Open Vision Banana: fully open recipe beats Vision Banana with 73.9 mIoU on Cityscapes
FunAILab · hf · 2026-11-13
Researchers release Open Vision Banana (OVB), a fully open, reproducible recipe for turning pretrained image generators into unified multi-task perception systems, filling the gap left by the closed Vision Banana.
- Built on 4B/9B open-weight base models with 2.1M–3.8M training pairs from public datasets
- Surpasses Vision Banana's reported Cityscapes zero-shot mIoU (73.9 vs 69.9); rivals SOTA affine-invariant depth estimators
- Identifies key design choices (color encoding, prompt design, parameter adaptation, task balancing) that interact non-trivially
- Introduces legend ensembling, a test-time scaling method that fuses segmentations under multiple color legends by majority vote, boosting mIoU by up to 8.7 points on Cityscapes and 15.0 on COCO-Object without retraining
- Releases data pipeline, training/eval code, and all intermediate and ablation checkpoints
More from Multimodal
- Dev one-shots ASMR-ifying any video using ElevenLabs sound effects — austin_malerba · 2026-11-13
- PlaygroundArt.AI festival clip wows Reddit with lifelike motion and hair detail — shahryar2025 · 2026-11-13
- Full Restaurant Brand Visuals for ~24 Higgsfield Credits: A Step-by-Step Workflow — PrajwalTomar_ · 2026-11-13
- Structured SillyTavern-to-ComfyUI prompts: interaction beats pose for Krea 2 and H3 — Meinungsbild · 2026-11-13
- Gemini 4 Hydrogen builds 400K-voxel Chinese mountain temple in Blender, zero manual input — iamfakhrealam · 2026-11-13
- Open-source browser-based bbox layout builder for MOSS-Image, fully offline with JSON export — Rex-Racer9 · 2026-11-13