ZipSplat
Scene reconstruction · A few photos or a short video of a scene
About ZipSplat
ZipSplat predicts a 3D Gaussian scene and cameras from a small set of images or video frames. Its documented pipeline does not require supplied camera poses or intrinsics, and refines the initial predictions with TokenBA.
The output opens in a browser splat viewer, alongside a fly-through along the predicted cameras. Controls for retaining fewer Gaussians let you inspect the trade-off between scene size and visible detail. Token-group colors provide another way to examine how the scene is represented.
How to use ZipSplat
- Upload overlapping scene photographs or a short video, or choose an example.
- Run reconstruction and inspect the initial view and camera fly-through.
- Adjust the retained-Gaussian or token visualization controls to compare the scene representation.
What to know before you start
- A Gaussian scene is a rendering representation, not an ordinary editable triangle mesh.
- The published model weights use a non-commercial license. Review the model’s terms before reuse.
The provider controls availability, sign-in, processing, and downloads. Save any result before reloading or closing the demo.
Must I know the camera positions beforehand?
No. ZipSplat predicts cameras from the input views as part of its reconstruction workflow.
