FastGS-Adaptive

Structure-Aware View-Consistent Densification
for Accelerated 3D Gaussian Splatting

Hana L. Goshu1 Tadesse G. Wakjira2 Mostafa Fouda3,4 Kin-Man Lam1
1The Hong Kong Polytechnic University  ·  2Kennesaw State University  ·  3Idaho State University  ·  4Tohoku University
Paper soon Code Compare Results
28.09PSNR dB
0.823SSIM
0.211LPIPS
8/9scenes improved

Novel-view rendering on Mip-NeRF 360. Ground truth left, FastGS-Adaptive right.

Drag to compare

FastGS-Adaptive
FastGS-Big
FastGS-BigOurs
FastGS-Adaptive architecture
AESLAdaptive Edge-Structured Loss — modulates the per-pixel photometric penalty with a Sobel edge prior cached once per camera.
PSAPerceptual Structure-Aware densification — replaces the thresholded residual in the scoring rule with an edge-gated combination of photometric and local SSIM error.

Results

Best PSNR, SSIM and LPIPS on Mip-NeRF 360; best PSNR on all three benchmarks.

Method Mip-NeRF 360Deep BlendingTanks & Temples
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
3DGS27.53.812.22129.71.903.24123.71.850.170
Mini-Splatting27.32.821.21729.99.907.24423.46.844.181
Speedy-Splat26.91.781.29529.42.898.27223.38.816.242
Taming-3DGS27.48.794.26129.50.894.27823.89.833.214
DashGaussian27.73.817.21829.65.906.24624.00.853.178
FastGS27.56.797.26130.03.901.27024.15.839.210
FastGS-Big27.93.820.21630.12.907.24324.39.855.175
FastGS-Adaptive 28.09.823.211 30.31.912.244 24.49.855.180

Baseline figures are those reported in the corresponding publications.

Per-scene PSNR gain

Per-scene PSNR gain over FastGS-Big on Mip-NeRF 360. Eight of nine scenes improve.

More scenes

kitchen
kitchen
bonsai
bonsai
bicycle
bicycle

BibTeX

@inproceedings{goshu2026fastgsadaptive,
  title     = {FastGS-Adaptive: Structure-Aware View-Consistent Densification
               for Accelerated 3D Gaussian Splatting},
  author    = {Goshu, Hana L. and Wakjira, Tadesse G. and Fouda, Mostafa and Lam, Kin-Man},
  year      = {2026}
}