Vec2Face+ for Face Dataset Generation

Haiyu Wu¹   Jaskirat Singh²   Sicong Tian³

Liang Zheng²   Kevin W. Bowyer¹  

¹University of Notre Dame
²The Australian National University
³Indiana University South Bend

This is the official implementation of Vec2Face+, an ID and attribute controllable face dataset generation model. This work includes:

 ✅ a more training-efficient and attribute controllable model than Vec2Face.
 ✅ a set of synthetic FR datasets that first-time achieve a higher average accuracy than CASIA-WebFace
 ✅ suggesting the importance of intra-class identity consistency
 ✅ reporting the failure of the existing synthetic datasets on twin verification
 If you find Vec2Face+ useful for your research, please give a ⭐ and consider citing us 😄

@article{wu2025vec2face+,
  title={Vec2Face+ for Face Dataset Generation},
  author={Wu, Haiyu and Singh, Jaskirat and Tian, Sicong and Zheng, Liang and Bowyer, Kevin W},
  journal={arXiv preprint arXiv:2507.17192},
  year={2025}
}

News/Updates

• [2024/07/28] 🔥 The generated HSFace datasets are available now!
• [2024/07/28] 🔥 Our paper is on Arxiv now!
• [2025/07/27] 🔥 We release Vec2Face+!

🔧 Installation

conda env create -f environment.yaml
conda activate vec2face-plus

Download model weights

1. The weights of the Vec2Face model and estimators used in this work can be manually from HuggingFace or using python:

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="weights/6DRepNet_300W_LP_AFLW2000.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="weights/arcface-r100-glint360k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="weights/magface-r100-glint360k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="vec2face_plus/main_model.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="vec2face_plus/pose_model.pth", local_dir="./")

or

python download_vec2face_weights.py

2. The weights of the FR models trained with VFace (10k, 20k, 100k, 200k) can be downloaded using python:

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="fr_weights/vface10k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="fr_weights/vface20k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="fr_weights/vface100k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="fr_weights/vface300k.pth", local_dir="./")

Download datasets

The dataset used for Vec2Face+ main model training can be downloaded from manually from HuggingFace or using python:

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="lmdb_dataset/WebFace4M/WebFace4M.lmdb", local_dir="./")

The dataset used for Vec2Face+ pose model training can be downloaded from manually from HuggingFace or using python:

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="WebFace4M/WebFace4M_with_landmarks.lmdb", local_dir="./")

The generated synthetic datasets VFace10k, VFace20k, VFace100K can be downloaded using python:

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="VFaces/vface10k.lmdb", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="VFaces/vface20k.lmdb", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face_plus", filename="VFaces/vface100k.lmdb", local_dir="./")

Tip: If you want to convert .lmdb to datasets (images), please refer to lmdb2dataset.py.

Model Architecture

⚡Image generation

Generating images with reference image, center features, or AttrOP can be found here. It is the same as the original Vec2Face.

Generating with pose condition

python pose_control.py \
--batch_size 64 \
--image_file "path/to/your/image/file" \
--example 1 \
--name "test_examples" \
--use_lora \
--lora_r 8 \
--model_weights "./vec2face_plus/pose_model.pth" \
--pose_file "./landmarks/landmark_examples.txt"

Training

The main model training is the same as the original Vec2Face, which can be found here.

Vec2Face+ pose model training

We only provide the WebFace4M dataset (see here), if you want to use other datasets, please referring the prepare_training_set.py to convert the dataset to .lmdb.

Once the dataset is prepared, run with the following command for training:

torchrun --nproc_per_node=4 pose_training.py \
--batch_size=64 --model=vec2face_vit_base_patch16 \
--train_source=./WebFace4M/WebFace4M_with_landmarks.lmdb \
--output_dir=./workspace/landmark_guided_pose_cond \
--lr 1e-4 --warmup_epochs 1 --cosine_lr --min_lr 1e-6 \
--model_weights=vec2face_plus/main_model.pth \
--use_amp --use_lora --lora_r 8 --epochs 10

FR model training

We borrowed the code from SOTA-Face-Recognition-Train-and-Test to train the model. Please make sure the random erase is enabled by adding the following line to the configuration file:

config.rand_erase = True

Please follow the guidance of SOTA-Face-Recognition-Train-and-Test for the rest of training process.

Performance

Datasets in 0.5M scale

This table compares the existing synthetic dataset generation methods on five standard face recognition test sets. The model trained with HSFace10K has better performance on CALFW than that trained with real dataset.

Training sets	# images	LFW	CFP-FP	CPLFW	AgeDB	CALFW	Avg.
IDiff-Face	0.5M	98.00	85.47	80.45	86.43	90.65	88.20
DCFace	0.5M	98.55	85.33	82.62	89.70	91.60	89.56
Arc2Face	0.5M	98.81	91.87	85.16	90.18	92.63	91.73
DigiFace	1M	95.40	87.40	78.87	76.97	78.62	83.45
SynFace	0.5M	91.93	75.03	70.43	61.63	74.73	74.75
SFace	0.6M	91.87	73.86	73.20	71.68	77.93	77.71
IDnet	0.5M	92.58	75.40	74.25	63.88	79.90	79.13
ExFaceGAN	0.5M	93.50	73.84	71.60	78.92	82.98	80.17
SFace2	0.6M	95.60	77.11	74.60	77.37	83.40	81.62
Langevin-Disco	0.6M	96.60	73.89	74.77	80.70	87.77	82.75
ID^3	0.5M	97.68	86.84	82.77	91.00	90.73	89.80
CemiFace	0.5M	99.03	91.06	87.65	91.33	92.42	92.30
HSFace10K	0.5M	98.87	88.97	85.47	93.12	93.57	92.00
VFace10K (ours)	0.5M	99.35	93.56	88.03	94.33	94.17	93.89
CASIA-WebFace (Real)	0.49M	99.38	96.91	89.78	94.50	93.35	94.79

Dataset scaling up to 300K identities and 15M images

This is the uniqueness of the proposed Vec2Face, which can easily scale the dataset size up.

Datasets	# images	LFW	CFP-FP	CPLFW	AgeDB	CALFW	Avg.
HSFace10K	0.5M	98.87	88.97	85.47	93.12	93.57	92.00
VFace10K (ours)	0.5M	99.35	93.56	88.03	94.33	94.17	93.89
HSFace20K	1M	98.87	89.87	86.13	93.85	93.65	92.47
VFace20K (ours)	1M	99.40	94.23	89.10	95.58	94.67	94.60
HSFace100K	5M	99.25	90.36	86.75	94.38	94.12	92.97
VFace100K (ours)	4M	99.52	95.09	90.23	95.17	94.38	94.88
HSFace200K	10M	99.23	90.81	87.30	94.22	94.52	93.22
HSFace300K	15M	99.30	91.54	87.70	94.45	94.58	93.52
VFace300K (ours)	12M	99.57	94.80	90.43	95.17	94.70	94.93
CASIA-WebFace (Real)	0.49M	99.38	96.91	89.78	94.50	93.35	94.79

Acknowledgements

• Thanks to the WebFace4M creators for providing such a high-quality facial dataset❤️.
• Thanks to Hugging Face for providing a handy dataset and model weight management platform❤️.
• The main code is borrowed from Vec2Face.