CAFE (Counterfactual Attribute Factuality Evaluation) is a benchmark for evaluating concept-faithful grounding in promptable segmentation models. Given a counterfactually edited image and a text prompt, a model must determine whether the queried concept is semantically valid for the target region and, if so, produce a precise segmentation mask.
