Langevin-Dispersion-10k

Face recognition models are trained on large-scale datasets, which have privacy and ethical concerns. Lately, the use of synthetic data to complement or replace genuine data for the training of face recognition models has been proposed. While promising results have been obtained, it still remains unclear if generative models can yield diverse enough data for such tasks. In this work, we introduce a new method, inspired by the physical motion of soft particles subjected to stochastic Brownian forces, allowing us to sample identities distributions in a latent space under various constraints. We introduce three complementary algorithms, called Langevin, Dispersion, and DisCo, aimed at generating large synthetic face datasets. With this in hands, we generate several face datasets and benchmark them by training face recognition models, showing that data generated with our method exceeds the performance of previously GAN-based datasets and achieves competitive performance with state-of-the-art diffusion-based synthetic datasets. While diffusion models are shown to memorize training data, we prevent leakage in our new synthetic datasets, paving the way for more responsible synthetic datasets. Project page: https://www.idiap.ch/paper/synthetics-disco

Overview

Training: Face Recognition models trained on Synthetic Face Recognition Datasets (Langevin-Dispersion-DisCo)
Backbone: IResnet50
Parameters: 43.6M
Task: Face Recognition models trained on Synthetic Face Datasets
Framework: Pytorch
Output structure: Batch of face embeddings (ie, features)

Evaluation of Models

For evaluation, we train face recognition models from scratch using the synthetic datasets and then evaluate trained models on benchmarking datasets. Performance benchmarks of face recognition models trained with our dataset and other datasets in the literature:

Dataset Type	Dataset Name	Generator	No. IDs	No. Samples	LFW	CPLFW	CALFW	CFP	AgeDB
Real images	MS-Celeb-1M	N/A	85,000	5,800,000	99.82	92.83	96.07	96.10	97.82
Real images	WebFace-4M	N/A	206,000	4,000,000	99.78	94.17	95.98	97.14	97.78
Real images	CASIA-WebFace	N/A	10,572	490,623	99.42	90.02	93.43	94.97	94.32
Computer Graphics	DigiFace-1M	Rendered mesh	109,999	1,219,995	90.68	72.55	73.75	79.43	68.43
Diffusion-based	DCFace-0.5M	custom trained	10,000	500,000	98.35	83.12	91.70	88.43	89.50
Diffusion-based	DCFace-1.2M	custom trained	60,000	1,200,000	98.90	84.97	92.80	89.04	91.52
Diffusion-based	IDiff-Face (Uniform)	custom trained	10,049	502,450	98.18	80.87	90.82	82.96	85.50
Diffusion-based	IDiff-Face (Two-Stage)	custom trained	10,050	502,500	98.00	77.77	88.55	82.57	82.35
GAN-based	Synface	StyleGAN2	10,000	999,994	86.57	65.10	70.08	66.79	59.13
GAN-based	SFace	StyleGAN2	10,572	1,885,877	93.65	74.90	80.97	75.36	70.32
GAN-based	SFace2	StyleGAN2	10,572	1,048,255	94.03	73.2	80.33	74.87	72.98
GAN-based	Syn-Multi-PIE	StyleGAN2	10,000	1,800,000	78.72	60.22	61.83	60.84	54.05
GAN-based	GANDiffFace	StyleGAN3	10,080	543,893	94.35	76.15	79.90	78.99	69.82
GAN-based	IDnet	StyleGAN2	10,577	1,057,200	84.48	68.12	71.42	68.93	62.63
GAN-based	ExFaceGAN	GAN-Control	10,000	599,944	85.98	66.97	70.00	66.96	57.37
GAN-based	Langevin-Dispersion [ours]	StyleGAN2	10,000	650,000	94.38	65.75	86.03	65.51	77.30
GAN-based	Langevin-DisCo [ours]	StyleGAN2	10,000	650,000	97.07	76.73	89.05	79.56	83.38
GAN-based	Langevin-DisCo [ours]	StyleGAN2	30,000	1,650,000	98.97	81.52	93.95	83.77	93.32

Comparison of the existing synthetic face datasets present in the literature with the best performing datasets created in this work. In addition to the number of identities and number of images, we present the recognition accuracy obtained by training an FR model on each dataset and evaluating on standard face recognition benchmarking datasets.

Running Code

Minimal code to instantiate the model and perform inference:

  # Inferece (Face Recognition)
  from face_alignment import align
  from inference import load_pretrained_model, to_input

  checkpoint = 'model_checkpoint.ckpt'
  model = load_pretrained_model(checkpoint, architecture='ir_50')
  path = 'path_to_the_image'
  aligned_rgb_img = align.get_aligned_face(path)
  bgr_input = to_input(aligned_rgb_img)
  feature, _ = model(bgr_input)

License

CC BY-NC-SA 4.0

Copyright

https://www.idiap.ch/paper/synthetics-disco/

Please refer to the link for information about the License & Copyright terms and conditions.

Citation

If you find our work useful, please cite the following publication:

@inproceedings{geissbuhler2025synthetic,
  title={Synthetic Face Datasets Generation via Latent Space Exploration from Brownian Identity Diffusion},
  author={David Geissb{\"u}hler and Hatef Otroshi Shahreza and S{\'e}bastien Marcel},
  booktitle={Forty-second International Conference on Machine Learning},
  year={2025}
}

Idiap
/

Langevin-Dispersion-10k