Reproduce “Open Vision Banana: A Fully Open Recipe for Generative Visual Perception” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Arpit Jadon, Valentinos Pariza, Jona Ruthardt, Ryousuke Yamada, A. Sophia Koepke, Andrei Bursuc, Yuki M. Asano

Open Vision Banana adapts open-weight image generators for six perception tasks: deterministic decoders turn generated RGB images into labels, masks, depth, or surface normals. For segmentation, legend ensembling changes class-color assignments across repeated predictions and combines decoded maps by majority vote. Its zero-shot 9B model reaches 73.9 mIoU on Cityscapes validation, versus Vision Banana’s reported 69.9; the comparison is not controlled.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.