Search papers, labs, and topics across Lattice.
LinearMask-GS addresses the premature mask saturation in 3D Gaussian Splatting (3DGS) compression by replacing the standard Gumbel-Sigmoid activation with a linear increment function during mask optimization. Steep sigmoid activations traditionally polarize mask values into an unrecoverable bimodal distribution before relative primitive importance can stabilize, degrading downstream pruning. By maintaining a unimodal, mid-confidence distribution throughout training, this method yields a 3.6x Gaussian reduction over vanilla 3DGS on Mip-NeRF 360 while simultaneously improving rendering fidelity on outdoor scenes by +0.38 dB PSNR.
Standard Gumbel-Sigmoid pruning fails in 3DGS because it forces binary decisions before importance ranks can stabilize鈥攕wapping it for a simple linear activation cuts primitive count by up to 3.6x while actually increasing rendering quality.
3D Gaussian Splatting (3DGS) enables real-time novel view synthesis but produces millions of primitives through adaptive densification, leading to significant storage overhead. Learned-mask pruning methods such as LP-3DGS address this by assigning each Gaussian a learnable mask to identify and prune redundant primitives. However, we identify a limitation of this paradigm: the steep slope of the Gumbel-Sigmoid activation drives mask values to the extremes within the short mask-training window, before the importance ranking has stabilized, producing a sharply bimodal distribution from which that ranking can no longer be reliably recovered. We propose LinearMask-GS, which replaces Gumbel-Sigmoid with a linear increment activation that keeps mask values in a mid-confidence regime throughout mask training, producing a stable, unimodal mask distribution whose ranking tracks importance. On Mip-NeRF 360, our method achieves 3.6x and 1.6x Gaussian reductions over 3DGS and LP-3DGS, respectively, while maintaining or improving rendering quality. For outdoor scenes, it yields a 1.6x reduction (from 2.18M to 1.36M) with notable gains in PSNR (+0.38 dB), SSIM (+0.025), and LPIPS (-0.029).
