This approach enables input-specific model construction without additional training and offers scalable, interpretable, and privacy-preserving adaptation.

SemLA dynamically computes a weighted average over LoRA-based adapters from the most relevant source domains, where weights are determined by the semantic similarity between each source domain and the target input, as measured in the CLIP embedding space.

We propose Semantic Library Adaptation (SemLA), a training-free method for test-time domain adaptation in open-vocabulary semantic segmentation.

This is great, thank you for sharing!