RadianceKit 1.4 — COLMAP import, Continue Training, and a careful cleanup

RadianceKit turns photos and video into Gaussian splats on the Mac. The 1.4 update does two loud things and a lot of quiet ones — and the quiet ones are where most of the work went.

(If you are new to splatting, the primer covers the basics first.)

Meet the pipeline where it already is: COLMAP import

Until now RadianceKit ran the whole pipeline itself, from photos through Structure-from-Motion to training. 1.4 adds COLMAP workspace import: drop a folder with a sparse reconstruction and an images/ directory, and RadianceKit skips SfM entirely — it loads the cameras and the sparse point cloud directly. That means anyone coming from RealityCapture, Metashape, Meshroom or Postshot can train on reconstructions they already have, in seconds rather than re-solving from scratch. A dialog asks whether to use the existing data or re-run photogrammetry, so the old path is still one click away.

Don't stop at the finish line: Continue Training

The other headline feature is Continue Training — extend a finished run by another 5,000, 10,000 or 20,000 iterations using the existing engine state, with the Adam moments preserved and the learning rate locked at its final value. Before this, "not quite sharp enough" meant starting over. Now it means a few more minutes on top of what you already have. There is also a one-click sample scene and a sample-photo download, so a first-time user can see a finished splat — or run the full pipeline — without owning a capture.

The unglamorous half

The rest of 1.4 is the careful, boring work that makes a tool trustworthy. A Manage Storage window finally shows where all the disk space went — training logs, exports, scenes, captures — with delete and reveal-in-Finder per item. An 8 GB memory guard detects low-RAM Macs and quietly drops to safer presets and resolutions instead of thermally shutting down or running out of memory mid-train.

And I spent real time on over-training. Benchmarks showed the last stretch of a 40,000-iteration run was adding Gaussians while barely improving the loss — growth without gain, correlated with colour-saturation outliers. So the Quality preset is now 35,000 iterations, and the mid-training compaction steps scale to the run length instead of being hard-coded. Fewer splats, the same result, a smaller file. There is also an experimental sky-dome mode that freezes injected background Gaussians so they stop drifting during training.

Where to look

The full changelog, SDK and capture notes are on radiancekit.de; the app is on the Mac App Store. And what you train here, you can open and tidy in RadianceView.