A biometric research competition has found that adapting a large foundation model with synthetic identity data can improve face-recognition performance without using photographs of real people for training.
The Competition on Adapting Foundation Models for Face Recognition Using Synthetic Training Data was organized for the 2026 International Joint Conference on Biometrics. A paper summarizing the results reports eight valid submissions from four teams across Full Data and Limited Data tracks.
Every team started with the CLIP ViT-L/14 vision-language model. Participants then adapted it for face recognition using training images generated exclusively with IDPERTURB, a method for creating synthetic identities and variations in their appearance. The full-data track allowed larger training sets, while the limited-data track was designed to test more resource-constrained approaches.
Submissions were assessed on LFW, CFP-FP, AgeDB-30, CALFW, CPLFW, IJB-B, IJB-C and TinyFace, covering both verification and identification tasks. The organizers also evaluated performance across four demographic groups in the Racial Faces in the Wild dataset, adding a fairness dimension to the overall ranking.
The adapted systems substantially improved on the off-the-shelf CLIP model, and several exceeded the competition baseline. DMSTI-Neurotechnology led the Full Data track using full fine-tuning with Sub-Center ArcFace, a loss function designed to improve separation between identities. Idiap-BSP won the Limited Data track with rank-stabilized low-rank adaptation, which changes a smaller set of model parameters.
The work adds to continuing efforts to build face-recognition systems without assembling large training collections of real faces. That goal has become more important as the academic community has debated the consent and governance of facial datasets. Synthetic data can reduce direct exposure of real identities, but it does not remove the need to test for demographic imbalance, unrealistic image variation or weaknesses that appear in operational conditions.
Benchmark results also do not establish performance in a deployed identity system. Real-world cameras, image compression, aging, occlusion and deliberate manipulation can affect matching. Separate research has proposed methods for detecting altered identities in facial images, illustrating the range of conditions a production system may need to handle.
The paper has been accepted for IJCB 2026 and is available as an arXiv preprint. The authors present the competition as evidence that foundation models can be specialized for biometric recognition with synthetic-only adaptation, while the two-track design shows that competitive gains are possible under different data and computing constraints.
Sources: arXiv
–
By Ji-seo Kim
