At CVPR 2026 the 3D Gaussian Splatting field stopped chasing prettier renders and started chasing industry. Across eight representative papers, the work closes the pipeline from feed-forward, second-scale generation to GPU-operator rebuilds, noise-robust SLAM and Neural ODE physics, turning 3DGS from a 3D photo album into the physical base of embodied AI and world models.
Making assets on demand
EcoSplat, from KAIST and Flawless AI, adds an importance scheduler so a model picks the top-K gaussians for a compute budget; at 5 per cent of points, about 78,000, it still renders at 24.72 dB. SparseSplat, from Fudan and ShanghaiTech, uses Shannon entropy to spend points where detail lives, hitting 24.20 dB with 150,000 gaussians, just 22 per cent of a 6.88 million baseline, at 600 to 1,100 frames per second.
Rendering and architecture
CaT-GS, from Shanghai Jiao Tong and UIUC, borrows game-engine inter-frame caching and hits a 10x speedup on a 7 million-gaussian city set, over 200 FPS at 1080p. EDGS, from LMU Munich, drops Gaussian densification entirely and reaches target quality in 15 per cent of the training time. TokenGS, from NVIDIA, decouples prediction from pixels using learnable tokens, giving pose-noise robustness and on-device test-time tuning.
From seeing to acting
The application papers are the payoff. SGAD-SLAM corrects live depth noise for robot navigation. ParticleGS, from Zhejiang University, treats each gaussian as a physical particle and uses a Neural ODE to extrapolate collision and deformation. Video2Robo, from Beijing Institute of Technology, turns one phone video into a 3DGS twin and synthesises robot training data in hours instead of weeks.
Notably, Chinese universities and startups lead much of this wave, from Fudan and SJTU to embodied-AI firms, putting the region at the front of spatial computing infrastructure.
Editor’s note: This is an adapted translation of the original LeiFeng Network report. It has been trimmed and restructured for readability for an international business audience.