How to Turn a Video into a 3D Model (Gaussian Splat)

Yes, a short video is enough

You do not need a 3D scanner or a turntable to make a 3D model from video. A steady 15 to 60 second clip from a normal smartphone, filmed while you walk around a still subject, contains hundreds of slightly different viewpoints. Software can work out where the camera was for each frame and then build a 3D scene that you can orbit, zoom and export.

The result is not a mesh for 3D printing. It is a Gaussian splat: a cloud of millions of small, soft, coloured shapes that together reproduce how the subject looked from every angle you filmed. That makes it excellent for realistic views of objects, rooms and outdoor spots, and less suited to measurements or CAD.

The examples on our home page were each made from a single phone video between 13 and 20 seconds long.

How to film a video for 3D

The single biggest factor in the quality of a 3D model from video is how you film it. Good footage beats any setting you can change later.

• Pick a subject that does not move: an object on a table, a plant, a statue, a corner of a room.
• Walk around it in a slow, steady arc. Do not stand still and pan the camera; the software needs your position to change.
• Keep the subject in frame and at a similar distance. Overlap between neighbouring frames is what lets the cameras be matched.
• Move slowly to avoid motion blur, and prefer even daylight or soft indoor light.
• If you can, do a second pass from a higher or lower angle so the top and the sides are covered.
• Avoid mirrors, glass, glossy surfaces, people walking through the shot and large blank walls. These confuse camera matching.

What happens to your video, step by step

Turning a video into a 3D model happens in three stages, and SplatLoom runs all of them on your own computer.

First, frames are extracted. Instead of using every frame, SplatLoom samples up to six frames per second, which is plenty of overlap without duplicating near-identical images. Blurry and duplicate frames are flagged.

Second, the camera position of every frame is solved with structure-from-motion, using the open-source COLMAP engine. This produces a sparse 3D point cloud and a precise path of where your phone travelled. If not enough frames connect into one scene, SplatLoom stops early and tells you to film more overlapping angles rather than spending time on a poor result.

Third, a Gaussian splat is trained on your graphics processor with the open-source Brush engine. Starting from the sparse points, it repeatedly renders the scene from each camera, compares it with the real frame and adjusts millions of small Gaussians until the views match. In our tests this used every other frame for training, which kept quality while saving time.

How long does it take?

Processing time depends on the length of the video, its resolution and your hardware. In our own tests on an Apple M5 Pro MacBook, phone videos of 13 to 20 seconds took about 9 to 14 minutes from import to a finished scene, with most of that time spent on training.

Longer videos and photo sets with more images take longer. Because everything runs locally there is no upload or queue, and you can keep working while it processes. These numbers are measurements from one machine, not a guarantee.

Checking and cleaning up the result

When processing finishes, the scene opens in a 3D viewer from the angle you filmed. Rotate around it and look at the edges: areas you never filmed will be empty or blurry, and the sky or distant background may look soft. That is normal for any 3D model made from video.

To tidy things up you can crop the scene to a box or sphere around your subject, select and delete stray regions with a rectangle, or remove thin spikes with one click. Every edit can be undone, and the original result is always kept.

Exporting your 3D model

SplatLoom exports in four formats, so you can use the result in different places:

• PLY: the standard Gaussian splat file, which other splat tools and viewers can open.
• MP4: a 12-second 1080p orbit video, ready to post.
• PNG: a still image from the angle you choose.
• HTML: a self-contained web page with an interactive viewer that works offline.

You can also open any PLY in our free online viewer, which runs entirely in your browser.

Common problems and fixes

If the camera positions cannot be solved, the video usually lacks overlap or has too much blur. Film again more slowly, and walk around the subject instead of panning.

If the scene looks smeared, there were probably too few viewpoints of that area. Add a second pass from another height.

If you see floating blobs or spikes in empty space, crop the scene to your subject first; that removes most of them without touching details.

If reflective or transparent parts look wrong, that is a current limit of the technique. Matte objects reconstruct far better.

Making a 3D model from video with SplatLoom

SplatLoom is a desktop app for Mac (Apple Silicon, macOS 15 or later) and Windows 10/11 that turns a video or a set of photos into a Gaussian splat on your own computer. Pick the video, click Generate, then view, tidy up and export. There is no account, no upload and no subscription: it is a one-time purchase of US$19.90 that includes both the Mac and Windows versions.