ACE-GS: Acing the Trade-off with Accurate, Compact and Efficient 3D Gaussian Splatting

The paper introduces ACE-GS, a progressive optimization framework designed for 3D Gaussian Splatting that enhances rendering quality while reducing computational demands. Key innovations include a momentum consistency-guided densification strategy for efficient primitive management and a sensitivity-driven sparsification mechanism that compresses the scene representation. ACE-GS achieves up to 3.7 times faster training than Speedy-Splat, converging in 3 to 5 minutes, and improves structural similarity and PSNR by up to 0.89 dB, establishing a new benchmark for high-fidelity novel view synthesis.