Researchers explore facial recognition with foundation models, synthetic data

In the era of convolutional neural networks, the most accurate face biometrics algorithms tend to be purpose-built. Researchers competing to see how close they can get with foundation models and synthetic data as part of a competition have produced results that indicate choices around adaptation method and loss function can deliver significant performance differences.

The Competition on Adapting Foundation Models for Face Recognition Using Synthetic Training Data (AFMFR) is part of the 2026 International Joint Conference on Biometrics (IJCB 2026), on September 1 to 4 in Rome.

The competition was organized by researchers at the Fraunhofer Institute for Computer Graphics Research IGD, the Technical University of Darmstadt and the University of Ljubljana, Slovenia.

Participants submitting valid entries were affiliated with Vilnius University (DMSTI), the Biometric Security and Privacy (BSP) team at the Idiap Research Institute, the Indian Institute of Technology, Jammu and ArogyaPandit Private Limited and the Southern University of Science and Technology, China.

Researchers adapted the CLIP ViT-L/14 foundation model to perform facial recognition for the competition. The performance of the submitted models was compared with eight different benchmarks, including LFW, AgeDB-30, IJB-B and IJB-C. and fairness was compared across four demographic groups.

Four teams submitted a total of eight entries across two tracks; one in which large-scale synthetic identity data was used for model adaptation, and a “Limited Data Track.”

Unsurprisingly, the models adapted with synthetic data delivered significantly better performance than the off-the-shelf model. The competition organizers suggest the findings indicate synthetic data diversity could potentially play a valuable role in adapting foundation models for generalization.

A recent study from the University of Luxembourg showed synthetic data can also be used with some effectiveness in the evaluation of facial recognition algorithms.

Neurotechnology, Idiap teams beat the baseline

The DMSTI-Neurotechnology team scored the best accuracy against all small scale benchmarks in the Full Data Track and the top average accuracy, at 95.51 percent. The same team also led across IJB-B and IJB-C comparisons, far surpassing the baseline and “demonstrating superior robustness in low-FAR Regimes.” Idiap-BSP led in TinyFace Rank-1 and Rank-5 accuracy.

Idiap-BSP also led the Limited Data Track in every IJB and TinyFace benchmark, as well as average accuracy on small-scale benchmarks, at 94.52 percent.

The same two teams led in demographic fairness across the two scenarios, respectively, but systematic performance gaps between different demographic groups persisted.

DMSTI- Neurotechnology used the Sub-Center ArcFace loss function to achieve its top results, while  Idiap-BSP used rank-stabilized LoRA adaptation for fine-tuning.

A competition on document forgery detection organized by Hochschule Darmstadt, Fraunhofer IGD and Facephi will also be held at IJCB 2026.

Early registration for IJCB 2026 ends August 16.

Article Topics

biometric research  |  biometrics  |  facial recognition  |  Fraunhofer Institute IGD  |  Idiap  |  IJIS  |  Neurotechnology  |  synthetic data  |  synthetic faces