VkSplat: High-Performance 3DGS Training in Vulkan Compute

VkSplat: High-Performance 3DGS Training in Vulkan Compute

VkSplat:基于 Vulkan 计算的高性能 3DGS 训练框架

Abstract: We present VkSplat, a high-performance, cross-vendor 3D Gaussian Splatting (3DGS) training pipeline implemented fully in Vulkan compute, addressing performance and compatibility limitation of existing training pipelines.

摘要: 我们提出了 VkSplat,这是一个完全基于 Vulkan 计算实现的高性能、跨厂商 3D 高斯泼溅(3DGS)训练流水线,旨在解决现有训练流水线在性能和兼容性方面的局限性。

With various optimizations, we achieve $3.3\times$ speed and $33%$ VRAM reduction over CUDA+PyTorch baseline, maintaining quality, and demonstrating compatibility across GPU vendors.

通过多种优化手段,我们在保持渲染质量的同时,相比 CUDA+PyTorch 基准实现了 3.3 倍的加速和 33% 的显存占用降低,并展示了其在不同 GPU 厂商硬件间的良好兼容性。

To the best of our knowledge, this is the first fully-Vulkan-based 3DGS training pipeline that achieves state-of-the-art performance.

据我们所知,这是首个实现最先进(SOTA)性能的完全基于 Vulkan 的 3DGS 训练流水线。


Paper Details:

  • Authors: Jingxiang Chen, Mohamed Ibrahim, Yang Liu
  • Subject: Computer Vision and Pattern Recognition (cs.CV)
  • arXiv ID: 2605.00219
  • Date: 30 Apr 2026

论文详情:

  • 作者: Jingxiang Chen, Mohamed Ibrahim, Yang Liu
  • 学科: 计算机视觉与模式识别 (cs.CV)
  • arXiv ID: 2605.00219
  • 日期: 2026 年 4 月 30 日