Turn a Smartphone Video Into a 3D Gaussian Splat

You do not need a drone, a camera rig, or a photogrammetry turntable to build a fully interactive 3D scene. You need a phone, about five minutes of video, and a capture routine that gives the reconstruction enough angles to work with.

That last part is where most people lose the scene. The phone footage is fine. The coverage is not.

See the finished result:

Click to explore in 3DInteractive 3DFrontyardKobid Upadhyay·Captures Studio

What You Are Actually Making

Gaussian Splatting takes a set of overlapping views of a scene and reconstructs it as millions of tiny colored, oriented blobs. Point a camera at those blobs from any angle and you get something that looks photographic rather than modeled. No polygons, no texture maps, no baked lighting.

The reconstruction figures out where your camera was for every frame it uses. It can only do that if consecutive frames share enough visual overlap, and it can only reconstruct a surface it has actually seen. Both of those facts drive everything below.

What You Need

• Any reasonably modern smartphone
• Roughly five minutes of video
• A subject you can physically walk a full circle around
• Even, stable lighting

That last one matters more than phone specs. Harsh moving shadows and blown out highlights confuse the reconstruction far more than a slightly older sensor does.

Step 1: Adjust Your Camera Settings

Do this before you record anything. Changing settings mid capture is one of the fastest ways to ruin a scan.

1. Open the camera app and give yourself a second to frame the subject.
2. Switch to video mode. Photo mode burst captures are possible but video is simpler and gives you far more frames.
3. Select ultrawide, usually labeled 0.6x. The wider field of view means each frame contains more of the scene, which gives the reconstruction more shared features between consecutive frames. It also means you can stand closer without cropping the subject.
4. Set the highest resolution and frame rate available. 4K at 60fps is the target. More pixels means more detail to reconstruct from. More frames per second means more usable stills and better tolerance for slightly quick movement.
5. Lock autofocus and autoexposure. Press and hold on your subject until the lock indicator appears. This is the single most important setting on this list. If focus hunts or exposure shifts as you walk, the reconstruction sees the same surface as two different surfaces.
6. Adjust the exposure slider until the image looks good. Slightly under is safer than slightly over. You can recover shadow detail. You cannot recover a blown highlight.
7. Start recording and do not touch the screen again until you are finished.

Step 2: Chest Level Scan

Hold the phone at chest height, parallel to the ground, pointed at the center of your subject.

Walk the complete perimeter in a counterclockwise direction. Keep your pace steady and keep the camera aimed inward the entire time. Think of yourself as orbiting rather than panning.

This pass establishes the main loop the reconstruction uses to solve camera positions. If this one is sloppy, nothing after it will save the scene.

Keep in mind: walk smoothly, do not stop and start, and do not let the subject drift out of frame.

Step 3: Overhead Scan

Raise the camera above head level and tilt it downward toward the subject.

Walk the same perimeter, same direction, same steady pace. Because you are looking down, this pass captures top surfaces the chest level loop could never see: the tops of objects, upward facing planes, anything sitting flat.

You are shooting slightly blind here since the screen is above your eyeline. That is fine. Trust the angle and keep your walking radius consistent.

Step 4: Low Angle Scan

Drop the camera to roughly knee level and tilt it upward.

One more full loop around the perimeter. This is the pass people skip most often, and it is exactly why undersides, overhangs, and the bottoms of objects come out soft or missing entirely in amateur splats.

Three loops at three heights gives the reconstruction a vertical spread of viewpoints, which is what lets it resolve depth properly instead of guessing.

Step 5: Details Scan

Now break the perimeter discipline and move in closer.

Pick the areas that actually matter: an interesting texture, a face, an intricate object, whatever a viewer will zoom into. Move around those areas in a hemisphere, covering the same three height angles you used for the wide loops. Low, level, high.

Move slowly. Detail comes from dense overlapping coverage, not from holding the camera still. A slow deliberate sweep gives you dozens of usable frames of the same surface from marginally different positions, which is precisely what the reconstruction wants.

Around five minutes of total footage across all passes is a good target.

Step 6: Upload to Captures Studio

1. Go to captures.studio.
2. Click Create 3D.
3. Choose your video file.
4. Give the capture a name.
5. Hit upload.

Processing is automatic and runs in the cloud, so your machine is not doing the heavy lifting and you can close the tab. Come back when it is done.

Step 7: Generate and Share

Once reconstruction finishes you get an interactive scene you can orbit, zoom, and share with a link. Anyone with a browser can open it. No plugin, no app, no download.

That is the whole appeal. You walked around something with a phone for five minutes, and now anyone in the world can look at it from any angle.

Common Mistakes Worth Avoiding

Moving too fast. Motion blur destroys feature matching. If your footage looks smeary when you scrub through it, you were walking too quickly for the light you had.

Forgetting to lock exposure. Every automatic adjustment is a discontinuity the reconstruction has to reconcile.

Only shooting one height. A single chest level loop produces a scene that looks convincing from exactly one elevation and falls apart the moment someone tilts the view.

Changing distance mid loop. Keep a consistent radius on the perimeter passes. Save the close work for the details scan.

Reflective or transparent subjects. Glass, mirrors, and polished chrome do not have a fixed appearance from different angles, so the reconstruction cannot place them. Matte surfaces are your friend.

Moving subjects. Anything that shifts between passes will smear. Wait for people to clear the frame, or accept the ghosting.

Quick Reference