DIVER:Diving Deeper into Distilled Data via Expressive Semantic Recovery

DIVER proposes a dual-stage dataset distillation framework addressing overfitting to specific architectures by recovering expressive semantics, improving generalization across heterogeneous architectures compared to single-stage distillation approaches.