Learn2Splat: Extending the Horizon of Learned 3DGS Optimization

Researchers introduced Learn2Splat, a learned optimizer for 3D Gaussian Splatting.

It uses meta-learning, checkpoint buffering, and rollout strategies to avoid long-horizon performance collapse.

The method improves early reconstruction and novel view synthesis quality while staying stable over longer training runs.

Results show consistent behavior across both sparse and dense view settings, suggesting better generalization.