Overview

FreeTimeGS introduces a novel 4D representation allowing Gaussian primitives to appear at arbitrary time and locations for dynamic scene reconstruction. Each Gaussian has motion and temporal opacity functions, providing flexibility to model complex motions without canonical space deformation fields.

Why it matters

Overcomes limitations of canonical-space methods that struggle with complex motions by eliminating the need for difficult deformation field optimization. Achieves real-time dynamic view synthesis with superior quality on scenes with fast, complex movements.

Key trade-offs / limitations

  • Requires lengthy reconstruction process for each dynamic scene
  • Currently supports only novel view synthesis (no relighting capabilities)
  • May need generative priors for optimization-free reconstruction in future
  • High temporal redundancy without motion function constraints
Link arXiv:2506.05348