RLSplat makes feed-forward Gaussian splatting scale with the budget

Latency and peak memory fall with the Gaussian budget, unlike prune-after-dense methods that stay nearly constant.

RLSplat is a budget-controllable feed-forward 3D Gaussian Splatting method whose representation size, inference latency, and peak memory decrease linearly as the Gaussian budget shrinks. It allocates the exact requested count across local scene regions instead of predicting a dense set and then pruning it. Decoding attends only to nearby Gaussians found by hierarchical neighbor search. In zero-shot tests the authors say it beats prior budget-controllable methods at every budget.