GaussAnything: Semantic Intent-Driven Refinement of Evolving Gaussian Scenes for Standalone VR

Standalone VR headsets are still bottlenecked by memory and compute, so GaussAnything targets a practical pain point: how to keep reconstructed 3D scenes responsive when users inspect specific objects. The system reallocates a fixed Gaussian budget toward the queried object, preserving global context while staying within a steady ~10 ms GPU frame cost. It’s a useful signal for VR rendering pipelines that need intent-aware detail, not just visibility-based LOD.