Training eye trackers without collecting eye data

Eye tracking requires expensive, device-specific data collection for each new AR/VR headset. GazePrior learns a 3D prior of how human eyes look across different people, gazes, and lighting—then uses it to reconstruct and re-render existing eye-tracking datasets from any target device's cameras. The result: synthetic data with photorealistic quality and accurate ground truth, all without new collection. Models trained this way outperform previous zero-shot approaches on accuracy and robustness.