Dataset Distillation via Drifting Compresses ImageNet Classes to Just 10 Images

giannis_daras · x · 2026-11-16

A NeurIPS 2026 paper, Dataset Distillation via Drifting, distills ImageNet's 1,300 images per class down to just 10 synthetic images optimized from noise. The authors frame the result as "the 10 most informative images that distill the concept" of each class; researcher AlexGDimakis notes it can be seen as a new type of art.

Original post →

More from Research

Research channel →