Image segmentation is usually addressed by training a model for a fixed setof object classes. Incorporating additional classes or more complex querieslater is expensive as it requires re-training the model on a dataset thatencompasses these expressions. Here we propose a system that can generate imagesegmentations based on arbitrary prompts at test time. A prompt can be either atext or an image. This approach enables us to create a unified model (trainedonce) for three common segmentation tasks, which come with distinct challenges:referring expression segmentation, zero-shot segmentation and one-shotsegmentation. We build upon the CLIP model as a backbone which we extend with atransformer-based decoder that enables dense prediction. After training on anextended version of the PhraseCut dataset, our system generates a binarysegmentation map for an image based on a free-text prompt or on an additionalimage expressing the query. We analyze different variants of the latterimage-based prompts in detail. This novel hybrid input allows for dynamicadaptation not only to the three segmentation tasks mentioned above, but to anybinary segmentation task where a text or image query can be formulated.Finally, we find our system to adapt well to generalized queries involvingaffordances or properties. Code is available athttps://eckerlab.org/code/clipseg.