---
license: mit
library_name: pytorch
pipeline_tag: video-classification
tags:
  - deepfake-detection
  - face-forensics
  - person-of-interest
  - facial-dynamics
  - temporal-transformer
---

# MOTIF — Person-of-Interest Deepfake Detection

Inference assets for **MOTIF: Person-of-Interest Deepfake Detection Beyond 3DMM Coefficients**, accepted at IEEE WIFS 2026.

**Paper:** [arXiv:2610.09830](https://arxiv.org/abs/2610.09830)  
**Authors:** Giovanni Affatato, Sara Mandelli, Paolo Bestagini, Stefano Tubaro.  
**Code:** [polimi-ispl/MOTIF](https://github.com/polimi-ispl/MOTIF)

MOTIF enrolls a person from genuine reference videos and scores a query video without retraining. Two independently trained temporal transformers process global 3D morphable-model coefficients and temporally centered reconstructed mouth motion in 75-frame windows with 50% overlap. Per-person whitening and leave-one-reference-video-out calibration produce branch scores; their arithmetic mean is the final authenticity score. Higher scores indicate greater consistency with the enrolled person.

The models were trained on real videos only, without manipulated videos or person-of-interest-specific training data. This repository contains inference checkpoints and normalization statistics, not training code or data.

## Usage

Requires an NVIDIA GPU and Python 3.10 or newer.

```bash
python -m pip install git+https://github.com/polimi-ispl/MOTIF
motif extract-reference --reference ref1.mp4 ref2.mp4 ref3.mp4 -o person.npz
motif score --reference-set person.npz --test query.mp4
```

The first run downloads these files automatically. Ten independent genuine reference videos match the paper protocol; at least two usable references and enough valid windows are required for leave-one-out calibration.

## Files

| File | Purpose | SHA-256 |
| --- | --- | --- |
| `motif_global.pth` | Global temporal transformer | `fd93ba656c7d2ad71481659b9cab723343a285d55174eeb277c15dc27312b99e` |
| `motif_mouth.pth` | Mouth temporal transformer | `e900ab14013072a0da2164d47b44e1635e56ec4a69682a7951a138b84203fa5a` |
| `norm_stats_all.npz` | Global normalization statistics | `7e3796edd3c452a98eab2b0e069903c70eb15b2488423053a7d1aed1f0e211d9` |
| `norm_stats_mvrecon.npz` | Mouth reconstruction statistics and projection bases | `7b506ff093ff48bb467e8c1c567f216a7e9e4c91da41be1e0e3802d9ee148c7c` |

## Third-party weights

Face-analysis assets (`retinaface_resnet50_2020-07-20_old_torch.pth`, `large_base_net.pth`, and `net_recon.pth`) are downloaded directly from the upstream [3DDFA-V3 Hub repository](https://huggingface.co/datasets/Zidu-Wang/3DDFA-V3). They are not mirrored here and retain their respective licenses and provenance (RetinaFace, HRN, and 3DDFA-V3).

## License

The MOTIF checkpoints and normalization assets are released under the MIT license. Third-party assets retain their own terms. This model license is separate from the paper's license.

## Citation

```bibtex
@inproceedings{affatato2026motif,
  title = {{MOTIF}: Person-of-Interest Deepfake Detection Beyond {3DMM} Coefficients},
  author = {Affatato, Giovanni and Mandelli, Sara and Bestagini, Paolo and Tubaro, Stefano},
  booktitle = {2026 IEEE International Workshop on Information Forensics and Security (WIFS)},
  year = {2026},
  eprint = {2610.09830},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2610.09830}
}
```