on: Nov. 12, 2026
in: NeurIPS

Dataset Distillation via Drifting

George Cazenavette
Angelina Quan
Giannis Daras
Antonio Torralba
Vincent Sitzmann

Project Page
Paper
Data
Code

@inproceedings{cazenavette2026drifting,
    title = { Dataset Distillation via Drifting },
    author = { Cazenavette, George and 
               Quan, Angelina and 
               Daras, Giannis and 
               Torralba, Antonio and 
               Sitzmann, Vincent },
    year = { 2026 },
    booktitle = { NeurIPS },
}

Copy to Clipboard

In this work, we distill large datasets down to just a few synthetic samples per class. We do this via drifting in the feature space of an arbitrary encoder. These distilled images reach state-of-the-art results in the traditional from-scratch classification task as well as a newly proposed set of feature-space tasks. We also show preliminary results generalizing this method to video and audio modalities.