Spirula Studio v2026.10.6 accepts laser scans in E57, LAS or PLY format as dataset inputs, using them to give reconstructions metric scale and depth and normal maps.
Until this release, the GPL-3.0 application built gaussian splatting datasets from photos and video alone, with structure from motion producing poses at an arbitrary scale and a monocular step estimating depth and normals. Partitioned training for large scenes arrived in v2026.9.30 six days earlier.
After reconstruction, a `spirula lidar` step fits rotation, translation and scale to move the model into the scan's frame, then refines it with point-to-plane ICP against the scan thinned to 2 million points. The scan, thinned to 500,000 points by default, becomes the seed points, keeping the model's own points only where the scan has none, such as sky, and depth and normal maps are rendered from it. Spirula Studio's release notes describe LiDAR support as "an experimental feature" tested on a limited set of public datasets.
On an FJD Trion P2 E57 with 51 panoramas plus Insta360 X5 footage, the design notes put the aligned model 1.3 cm from the scan surface after ICP. Held out PSNR after 7,000 steps was 17.27 dB with the scan's seeds, depth and normals and 17.37 dB on the unaligned model with its own points, inside that capture's roughly 1 dB run to run noise.
A separate `spirula e57` command, contributed by Valentin Heun, turns an E57 that carries the scanner's own registered photographs straight into a transforms.json dataset with no structure from motion. Pinhole images become PINHOLE cameras and full panoramas EQUIRECTANGULAR ones.
The trainer already accepted a region of interest through `--roi-region`, but only from a partition part or a hand written JSON file. A new editor builds one from boxes, ellipsoids, cylinders and extruded outlines, each added, cut or overlapped, and training picks up the saved region automatically.
Structure from motion now uses GPS during mapping and reads DJI Avata 360 telemetry and factory lens calibration. Input adds HEIC, TIFF and DWAA or DWAB compressed EXR, and feature points and training can take separate masks. Fixes cover occasional training divergence and meshing crashes.
Spirula Studio v2026.10.6 is available now on GitHub.
