Reproduce “Image Generators are Generalist Vision Learners” on OpenResearch

About this paper

Valentin Gabeur, Shangbang Long, Songyou Peng, Paul Voigtlaender, Shuyang Sun, Yanan Bao, Karen Truong, Zhicheng Wang and 17 more

Vision Banana uses instruction-tuned Nano Banana Pro to produce RGB visualizations that decode into segmentation masks and depth measurements, mixing a low ratio of vision-task data with image-generation data. It reports 69.9 mIoU versus SAM 3’s 65.2 on Cityscapes, and 0.929 versus 0.918 depth accuracy across four datasets. Results support competitive performance, but do not establish universal superiority or isolate pretraining as the cause.

Key insights

• Image generators are intrinsically "generalist vision learners," and generative vision pretraining can provide robust, transferable visual understanding capabilities analogous to LLMs.
• Image generation can serve as a universal interface for a wide array of computer vision tasks by expressing diverse outputs as decodable RGB images.
• Generative models can inherently manage ambiguities in vision tasks, learning the full data distribution and producing more robust predictions compared to some discriminative approaches.

Summary by alphaXiv.