Arash Keshmirian has released Gauss Cannon v1.3.0, which turns the Blender add-on into a COLMAP dataset generator with depth and normal maps.
Gauss Cannon v1.2.0 rebuilt the point cloud raytracer in May, and until now the add-on exported a transforms.json file with separate modes for LichtFeld Studio, Postshot and Brush. Version 1.3.0 removes that export and writes only a COLMAP sparse model, which LichtFeld Studio, Postshot, Brush and other COLMAP based gaussian splatting trainers read directly.
Export COLMAP Model writes cameras.bin, images.bin and points3D.bin to sparse/0 next to the images folder. The cameras.bin file stores PINHOLE intrinsics at the rendered image size, accounting for resolution percentage, lens shift, non-square pixels and sensor fit, and an animated lens gets one camera per distinct set of intrinsics. The images.bin file holds one world to camera pose per rendered frame, and points3D.bin stays empty until Generate Point Cloud runs.
Gauss Cannon v1.3.0 also renders depth and normal maps for supervised training. Depth maps are 16 bit single channel PNGs of planar z depth times one scale for the whole scene, normal maps are 8 bit PNGs of camera space normals in OpenCV axes, and geometry_maps.json records the scale and encodings. LichtFeld Studio and Spirula Studio both read the format. LichtFeld Studio 0.5.4 added normal guided training last week, and the losses are enabled there with --use-depth-loss, --use-normal-loss and --normal-loss-space camera-opencv. The panel warns when Ray Density is below 32, the level Gauss Cannon recommends for LichtFeld depth supervision.
Gauss Cannon v1.3.0 requires Blender 5.0, up from 4.2, because the maps rely on Blender 5 compositor node groups and its bundled OpenImageIO. The pointcloud.ply output is now always Z-up, so workflows that relied on the old Y-up PLY will see a different orientation. Camera poses now match the render when the camera carries delta transforms, F-curve modifiers, drivers or NLA strips.
The pass setup is saved in the .blend, so blender -b scene.blend -a renders the passes without the add-on installed, and Convert Rendered Passes in the add-on's panel then processes the frames in parallel. Render output paths are now stored relative to the .blend. A studio can render a synthetic training set across farm nodes that never have Gauss Cannon installed, then convert the passes on a machine that does.
Gauss Cannon v1.3.0 is available now on GitHub under the GPL-3.0 license.
