Primitive Memory Networks with Qualified Prompts for Compositional Zero-Shot Learning

Compositional Zero-Shot Learning (CZSL) recognizes unseen attribute-object compositions by transferring primitive knowledge learned from seen pairs. Existing CLIP-based approaches often rely on simple textual prompts and limited visual representatives, which may inadequately characterize primitives with diverse semantics and intra-class appearances. To address these limitations, we propose Primitive Memory Networks with Qualified Prompts (PMNQP) to provide complementary semantic and visual evidence for CZSL. Specifically, qualified prompts incorporate multiple primitive-related descriptions generated once by a large language model offline, without requiring LLM access during model training or inference. Furthermore, primitive memory networks maintain a fixed-capacity bank of selected visual exemplars for each attribute and object, update slot membership online during training, and perform query-dependent readout from the fixed banks at inference. Together, these complementary semantic and visual pathways address the limited context provided by coarse primitive prompts and the inadequate coverage of intra-class variation by a single visual representation. Extensive experiments on three benchmarks show that PMNQP achieves competitive performance under both closed-world and open-world CZSL settings. Code is available at https://github.com/YtJin-git/PMNQP.