VERITAS model mirror: GenD + DFDC-B7 face-swap detectors (ONNX)
This repository mirrors ONNX exports of two published, pretrained, MIT-licensed deepfake detectors. It exists so that the setup of VERITAS, a student deepfake-forensics lab built for the INFINIA 2026 school science fair (team: Khais Mohammed and Prithvi), can download verified files instead of rebuilding them from the original PyTorch checkpoints on every laptop.
We did not train these networks. All credit for the weights goes to their original authors (below). The files were exported to ONNX by the VERITAS team; numerically checked against the originals.
Files
| file | what it is | size (bytes) | SHA-256 |
|---|---|---|---|
gend_clip.onnx |
GenD, CLIP ViT-L/14 vision tower + GenD head, fp32 (exact published weights) | 1,212,926,317 | 371b05b7c3e66d6998313406a4d1fb5610cb195fe87634c1b1c842981364889f |
gend_clip_int8.onnx |
the same network, dynamic int8 quantisation (ONNX Runtime), an optional low-RAM fallback | 308,306,010 | 4234075cd6dc79e8be489944b057ba062bd4431d7417e18df1f415628c529e9e |
dfdc_b7.onnx |
DFDC first-place EfficientNet-B7, one of its 7 published checkpoints, fp32 | 254,811,737 | 6c14e1dae05a7ed999c971e581f620f7fe9f677d9e605307c006978a818bd7f1 |
Nothing else is mirrored here: no face detector (VERITAS downloads InsightFace's SCRFD det_10g.onnx from the
official InsightFace release, because its licence is non-commercial research only), no other model and no images,
audio or video.
The original models
GenD - "Deepfake Detection that Generalizes Across Benchmarks", Andrii Yermakov, Jan Cech, Jiri Matas, Mario
Fritz, WACV 2026. Code: https://github.com/yermandy/GenD . Weights: https://huggingface.co/yermandy/GenD_CLIP_L_14
(revision 891ce014a0308386c4d7d25b3dcf436a22db5504, model.safetensors, SHA-256
d76f0bdfd74a29fe1b1c1b84a80ac92486993e426878e8c7a3944281fbb96833). Licence: MIT. Trained by its authors on the
FaceForensics++ training split (FaceForensics++ terms of use apply to that data). Architecture: the CLIP ViT-L/14
vision tower (configuration of openai/clip-vit-large-patch14, MIT); every weight comes from GenD's checkpoint.
DFDC-B7 - the 1st-place solution of the Deepfake Detection Challenge (2020) by Selim Seferbekov.
Code and weights: https://github.com/selimsef/dfdc_deepfake_challenge (release 0.0.1, checkpoint
final_999_DeepFakeClassifier_tf_efficientnet_b7_ns_0_23, SHA-256
849036dc211387420412feb68c7451e98a948072b94e213e1104e2f2bf7791ad). Licence: MIT. Trained by its author on the DFDC
training set (DFDC research terms apply to that data). Architecture: tf_efficientnet_b7_ns from timm.
How the files were made and checked
scripts/export_faceswap_onnx.py in the VERITAS repository downloads the original checkpoints from the pinned URLs
above (SHA-256 checked), rebuilds the networks, exports them with torch.onnx (torch 2.15.0, onnx 1.23.2,
timm 1.0.31) and checks the result on fixed inputs against reference logits recorded from the PyTorch originals
(tolerance 2e-3 logit units). In our benchmark the ONNX outputs matched PyTorch within 2.4e-6 (GenD) and 5e-7 (B7)
logit units. With the locked torch version the rebuild is byte-identical to the files here. The int8 file was made
from gend_clip.onnx with ONNX Runtime dynamic quantisation (quantize_dynamic: int8 per-channel weights,
MatMul layers only); it is not numerically identical to fp32, and in our
tests its decisions jittered on borderline clips, so VERITAS uses fp32 by default.
Inputs and outputs
| file | input x |
output y |
|---|---|---|
gend_clip*.onnx |
float32 [1, 3, 224, 224]: a face aligned with 5 SCRFD landmarks (GenD's alignment, scale 1.3), RGB, CLIP preprocessing (resize 224 bicubic, centre crop, mean [0.48145466, 0.4578275, 0.40821073], std [0.26862954, 0.26130258, 0.27577711]) |
[1, 2] logits; softmax index 1 = P(fake) |
dfdc_b7.onnx |
float32 [1, 3, 380, 380]: an MTCNN face box with 1/3 margin, longest side resized to 380 and centred on black, RGB / 255, ImageNet mean [0.485, 0.456, 0.406] / std [0.229, 0.224, 0.225] |
[1, 1] logit; sigmoid = P(fake) |
The full pre-processing, face detection and score fusion are in server/forensics/face_swap.py of the VERITAS repository.
How VERITAS uses them, and the limits
VERITAS benchmarked six published face-swap detectors on 75 research clips (FaceForensics++ test split, Celeb-DF v2 test list, Google DFD) and integrated this pair. Each network scores 8 evenly spaced video frames; the thresholds were chosen on the 45 FaceForensics++ test clips only. The test is one-sided in VERITAS: it may raise an alarm, but it never vouches that a face is real, because both networks call fully generated (GAN) faces "real".
Please read the numbers with care (details in docs/MODEL_SELECTION.md of the repository):
- The pair was chosen after seeing its results on the Celeb-DF + DFD clips (datasets neither network was trained on), so its 0.998 AUC there (whole engine, 15 real + 15 fake clips) is optimistic. GenD's authors also used Celeb-DF++ validation videos to pick their checkpoint, so only Google DFD is fully unseen.
- It raises false alarms on grainy, heavily compressed real video: on its own it fired on 21 of 39 real clips degraded like a cheap webcam.
- It is meant for videos, not single photos.
These files are provided as-is for research and education, under the MIT licences of the original works.
Licences
- GenD weights: MIT (c) the GenD authors. ONNX exports of them (
gend_clip.onnx,gend_clip_int8.onnx): MIT. - DFDC-B7 weights: MIT (c) 2020 Selim Seferbekov. ONNX export (
dfdc_b7.onnx): MIT. - The training data of both networks (FaceForensics++, DFDC) has its own research terms; no data is included here.