For multimodal classifiers, finding counterfactual explanations requires looking at all modalities together—CRANE does this by retrieving real training examples that are similar across text, images, and other inputs simultaneously, rather than explaining each modality independently.
CRANE is the first post-hoc method for generating counterfactual explanations for multimodal classifiers. Instead of explaining each modality separately, it retrieves real training examples from the target class that are jointly close across all modalities, ensuring explanations are plausible and grounded in actual data.