What it is
Splats is our pipeline for capturing real places in 3D. You film an object, a room or an outdoor spot. The pipeline works out where the camera was for each frame, then builds a "Gaussian splat": millions of tiny soft colored blobs that together look like the scene from any angle. The result opens in a web browser where you can orbit and walk around it. The gallery is live; the capture side is still a workbench.
Try it
Open the gallery and pick a scene. Drag to orbit, scroll to zoom, and move around until it clicks that this is a real place. Several entries compare methods on the same scene: one built the classic way, one from a raw point cloud, one fully trained. Flip between them to see what each step adds. All scenes are real captures. A desktop with a decent graphics card is smoothest; phones work but can stutter on bigger scenes.
How it works
- Shoot a slow, steady video walking around the subject.
- The pipeline pulls the sharpest frames.
- It estimates each camera position, using either the classic COLMAP tool or a newer neural method (LingBot-Map), or both.
- A rented cloud GPU trains the splat from those frames and positions.
- The finished file is added to the web gallery.
What exists today
- Scripts for frame extraction, camera-pose conversion, refinement and export
- Step-by-step GPU notebooks for several scenes, with quality scores
- A live web viewer and a 13-entry gallery, including side-by-side method comparisons
- Lab notes on what failed and why (for example, a plain hallway where the classic method broke down and the neural one worked)
What isn't there yet
- You can't upload your own video; only our scenes are viewable
- Not a one-command tool; each scene needs a GPU session and manual steps
- Local training without a GPU is roughly 100 times slower
- The viewer loads an unpinned library version, which could break without warning
- Many experimental notebook copies to tidy
Under the hood
- Python and shell scripts with ffmpeg, COLMAP and OpenSplat
- LingBot-Map (open source, Apache 2.0) for neural camera poses
- Training runs in Google Colab notebooks on a rented GPU
- The viewer is static HTML using gsplat.js, driven by a small scene list file
