Reconstruct and track a face across an image sequence

FaceAnything is built for 4D face reconstruction: it takes a face image sequence or short video and predicts how the face geometry changes over time. Instead of only returning a stylized render, the workflow gives you several inspection views, including depth, normals, canonical face-coordinate maps, and point clouds with colored tracks.

That makes it useful when you care about motion and correspondence, not just the final look. For example, you can upload a short head-turn clip, compare the depth video against the original frames, then scrub through the point-cloud output to see whether the same facial regions stay trackable across the sequence.

Start with the Hugging Face Space and upload a short, clean face video or an ordered set of face frames. A good first test is 10 to 20 frames with one visible face, steady lighting, and some head motion. Watch whether the depth, normals, and point tracks stay stable as the face moves.

For local use, the public GitHub repo includes an install script and `run_inference.py`, while the model checkpoint is available from the Hugging Face model repo. This is not a light laptop workflow: local reproduction expects CUDA and a roughly 15 GB checkpoint, so use the browser demo as the first filter.

This is a GPU-heavy research workflow, not a casual webcam filter. Results are easiest to judge on clean, face-forward clips, and local setup requires a CUDA environment plus the large checkpoint.

• Inspect face motion as depth maps, surface normals, and point-cloud tracks.
• Test whether a short face clip has enough consistency for reconstruction work.
• Generate visual debugging outputs for facial tracking or capture pipelines.
• Compare joint sequence processing against frame-by-frame reconstruction.
• Build a local research workflow when you need repeatable face geometry outputs.

Try the demo

View model page