Arcturus Vision Converts Steam Frame Captures to 3DGS
Michael Rubloff
Arcturus Vision has published an official script that converts a recording from its $149 Steam Frame camera into a gaussian splat off the headset. The camera records stereo spatial video and performs no reconstruction of its own.
The conversion is distributed as the AV1 to 3DGS scripts and run as avcam_to_3dgs.py, taking a synced camera MP4 and writing a scene.sog. It extracts and tone maps the synchronized stereo images, uses the recording's own calibration and trajectory to build an aligned COLMAP scene, trains at half resolution, and limits the result to 1.5 million Gaussians. Adding --training-image-scale 1 raises detail at the cost of training time and GPU memory, --headless suppresses the training interface, and --resume reuses a folder that already holds a valid COLMAP reconstruction.
Running it requires a Linux workstation with an NVIDIA GPU at compute capability 7.5 or newer, driver 570 or newer, uv, and both COLMAP and LichtFeld Studio on PATH. Arcturus Vision documents that limitation plainly, stating that the workflow is generally compatible with Windows and that LichtFeld Studio supports Windows, but that the scripts have not been adapted for it. Somebody who buys the camera and owns a Windows gaming PC has no supported route to a splat. Hardware video decoding and GPU COLMAP are selected automatically where supported, with CPU fallback warnings otherwise.
The company showed a CUDA splatting engine of its own in 2023 that trained to 7,000 steps in 68 seconds inside 5 GB of VRAM.
Gaussian XR, the free Steam Frame viewer, reads PLY and SOG and carries eight example scenes, so a scene.sog produced by the script loads the same way a model trained elsewhere does. The camera itself sells for $149 and attaches to the headset's front expansion port. Arcturus Vision Sync pulls recordings off the headset on Windows or Mac and converts them for Apple Vision Pro and Meta Quest.
New to Gaussian Splatting? Start here
