Turbo-GS：稀疏像素渲染与高斯预算控制

English translation pending — this page shows the Chinese original.

一句话结论

CVPR 2026正式主会在已知位姿的逐场3DGS中以四偏移交替的四分之一像素训练、收敛预算与颜色梯度增密缩短4K优化；A100上五个4K Mip-NeRF360场景均值24分钟/26.80 PSNR，对原3DGS 113分钟/26.75，但模型存储680MB对412MB，资源成本须按终点拆账（主文p.15459 Table1；补充p.4 Table11）。

输入、算法与限制

输入已标定RGB、相机及SfM初始化，输出逐场优化高斯和新视角渲染。d=2四种随机采样偏移令每轮像素和tile交集开销下降，低分辨率投影按d缩放并从原图原位置取RGB监督。收敛曲线拟合控制目标高斯预算，位置梯度弱而颜色梯度强时补充增密（颜色门限为位置门限0.01倍，20%概率激活），可选Speedy-Splat敏感度60%剪枝。底座的Taming CUDA反传也是重要加速来源（主文§4 pp.15457–15459；补充p.1 Alg.1）。四偏移组合的边缘重建仍有差异，完整视图薄结构质量需单独验。

证据账、消融与适用范围

4K Table1 p.15459：Turbo-3DGS五场24min/26.80 PSNR/.327 LPIPS/680MB/峰2.80M高斯；3DGS 113min/26.75/.410/412MB/1.74M，3DGS-accel 40min/26.52/.364/383MB/1.62M；EyefulTower两场Turbo 27min/31.07 vs 3DGS 200min/31.18，PSNR稍低。Memory是模型存储，峰显存未在主表同口径列出。
1K Table2 p.15460：m360 Turbo 12.18min/27.39/.205 LPIPS/861MB，原3DGS 30.08min/27.45/.217/640MB；加敏感度剪枝后12.10min/27.06/.253/287MB。Scaffold 23.85min/27.71，与Turbo-Scaffold 6.86min/27.43；4K Scaffold-accel只训10K步，同30K的原Scaffold不可直作等步基准。
Table3 p.15460：同加速版基线31min/26.52，单DR 14min/26.66、单CA 12min/26.58、单CG 16min/26.75；Scaffold单DR 12→7min却使PSNR25.81→25.58。这是逐项消融，不是多项累积曲线。报告为作者实验，A100单卡，未含获取影像/SfM或跨种子方差；4K m360有五场，不能以全九场外推（补充Table11）。

原始链接与证据

CVF 11页主文和5页补充全文、页码及SHA见raw/ingest/2026-09-25-turbo-gs/analysis.md与retrieval-audit.json。项目仓库inspirelt/Turbo-GS可访问，GitHub API许可字段缺失，商用前审第三方授权；未本地运行CUDA指标。

相关页面

3D Gaussian Splatting: Representation, Optimization, Geometry, and Dynamic Scenes；FastGS：多视误差驱动的增密、剪枝与训练加速、Mip-Splatting：跨尺度采样与三维高斯抗混叠。按相同场景/图像分辨率/GPU/步数/目标PSNR-LPIPS及真实峰显存测试，同底座Taming与无稀疏采样做成对消融。

{ "id": "2026-09-25-turbo-gs", "type": "source", "title": "Turbo-GS：稀疏像素渲染与高斯预算控制", "status": "reviewed", "note_status": "reviewed", "created": "2026-09-25", "updated": "2026-09-25", "source_type": "paper", "venue": "CVPR 2026", "authors": "[Ankit Dhiman, Tao Lu, R Srinath, Emre Arslan, Angela Xing, Yuanbo Xiangli, Venkatesh Babu Radhakrishnan, Srinath Sridhar]", "published_at": "2026-06-01", "ingested_at": "2026-09-25", "canonical_links": [ "https://openaccess.thecvf.com/content/CVPR2026/html/Dhiman_Turbo-GS_Accelerating_3D_Gaussian_Fitting_for_High-Resolution_Radiance_Fields_CVPR_2026_paper.html", "https://openaccess.thecvf.com/content/CVPR2026/papers/Dhiman_Turbo-GS_Accelerating_3D_Gaussian_Fitting_for_High-Resolution_Radiance_Fields_CVPR_2026_paper.pdf", "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Dhiman_Turbo-GS_Accelerating_3D_CVPR_2026_supplemental.pdf", "https://github.com/inspirelt/Turbo-GS" ], "raw_entry": "raw/ingest/2026-09-25-turbo-gs/", "topics": "[topics/3d-gaussian-splatting]", "entities": [], "claims": [], "questions": [], "tags": [ "near-cvpr-2025", "3d-vision", "gaussian-splatting", "efficiency", "primary-source" ] }

Turbo-GS稀疏像素渲染与高斯预算控制