Spirula Studio

Spirula Studio is a 3D Gaussian Splatting trainer written in C++ and, in its README's own wording, developed and maintained almost entirely by one person. The repository was created on May 4, 2024 under the name spirulae-splat and was renamed in 2026. It is licensed GPL-3.0 and free to download, and it releases roughly every one to two weeks. Recent releases have concentrated on the structure from motion front end and on training quality. v2026.9.24, released on September 24, 2026, adds editing for 3DGS models, sparse reconstructions and meshes and exports image and video renders, and the release notes label both the editing and the rendering experimental. The same release adds an MCP server for agentic workflows, manual refinement of masks, saving and loading of hand edited masks for fixed areas of frames, .insp and .lrv input, and extraction of every frame of a video when fps is set to 0. v2026.9.20 added warm start training from an existing 3DGS PLY file, adaptive frame extraction from video, presets and batch processing for dataset creation and meshing, a choice of GPU for dataset creation and meshing, and training on RGBA images. v2026.9.13, released on September 13, 2026, recovers metric scale and orientation from telemetry metadata carried in images and video, resumes a partially failed or cancelled dataset creation run, adds a rig constraint for structure from motion that improves robustness and accuracy on dual fisheye systems, and creates datasets from GoPro .360 video. The same release fixes rotated and upside down reconstructions caused by EXIF orientation, clears occasional fogs in perspective EWA projection renders in the native GUI, and adds dataset centering options that address floating point precision for datasets in the ECEF frame. v2026.9.8 added LoMa feature detection and matching in structure from motion, metric scale recovery from EXIF GPS data, an improved dynamic range for the HDR preset, a centered-object preset, and a pseudorandom background mode that discourages background transparency, alongside a fix for ALIKED feature extraction failing on high resolution images. Earlier in the run, v2026.9.2 fixed crashes when training high resolution images with masks and errors loading datasets with Unicode file paths on Windows, and v2026.8.20 brought a general improvement in splat quality and a visualisation of the pixel error map used to allocate splats during training.

The application takes raw photos or video through to a Gaussian Splat and then to a textured mesh in one self contained binary. Structure from motion, frame extraction from video, AI masking, batch training and meshing all run inside the app and are reachable from either the desktop GUI or the command line. There is no Python, no PyTorch and no separate COLMAP install. The AI masking uses Meta SAM checkpoints, which are downloaded on first use rather than bundled with the binary. Splatara rebuilt the desktop pipeline of its SplataraScan app on Spirula in September 2026.

Two backends are available. Vulkan is the recommended one and covers NVIDIA, AMD, Intel and Apple Silicon on Windows, Linux and macOS, with MoltenVK statically linked on the macOS build. A legacy CUDA backend covers most NVIDIA GPUs on Windows and Linux only. Both backends train and mesh identically, the project puts the speed difference between them within a few percent depending on driver, and Vulkan is slightly more VRAM efficient in some cases. Native SfM, video frame extraction and AI masking are Vulkan only. macOS training was validated on August 14, 2026.

Quantized training fits up to 10 million SH3 full spherical harmonic Gaussians into 8 GB of VRAM, a figure the project demonstrates in a screenshot of the GUI training 4K images on a laptop GPU. The training strategy combines MCMC, IGS+ and MRNF. A modified bilateral grid and PPISP handle exposure and white balance correction without the colour shift or darkening those corrections usually introduce. Datasets from 360 and fisheye cameras, including equirectangular imagery, load and train with no undistortion step.

Binaries for Windows, Linux and macOS come from the GitHub releases page and unzip to a double clickable GUI. On remote and cloud GPUs the CLI serves a viewer on an HTTP port during training, which forwards over ssh so progress can be watched in a browser. A WebGL2 and WebAssembly viewer is published alongside the project. The interface ships in 13 languages, covering Japanese, Simplified and Traditional Chinese, Korean, German, French, Spanish, Portuguese, Italian, Dutch, Russian and Turkish alongside English.