RLSplat beats FF3DGS baselines with 4× fewer Gaussians
TL;DR
- RLSplat surpasses the strongest prior budget-controllable FF3DGS baseline in zero-shot evaluation even when using 4× fewer Gaussians.
- At low budgets, the method runs up to 4.1× faster and uses 9.6× less peak memory than prior budget-controllable approaches.
- Instead of pruning a dense prediction, RLSplat learns a Gaussian allocation policy and uses a hierarchical neighbor search for linear-cost decoding.
Feed-forward 3D Gaussian Splatting rebuilds a scene in a single pass, but when a device sets a tight Gaussian budget, prior methods still predict a dense set first and prune it down, paying nearly constant latency and peak memory regardless. RLSplat, an arXiv preprint posted October 20, drops the pruning step: a learned allocation policy distributes the budget across local scene regions so that "exactly the requested number of Gaussians is generated," and each Gaussian "attends only to nearby Gaussians found via an efficient hierarchical neighbor search," so decoding cost scales with the budget rather than the dense prediction.
In zero-shot evaluation, the authors report RLSplat "outperforms prior budget-controllable methods at every budget, surpassing the strongest baseline even with 4× fewer Gaussians, while running up to 4.1× faster with 9.6× less peak memory at low budgets." The abstract publishes no per-scene PSNR or benchmark names, and reference code is listed on GitHub under mlvlab/RLSplat.
Originally reported by paper
Read the original article →Original headline: RLSplat Matches 3DGS Quality With 4× Fewer Gaussians, 4.1× Faster and 9.6× Less Memory