--- license: cc-by-nc-4.0 license_name: cc-by-nc-4.0-plus-upstream license_link: LICENSE tags: - deepfake-detection - media-forensics - deepfakebench - verity library_name: pytorch --- # Verity detector checkpoints Mirror of the pretrained checkpoints used by **Verity**, an open deepfake examination pipeline for investigators and researchers. Verity downloads the files below on first run (`verity weights pull`) and verifies each one against the SHA-256 in `manifest.json` before it is used. **This repo re-hosts other people's work.** Every file keeps the license its authors released it under; the repo-level tag is the most restrictive one (CC BY-NC 4.0). Nothing here may be used commercially. See `NOTICE.md` for per-file attribution and `LICENSE` for the license texts. Checkpoints that upstream authors did not license for redistribution (NPR, MLEP) are **not** in this repo; Verity fetches those from the authors' repositories directly. Assembled 2026-10-21 from Verity's `detectors/weights_manifest.yaml`. ## Files | Path | Size | License | What it is | SHA-256 | |---|---|---|---|---| | `aasist/AASIST.pth` | 1.3 MB | MIT | AASIST audio anti-spoofing checkpoint (ASVspoof2019 LA) | `51d2d9cf0738172f…` | | `backbones/hrnetv2_w48_imagenet_pretrained.pth` | 310.6 MB | MIT | HRNet-W48 ImageNet backbone (Face X-ray; no detector head) | `0efec102d97f2ef5…` | | `backbones/xception-b5690688.pth` | 91.7 MB | BSD-3-Clause | ImageNet Xception backbone (Cadene port, BSD-3-Clause; used by IID; no detector head) | `b56906886cbaf573…` | | `community_forensics/model.safetensors` | 87.3 MB | MIT | Community Forensics ViT-S/16-384 detector (~4,800 generators, CVPR 2025) | `b89f36275f3bf5e2…` | | `deepfakebench/I3D_8x8_R50.pth` | 112.5 MB | CC-BY-NC-4.0 | Kinetics-pretrained I3D 3D-ResNet-50 backbone (no detector head) | `c9e4e6c8e86ce231…` | | `deepfakebench/capsule_best.pth` | 15.7 MB | CC-BY-NC-4.0 | Capsule-Forensics (FF++ c23) | `40e96997c2248aeb…` | | `deepfakebench/cnnaug_best.pth` | 85.3 MB | CC-BY-NC-4.0 | CNN-Aug / ResNet-34 (FF++ c23) | `2ffdc665b9bd723a…` | | `deepfakebench/core_best.pth` | 87.8 MB | CC-BY-NC-4.0 | CORE (FF++ c23) | `3d34dca33ab66ee0…` | | `deepfakebench/effnb4_best.pth` | 71.0 MB | CC-BY-NC-4.0 | EfficientNet-B4 (FF++ c23) | `74cbd82d7be1bf5e…` | | `deepfakebench/f3net_best.pth` | 90.4 MB | CC-BY-NC-4.0 | F3Net (FF++ c23) | `deee7d25bfc1b8b5…` | | `deepfakebench/ffd_best.pth` | 87.8 MB | CC-BY-NC-4.0 | FFD (FF++ c23) | `445f5b70f115bf77…` | | `deepfakebench/meso4Incep_best.pth` | 0.1 MB | CC-BY-NC-4.0 | MesoInception-4 (FF++ c23) | `16a2d7cec846e5bc…` | | `deepfakebench/meso4_best.pth` | 0.1 MB | CC-BY-NC-4.0 | MesoNet-4 (FF++ c23) | `a362ebbef2658daa…` | | `deepfakebench/recce_best.pth` | 191.5 MB | CC-BY-NC-4.0 | RECCE (FF++ c23); reconstruction decoder untrained in this release | `1349fd80d75cdb2d…` | | `deepfakebench/spsl_best.pth` | 87.8 MB | CC-BY-NC-4.0 | SPSL (FF++ c23) | `7305d660c68aa6ac…` | | `deepfakebench/srm_best.pth` | 222.1 MB | CC-BY-NC-4.0 | SRM (FF++ c23) | `666e8fc7f9938f54…` | | `deepfakebench/ucf_best.pth` | 188.0 MB | CC-BY-NC-4.0 | UCF (FF++ c23) | `4afa2517feeb4730…` | | `deepfakebench/xception_best.pth` | 87.8 MB | CC-BY-NC-4.0 | Xception (FF++ c23) | `dd61186f9ac8e99a…` | | `lipforensics/lipforensics_ff.pth` | 144.1 MB | MIT | LipForensics (FF++-trained) mouth-motion detector | `4b7790bc8e02d0c2…` | | `safe/checkpoint-best.pth` | 5.8 MB | Apache-2.0 | SAFE wavelet-HH synthetic-image detector (KDD 2025) | `b3f5ecfb46a154ed…` | | `univfd/fc_weights.pth` | 0.0 MB | MIT | UnivFD linear probe on frozen CLIP ViT-L/14 features (768 -> 1) | `477100745713bcc9…` | ## Sources ### AASIST (clovaai/aasist) - Upstream: https://github.com/clovaai/aasist - License: MIT - Copyright (c) 2021-present NAVER Corp. - Citation: Jung, J.