TraceAnything

Why we included this project

TraceAnything is a research codebase that turns video into a 4D trajectory field. Give it a video, an image pair, or an unstructured set of frames, and a single forward pass estimates a continuous parametric 3D path for every pixel over time. Point tracking, dynamic 3D reconstruction, and motion analysis all build on that output. The repo ships pretrained weights plus an interactive 3D viewer, so you can inspect the trajectories yourself instead of trusting the paper's figures. It runs on plain PyTorch with a single checkpoint, though you will want a GPU with at least 48 GB of VRAM, and the weights are licensed for research and non-commercial use only, which is worth checking before you plan anything beyond a prototype.