BackbonePANSegmentationRegressionImageEmbeddingP5P4P3man, horse,dog, cat, …Text promptTextEncodermanhorsedogTextualEmbedding𝑃AuxiliaryNetwork𝑓𝜃Visual promptActivationBranchSemanticBranchPrompt-aware weightAggregationPromptEmbeddingSemantic-Activated Vis. Prompt EncoderRe-parameterizable Region-Text AlignmentLabelRe-parameterizationInferenceTransferringClassificationOffline for training/inferenceLazy Region-Prompt ContrastSpecialized EmbeddingAnchor PointsPrompt freeBuilt-in VocabularyRetrievalSemanticfeature