donalleniii/lingbot-desktop-mac

Lingbot-Desktop is a Mac app that turns a video into an explorable 3D scene. You drag a video into the window, wait a few minutes while it processes, and end up with a 3D point cloud you can orbit around, along with the camera path that was solved for the video. Everything runs locally on Apple Silicon chips, with no cloud service and no remote GPU involved. The actual 3D reconstruction is done by a separate research project called Lingbot-Map, built by the Robbyant team, based on a published paper about streaming 3D reconstruction. This project is a Mac wrapper around that work: it provides the drag and drop interface, a FastAPI backend, an installer, and a custom viewer with features like automatic upright alignment, a playback strip, and buttons to save the result as a point cloud or a splat format file. It is important to note the output is a point cloud, not Gaussian splatting, even though a splat format file can be exported. Installation is a single command that downloads and runs an installer script, which asks before installing anything, sets up a Python virtual environment, installs Apple Silicon PyTorch, clones the Lingbot-Map project at a pinned version, and downloads a roughly 4.6 gigabyte model checkpoint from Hugging Face. A more manual, step by step installation path is also documented for anyone who prefers to review each step first. Once installed, a double clickable launch file starts a local server and opens the viewer in a browser window. Performance depends heavily on your Mac's hardware. On the developer's own machine, a high end Apple Silicon MacBook Pro, processing runs at roughly 0.3 to 1.0 frames per second, meaning a short clip takes a few minutes rather than being instant, and the README is upfront that this is not real time and that weaker machines will be slower. It also gives a practical filming tip: moving the camera through a scene works far better than pointing it and rotating in place. The wrapper code itself is released under the Apache 2.0 license, but the downloaded model checkpoint carries its own separate license terms that may be more restrictive, so the README asks users to check those before use.