How should you choose between Splatware training options?
Start with the final use case, then choose the training option that balances fidelity, speed, model size, and output needs.
A Gaussian Splatting workflow has several goals that can compete with each other. More Gaussians can improve detail, but a public viewer also needs performance. Fast training is useful for experiments, but final client work may need more quality.
The right choice depends on what the model is supposed to do. A product demo, real estate tour, aerial reconstruction, simulation scene, and research dataset should not all use the same assumptions.
How many Gaussians should a Splatware model use?
Use enough Gaussians to preserve important detail, but do not choose a high count only because it sounds better.
Gaussian count decision guide
| Gaussian count direction | Best for | Tradeoff |
|---|---|---|
| Lower count | Tests, drafts, small scenes, quick previews. | Faster and lighter, but less detail. |
| Medium count | Balanced public viewers and general commercial use. | Good compromise between quality and performance. |
| High count | Detailed products, premium scenes, demanding real estate, aerial captures. | Better fidelity, larger files, more viewer load. |
| Up to 10 million Gaussians | Production workflows where maximum detail is needed and supported. | Highest fidelity direction, but must be tested for performance. |
| Custom count | Teams that know their target device and quality needs. | Requires judgment and review. |
Which training direction fits each input type?
Different inputs need different training priorities.
Input-based model choice
| Input | Training priority | Why |
|---|---|---|
| Smartphone video | Stability and quality review | Phone footage can work well but often varies in sharpness. |
| DSLR or mirrorless images | High-fidelity detail | Sharp photos can support strong product and scene reconstruction. |
| 360 video | Spatial completeness and viewer performance | Large scenes need both coverage and browser usability. |
| INSV footage | Preparation plus spatial reconstruction | Camera-specific footage may need export or direct workflow handling. |
| Drone footage | Large-scene fidelity | Aerial scenes need detail over broad areas. |
| COLMAP dataset | Advanced dataset training | Prepared poses and structure can support controlled pipelines. |
| Mesh input | Mesh-to-splat conversion | Useful when existing geometry should become splat-style output. |
Which Splatware outputs matter for advanced users?
Advanced users should consider not only the splat itself, but also point cloud, dense mesh, camera path, background audio, and downloads.
Advanced output comparison
| Output | Best for | Why it matters |
|---|---|---|
| Splat model | Photorealistic viewing and sharing. | Primary 3DGS result. |
| Point cloud | Inspection, analysis, and downstream processing. | Helps users understand reconstructed structure. |
| Dense mesh | Physics, collision, simulation, and geometry workflows. | Adds surface-style data where splats alone are not enough. |
| Camera path | Capture review, playback, tours, and technical workflows. | Preserves movement and viewpoint context. |
| Background audio | Immersive walkthroughs and scene presentation. | Adds context to tours and recorded environments. |
| Downloads | Ownership, portability, and integration. | Allows advanced use outside the hosted viewer. |
Should you choose speed or quality?
Choose speed for validation and quality for final outputs.
Speed vs quality decision
| Situation | Choose | Reason |
|---|---|---|
| Testing whether a capture worked | Speed | You need feedback quickly. |
| Client-facing product model | Quality | The model affects trust and presentation. |
| Real estate draft tour | Balanced | You need spatial review without overbuilding the draft. |
| Final property or product asset | Quality | The output will be judged by buyers or clients. |
| Aerial reconstruction | Quality with performance testing | Large scenes need detail but must remain viewable. |
| Public website embed | Balanced | Load time matters as much as fidelity. |
What are common mistakes when choosing Splatware model settings?
The biggest mistake is choosing maximum settings for everything or using draft settings for final work.
Model selection mistakes
| Mistake | Why it hurts | Better approach |
|---|---|---|
| Always choosing maximum Gaussian count | Can create unnecessary file size and viewer load. | Match count to use case. |
| Using fast tests as final output | May look unfinished for clients. | Retrain with higher fidelity when needed. |
| Ignoring input quality | No setting can fully fix bad capture. | Improve capture before increasing model size. |
| Ignoring downstream outputs | You may later need mesh, point cloud, or downloads. | Decide outputs before training. |
| Not testing the viewer | High quality is useless if the link performs poorly. | Test on target devices. |
| Forgetting the audience | Internal and public models need different tradeoffs. | Choose settings for the viewer, not just the creator. |
Frequently asked questions
What Splatware model should I use for a first test?
Use a faster or lighter setting for a first test. The goal is to confirm that the capture has enough coverage and quality before spending time on a final high-fidelity model.
What Splatware model should I use for product visualization?
Use a high-fidelity setting that preserves detail, edges, labels, material appearance, and important surfaces. Product splats are judged on clarity and trust.
What Splatware model should I use for real estate?
Use a balanced setting that preserves spatial clarity while keeping the viewer smooth enough for buyers or clients to open in a browser.
What Splatware model should I use for drone capture?
Use a large-scene, high-fidelity direction with strong coverage and performance testing. Aerial reconstructions often need detail across broad areas.
Does more Gaussians always mean better?
Not always. More Gaussians can improve fidelity, but they can also increase file size and viewer load. The best count depends on the use case and target devices.
Can Splatware output up to 10 million Gaussians?
Splatware can support production workflows with configurable Gaussian counts, including up to 10 million Gaussians when the workflow and plan support it.
Can I download the output of a Splatware model?
Yes. Splatware is designed to provide access to generated model outputs and related assets for advanced workflows.
When do I need a dense mesh in addition to a splat?
Use a dense mesh when you need geometry-oriented workflows such as physics, collision, simulation, measurement support, or downstream 3D processing.