Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers

Researchers propose a two-stage token selection method for visual geometry transformers to cut global attention cost.

The method first picks diverse, informative frames, then prunes redundant tokens within those frames using layer-aware sparsification.

Attention entropy guides the pruning process, helping preserve useful information while removing excess tokens.

On large scenes, the approach delivers over 85% speedup and can match or improve reconstruction quality.