State-of-the-art face recognition models are trained on millions of real human face images collected from the internet. DigiFace-1M aims to tackle three major problems associated with such large-scale face recognition datasets.

• Ethical issues - Many existing datasets are obtained by collecting web images without explicit consent. Our digital faces are created using a generative model built from high quality head scans of a small number of individuals obtained with consent.
• Labeling noise - Web images collected by searching the names of celebrities often contain errors. Our synthetic data has guaranteed correctness of labels.
• Data bias - Face recognition models are generally trained and tested on celebrity faces, many of which are taken with strong lighting and make-up. They also have imbalanced racial distribution. Our synthetic data generation pipeline allows us to control the distribution of the data and ensure a fair dataset.

We build on the synthetic face generation framework of Wood et al. to create a dataset of over one million synthetic face images. This synthetics dataset helps us to overcome three primary shortcomings of existing large-scale face recognition datasets:

We define identity as a unique combination of facial geometry, texture, eye color and hair style. For each identity, we sample a set of accessories including clothing, make-up, glasses, face-wear and head-wear.

While hair style can change for an individual, most people maintain similar hair style (for both facial and head hair) which makes hair style an important cue for the person's identity. Therefore, for the same identity, we randomize only the color, density and thickness of the hair (top row) and avoiding the impression of changing identity (bottom row). This simulates aging to some extent as hair typically becomes grayer, sparser and thinner during aging. The hair style is only changed when the added head-wear is not compatible with the original hair style.

After sampling the identity and the accessories, we can render multiple images by varying the pose, expression, environment (lighting and background) and camera.

DigiFace-1M is split into two parts:

• 720K images with 10K identities (72 images per identity). For each identity, 4 different sets of accessories are sampled and 18 images are rendered for each set.
• 500K images with 100K identities (5 images per identity). For each identity, only one set of accessories is sampled.

SynFace is the current state-of-the-art for face recognition model trained on synthetic faces. They used DiscoFaceGAN to generate 500K synthetic faces of 10K unique identities. We significantly outperform SynFace across all datasets, suggesting that our rendered synthetic faces are better than GAN-generated faces for learning face recognition. This is likely because GAN-generated images do not enforce identity or geometric consistency, and are not effective at changing accessories. The GAN models also have unresolved ethical and bias concerns as they are typically trained on large-scale real face datasets.

When a small number of real face images are available, we can use them to fine-tune the network that is pre-trained on our synthetic data. Such fine-tuning significantly improves the accuracy across all datasets.

However, there remains a substantial accuracy gap from the state-of-the-art methods that are trained on large-scale real face datasets. This gap can be reduced by adopting better data augmentation or by improving the realism of the face generation pipeline. We leave this as future work.