Dataset Distillation via Drifting: 10 images per class from ImageNet, new SOTA and ECCV award
giannis_daras · x · 2026-11-14
The NeurIPS 2026 paper "Dataset Distillation via Drifting" distills ImageNet's 1,300 images per class down to just 10 optimized from noise. A follow-up winning ECCV 2026 Art Gallery award asks whether a small set of images can optimally cover a dataset in a pre-trained feature space, co-opting Drifting Generative Models (Minyang et al.) for dataset distillation—achieving visually stunning results and a new SOTA with unprecedented classification accuracy.
Related event: Dataset Distillation Hits SOTA: 10 Images Per Class(2 posts)→
More from Multimodal
- NVIDIA's Déjà View: 117M model loops one transformer block for 3D from photos — NVIDIAAI · 2026-11-14
- MiniMax H3 Image's ~50GB stack runs on 8GB VRAM at 57s/image — how is there no OOM? — wreck_of_u · 2026-11-14
- Pocket-book language learning game runs LFM3-Omni locally via WebGPU — helloiamleonie · 2026-11-14
- Region tokens: an hour of video for fewer tokens than 20 seconds of patches — Cohere · 2026-11-14
- AI video of a tennis player getting possessed by a demon goes viral — StevenWintower · 2026-11-14
- Animated AI short retells Caesar's famous double walls at Alesia — EpicFoolss · 2026-11-14