How do you make a Gaussian splat from photos or a video?
You need a set of overlapping photos of the subject from many angles, or a video that moves around it, and a splatting method that suits the scene. For most scenes, standard Gaussian splatting with 10 to 500 photos is the right default. If you only have a handful of images, use a method built for sparse input. If you only have one photo, a single-image model can still give you a usable splat for nearby viewpoints.
OTOY Studio runs all of these as nodes, so you don't need a local GPU or a COLMAP install. Connect photos, a ZIP of an image sequence, or a video, and run.
Which method to use
| Node | What it does | Credits per run |
|---|---|---|
| Gaussian Splat | Standard 3D Gaussian Splatting (COLMAP), the general-purpose choice for scenes, walkthroughs, and product capture. Needs 10-500 overlapping photos, or a video. Outputs a viewable PLY splat. ~15-25 min. | 10 |
| InstantSplat | MASt3R neural reconstruction (no COLMAP) for sparse or tricky scenes where feature matching struggles. 3-35 images (all the same resolution), or a video. Outputs a viewable splat. ~10-20 min. | 22 |
| 2D Gaussian Splat | Surface-aligned splatting for geometry, extracts a clean triangle mesh (TSDF) and per-view depth maps. Needs 10-500 overlapping photos, or a video. Outputs a mesh + depth, not an inline splat. ~15-25 min. | 25 |
| Neural Harmonic Textures | Highest-fidelity splatting, a per-gaussian neural decoder captures reflections, glass, and other view-dependent materials that standard SH can't. Needs 10-500 overlapping photos, or a video. ~15-25 min. | 29 |
| Sharp (1-Photo 3D) | Instant single-photo 3D, a feed-forward net predicts ~1.2M gaussians in ~2s. Exactly 1 image; produces a 2.5D splat for nearby viewpoints. Best for quick previews and product shots. | 2 |
| TripoSplat (1-Photo 3D) | Generative single-photo 3D, a flow-matching model synthesizes a ~262K-gaussian splat from one image in ~1-2 min. Exactly 1 image; no multi-view or training. Tune with guidance/steps for fidelity. | 2 |
| Generate World | Generates an explorable world as a Gaussian splat from an image or video. | 17 |
Credits are list price at 1 credit = €0.01. The node shows the exact cost before you run it.
Shooting tips that matter more than the method
- Overlap. Each photo should share most of its view with the next. Walk around the subject in small steps rather than big jumps.
- Keep exposure and focus locked. Changing exposure between frames shows up as flicker and floaters.
- Avoid motion. People walking through, moving leaves and reflections that shift will all turn into noise. Neural Harmonic Textures handles reflections and glass better than standard splatting.
- From video, move slowly and avoid motion blur. A few hundred sharp frames beat a thousand blurry ones.
After the splat
Splat outputs in OTOY Studio can be edited in the built-in SuperSplat editor, which saves the result as a new node and keeps the original. You can render flythrough videos and stills straight from the editor, or drop the splat into a scene with other models and animate a camera. Splats download as .ply, and edited splats export as SPZ. You can also upload your own .ply splats (up to 500MB) to use in scenes. OctaneRender 2026.1 can render Gaussian splats path-traced alongside regular geometry.
A rendered flythrough can then be the reference for an AI video model, which is a reliable way to get a real camera move through a real space. See AI video with 3D control.
Last verified 2026-09-28. Prices and model lists come from OTOY Studio's billing configuration, exported 2026-09-28.