FUSE-GS: Frequency-Aware Pruning and Scale-Driven Rendering for Energy Efficient 3D Gaussian Splatting

Abstract

Classical image compression techniques exploit the importance differences existing between low-frequency and high-frequency of image data. In 3D Gaussian Splatting (3D-GS), pruning approaches have been widely employed to reduce computational complexity, but those are applied to uniform criteria without considering frequency characteristics, leaving compressible low-frequency regions over-represented. Furthermore, within the standard tile-wise rendering pipeline, enlarged Gaussians resulting from pruning increase alpha computation costs and introduce redundant processing across multiple tiles. In this paper, we propose FUSE-GS, an algorithm-architecture co-design exploiting frequency-domain analysis for pruning as well as efficient hardware architecture for larger Gaussians. FUSE-GS uses DCT for computing per-Gaussian frequency scores to identify contributions of low-frequency content. Then, it employs an opacity reset mechanism enabling gradient-based pruning of smooth regions into fewer and larger Gaussians. To efficiently render these enlarged Gaussians, we introduce a Gaussian-wise rendering architecture with two key units: a Block Filtering Unit (BFU) for block-level identification of active regions and a Delta Blending Unit (DBU) for reducing alpha computation via shared delta calculations. Experimental results show that FUSE-GS achieves 1.7× improvement in energy efficiency compared to state-of-the-art gaussian-wise 3D-GS accelerators.