Spirula Studio v2026.9.24 adds editing for 3DGS models, sparse reconstructions and meshes.
Until this release the GPL-3.0 application was a trainer, taking photos or video through frame extraction, AI masking and structure from motion to a trained Gaussian splat and then a mesh, on Vulkan or CUDA backends. Spirula Studio v2026.9.24 also exports image and video renders. Over the last week it has rocketed in popularity. I have been using it often and it does an amazing job, especially on 360 video reconstructions.
That single binary pipeline is what Splatara moved SplataraScan's desktop processing onto earlier this week, replacing COLMAP and RealityScan and shipping its Windows companion as a 240.7 MB archive.
The release also adds an MCP server, described in the notes as being "for agentic workflows." The notes do not say which operations the server exposes or how it is launched.
Masking gains manual refinement, and hand edited masks covering fixed areas of frames can now be saved and reloaded. Input support extends to the insp and lrv formats, and entering 0 for fps extracts every single frame of a video. Training gains explicit control over the color space of the input point cloud, and model checkpoints can download from a ModelScope mirror.
Structure from motion is tuned for sequential datasets with repetitive features, detects fisheye borders more reliably, and aligns scenes better after sparse reconstruction, especially when no telemetry data is available. The notes describe the SfM changes as intentional and ask users to report any significant regression in accuracy or speed against earlier versions.
The update lands four days after v2026.9.20, which added warm start training from an existing 3DGS PLY file and adaptive frame extraction from video.
Spirula Studio v2026.9.24 is available now on GitHub.
