[ICCV 2025] FaceLift: Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads

💡 Tips for Best Results

• Works best with near-frontal portrait images.
• The provided checkpoints were not trained with accessories (glasses, hats, etc.). Portraits containing accessories may produce suboptimal results.
• If face detection fails, try disabling auto-cropping and manually crop to square.
• Inference complete when the turntable video is generated, the interactive 3D Gaussian might take several seconds to load.
• External Views mode: Upload 6 views in this order: Front → 45° left → Profile left → Back → Profile right → 45° right
• Strip mode: Upload a single 3072×512 image containing all 6 views side by side in the same order.