Table 1. Quantitative comparison on Neural 3D dataset. In the Colmap column, SA denotes 'Sparse point cloud for All frames' and D0 denotes 'Dense point cloud for the 0th frame'. Following the original STG paper, which reports training six models for every 50 frames, we provide results for both the multi-model approach and a single-model approach trained on the full 300-frame sequence.

Method	Colmap	Preproc. Time ↓	PSNR ↑	SSIM ↑	LPIPS ↓	Train Time ↓	FPS ↑	Storage ↓	Frames
4DGS	D0	6 mins	28.72	0.9306	0.1528	33 mins	98	40.3	300
STG	SA	25 mins	31.75	0.9473	0.1423	2h 43mins	683	127.5	50×6
STG	SA	25 mins	31.46	0.9432	0.1474	29 mins	532	54.0	300
TaylorG	SA	25 mins	29.80	0.9558	0.1597	9 hours	125	205.7	300
Swift4D	D0	18 mins	29.93	0.9383	0.1370	19 mins	273	141.2	300
DeGauss	D0	6 mins	30.16	0.9357	0.1430	1h 27mins	95	117.5	300
OURS-35K	✗	4 sec	32.35	0.9480	0.1295	10 mins	766	23.1	300
OURS-45K	✗	4 sec	32.72	0.9502	0.1221	14 mins	755	23.7	300