MaskGaussian++: Probabilistic Masks for General 3D Gaussian Representation

While 3D Gaussian Splatting (3DGS) has demonstrated remarkable performance in novel view synthesis and realtime rendering, the high memory consumption due to the use of millions of Gaussians limits its practicality. To mitigate this issue, improvements have been made by pruning unnecessary Gaussians, either through a hand-crafted criterion or by using learned masks. However, these methods deterministically remove Gaussians based on a snapshot of the pruning moment, leading to sub-optimized reconstruction performance from a long-term perspective. To address this issue, we introduce MaskGaussian++, which models Gaussians as probabilistic entities rather than permanently removing them, and utilize them according to their probability of existence. To achieve this, we propose a masked-rasterization technique that enables unused yet probabilistically existing Gaussians to receive gradients, allowing for dynamic assessment of their contribution to the evolving scene and adjustment of their probability of existence. Beyond vanilla 3DGS, this probabilistic masking is a model-agnostic mechanism for the broader Gaussian family: we adapt it to neural (anchor-based) Gaussians with a shared anchor mask, and to distributed multi- GPU Gaussians with a distributed masked-rasterizer, through which we further uncover a log-linear scaling law of Gaussian Splatting spanning up to 90 million primitives, and to dynamic 4D Gaussians with a single time-invariant per-primitive mask that transfers across two structurally different 4D representations. Extensive experiments demonstrate the superiority of the proposed method: on vanilla 3DGS it prunes over 60% of Gaussians on average with only a 0.02 PSNR decline, halves the storage of neural Gaussians, scales $2\times$ more efficiently than Grendel-GS in large-scale distributed training, and compressing dynamic 4D Gaussians several-fold at on-par or better quality.