02 · Training · 3D Gaussian Splatting · controlled ablation

FastGS Training Efficiency

Once the representation looked promising, the next question was how expensive it was to create. This study asks how much 3DGS training cost can be reduced before held-out quality degrades materially.

BonsaiGardengsplatA100training scheduledensification ablation

Open study branch on GitHub

Why this came next

A good representation is not yet a practical one if creating it is unnecessarily expensive.

3D Gaussian Splatting grows its scene representation during training through densification. More Gaussians can increase detail, but they also increase training time, memory consumption, and checkpoint size.

After the NeuralScene comparison, I wanted to know whether all of that model growth was necessary. I therefore varied the optimization budget and when densification stopped, then measured the consequences rather than assuming that longer training was always better.

Research question

How much can 3DGS training time be reduced through budget and densification-schedule changes before held-out quality degrades materially? Six Bonsai presets were used to identify candidate operating points. I then repeated a smaller set on Garden to test whether the same behavior transferred across scene complexity.

Bonsai ablation

Preset	Train sec	Checkpoint MiB	PSNR	SSIM	LPIPS	Gaussians
baseline 7K	421.79	212.53	29.6619	0.9246	0.1527	944,291
early 5K	299.69	207.18	29.5314	0.9257	0.1562	920,498
early 3.5K	209.96	159.16	27.8726	0.9046	0.1863	707,159
densify stop 3K / 5K	299.95	188.85	30.1942	0.9302	0.1507	839,074
sparse refine 5K	295.29	122.97	29.6309	0.9246	0.1615	546,368
densify stop 3K / SH fast 5K	299.69	187.12	30.2438	0.9312	0.1498	831,388

Bonsai: about 29% less training time without a held-out quality loss.

The balanced 5K schedule reduced training time by about 28.95%, Gaussian/checkpoint scale by about 11.96%, and peak memory by about 14.39% relative to the 7K baseline, while PSNR increased by 0.58 dB. For this scene, the longer baseline was not automatically the better operating point.

Garden replication

Preset	Train sec	Checkpoint MiB	PSNR	SSIM	LPIPS	Gaussians
baseline 7K	999.32	591.49	25.0484	0.7284	0.2889	2,628,056
balanced 5K	699.49	298.24	24.5905	0.6878	0.3536	1,325,117
compact / sparse 5K	692.37	210.57	24.3232	0.6751	0.3750	935,554

What transferred

The same shorter schedule again reduced training time by about 30%. Model growth, checkpoint size, and memory also fell substantially.

What did not transfer

Garden paid a measurable quality penalty. The balanced schedule reduced PSNR by about 0.46 dB and increased LPIPS. The same training shortcut was therefore not equally safe across scenes.

What I learned

The useful result was not a universal recipe. It was the failure of one recipe to generalize.

Bonsai suggested that the baseline schedule was doing more optimization and densification than this scene needed. Garden showed that a more complex scene could still benefit from the cost reduction while paying for it in reconstruction quality.

That changed the framing from “find the fast setting” to “find a scene-aware operating point.” A practical system would need to account for content complexity instead of hard-coding one training schedule.

What this led to next

03 · SplatStream Lab

Training became cheaper, but the resulting representation was still large. That shifted the bottleneck from creation cost to delivery cost: how much could the trained Gaussian scene be reduced before visual quality collapsed?

Continue to the compression study →

Research boundary

The study uses a single run per configuration, so the results are descriptive rather than statistical. The 5K schedules are empirical operating points for these scenes, not universal prescriptions. Checkpoint size includes training-state overhead and should not be interpreted as a deployable streaming payload.