PrePARE: Pre-AA Token Pruning for Frozen Multi-View Geometry Transformers
PaceVGGT is a novel pre-Alternating-Attention pruning framework for frozen Visual Geometry Transformers (VGGT). It reduces computational cost by moving token reduction to the DINO-to-AA interface, targeting quadratic scaling issues. This method uses a geometry-aware scoring criterion distilled from internal attention signals to selectively prune tokens while preserving critical information for 3D reconstruction tasks.