spirula-studio README

The project's README from github.com/harry7557558/spirula-studio. As one markdown file: README.md.

Spirula Studio

GPLv3 License  GitHub Releases  Platform

Download • Build from Source • Gallery • Web Viewer

Spirula Studio trains 3D Gaussian Splatting models – from raw photo/video to splat to textured mesh – in one self-contained binary. No Python/PyTorch, no separate COLMAP install. Runs on NVIDIA, AMD, Intel, and Apple GPUs via Vulkan, trains 10M full-SH Gaussians in 8 GB VRAM, and has native support for fisheye and 360° cameras.

Spirula Studio - Open Source 3D Gaussian Splatting Pipeline

Dataset credit: Garage by Simon Bethke (CC BY-SA 4.0); Flight Systems and Control Lab at UTIAS; MegaDepth-X; Mip-NeRF 360.

Features

News

Download

Binaries for Windows, Linux, and macOS can be downloaded from Releases page. Simply select the one for your platform, download and unzip, and double click to open the GUI.

If you are training on remote/cloud GPUs, you may use the CLI – Run spirula --help for details. By default, spirula train command will serve a viewer on an HTTP port, one you can forward over ssh and view training progress in your web browser.

Build from source

To build from source, Spirula Studio provides two backends:

Both provide the same training and meshing functionality. CUDA backend may be faster or slower than Vulkan depending on GPU driver, with difference generally within a few percents. Vulkan backend can be slightly more VRAM efficient in some cases.

Backend GPU/Vendor Support Platform Support Dependencies Additional Features
Vulkan NVIDIA, AMD, Intel, Apple Silicon Windows, Linux, macOS Vulkan/MoltenVK, CMake/Ninja Native support for SfM, frame extraction from videos, and AI masking
CUDA Most NVIDIA GPUs Windows, Linux CUDA, CMake/Ninja -
Details for building the Vulkan backend

Make sure you have Vulkan SDK installed. On macOS, MoltenVK is automatically fetched by CMake. Clone the repository and run the commands:

Windows with MSVC:

cd spirula-studio\
.\build_develop.bat -DSS_BACKEND=vulkan -DSS_ENABLE_PATENTED=ON

If it builds successfully, you get build_vulkan\spirula.exe.

Windows with GCC/Clang:

cd spirula-studio\
cmake -G Ninja -B build_vulkan -DCMAKE_BUILD_TYPE=Release -DSS_BACKEND=vulkan -DSS_ENABLE_PATENTED=ON -DCMAKE_MAKE_PROGRAM=Ninja
cmake --build build_vulkan -j

Pass -DCMAKE_C_COMPILER and -DCMAKE_CXX_COMPILER to the first cmake command if needed.

If it builds successfully, you get build_vulkan\spirula.exe.

Linux:

cd spirula-studio/
bash build_develop.bash -DSS_BACKEND=vulkan -DSS_ENABLE_PATENTED=ON

If it builds successfully, you get build_vulkan/spirula binary.

macOS:

cd spirula-studio/
bash build_develop.bash -DSS_BACKEND=vulkan -DSS_ENABLE_PATENTED=ON
cmake --build build --target macos_app
cmake --build build --target macos_dmg

macOS has only the one backend, so it builds into build/ rather than into a per-backend tree. If it builds successfully, you get build/spirula binary similar to Linux. Additionally, it wraps that binary in a double-clickable build/Spirula Studio.app, as well as disk image build/Spirula Studio.dmg. MoltenVK is statically linked by default and will run on a Mac without dependency installed.

Notes regarding third-party licensing

-DSS_ENABLE_PATENTED=ON enables decoding video on the GPU instead of shelling out to ffmpeg (about 15x faster frame extraction, and without need to install ffmpeg). However, AVC/HEVC bitstream parsers carry third-party patent exposure. If you turn this on, you are responsible for ensuring compliance with local patent laws regarding AVC/HEVC playback.

