Vai al contenuto principale

Image to Gaussian Splat: Turn One Image into PLY or SPLAT

A practical image to Gaussian splat workflow: prepare one source image, choose the TripoSplat settings, run the job, and inspect the PLY or SPLAT output before you hand it off.
Pubblicato9 ott 2026
Aggiornato9 ott 2026

Image to Gaussian splat is the task of turning a single still image into a Gaussian splat file - a scene made of many small 3D Gaussians - rather than a polygon mesh. On TripoSplat that job runs in the hosted AI 3D Generator: you add one image, set the Gaussian budget and inference controls, run the TripoSplat lane, and download the result as PLY or SPLAT.

This page is the workflow itself. For the concept behind it, read 3D Gaussian splatting.

What this task is, and what it is not

The input is one image. There is no multi-view capture and no video; the reconstruction has to infer depth from that single view. That makes source quality the biggest lever you control.

It is not photogrammetry. Nothing is measured, and no camera poses are solved. The output is a viewable Gaussian scene, so treat the parts of the object you could not see as estimates.

Step by step

  1. Prepare the source image. A single opaque subject, centred, fully in frame, lit evenly, on a background that differs from the subject. Shoot or crop so nothing important is cut off. If the file is heavy, the free image compressor shrinks it locally first, and the image cropper fixes framing without uploading.
  2. Open the AI 3D Generator. Upload a JPEG, PNG, WebP, GIF, or AVIF file up to 10 MB, or start from a sample.
  3. Set the Gaussian budget and inference controls. The TripoSplat lane exposes Gaussian count, inference steps, and guidance scale, plus the output format. Raise the budget when you want more surface detail; lower it when you want a cheaper, lighter file.
  4. Pick the output format. PLY is the default and the safe choice for desktop viewers. Choose SPLAT when your target viewer expects the compact splat format.
  5. Run the job. Guests can try a single lightweight preview with fewer Gaussians. A signed-in TripoSplat run is priced 48-72 credits depending on the settings.
  6. Inspect the result. Preview it in the workspace, then open the downloaded file in the free PLY and splat viewer to check it locally. If you actually need a mesh, use the 3D mesh generator guide instead of converting the splat.

When to use this workflow

  • You have one good photo or render and want a Gaussian scene fast.
  • You want a lightweight, viewable asset for a turntable, viewer, or preview.
  • You are validating whether a subject is even worth a full pipeline pass.

If the object must be printable, measured, or rigid, a splat is the wrong target - start from a mesh.

Common mistakes

  • Feeding a cluttered scene. Extra objects get reconstructed too. Isolate the subject.
  • Expecting the back of the object. Single-view reconstruction guesses hidden faces.
  • Maxing the budget by reflex. A higher Gaussian count and step count costs more credits and produces a heavier file; start balanced.
  • Skipping the local check. Look at the file in a viewer before handing it to anyone.
  • Confusing the formats. PLY and SPLAT are splat containers; OBJ, STL, GLB, and FBX are mesh containers. Converters like GLB to STL only make sense on the mesh side.

FAQ

How many images does image to Gaussian splat need?

One. The hosted lane reconstructs from a single image, so more images do not help here.

What file types can I upload?

JPEG, PNG, WebP, GIF, and AVIF, up to 10 MB.

How much does a run cost?

A signed-in TripoSplat run is priced 48-72 credits depending on the Gaussian budget and settings, and the workspace shows the exact price before you submit. Guests can run one lightweight preview first. See pricing for credit packs.

Can I get a mesh instead?

Yes - that is a different lane. See the 3D mesh generator guide, or the image to 3D print guide if the end goal is a print.

Ready to try it? Open the AI 3D Generator with one clean image.

Avvisi sulle nuove funzioni

Un'email quando esce un nuovo prodotto o una nuova funzione. Niente spam, disiscriviti quando vuoi.