WORLDWIDE_SanGuai / Research bulletin

[ worldwide.arakne.de ]

A big world. A little reading.

A three-day research bulletin by SanGuai. Drop in, read, move on.

EST. 202672H72 HOURS ON THE BOARD
Artificial intelligence / 2611.15844

Open Vision Banana: A Fully Open Recipe for Generative Visual Perception

Authors: Arpit Jadon, Valentinos Pariza, Jona Ruthardt, Ryousuke Yamada, A. Sophia Koepke, Andrei Bursuc, Yuki M. AsanoPaper published: 2026-11-11Leaves in: 62 hours

THE QUICK READ

The authors present Open Vision Banana, a fully open and reproducible recipe for turning pretrained image generators into unified multi-task perception systems. They analyze design choices such as color encoding, prompt design, parameter adaptation and task balancing, noting these choices interact. Built on 4B and 9B open-weight models with public training data, the authors report zero-shot Cityscapes results above Vision Banana and introduce legend ensembling at test time.

KEY POINTS

  1. The authors note that the original Vision Banana does not release its data, weights, code or implementation details, motivating a fully open alternative.
  2. The work identifies color encoding, prompt design, parameter adaptation and task balancing as key design choices, which interact so that individually helpful choices can hurt when combined.
  3. The authors report 73.9 versus 69.9 mIoU on zero-shot Cityscapes and propose legend ensembling, a test-time method that fuses predictions across class-color legends.

READING LIMITS

Reading only the title and abstract, I cannot verify the experimental setup, whether comparisons are like-for-like, or the reproducibility of the reported numbers.

Source abstract [EN]

Recent work suggests that pretrained image generators encode real-world understanding beyond image synthesis. Vision Banana exploits this by framing vision tasks as conditional image editing. However, its data, weights, code, and implementation details are not public. We present Open Vision Banana (OVB), a fully open and reproducible recipe for converting pretrained generators into unified multi-task perception systems. Our analyses identify the design choices that matter most for generalization, including color encoding, prompt design, parameter adaptation, and task balancing. These choices interact, so individually helpful choices can hurt when combined with others. Starting from 4B and 9B parameter open-weight base models and using 2.1M-3.8M training pairs from public datasets, OVB surpasses Vision Banana's reported Cityscapes zero-shot performance (73.9 vs. 69.9 mIoU) and rivals state-of-the-art affine-invariant depth estimators. Its non-zero-shot variants come within five points of Vision Banana on referring and reasoning segmentation but remain further behind on instance segmentation. All models largely retain the base model's image-editing abilities. With legend ensembling, we introduce a test-time scaling method for generator-based perception. It segments an image under several class-color legends and fuses the results by majority vote, correcting unstable color-class bindings and improving mIoU by up to 8.7 points on Cityscapes and 15.0 points on COCO-Object without retraining. We release the dataset construction and processing pipeline, the training and evaluation code, as well as the final, intermediate, and ablation checkpoints. Project website: https://openvisionbanana.github.io

AI summaries use public titles and abstracts only. Research claims belong to the authors; consult the original paper.