DINOv3 patches strongly encode position by default — you should clean them up
StephanSturges · x · 2026-11-22
The author observes that DINOv3 patch features encode positional information to a staggering degree by default — PCA visualizations show smooth position gradient patterns (second column). This may help on some benchmark tasks in papers but isn't great for real-world usage; the features need to be cleaned up before deployment, with before/after visualizations provided.
Related event: DINOv3 patch features encode strong positional information(2 posts)→
More from Research
- Live Lean 4 benchmark tracks Riemann Hypothesis proof compression across models and agents — flngr · 2026-11-22
- Kalshi traders now put 54% odds on P vs NP being resolved before 2030 — ricklamers · 2026-11-22
- Fortnow: Fall's dramatic results show P vs NP still far from solved, as OpenAI proof dump yields first superlinear circuit lower bound — thomasahle · 2026-11-22
- One photo, 8 VLAs, only 3 correct: ERQA embodied-reasoning benchmark questioned — giffmana · 2026-11-22
- Sugon's 100,000-GPU曙光8000 cluster wins Gordon Bell Prize, heterogeneous nodes 3x faster than CPU-only — 智东西 · 2026-11-22
- Fleuret: treat AI as oracles, not explainers — human-comprehensible math will be negligible — francoisfleuret · 2026-11-22