Arrival.Space has switched on video to gaussian splat conversion inside Splat.js, adding frame picking, retraining and an experimental avatar toggle.
Splat.js already carried the reconstruction code. Video input existed in the library as extractSharpFrames and was switched off in the app until frame selection reached the project's quality bar. The app bundle now imports extractSharpFrames and isVideoFile from its video module, the entry control reads "Choose photos or a video", and the file input accepts mov, mp4, m4v, webm and mkv. The page takes 20 to 200 overlapping photos of one place, a walk-around video, or a capture recorded in the tab.
Frame picking arrives as a review card. Splat.js scores every decoded frame for sharpness, motion and cuts, draws those scores as a timeline with the picks and blur dips marked, and shows a filmstrip of the selection. The time range is draggable at either end, so only part of a long take is reconstructed, and the card states the frame ceiling the device can hold.
The experimental path is a "This is a person" checkbox on the Start training card, hidden until the capture is the user's own. Arrival.Space describes an orbit at arm's length with the arms slightly out, cutting the person out of every frame before training, then isolating the splat, fitting it to a body model and placing it on the user's arrival.space account. A separate Avatar settings group carries person crops off, on with a room pose, or on with the head re-posed to the face, alongside a horizontal spherical harmonics switch.
Video ingest is not itself new to Splat.js. The benchmark set already included a handheld camp site walk at 113 frames and 1,000,000 splats. Arrival.Space 0.84 added Video2Splat, a Pro tier feature that converted video on the vendor's servers. The Splat.js path runs camera solving, splat training and export locally in the tab, and eight of the nine benchmarks on the page close "No COLMAP, no server."
The splat.js repository is MIT licensed, meaning it can be used commercially.
