LipSync-A

Overview

LipSync-A is a multi-generator talking-head dataset for LipSync detection and source attribution. It is introduced in Ariadne’s Thread of LipSync: Unraveling Forgeries via Inconsistency between Lip Motions and Head Poses (ICML 2026).

LipSync forgeries keep the source face and fabricate speech that was never spoken. LipSync-A labels every clip with its generator so a model can both detect the forgery and trace which method produced it.

This release contains 21 generators and 10,573 forged videos. 9,474 of them have a paired real clip. Identity and driving signals come from VFHQ, VoxCeleb, VCTK, HDTF, and LaPa.

Dataset structure

Each generator is one top-level folder. The folder name is the attribution label.

Path	Content
1_fake/	Forged talking-head video
2_identity/	Identity image or video
3_audio/	Driving audio (audio-driven methods)
0_real/	Paired real video, same stem as 1_fake
selection.json	Sample list and source metadata

0_real is the identity-source clip for audio-driven methods, and the driving video for video-driven methods. LaPa / MEAD stills and commercial Wan outputs have no recoverable real video, so those samples have no 0_real.

Methods

The table below is the composition of this release: each folder is one generator, labeled by architecture family, driving paradigm, and the data used to train the original model.

Generator	Category	Paradigm	Training Data
X2Face (Wiles et al., 2018)	Geometric	Video-driven	VoxCeleb
TPSM (Zhao & Zhang, 2022)	Landmark	Video-driven	VoxCeleb
FaceVid (Wang et al., 2021)	Landmark	Video-driven	VoxCeleb
LIA (Wang et al.)	Latent Space	Video-driven	VoxCeleb
DINet (Zhang et al., 2023)	CNN	Audio-driven	HDTF, MEAD
DaGAN (Hong et al., 2022)	GAN	Video-driven	VoxCeleb
MakeItTalk (Zhou et al., 2020)	GAN	Audio-driven	VoxCeleb, ObamaSet
Wav2Lip (Prajwal et al., 2020)	GAN	Audio-driven	LRW, LRS2
TalkLip (Wang et al., 2023)	GAN	Audio-driven	LRS
SadTalker (Zhang et al., 2023)	VAE	Audio-driven	VoxCeleb
V-Express (Wang et al., 2024)	VAE	Audio-driven	VFHQ
IP_LAP (Zhong et al., 2023)	Transformer	Audio-driven	LRS2
EAT (Gan et al., 2023)	Transformer	Audio-driven	MEAD, LRW
DreamTalk (Ma et al., 2023)	Diffusion	Audio-driven	MEAD, HDTF
Sonic (Ji et al., 2025)	Diffusion	Audio-driven	VFHQ, CelebV-Text
KDTalker (Yang et al., 2025)	Diffusion	Audio-driven	HDTF
OmniSync (Peng et al., 2025)	Diffusion	Audio-driven	Web-Collected
InfiniteTalk (Yang et al., 2025)	Diffusion	Audio-driven	Internal
LatentSync (Li et al., 2024)	Diffusion	Audio-driven	VoxCeleb2, HDTF
SkyReels	Diffusion	Audio-driven	Internal
WavSV (Wan, commercial)	Diffusion	Audio-driven	Internal

Folder names match the attribution labels: face_vid, wav2lip, talklip, and dreamtalk correspond to FaceVid, Wav2Lip, TalkLip, and DreamTalk.

Citation

@inproceedings{she2026ariadne,
  title     = {Ariadne's Thread of LipSync: Unraveling Forgeries via Inconsistency between Lip Motions and Head Poses},
  author    = {She, Tianyi and Liu, Jiawei and Liu, Weifeng and Zhao, Hanqing and Zhang, Weiming and Chen, Kejiang},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026}
}

Ethical Statement

When using this dataset, comply with applicable laws and regulations. Do not use it to create misleading or deceptive media.