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.
Dataset credit: Garage by Simon Bethke (CC BY-SA 4.0); Flight Systems and Control Lab at UTIAS; MegaDepth-X; Mip-NeRF 360.
Features
-
Cross vendor support via Vulkan compute – Runs on NVIDIA, AMD, Intel, and Apple GPUs
-
One strategy combining advantages of MCMC/IGS+/MRNF – Sharper results, fewer floaters, from objects to large scenes
-
Extreme VRAM efficiency with quantized training – Up to 10 million SH3 Gaussians in 8GB VRAM
-
Native 360° camera and equirectangular support – Load a dataset and train, no undistortion needed
-
Modified Bilateral grid and PPISP for exposure/WB correction – Improving quality without unwanted color shift or darkening
-
Built-in lightning-fast SfM, AI masking, frame extraction from videos – No need to wait for COLMAP or run separate scripts
-
Depth/normal, meshing, skybox, linear color... And more.
News
-
September 10, 2026: Metric scale – The dataset creation module now uses telemetry metadata in common video and image formats to recover metric scale and orientation, addressing the popular report that reconstruction results are too large/small or tilted.
-
September 3, 2026: LoMa feature support – The SfM module now supports LoMa for feature detection and matching on difficult datasets.
-
August 14, 2026: macOS support – Support for training on macOS/Apple Silicon has been validated. The app can now be downloaded from Releases page.
-
August 8, 2026: Multilingual support – Multilingual support has been added, available to both GUI and CLI. Supported languages: English, 日本語, 简体中文, 繁體中文, 한국어, Deutsch, Français, Español, Português, Italiano, Nederlands, Русский, Türkçe.
-
August 8, 2026: End-to-end workflow – The Vulkan backend now has components to extract frames from video, AI masking, native SfM, meshing, and batch processing, accessible from both GUI and CLI.
-
July 22, 2026: Cross-vendor support – A Vulkan backend has been added, which works on NVIDIA, AMD, and Intel GPUs.
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:
-
Vulkan (Recommended): The cross-platform and cross-vendor option. Most tested. Works on all major GPUs. Faster to build and produces smaller binary.
-
CUDA: Legacy option for CUDA-capable NVIDIA GPUs.
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=ONIf 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 -jPass -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=ONIf 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_dmgmacOS 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=cudaIf it builds successfully, you get build_cuda\spirula.exe.
Linux:
cd spirula-studio/
bash build_develop.bash -DSS_BACKEND=cudaIf it builds successfully, you get build_cuda/spirula binary.
Gallery
You can find some professional-quality splats trained by Spirula Studio from Megascapes Library and their 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.