SAS: Semantic-aware Sampling for Generative Dataset Distillation

In three linesSAS introduces semantic-aware dataset distillation leveraging CLIP as a semantic prior to improve compressed dataset quality. Three scoring functions evaluate class relevance, inter-class separability, and intra-set diversity. A two-stage strategy filters discriminative samples then dynamically selects to reduce redundancy while preserving semantic coverage.

Summary generated by Claude — human-verified