Which files can I upload?
E57 from terrestrial and mobile scanners, LAS and LAZ, PLY point clouds, and Riegl drone surveys. Scans with panoramic imagery and station positions get photoreal training; scans without them get a clean scene coloured from the points.
How are people and moving objects removed?
Where a scanner station looked straight through a spot that another station recorded as solid, that spot was a passing object, so it is removed. The scanner's photographs are checked as well, which catches people the laser only grazed. You can always review the result before sharing it.
Does it work on buildings with several floors?
Yes. Floors are split automatically and each floor is divided into rooms. Large buildings are processed floor by floor and merged into one model that streams in the browser, with a floor selector for viewers.
What are the semantic labels for?
They tell the pipeline what each part of the scene is, so it can treat people, glass and structure differently. They also travel with the model, so you can filter or search the scene by class.
Can I get a mesh instead of a splat?
Every job can also produce a textured mesh from the same capture, in GLB or glTF. OBJ and STL are available for CAD and printing workflows.
How long does it take?
A room-scale scan is usually ready in under an hour. Whole buildings take longer, scaling with the area and the number of floors; you can follow every stage live.
Do I need credits before I upload?
Uploading is free. After floor analysis, the selected rooms and quality determine the credits needed to start processing. Good costs 1 credit per room and Best costs 2. Credits are charged only after a successful result; failed jobs are never charged.
How long are my files kept?
e57.gs is processing, not permanent storage. Retention follows the limits shown in the service, so keep your own copies of source files and outputs. Share links can be opened by anyone who has the link.
Is e57.gs a converter or a training service?
It is an online E57 to Gaussian Splat converter whose full-training path runs 3D Gaussian Splatting optimization. It does more than map each input point to one fixed-color Gaussian: compatible scanner imagery and poses provide observed views for novel view synthesis.
How can I inspect the result's appearance data?
After a successful conversion, inspect the downloaded Gaussian PLY header. A degree-3 result contains 45 f_rest_* properties in addition to the base color coefficients. This confirms the expected field structure, but not optimization by itself. You can also orbit reflective or glossy areas in the browser and look for appearance that changes with the viewpoint.
What happens if an E57 has no usable imagery or camera poses?
Geometry and per-point RGB can still produce a viewer-ready geometry-based splat, but fixed appearance is not the same as full novel-view training. It cannot recreate reflections, hidden surfaces or photographic observations that were never captured.
What output formats are available?
Successful jobs expose the Gaussian Splat assets produced by the converter, commonly PLY and/or SPLAT depending on the completed result. The Gaussian PLY can preserve optimized appearance attributes for compatible tools.
What is the difference between PLY and SPLAT?
PLY is a flexible container that can store point or Gaussian attributes and is useful for compatible editing and interchange tools. SPLAT is a compact viewer-oriented representation; feature support varies by viewer.
Why does training take longer than direct format conversion?
Direct point remapping is mainly bounded by file parsing and disk I/O. Full 3DGS training runs iterative GPU optimization, adaptive densification and pruning. A typical full-training job takes about 12–40 minutes, depending on the source data and available compute.