Variants. We integrate our method with four prompt-based approaches: Ours uses CLAP with hand-
crafted prompt templates, CoOp + Ours with soft prompts [12], MaPLe + Ours uses multi-modal soft
prompts [20], and AttriCLIP + Ours uses CLAP4CLIP with instance-conditioned soft prompts [19].
Ours w/o Variational Inference (VI) is the deterministic variant of Ours depicted in App. Fig. 7. We
leave the details of the training and the hyperparameters of our models in App. A.2.
