File size: 7,686 Bytes
c6efbdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
---
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.