Study Finds Higher Input Resolution Dilutes Vision Transformer Attention, Proposes Training-Free Rescaling
A new arXiv paper shows that running a vision transformer at a higher resolution than it was trained on spreads each attention row across far more patches, weakening attention, and proposes a training-free rescaling method to address it.
Read the full story at arXiv cs.CV ↗
Timeline · 1 report
- 2026-11-17 05:00 · arXiv cs.CV
Spread Too Thin: How Higher Resolution Dilutes a Vision Transformer's Attention, and a Training-Free Rescaling