Which Trainer Should I Use?
A quick-decision guide distilled from the trainer benchmarks. If you want the full evidence, read Indoor, Outdoor, and Indoor vs Outdoor. This page is the short answer.
What 'quality' means here
These benchmarks did not compute perceptual fidelity metrics (PSNR / SSIM / LPIPS).
All five pipelines were trained to 30,000 iterations with default hyperparameters, and were
compared on structural and qualitative grounds: Gaussian count, exported .ply size,
training time, raw-artifact distribution, and how cleanly each model could be post-processed
in SuperSplat. "Quality" below refers to structural compactness and cleanliness, not measured
image fidelity. All numbers come from a single RTX 4060 / Windows 11 machine, one indoor
and one outdoor 151-frame DL3DV scene.
TL;DR
| If you want… | Use | Why |
|---|---|---|
| The fastest, lightest result | Nerfstudio | Shortest training time (25–30 min) and by far the smallest output (40–47 MB); stays compact with little cleaning. |
| The best speed-to-density balance | gsplat | Densest reconstruction of all five, yet the fastest of the high-density trainers (45–50 min). |
| A predictable, fixed-budget model | LichtFeld Studio | MCMC densification caps the model at exactly 1 M Gaussians in both scenes; fast and consistent. |
| The least post-processing effort | OpenSplat | Most stable across scenes — raw output already sits mostly inside the kept envelope (~21% removed by cleaning). |
| The canonical reference | Inria gaussian-splatting | The original 3DGS implementation; structurally stable but the slowest. |
Numbers at a glance
Results per scene (RTX 4060, 30k iterations, default hyperparameters):
| Trainer | Gaussians (in / out) | Size MB (in / out) | Train min (in / out) | Cleaning Δ (in / out) |
|---|---|---|---|---|
| Nerfstudio | 170k / 198k | 40 / 47 | 30 / 25 | −25% / −13% |
| gsplat | 1.27M / 1.03M | 285 / 232 | 50 / 45 | −45% / −46% |
| OpenSplat | 511k / 589k | 121 / 139 | 60 / 50 | −21% / −21% |
| LichtFeld Studio | 1.00M / 1.00M | 237 / 237 | 60 / 50 | −20% / −58% |
| Inria | 956k / 777k | 226 / 184 | 120 / 60 | −46% / −52% |
in = indoor, out = outdoor. Cleaning Δ = percentage of Gaussians removed by SuperSplat post-processing.
Trade-offs in detail
Training speed
- Nerfstudio is the fastest in absolute terms (25–30 min).
- gsplat is the fastest trainer that still produces a dense model (45–50 min).
- Inria is the slowest — 120 min indoors, roughly double the others on this hardware. Treat it as a reference baseline rather than a production choice.
Output size and density
- gsplat produces the densest, heaviest models; LichtFeld Studio is close but fixed at a 1 M Gaussian budget.
- Nerfstudio is an order of magnitude lighter than the rest — useful when storage, download bandwidth, or runtime (e.g. XR rendering) is a constraint.
- Storage cost per 100k Gaussians is nearly identical across tools (~22–24 MB), so file size tracks Gaussian count rather than encoding efficiency.
Cleanliness of the raw model
- OpenSplat loses only ~21% to cleaning in both scenes — its raw output is already well-contained.
- Inria and gsplat carry ~45–46% removable peripheral splats indoors; much of their raw size is clutter.
- LichtFeld Studio is the most scene-dependent: only ~20% removable indoors but ~58% outdoors.
Hardware
All results are from a single RTX 4060 — a mid-range consumer GPU — so none of these trainers requires datacenter hardware for 151-frame scenes. All five are command-line / framework-level tools targeting Windows or Linux. If you want a one-click GUI experience, see the commercial trainers covered in Overview / Trainers.
Indoor vs outdoor
- Indoor scenes are forgiving: tight boundaries make cleaning straightforward and aggressive pruning safe. Most tools converge to compact models regardless of starting density. Optimize for speed and size — Nerfstudio or gsplat are strong picks.
- Outdoor scenes are harder: far-field sky and vegetation are legitimate content, so cleaning must stay conservative. OpenSplat and Nerfstudio keep a large share of their Gaussians as valid background (low cleaning Δ), while LichtFeld Studio and Inria carry a lot of removable peripheral clutter. For unconstrained scenes, favor a trainer whose raw output is already well-bounded and expect more manual cleanup.
Note on the XR pipeline
For the downstream XR-viewer evaluation, the indoor scene was used as the reference, because its bounded geometry gives a more controlled and consistent benchmark across pipelines.
Bottom line
- Default pick for most users: gsplat — fast, dense, well-supported, and the CUDA backend behind Nerfstudio's Splatfacto model.
- When size or runtime is king (e.g. XR streaming): Nerfstudio.
- When you want a fixed, predictable output budget: LichtFeld Studio.
- When you want the least post-processing effort, especially outdoors: OpenSplat.
- When you need the canonical 3DGS reference: Inria gaussian-splatting — accept the longer training time.
Setup instructions for each trainer are in the How-To guides.