Splat Pipeline

Every splat on this site came out of this, including the one above, trained end to end through this exact pipeline. It's the boring infrastructure that makes the fun stuff possible: point it at a video file or a folder of photos, and it hands back a trained .ply you can drop straight into a browser.

make all VIDEO=~/Movies/clip.mp4
make photos-quality PHOTOS_DIR=~/Pictures/scene
make hdr-photos PHOTOS_DIR=~/Pictures/brackets

What's actually happening

COLMAP solves camera poses — where the camera was for every frame — which is the part that quietly determines whether the final result is crisp or mush. Then the Gaussians get trained against those poses, natively on Apple Silicon via MLX rather than falling back to CPU or renting a GPU.

The parts that took the most iteration weren't the training, they were everything around it:

• Frame selection. Dumping every frame of a video into COLMAP is slow and often worse — near duplicates add time without adding parallax. Sampling matters.
• Resolution policy. Fine detail like grass and foliage falls apart if you downscale before feature extraction, so the quality path imports at full resolution and runs COLMAP features at 3200px.
• HDR brackets, for scenes where a single exposure clips the windows or crushes the shadows.

The LiDAR variant

A second pipeline adds iPhone LiDAR depth as a regularizer during training. Pure photometric splatting produces beautiful detail and terrible flat walls — they come out lumpy, because nothing in the loss function knows a wall is planar. Feeding in depth, plus a planar pruning pass, fixes the walls without costing the detail elsewhere. On a 16 GB M2 that combination is what makes the difference between a scan that looks like a room and one that looks like a cloud.