These are Gaussian splats — 3D scenes reconstructed from ordinary handheld video. No scanner, no depth sensor. Frames get pulled from the video, the software works out where the camera was for each one, and then it fits a few hundred thousand fuzzy coloured blobs until the scene renders back to match the photos.
Working out the camera positions is the hard part, so the badges name which method was used. COLMAP is the classic approach — it matches visual features between frames, and it struggles on blank walls and low light. LingBot is a learned model that estimates positions directly and holds up much better indoors. Hybrid starts from LingBot's guess and then lets COLMAP refine it.
Where a score is shown, it's PSNR in decibels: one frame is held back from
training, then the finished scene is rendered from that exact camera angle and compared to the
real photo. Higher is closer. Roughly — 30 dB is very good, 18 dB is
recognisable but soft, 14 dB means the scene is fighting you.
A few cards are tagged raw LingBot cloud. Those are the network's untrained first guess — its depth estimate unprojected straight into points, before any Gaussian optimization. They're parked next to their trained versions so you can see exactly what the optimizer adds: it sharpens surfaces, fills gaps, and clears floaters, but the rough shape is already there in the raw cloud.