Haiyu Wu1 Jaskirat Singh2 Sicong Tian3
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:
✅ that generates face images purely based on the given image features
✅ that achieves state-of-the-art performance in five standard test sets among synthetic datasets
✅ that first achieves higher accuracy than the same-scale real dataset (on CALFW)
✅ that can easily scale the dataset size to 10M images from 200k identities
Please ⭐ if you find it is helpful😄
- [2024/09/15] 🔥 The generated HSFace datasets are available now!
- [2024/09/05] 🔥 Our paper is on Arxiv now!
- [2024/09/02] 🔥 We release Vec2Face demo!
- [2024/09/01] 🔥 We release Vec2Face and HSFace datasets!
conda env create -f environment.yaml
conda activate vec2face- 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", filename="weights/vec2face_generator.pth", local_dir="./")or
python download_vec2face_weights.py
- The weights of the FR models trained with HSFace (10k, 20k, 100k, 200k) can be downloaded using python:
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="fr_weights/hsface10k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="fr_weights/hsface20k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="fr_weights/hsface100k.pth", local_dir="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="fr_weights/hsface200k.pth", local_dir="./")The dataset used for Vec2Face 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="./")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="lmdb_dataset/WebFace4M/50000_ids_1022444_ims.npy", local_dir="./")The generated synthetic datasets HSFace300k can be downloaded at Gdrive and 百度云 (code:vc2f), HSFace10k and HSFace20k can be downloaded using python:
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="HSFaces/hsface10k.lmdb", local_dir="./", repo_type="dataset")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="HSFaces/hsface20k.lmdb", local_dir="./", repo_type="dataset")For HSFace100k and HSFace200k, they are the first 100k and 200k folders in the HSFace300k. You can conveniently use the indices mask to train the model with either of them. The mask files can be downloaded using python:
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="HSFaces/hsface100k_mask.npy", local_dir="./", repo_type="dataset")
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="HSFaces/hsface200k_mask.npy", local_dir="./", repo_type="dataset")and add another line in the config file for FR model training:
config.mask="./HSFaces/hsface100k_mask.npy"Tip: If you want to convert .lmdb to datasets (images), please refer to lmdb2dataset.py.
Putting reference images in a folder or collecting image paths in a .txt file for preparation. Then run following code:
python image_generation_with_reference.py \
--image_file "path/of/the/image/file or folder" \
--model_weights weights/vec2face_generator.pth \
--batch_size 5 \
--example 10 \
--name images-of-references
Note that, the input images should be cropped and aligned. If they are not, please use face detectors (e.g., img2pose) to crop the images first. We don't suggest you to modify the code to use the embedding extracted from insightface, because it takes forever to run. (Trust me, I have tried.)
Before generating images, the identity vectors need to be created/calculated and saved in a .npy file. We provide an example for you, but you can create your own center features (see issue #2).
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="BooBooWu/Vec2Face", filename="center_feature_examples.npy", local_dir="./")Image generation with sampled identity features:
python image_generation.py \
--model_weights weights/vec2face_generator.pth \
--batch_size 5 \
--example 1 \
--start_end 0:10 \
--name test \
--center_feature center_feature_examples.npy
Image generation with target yaw angle:
python pose_image_generation.py \
--model_weights weights/vec2face_generator.pth \
--batch_size 5 \
--example 1 \
--start_end 0:10 \
--center_feature center_feature_examples.npy \
--name test \
--pose 45 \
--image_quality 25
We only provide the WebFace4M dataset (see here) and the mask that we used for training the model, if you want to use other datasets, please referring the prepare_training_set.py to convert the dataset to .lmdb. Please refer to issue #3 for details.