Masking needs a SAM checkpoint, which the GUI downloads on first use and caches. The checkpoints are Meta's models under Meta's licenses – SAM 2.1 is Apache-2.0, SAM 3 is under Meta's own, non-standard license. They are never bundled, and the GUI shows the terms before fetching anything. On the command line, point --model at a file you downloaded yourself.

Details for building the CUDA backend

Make sure you have a recent version of CUDA installed. On Windows, you also need MSVC compiler compatible with your CUDA version. Clone the repository and run the commands:

Windows:

cd spirula-studio\
.\build_develop.bat -DSS_BACKEND=cuda

If it builds successfully, you get build_cuda\spirula.exe.

Linux:

cd spirula-studio/
bash build_develop.bash -DSS_BACKEND=cuda

If it builds successfully, you get build_cuda/spirula binary.

You can find some professional-quality splats trained by Spirula Studio from Megascapes Library and their SuperSplat page.

Collection of splats created by the users of Spirula Studio can also be found on SuperSplat page.

Some splats created by the author of Spirula Studio can also be found on my SuperSplat page.

Trivia

Spirula Studio (formerly spirulae-splat) is named after the now-inactive project spirulae, which was named after the deep-ocean cephalopod mollusk.

Spirula Studio is developed and maintained almost entirely by one person. Issues and PRs welcome – I sometimes respond late, but rest assured that I do review them all.

spirula-studio (harry7557558/spirula-studio)

Searching Database…

1.8K ·GitHub
Spirula-studio
What is this?Spirula Studio is a 3D Gaussian Splatting trainer: it turns raw photos or video into splats and textured meshes, all from one self-contained binary.
Is it a Python tool I pip-install, or a standalone app?A standalone desktop app, not a Python library: one self-contained binary with a double-click GUI plus a spirula CLI. No Python, PyTorch, or separate COLMAP install needed.
How do I get it running?Grab a prebuilt binary for Windows, Linux, or macOS from the Releases page, unzip, and double-click to open the GUI; the CLI (spirula --help) suits remote training.
Any gotchas around AI masking or building from source?Masking needs a SAM checkpoint the GUI downloads on first use (SAM 2.1 is Apache-2.0, SAM 3 is Meta's non-standard license). Building with -DSS_ENABLE_PATENTED=ON gives ~15x faster video decoding but leaves AVC/HEVC patent compliance to you.
Can I train on a remote or cloud GPU and still watch progress?Yes: spirula train serves a viewer on an HTTP port — forward it over SSH and follow training in your web browser while the CLI runs on the remote machine.
Will it run on my AMD or Intel GPU, or is NVIDIA required?No NVIDIA needed: the Vulkan backend trains on NVIDIA, AMD, Intel, and Apple GPUs — quantized training fits up to 10 million SH3 Gaussians in 8 GB VRAM. A legacy CUDA backend covers NVIDIA too.
Why is it called Spirula Studio?It was formerly spirulae-splat, named after the now-inactive spirulae project, which took its name from the spirula — a deep-ocean cephalopod mollusk.
What makes its splat quality competitive?One strategy combines advantages of MCMC, IGS+, and MRNF for sharper results with fewer floaters on objects up to large scenes, and a modified bilateral grid plus PPISP correct exposure and white balance without unwanted color shift or darkening.
My reconstruction comes out the wrong size or tilted — is there a fix?Yes: since September 2026 the dataset creation module reads telemetry metadata from common video and image formats to recover metric scale and orientation, addressing that popular report.
Who develops Spirula Studio?Almost entirely one person — issues and PRs are all reviewed, though replies can be slow. Professional-quality samples and user splats are showcased on SuperSplat pages.
What are similar projects?instant-ngp, colmap, Meshroom, ooosplat, nerfstudio, gaussian-splatting, Open3D, engine
Which topics is it in?3d-gaussian-splatting, photogrammetry, gaussian-splatting, neural-rendering, gpu-programming, 3d-reconstruction, computer-graphics, vulkan, rendering, graphics, app, quantization
Creator
harry7557558
Size
46 MB
Stars
1.8K
Updated
CommitOct 18Releasev2026.10.16Oct 16
Age
2y
Commits · 90 Days