CUPID — Person-of-Interest Deepfake Detection

Weights for CUPID: Reconstructing UV Texture Maps for Interpretable Person-of-Interest Deepfake Detection.

📄 Paper: arXiv:2606.20302

cupid_mae.pth is a ViT-Tiny Masked Autoencoder trained self-supervised on UV face textures of real videos (VoxCeleb2). The CUPID pipeline scores a test video by max cosine similarity between its CLS-token features and those of reference videos of the person of interest.

Usage

pip install git+https://github.com/polimi-ispl/CUPID
cupid extract-reference --reference ref1.mp4 ref2.mp4 -o poi.pt
cupid score --reference-set poi.pt --test test.mp4

Weights are downloaded automatically on first use.

Third-party weights

CUPID's UV-texture extraction uses four asset files from 3DDFA_V3 (CVPR 2024), which are NOT mirrored here. They are downloaded directly from the authors' repository at Zidu-Wang/3DDFA-V3 and remain subject to their respective licenses and provenance (RetinaFace weights from biubug6/Pytorch_Retinaface, large_base_net.pth from HRN, net_recon.pth from 3DDFA_V3, face_model.npy derived from the Basel Face Model, Exp_Pca, and Deep3D).

License

The CUPID checkpoint is released under the MIT license.

Citation

@article{affatato2026cupid,
  title   = {{CUPID}: Reconstructing UV Texture Maps for Interpretable
             Person-of-Interest Deepfake Detection},
  author  = {Affatato, Giovanni and Mandelli, Sara and Bestagini, Paolo and Tubaro, Stefano},
  journal = {arXiv preprint arXiv:2606.20302},
  year    = {2026},
}