3DDF

This gated research release contains the code and strict evaluation data used for 3D-generated face forgery detection experiments. It is organized as method-level archives so the Hub repository stays usable without committing hundreds of thousands of individual frame files.

Included Data

• NeRF: SyncTalk, ERNeRF, RADNeRF
• 3DGS: InsTag, DEGSTalk, TalkingGaussian
• EG3D-family: Real3D, MimicTalk, EMOPortraits
• External evaluation: GAGAvatar, GPAvatar
• Real reference pool: HDTF/TFHP

The strict release contains 233,863 cropped RGB frames, matching 81-point landmark arrays, and the corresponding 5,146 source videos. Invalid or unreferenced files are excluded using manifests/3ddf_strict.jsonl.

Layout

code/DeepfakeBench/       cleaned source snapshot and configs
data/processed/           RGB frame + landmark archives by method
data/raw/                 source video archives by method
manifests/                relative-path manifests and split metadata
SHA256SUMS                checksums for the complete release

Archives

Archive	Size	SHA-256
data/processed/DEGSTalk.tar.zst	1.146 GiB	f347a7019aebee38bc38c2f7357e8477b2c010a77e21f61018cc276ff00ad865
data/processed/EMOPortraits.tar.zst	1.197 GiB	8553abee263710bebdbd201643018661526a3fbe7f638610d4d3e131498ac8dc
data/processed/ERNeRF.tar.zst	2.413 GiB	90e97a62fbfe4c1158c581b0fb4d029d515974dfcc96607dc8fdbb1510a59649
data/processed/GAGAvatar.tar.zst	0.623 GiB	a0fa7abaa229fdaf09b83246b27c4e5d6374e48b2c1e3fe6a0bc980f755bfdc8
data/processed/GPAvatar.tar.zst	0.872 GiB	aae2a205a971d863d4192a24fdfec10d6fa6e277e97357e0c8d0305233ef167d
data/processed/InsTag.tar.zst	1.054 GiB	308c33c35e43de4bf40caaa8ac3cc29bef4f3978707b1d0774ad479acf1c4b94
data/processed/MimicTalk.tar.zst	1.314 GiB	7444bc72973a5cad4d6ad02eb8aff7dcb4d65e6cd030de9dd687770d9ea3b415
data/processed/RADNeRF.tar.zst	1.095 GiB	1438936ef7c091e5820c1ccf10e1dcee3ced30dd507df6b827b41278b48eecad
data/processed/Real.tar.zst	4.170 GiB	3e89ebe179f127c4fcb012a867dd4f45b693950bd695d33d99549f10e4620304
data/processed/Real3D.tar.zst	1.278 GiB	551290edf2dea4f8877372da11fab02d4787615271b7c0c638f6f906e48a8667
data/processed/SyncTalk.tar.zst	3.172 GiB	c71a30bec9f5bbb51d66276893c700989610b0dde541ea092ccae58493feff83
data/processed/TalkingGaussian.tar.zst	1.058 GiB	dff252936e1596ed2e435f7597d3487bfc0d7001ff936214708ec2e397cc281f
data/raw/DEGSTalk.tar.zst	0.937 GiB	4fd8c30ad1594e870b87d391c681157fad7ef383165d55370b4428eddce7e49e
data/raw/EMOPortraits.tar.zst	0.187 GiB	1e3b6e4fc2ea628d38415cc0179cdb8dcf56f4a82e648a6b4ef716286a9f9dba
data/raw/ERNeRF.tar.zst	0.568 GiB	fdbc913e61931cf03e1da331b45b317990f5d6860e6638eedafbaa190aa9f191
data/raw/GAGAvatar.tar.zst	0.369 GiB	0ea325be064b403fe06b03a01382704c2a5ec1f31504ff1ca9dcbffa18bec6cf
data/raw/GPAvatar.tar.zst	0.225 GiB	05e57f2b4b77bd8aec66e4dcd4778f5cda0aeff35481a7224c836e98113719e8
data/raw/InsTag.tar.zst	1.051 GiB	5b28e3ed7f665fa07814ce7135095a791074d41cbbc4b424c116aadd55a43314
data/raw/MimicTalk.tar.zst	0.159 GiB	995cc475b56b6094a2fe2db78494787dfd6b6a0a281d027a1de5964d6c5e0018
data/raw/RADNeRF.tar.zst	0.451 GiB	383b1b8b22e1d4be3153a2cfff1c238da6f6c1d6a8799317423d59b9f5598f5a
data/raw/Real.tar.zst	3.307 GiB	691d82bd32224aa5b24cc916e3e28a75974003540c7c3dd8c3d3a7927ae62ea3
data/raw/Real3D.tar.zst	0.167 GiB	9d65b79d519273dbb57daa25b35dc5805eb47d1a2108fd004515cee3fad723d1
data/raw/SyncTalk.tar.zst	1.369 GiB	866349ec048bc2fb96008907218fa359e5b6f787fa597ef502fec25bb4e97453
data/raw/TalkingGaussian.tar.zst	0.394 GiB	b120d744b31a054ad8872ed221886e3ee5c9f4bdd8648d8d4673a9599dff4c22

Each archive stores paths below 3DDF_Dataset/. Extract the selected archives into the packaged DeepfakeBench directory:

for archive in data/processed/*.tar.zst data/raw/*.tar.zst; do
  tar --zstd -xf "$archive" -C code/DeepfakeBench
done

The packaged manifests use ./3DDF_Dataset/... paths relative to code/DeepfakeBench. The original local absolute paths have been removed.

Training

The DeepfakeBench entry point is:

cd code/DeepfakeBench
python training/train.py \
  --detector_path training/config/detector/altfreezing.yaml \
  --ckpt_key 3DDF

Backbone weights and Python environments are not mirrored in this release. Use the upstream model sources and the dependency files under code/DeepfakeBench/.

Access And Responsible Use

This repository is gated. It is intended for non-commercial research on deepfake detection, robustness, and provenance. It must not be used for identity impersonation, harassment, surveillance, or creation of deceptive media. Recipients remain responsible for respecting the terms of every source dataset and generation method.

The DeepfakeBench software is distributed under its included CC BY-NC 4.0 license. That software license does not automatically relicense third-party data or model weights. Data components retain their original terms.

Integrity

Run the following from the repository root after download:

sha256sum -c SHA256SUMS

Release composition and byte counts are recorded in manifests/release_manifest.json.