From video to mesh in one binary, on any GPU

What it does

Spirula Studio ingests raw photos or video and trains 3D Gaussian Splatting models, then converts them into textured meshes. The whole workflow lives in one self-contained binary: it extracts video frames, runs structure-from-motion, handles masking, and manages training without calling out to Python, PyTorch, or a separate COLMAP install. It targets Windows, Linux, and macOS through a Vulkan compute backend, with a legacy CUDA path for NVIDIA-only setups.

The interesting bit

The project is essentially a ground-up rewrite of the 3DGS toolchain in C++ to escape the NVIDIA/CUDA/PyTorch gravity well. It fuses ideas from MCMC, IGS+, and MRNF into one training strategy and uses quantized training to fit up to 10 million SH3 Gaussians into 8 GB of VRAM. That efficiency, plus native support for fisheye and 360° equirectangular cameras, makes it unusually hardware-agnostic for a field that usually demands an RTX card and a conda environment.

Key highlights

• Cross-vendor GPU support via Vulkan compute (NVIDIA, AMD, Intel, Apple Silicon).
• End-to-end workflow in one binary: frame extraction, SfM, AI masking, training, and meshing.
• Quantized training claims up to 10 million SH3 Gaussians in 8 GB VRAM.
• Native fisheye and 360° equirectangular camera support without undistortion.
• Modified bilateral grid and PPISP for exposure and white-balance correction.

Caveats

• The CUDA backend is explicitly labeled “legacy”; Vulkan is the recommended and more fully featured route.
• GPU-accelerated video decoding carries third-party patent exposure for AVC/HEVC bitstream parsing if enabled at build time, and AI masking requires separately downloaded SAM checkpoints.
• The project is maintained almost entirely by one person, who notes they sometimes respond late to issues.

Verdict

Worth a look if you want to train 3D Gaussian splats on a laptop GPU or avoid wrestling with Python environment rot. Probably not for researchers who need to hack on PyTorch training loops directly.