affine-invariant depth estimators," while retaining the base model's image-editing ability. The gap is wider elsewhere. "Its non-zero-shot variants come within five points of Vision Banana on referring and reasoning segmentation but remain further behind on instance segmentation," the authors write; the specific instance-segmentation score is 37.4 cgF1 versus Vision Banana's 47.5 on SA-Co/Gold.

The 73.9 number depends on a test-time trick the authors introduce called legend ensembling. A single forward pass on Cityscapes only reaches 65.4 mIoU; running the model multiple times under different color-to-class legends and fusing the outputs by per-pixel majority vote adds 8.7 points, and the same method adds 15.0 points on COCO-Object.

The release is unusually complete for an academic reproduction: the data pipeline, training and evaluation code, and both final and intermediate checkpoints all ship alongside the paper. Training itself is not cheap, running roughly 73 hours on 16 GH200 GPUs for the 4B model and about 79 hours on 32 GH200s for the 9B.