kocasariumut/faceanything

Face Anything is a research project from academic authors, presented at the ECCV 2026 computer vision conference, that reconstructs a moving 3D face from a plain video, a folder of images, or even a single photo. Instead of needing special cameras or markers on someone's face, it takes ordinary footage and produces a detailed, moving 3D model of that face over time. The core idea is something the authors call canonical facial point prediction. Every pixel showing a face gets mapped to a shared, standardized coordinate system, so the model can track the same point on a face consistently across every frame of a video, even as the face moves, turns, or changes expression. This turns two normally separate problems, tracking points over time and building 3D shape, into one single approach. Running it produces a wide range of outputs: colored point-cloud videos, videos colored by depth or surface direction, files showing camera positions for each frame, and raw prediction data. A single showcase video combines several of these views into one orbiting camera "grand tour," and everything can also be exported as individual video files or 3D point cloud files. This tool requires real technical setup. It needs a computer with an NVIDIA graphics card that supports CUDA, plus specific versions of Python and PyTorch, and a roughly 15 gigabyte model checkpoint file that must be downloaded separately from Google Drive or Hugging Face. There's a one-line install script that handles most of this automatically. Background removal from footage happens automatically using a separate open-source tool, though users can supply their own masks or turn this off. Users can choose between two processing modes: one that looks at the whole sequence together for smoother, more consistent 3D results, and one that processes each frame independently for sharper detail but less consistency between frames. This is a research code release meant for people already comfortable with machine learning tools and command-line setups, not a plug-and-play consumer app.