Key ideas
Prompt-Driven Synthesis → CLIP space: Synthetically generated images from prompts are projected into CLIP’s shared embedding space, after which the PIN process is applied to refine alignment between text and image features.
Knowledge-based prompts: We incorporate knowledge-guided prompts so the text descriptions are steered toward domain-appropriate semantics, improving adaptation to the target domain.
POSIDA surpasses the baseline on both the Snow and Rain target domains.