Once the dataset is ready, modifying the following code to run the multi-node distributed training:
torchrun --nproc_per_node=4 --node_rank=0 --master_addr="host_addr" --master_port=3333 vec2face.py \
--rep_drop_prob 0.1 \
--use_rep \
--batch_size 32 \
--model vec2face_vit_base_patch16 \
--epochs 800 \
--warmup_epochs 5 \
--blr 4e-5 \
--output_dir workspace/pixel_generator/ \
--train_source ./lmdb_dataset/WebFace4M/WebFace4M.lmdb \
--mask lmdb_dataset/WebFace4M/50000_ids_1022444_ims.npy \
--accum_iter 1
If training on one node, run with the following command:
torchrun --nproc_per_node=4 vec2face.py \
--rep_drop_prob 0.1 \
--use_rep \
--batch_size 32 \
--model vec2face_vit_base_patch16 \
--epochs 800 \
--warmup_epochs 5 \
--blr 4e-5 \
--output_dir workspace/pixel_generator/ \
--train_source ./lmdb_dataset/WebFace4M/WebFace4M.lmdb \
--mask lmdb_dataset/WebFace4M/50000_ids_1022444_ims.npy \
--accum_iter 1
We borrowed the code from SOTA-Face-Recognition-Train-and-Test to train the model. The random erasing function could be added after line 84 in data_loader_train_lmdb.py, as shown below:
transform = transforms.Compose(
[
transforms.Resize((112, 112)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
transforms.RandomErasing()
]
)Please follow the guidance of SOTA-Face-Recognition-Train-and-Test for the rest of training process.
Since some people requested for the config file to reproduce the performance, here is the example of training with HSFace10K:
from easydict import EasyDict
config = EasyDict()
config.prefix = "arcface-r50-vec2face-hsface10k"
config.head = "arcface"
config.input_size = [112, 112]
config.embedding_size = 512
config.depth = "50"
config.batch_size = 128
config.weight_decay = 5e-4
config.lr = 0.1
config.momentum = 0.9
config.epochs = 26
config.margin = 0.5
config.fp16 = True
config.sample_rate = 1.0
config.num_ims = 500000
config.reduce_lr = [12, 20, 24]
config.train_source = "./HSFaces/hsface10k.lmdb"
config.val_list = ["lfw", "cfp_fp", "agedb_30", "calfw", "cplfw"]
config.mask = None
config.augment = True
config.mode = "se"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 |
| HSFace10K(Ours) | 0.5M | 98.87 | 88.97 | 85.47 | 93.12 | 93.57 | 92.00 |
| CASIA-WebFace (Real) | 0.49M | 99.38 | 96.91 | 89.78 | 94.50 | 93.35 | 94.79 |
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 |
| HSFace20K | 1M | 98.87 | 89.87 | 86.13 | 93.85 | 93.65 | 92.47 |
| HSFace100K | 5M | 99.25 | 90.36 | 86.75 | 94.38 | 94.12 | 92.97 |
| 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 |
| CASIA-WebFace (Real) | 0.49M | 99.38 | 96.91 | 89.78 | 94.50 | 93.35 | 94.79 |
We test the model performance on other four datasets, Hadrian (facial hair), Eclipse (face exposure), SLLFW (similar-looking), and DoppelVer (doppelganger).
| Datasets | Hadrian | Eclipse | SLLFW | DoppelVer |
|---|---|---|---|---|
| HSFace10K | 69.47 | 64.55 | 92.87 | 86.91 |
| HSFace20K | 75.22 | 67.55 | 94.37 | 88.91 |
| HSFace100K | 80.00 | 70.35 | 95.58 | 90.39 |
| HSFace200K | 79.85 | 71.12 | 95.70 | 89.86 |
| HSFace300K | 81.55 | 71.35 | 95.95 | 90.49 |
| CASIA-WebFace (Real) | 77.82 | 68.52 | 96.95 | 95.11 |
- 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❤️.
- Thanks to JiaquanYe for helping the training stability❤️.
If you find Vec2Face useful for your research, please consider citing us😄:
@article{wu2024vec2face,
title={Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors},
author={Wu, Haiyu and Singh, Jaskirat and Tian, Sicong and Zheng, Liang and Bowyer, Kevin W.},
journal={arXiv preprint arXiv:2409.02979},
year={2024}
}