**SUNDAE: Spectrally Pruned Gaussian Fields with Neural Compensation** **Authors:** Runyi Yang **Abstract:** This paper addresses the high memory consumption issue of 3D Gaussian Splatting (3DGS), a novel 3D representation method known for its fast rendering speed and high quality. The method proposes SUNDAE, a memory-efficient Gaussian field that combines spectral pruning and neural compensation. Spectral pruning models the relationships between Gaussian primitives using a graph signal processing approach, allowing for the pruning of redundant primitives while preserving essential signals. Neural compensation uses a lightweight neural network to mix splatted features, compensating for quality losses caused by pruning. The method demonstrates superior performance on various benchmarks, achieving high PSNR and SSIM scores with significantly reduced memory usage compared to vanilla 3DGS. **Key Contributions:** - A novel primitive pruning framework for Gaussian fields based on spectral graph pruning. - A feature splatting and mixing module to compensate for quality losses. - State-of-the-art results in terms of quality and speed with low memory footprint. **Methods:** 1. **Spectral Graph Pruning:** Constructs a graph to model the relationships between Gaussian primitives and uses a band-limited graph filter to prune redundant primitives. 2. **Neural Compensation:** Utilizes a lightweight neural network to integrate information from different primitives, compensating for quality losses. **Experiments:** - Evaluates SUNDAE on three datasets (Mip-NeRF360, Tanks&Temples, Deep Blending). - Compares performance in terms of PSNR, SSIM, FPS, and memory usage. - Ablation studies validate the effectiveness of spectral pruning and neural compensation. **Results:** - SUNDAE achieves high-quality rendering with significantly reduced memory usage, outperforming other methods in terms of efficiency and quality. **Conclusion:** SUNDAE effectively models the relationships between Gaussian primitives and compensates for quality losses, achieving a balance between rendering quality, speed, and memory efficiency.**SUNDAE: Spectrally Pruned Gaussian Fields with Neural Compensation** **Authors:** Runyi Yang **Abstract:** This paper addresses the high memory consumption issue of 3D Gaussian Splatting (3DGS), a novel 3D representation method known for its fast rendering speed and high quality. The method proposes SUNDAE, a memory-efficient Gaussian field that combines spectral pruning and neural compensation. Spectral pruning models the relationships between Gaussian primitives using a graph signal processing approach, allowing for the pruning of redundant primitives while preserving essential signals. Neural compensation uses a lightweight neural network to mix splatted features, compensating for quality losses caused by pruning. The method demonstrates superior performance on various benchmarks, achieving high PSNR and SSIM scores with significantly reduced memory usage compared to vanilla 3DGS. **Key Contributions:** - A novel primitive pruning framework for Gaussian fields based on spectral graph pruning. - A feature splatting and mixing module to compensate for quality losses. - State-of-the-art results in terms of quality and speed with low memory footprint. **Methods:** 1. **Spectral Graph Pruning:** Constructs a graph to model the relationships between Gaussian primitives and uses a band-limited graph filter to prune redundant primitives. 2. **Neural Compensation:** Utilizes a lightweight neural network to integrate information from different primitives, compensating for quality losses. **Experiments:** - Evaluates SUNDAE on three datasets (Mip-NeRF360, Tanks&Temples, Deep Blending). - Compares performance in terms of PSNR, SSIM, FPS, and memory usage. - Ablation studies validate the effectiveness of spectral pruning and neural compensation. **Results:** - SUNDAE achieves high-quality rendering with significantly reduced memory usage, outperforming other methods in terms of efficiency and quality. **Conclusion:** SUNDAE effectively models the relationships between Gaussian primitives and compensates for quality losses, achieving a balance between rendering quality, speed, and memory efficiency.