Fundamental AI Lab Open-Sources Vision Banana Recipe on FLUX.2-Klein, Hits 73.9 mIoU on Cityscapes

UTN's Fundamental AI Lab on Nov 13 released Open Vision Banana, a fully open reproduction of Google's proprietary Vision Banana unified perception model, built on public FLUX.2-Klein image generators at 4B and 9B scale. The 9B zero-shot checkpoint hits 73.9 mIoU on Cityscapes (versus Vision Banana's 69.9) and matches state-of-the-art depth estimators while retaining base image-editing ability. The release ships complete training data, code, weights and a novel 'legend ensembling' test-time method that adds up to 15 mIoU on COCO-Object without retraining.