Towards Universal Dataset Distillation via Task-Driven Diffusion

Figure 1. (a) We introduce the Universal Dataset Distillation Framework (UniDD), a novel dataset distillation framework supporting a range of vision tasks, including classification, object detection, and segmentation. UniDD incorporates diverse label types, such as class labels, bounding box coordinates, and pixel-level masks, facilitating the synthesis of task-specific datasets. (b) UniDD improves the performance of dataset distillation across diverse tasks on ImageNet [8], Pascal VOC [11, 12], and MS COCO [28].