Retrieve and Segment · interactive demo

Pick a test image and see the model's zero-shot open-vocabulary segmentation. Then click support images below to add them to the visual support set — RNS retrieves from them at test time, and the prediction updates instantly. All predictions are precomputed with the real model.