Spirula Studio is a self-contained desktop application that turns photos and video into 3D Gaussian Splatting scenes and textured meshes. It is meant for people who want to capture real objects and places in 3D — visual-effects and game artists, scanning hobbyists, researchers working with fisheye or 360-degree footage — without first assembling a Python machine-learning environment. Developers who maintain reconstruction pipelines will also care, because the whole chain ships as a single GPL-3.0 C++ program instead of a stack of loosely joined scripts.
What it does
The project covers the full path from raw footage to a finished asset. Its own summary is "photo/video to splat to textured mesh, in one self-contained binary," and the pieces it bundles to get there are the parts that usually live in separate tools:
- frame extraction from videos, so a phone clip can be an input directly
- a built-in structure-from-motion stage, described by the README as lightning-fast, replacing a separate COLMAP install
- AI masking, to remove unwanted subjects from the images that feed training
- Gaussian Splatting training itself, with exposure and white-balance correction
- export to a textured mesh, plus a web viewer for looking at results in a browser
Nothing here requires Python or PyTorch. That matters less for the quality of the output than for the hours normally lost before training even starts, which is the pitch the channel's video leads with.
How it works
The distinguishing engineering choice is the compute backend. Training runs on Vulkan compute as well as CUDA, so NVIDIA, AMD, Intel and Apple GPUs are all supported rather than just NVIDIA cards. Most splatting trainers are written against CUDA kernels and simply do not run anywhere else; Spirula Studio treats cross-vendor support as a first-class feature and advertises Windows, Linux and macOS builds.
The second choice is memory. Training uses quantization, which the project says lets it fit up to ten million SH3 (full spherical-harmonic) Gaussians into 8 GB of VRAM. That is the difference between needing a workstation card and running a large scene on a normal consumer GPU.
On the algorithm side, the README describes one densification and training strategy that combines ideas from MCMC, IGS+ and MRNF, aimed at sharper results and fewer floaters across both single objects and large scenes. For photometric consistency it uses a modified bilateral grid together with PPISP to correct exposure and white balance, with the stated goal of improving quality without introducing color shifts or darkening.
Camera handling is native rather than bolted on. Fisheye and 360-degree equirectangular datasets load and train as they are, with no undistortion pass beforehand — a real convenience if your capture rig is a consumer 360 camera.
Getting started
Prebuilt releases are published on the repository's releases page for Windows, Linux and macOS, so the fastest route is to download a binary and point it at a video or an image folder. The README also documents building from source for anyone who wants to compile against their own toolchain. A hosted web viewer and a gallery are linked from the top of the README, which is the quickest way to judge output quality before installing anything. The license is GPL-3.0, which is worth checking against your plans if you intend to embed the code in a product rather than just use the app.
When to use it / when not
Reach for it when your hardware is not an NVIDIA card, when you are shooting with fisheye or 360 rigs, when VRAM is your limiting factor, or when you want a mesh at the end instead of a splat file you then have to convert. It is also a sensible default when the people doing the capture are artists rather than engineers, since there is no environment to maintain.
It is a weaker fit if your work is modifying the training mathematics itself. A research group that wants to swap in a new loss function or densification rule will be more productive in a Python codebase where that is a few lines, rather than in a C++ and Vulkan application.
Alternatives
The conventional route is COLMAP for structure-from-motion followed by a Python and PyTorch Gaussian Splatting trainer. That stack is more hackable and closer to published research, at the cost of setup time, CUDA-only hardware assumptions and manual handling of unusual camera models. Spirula Studio folds the same stages into one program and trades scriptability for a pipeline that runs out of the box.
At 720 stars, GPL-3.0, active since 2024 and still being pushed to, this is worth serious attention from anyone doing 3D capture on non-NVIDIA hardware, anyone fighting VRAM limits, and anyone whose datasets come from 360 cameras. If you already have a working Python splatting setup and a large NVIDIA GPU, the case is weaker — but as a single download that takes a video in and hands a textured mesh back, it removes more friction from the workflow than any individual feature on its list.