Open Vision Banana: A Fully Open Recipe for Generative Visual Perception
OVB beats Vision Banana’s reported zero-shot Cityscapes score, with training code, data pipelines and checkpoints released.
The system turns pretrained image generators into multitask perception models, reaching 73.9 mIoU against Vision Banana’s reported 69.9. It also rivals leading affine-invariant depth estimators, though its non-zero-shot variants remain further behind Vision Banana on instance segmentation. A test-time method combines predictions made under different class-color legends, improving segmentation by up to 15.0 mIoU points without retraining. The models largely retain their original image-editing abilities.
HF Daily Papers' note
The system turns pretrained image generators into multitask perception models, reaching 73.9 mIoU against Vision Banana’s reported 69.9. It also rivals leading affine-invariant depth estimators, though its non-zero-shot variants remain further behind Vision Banana on instance segmentation. A test-time method combines predictions made under different class-color legends, improving segmentation by up to 15.0 mIoU points without retraining. The models largely retain their original image-editing abilities.
HF Daily Papers' note
score 5