Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation