How to Make a Gaussian Splat: A Step-by-Step Guide
Tutorial · 2026-06-05 · 9 min read · by SplatMart Team
Make your first Gaussian splat from a phone video, capture technique, the apps and tools to use, processing, and how to avoid floaters and blur.
To make a Gaussian splat, you capture a scene with overlapping photos or a slow video, then process it with a splatting app or tool that reconstructs it into a 3D model. You don't need special hardware to start, a phone and good technique are enough. Here's the full step-by-step.
What you'll need
- A camera, a modern smartphone is fine.
- A capture/processing tool such as SplatMart Leighton, Luma AI, Polycam, KIRI Engine, or Postshot, or open-source tools if you want to train it yourself on a GPU.
- A subject with even lighting and not much movement.
Step 1: Plan and light the scene
Pick a subject that holds still and light it evenly. Soft, diffuse light, overcast outdoors or bright indirect light indoors, gives the cleanest results. Avoid harsh direct sun, strong shadows, and shiny or transparent surfaces, which are the hardest things to capture well.
Step 2: Capture with full coverage
This is the step that makes or breaks a splat. Move slowly around the subject and photograph or film it from many angles and heights, low, mid, and high, with lots of overlap between viewpoints. Imagine wrapping the subject in a dome of camera positions and filling in every part of it.
Capture do's and don'ts
- Do move steadily and keep the subject in frame the whole time.
- Do capture more coverage than you think you need, especially edges and the top.
- Don't let people, pets, leaves, or traffic move through the shot, moving objects become floaters.
- Don't change zoom mid-capture, and avoid motion blur by moving slowly.
Step 3: Process into a splat
Upload your capture to your chosen app. Behind the scenes it solves the camera positions (Structure from Motion) and then trains the Gaussians to match your images. Cloud tools handle all of this for you; SplatMart's Leighton accepts the original video without manual frame extraction. If you train locally with open-source tools, you'll need a GPU and more setup. See the dedicated video-to-Gaussian-splat walkthrough.
Step 4: Review, clean up, and export
- Inspect the result for floaters, holes, and blur, re-shoot weak areas if needed.
- Crop the scene bounds tightly to cut background junk.
- Export as .ply or .splat so you can view it, use it in an engine, or share it. See our guide to splat file formats.
The faster alternative: buy a splat
Capturing a great splat takes practice, and some subjects (large, busy, or reflective scenes) are genuinely hard. If you need a polished asset for a project now, you can buy a professionally captured one instead. SplatMart is a marketplace of ready-to-use splats, browse and download one that fits your scene.
Frequently asked questions
How do I make a Gaussian splat from a video?
Record a slow, steady video that orbits a static subject from several heights, then upload it to a splatting tool such as SplatMart Leighton, Luma AI, Polycam, KIRI Engine, or Postshot. The tool solves the camera positions and trains the splat; cloud options do not require your own GPU.
Do I need a GPU to make a Gaussian splat?
Not if you use a cloud app, the processing runs on their servers. You only need a GPU if you train splats locally with open-source tools.
How long does it take to create a splat?
Capture takes a few minutes; cloud processing typically takes minutes to under an hour depending on the scene and service.
Why does my splat look blurry or full of floaters?
That almost always traces back to capture: not enough overlap or coverage, moving objects, harsh lighting, or motion blur. Re-shoot slowly with even light and fuller coverage, and crop the bounds tightly.