Benchmarks for gsplat

Benchmarks for gsplat#

2025-10-01

3 min read time

Applies to Linux

The ROCm/gsplat repository includes a standalone script that reproduces the official Gaussian Splatting results with matching performance on PSNR, SSIM, LPIPS, and the converged number of Gaussians.

To run the benchmark:

cd examples
pip install -r requirements.txt
# download mipnerf_360 benchmark data
python datasets/download_dataset.py
# run batch evaluation
bash benchmarks/basic.sh

3D Gaussian Splatting (3DGS) evaluation on MI300X#

The following table summarizes the average evaluation metrics, training memory usage, and training time.

Model	PSNR (dB)	SSIM	LPIPS	Train Memory (GB)	Train Time (s)
gsplat-7k	27.61	0.84	0.18	4.52	159.96
gsplat-30k	29.15	0.87	0.12	6.31	872.00

Performance metrics terms#

PSNR (Peak Signal-to-Noise Ratio): Ratio between the maximum possible signal power and the power of corrupting noise. Higher values indicate better reconstruction quality. Typically, values above 30 dB represent good quality.
SSIM (Structural Similarity Index): Measures the similarity between two images considering luminance, contrast, and structure. Ranges from -1 to 1, where 1 indicates perfect similarity. SSIM is more perceptually aligned than PSNR.
LPIPS (Learned Perceptual Image Patch Similarity): Uses a neural network (AlexNet or VGG) to compute perceptual similarity between images. Lower values indicate more perceptually similar images. It’s considered to better align with human perception than PSNR or SSIM.
Number of Gaussians and rendering time: Provides insights into model efficiency.

Scene-wise training memory and time#

The following tables show training time and memory for each scene. The goal is to train faster with less GPU memory.

Train memory (GB):

Model	Bicycle	Bonsai	Counter	Garden	Kitchen	Room	Stump
gsplat-7k	6.70	2.85	2.82	7.26	3.04	2.81	6.16
gsplat-30k	11.59	2.95	2.82	11.06	3.59	3.61	8.56

Train time (s):

Model	Bicycle	Bonsai	Counter	Garden	Kitchen	Room	Stump
gsplat-7k	147.44	161.15	169.35	168.50	175.19	155.99	142.11
gsplat-30k	1102.74	719.39	771.37	1067.28	814.53	758.39	870.30

Reproduced metrics per scene#

PSNR (dB):

Model	Bicycle	Bonsai	Counter	Garden	Kitchen	Room	Stump
gsplat-7k	23.69	30.15	27.57	26.59	29.42	29.97	25.90
gsplat-30k	24.93	32.26	29.19	27.63	31.54	31.75	26.78

SSIM:

Model	Bicycle	Bonsai	Counter	Garden	Kitchen	Room	Stump
gsplat-7k	0.67	0.93	0.89	0.83	0.91	0.90	0.73
gsplat-30k	0.76	0.94	0.91	0.87	0.93	0.92	0.77

LPIPS:

Model	Bicycle	Bonsai	Counter	Garden	Kitchen	Room	Stump
gsplat-7k	0.30	0.13	0.18	0.11	0.11	0.19	0.23
gsplat-30k	0.16	0.11	0.13	0.07	0.08	0.14	0.14

Number of Gaussians:

Model	Bicycle	Bonsai	Counter	Garden	Kitchen	Room	Stump
gsplat-7k	4.46M	1.55M	1.64M	4.83M	1.89M	1.48M	4.11M
gsplat-30k	7.78M	1.85M	1.30M	6.69M	2.12M	2.29M	5.73M

Summary#

Gsplat-7k trains faster with less GPU memory, while gsplat-30k achieves higher PSNR and SSIM metrics but requires more memory and time. This evaluation demonstrates the trade-offs between model size, training efficiency, and reconstruction quality on the Mip-NeRF 360 dataset.