Tools / Photo to 3D Splat
Photo to 3D Splat
Turn one photo of a person or a thing into a photo-real 3D Gaussian splat you can walk around: the photo's own face and clothes on the front, a full body on the sides and back.
How it works
The steps
- BiRefNet cuts the person (or thing) out of the photo.
- Idol turns the cut-out into 262,144 3D Gaussians in about 20 seconds - a whole body, including the sides and back it has to invent. On its own it is soft and the face drifts a little.
- MoGe-3 measures depth and surface direction for every pixel of the photo.
- The hybrid lines the photo up with Idol's body (outline fit, then depth), keeps the photo's own pixels wherever they face the camera, removes Idol's front layer under them and blends a thin seam. Idol keeps the sides and back.
- Out comes one .spz file of 3-6 MB, standing on the ground at the height you choose, plus a turntable preview from eight sides.
Examples
Turn them around
Three figures, each from one picture. "Hybrid" is the finished splat; "Idol only" is the body before the photo goes on the front. Compare the faces from the front and at 45°, then look at the back.
Drag to turn it, pinch or scroll to zoom.
What it is good for
Things that stand still
A splat looks like the photo from the front and holds together all the way round, and it is cheap enough to fill a square with people. It does not bend, so anything that moves needs a rigged mesh.
Every run
The turntable preview
Each run also draws a quick preview from eight sides: the hybrid on top, Idol only underneath.
Bench, 31 Oct 2026
Three ways compared
| Way | Front | 45° and side | Back | Verdict |
|---|---|---|---|---|
| A: photo lift (BiRefNet + MoGe-3, one Gaussian per pixel) | the exact photo | a thin relief past about 30° | a dark outline | good only for straight-on backdrops |
| B: Idol only | soft, the face changes a little | a believable body | a believable body | fine for crowds seen from afar |
| C: hybrid (Idol body + photo front) | the photo's face and clothes | photo face to about 45°, an edge from the side | Idol's back, no holes | the default |
Models and licences
What it runs on
| Step | Model or tool | Licence |
|---|---|---|
| Cut-out | BiRefNet | MIT |
| Body, sides and back | Idol (speridlabs), 262,144 Gaussians | Apache-2.0 (code and weights) |
| Idol's picture encoders | DINOv3 ViT-H+ (Meta) and the FLUX.2 VAE | DINOv3 licence; FLUX.2 VAE Apache-2.0 |
| Depth and surface direction per pixel | MoGe-3 ViT-L | MIT |
| Line-up, seam blend, .spz file | our hybrid | ours |
Splats are frozen: no skeleton, no lighting, no collisions - characters that walk, jump or fight go through Photo to 3D Character instead. Known limits: from the side the photo's face sits on Idol's head with a visible edge; a hand-held thing that Idol placed differently (the felt figure's phone) can show twice; the height is set by hand because a photo does not say how tall someone is.
Benchmarks and live pages