Researchers say ‘drifting’ squeezes each ImageNet class into 10 synthetic images

From 1,300 images to 10 per class: that’s the ImageNet compression the researchers report with Dataset Distillation via Drifting. Their project page says the synthetic images start as noise and are optimized in DINOv3’s feature space, achieving state-of-the-art dataset-distillation classification results. Coauthor Giannis Daras says their unnatural appearance is deliberate: each combines information from multiple real samples. Realism isn't the assignment.