-W., Heo, H.-S., Tak, H., Shim, H.-J., Chung, J.-S., Lee, B.-J., Yu, H.-J., Evans, N. "AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks." ICASSP 2022. - Note: 2019-era ASVspoof2019 LA baseline; generalises poorly to modern in-the-wild deepfake audio (Deepfake-Eval-2024 AUC 0.43). Shipped as a reference baseline. - Files: `aasist/AASIST.pth` ### HRNet-Image-Classification (HRNet/HRNet-Image-Classification) - Upstream: https://github.com/HRNet/HRNet-Image-Classification - License: MIT - Copyright (c) Microsoft Corporation - Citation: Wang, J. et al. "Deep High-Resolution Representation Learning for Visual Recognition." IEEE TPAMI 2020. - Files: `backbones/hrnetv2_w48_imagenet_pretrained.pth` ### DeepfakeBench (SCLBD/DeepfakeBench) - Upstream: https://github.com/SCLBD/DeepfakeBench - License: CC-BY-NC-4.0 - Copyright (c) 2023, The Chinese University of Hong Kong, Shenzhen - Citation: Yan, Z., Zhang, Y., Yuan, X., Lyu, S., Wu, B. "DeepfakeBench: A Comprehensive Benchmark of Deepfake Detection." NeurIPS 2023 (Datasets & Benchmarks). arXiv:2307.01426 - Note: Detector checkpoints were trained on FaceForensics++ (c23), a dataset released for research use. The DeepfakeBench LICENSE text names the source code; the released weights carry no separate statement and are treated as CC BY-NC 4.0 as well. - Files: `backbones/xception-b5690688.pth`, `deepfakebench/I3D_8x8_R50.pth`, `deepfakebench/capsule_best.pth`, `deepfakebench/cnnaug_best.pth`, `deepfakebench/core_best.pth`, `deepfakebench/effnb4_best.pth`, `deepfakebench/f3net_best.pth`, `deepfakebench/ffd_best.pth`, `deepfakebench/meso4Incep_best.pth`, `deepfakebench/meso4_best.pth`, `deepfakebench/recce_best.pth`, `deepfakebench/spsl_best.pth`, `deepfakebench/srm_best.pth`, `deepfakebench/ucf_best.pth`, `deepfakebench/xception_best.pth` ### Community Forensics (OwensLab/commfor-model-384) - Upstream: https://huggingface.co/OwensLab/commfor-model-384 - License: MIT - (c) Jeongsoo Park, Andrew Owens (Harvard Vision Lab) - Citation: Park, J., Owens, A. "Community Forensics: Using Thousands of Generators to Train Fake Image Detectors." CVPR 2025. arXiv:2411.04125 - Files: `community_forensics/model.safetensors` ### LipForensics (ahaliassos/LipForensics) - Upstream: https://github.com/ahaliassos/LipForensics - License: MIT - Copyright (c) 2021 Alexandros Haliassos - Citation: Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M. "Lips Don't Lie: A Generalisable and Robust Approach to Face Forgery Detection." CVPR 2021. - Note: The checkpoint is linked from the README (Google Drive); the MIT license covers the repository. - Files: `lipforensics/lipforensics_ff.pth` ### SAFE (Ouxiang-Li/SAFE) - Upstream: https://github.com/Ouxiang-Li/SAFE - License: Apache-2.0 - Citation: Li, O., Cai, J., Hao, Y., Jiang, X., Hu, Y., Feng, F. "Improving Synthetic Image Detection Towards Generalization: An Image Transformation Perspective." KDD 2025. arXiv:2408.06741 - Files: `safe/checkpoint-best.pth` ### UniversalFakeDetect (WisconsinAIVision/UniversalFakeDetect) - Upstream: https://github.com/WisconsinAIVision/UniversalFakeDetect - License: MIT - Copyright (c) 2025 Wisconsin AI and Vision Lab (WAIV) - Citation: Ojha, U., Li, Y., Lee, Y. J. "Towards Universal Fake Image Detectors that Generalize Across Generative Models." CVPR 2023. - Files: `univfd/fc_weights.pth` ## Use ```bash pip install -U "huggingface_hub[cli]" hf download ADPer/verity-weights --local-dir weights # or let Verity do it: VERITY_WEIGHTS_HUB=ADPer/verity-weights verity weights pull --trained ``` These models are statistical indicators, not proof. Their outputs must be interpreted by a qualified examiner alongside the explainability artifacts Verity produces, and never cited as a standalone determination.