Dino Forcing paper predicts DINO reps instead of co-denoising, halving training epochs on ImageNet
kastnerkyle · x · 2026-10-10
A paper from ENPC & Ecole Polytechnique (with Alexei Efros among the authors), Dino Forcing Flow Models: Do not denoise what you can predict, proposes a simpler alternative to co-denoising pretrained DINO representations in flow matching models: directly predict the representation and condition the model on its own prediction.
Key points:
- Removes the second denoising ODE and all representation-specific schedules, while keeping representation guidance benefits
- On ImageNet, outperforms latent-space SOTA with 2x fewer epochs than prior methods
- In pixel space, improves FID by more than 20% over comparable prior work
- Guiding principle: "do not denoise what you can predict"; code is openly available
More from Research
- Tree-like procedural 3D data is hard for LLMs to learn — and data is scarce — keenanisalive · 2026-10-10
- OmniSeek turns Omni-LLMs into evidence-seeking agents, +15.5 points on VideoHolmes — mohitban47 · 2026-10-10
- Study: Parametric neural control differentiates top neural network models of primate visual cortex — Dr_Alex_Crimi · 2026-10-10
- Architecture papers reading list highlights Mamba-3 non-commutative state tracking advance — zmkzmkz · 2026-10-10
- Dex-One2Many: Real2Sim2Real Turns One Human Video Into Robots Generalizing Across Configurations — furongh · 2026-10-10
- DINOv2 and Qwen3 Embedding Spaces Aligned Without a Single Image-Caption Pair — NandoDF · 2026-10-10