OpenSeeFace implements a neural network-based facial landmark detection system that tracks faces in real time at 30-60 fps on CPU. The model was originally trained with PyTorch and converted to ONNX format for efficient inference using onnxruntime, with four variants offering different speed-to-quality trade-offs. It outputs quasi-3D facial landmarks optimized for animating virtual avatars rather than purely numerical accuracy benchmarks. Integration options include Unity, Godot, and compatibility with virtual YouTuber workflows like VSeeFace and VTube Studio.