Dataset Distillation via Drifting compresses 1,300 ImageNet images per class into just 10
kastnerkyle · x · 2026-11-15
A new NeurIPS 2026 paper, "Dataset Distillation via Drifting," distills datasets rather than models: it compresses ImageNet's 1,300 images per class into just 10 synthetic images optimized from noise. The images look slightly unnatural by design — they entangle concepts from multiple real samples into one synthetic sample. Models trained on them perform well on downstream tasks like classification. The authors also demo an artsy use case: summarizing an artist's entire body of work into a single painting.
- Paper: "Dataset Distillation via Drifting" (GCazenavette, Giannis Daras et al.)
- Key result: 10 images per class can teach a model what a class looks like
Related event: NeurIPS paper distills ImageNet to 10 images per class with SOTA results(4 posts)→
More from Multimodal
- Meituan open-sources LongCat-Video-Avatar 1.6 for 10+ minute talking-avatar videos — pmttyji · 2026-11-15
- Change one thing at a time: why iterating AI images isn't design improvement — alifcoder · 2026-11-15
- Fable 5.2 builds a swimmable underwater world in a single 97 KB HTML file — kimmonismus · 2026-11-15
- 16-person consistency test: 32x32 blank input makes Minimax H3 Image shine — PhysicsAutomatic3074 · 2026-11-15
- AI meme: 'We only serve vegans' — the lion replies 'Perfect, I'll have two' — kingslyj · 2026-11-15
- Gemini Hydrogen produces a full result from a single prompt, demo shared — CodeByPoonam · 2026-11-15