Open Vision Banana: fully open recipe turns image generators into multi-task perception systems
NielsRogge · x · 2026-11-17
Yuki M. Asano's group released Open Vision Banana (OVB), a fully open, reproducible recipe that converts pretrained image generators into unified multi-task perception systems via conditional image editing. Key points:
- Identifies the design choices that matter for generalization—color encoding, prompt design, parameter adaptation, task balancing—and shows they interact non-trivially
- Built from 4B/9B open-weight base models with 2.1M–3.8M pairs from public datasets
- Beats Vision Banana's reported Cityscapes zero-shot mIoU (73.9 vs 69.9); rivals SOTA depth estimators; within 5 points on referring/reasoning segmentation, behind on instance segmentation
- Introduces legend ensembling, a test-time scaling method fusing segmentations under multiple class-color legends, adding up to 8.7 mIoU on Cityscapes and 15.0 on COCO-Object without retraining
Data pipeline, training/eval code and all checkpoints are public.
More from Multimodal
- ElevenLabs Mobile App Adds Customizable Video Captions — azed_ai · 2026-11-17
- Chromostereopsis prompt trick makes flat AI images pop with fake depth — LudovicCreator · 2026-11-17
- CLAP2 Audio Segmentation Demo Built on Last Week's Model Release — wightmanr · 2026-11-17
- AI-Generated Robert Pattinson as James Bond Images Go Viral for Film-Grade Realism — egeberkina · 2026-11-17
- Higgsfield Launches Short Drama Studio: One Script, a Full Season — CurieuxExplorer · 2026-11-17
- AI acting tests with MiniMax H3: one character, six subtle emotions — Acrobatic-Example315 · 2026-11-17