Gauss Cannon

Gauss Cannon is a Blender add-on that automates the laborious parts of preparing scenes for Gaussian Splatting training. Instead of manually placing and animating dozens of cameras, artists drop any mesh in as a template — the add-on instantiates a camera at each face center, orients it opposite the face normal, applies the chosen focal length and resolution, and uses raycasts to skip cameras that would end up inside geometry.

Outputs are complete training packages of rendered images, a binary COLMAP sparse model and a colored PLY point cloud. Version 1.3.0, released on October 14, 2026, made the COLMAP model the only export, writing cameras.bin, images.bin and points3D.bin under sparse/0 next to the images, and removed the earlier transforms.json export, so LichtFeld Studio, Postshot, Brush and other COLMAP based trainers load the output directly. Gauss Cannon is open source and maintained by Arash Keshmirian on GitHub.

Gauss Cannon is licensed GPL-3.0 and, from version 1.3.0, requires Blender 5.0.0 or later, because its depth and normal maps use Blender 5's compositor node groups and bundled OpenImageIO. Releases up to 1.2.0 supported Blender 4.2.0 or later. It installs from a release zip through the Install from Disk option in Blender's Get Extensions panel. Cameras export with PINHOLE intrinsics at the rendered image size, and cameras and points stay in Blender's Z up world frame, so the former Export Mode and Coordinate System settings are gone. Point clouds are ray traced from the selected meshes with BVH accelerated ray casting, and Generate Point Cloud writes points3D.bin alongside pointcloud.ply. Version 1.2.0 fixed point cloud performance on large Blender scenes. The repository now sits under the warpgatelabs organization on GitHub.

Version 1.3.0 also renders exact depth and normal maps for depth and normal supervised training. Depth is written as a single channel 16 bit PNG of planar z depth with one scale for the whole scene, normals as a three channel 8 bit PNG in OpenCV camera space, and a geometry_maps.json file records the scale and encodings, a layout that LichtFeld Studio and Spirula Studio read. The pass setup is saved in the .blend file, so a render farm can render the scene without the add-on installed before Convert Rendered Passes turns the output into maps. The same release renamed the Generate Cameras step to Generate Camera Path.