PoseDefCycleGAN: Identity-preserving face frontalization with deformable convolutions and pose-aware supervision

Face recognition systems have achieved impressive accuracy in controlled environments but continue to face challenges under extreme pose variations. To address this limitation, we propose a novel face frontalization framework, PoseDefCycleGAN, that combines the strengths of CycleGAN, deformable convolution, and pose-guided supervision. Our method leverages deformable convolution in the final layer of the generator to dynamically adapt the receptive field, enabling better reconstruction of complex facial geometries. Additionally, we incorporate a lightweight pose classification network to enforce pose-aware regularization, encouraging the generation of semantically consistent frontal images. The proposed model is trained using unpaired data and optimized with a combination of adversarial, cycle consistency, identity-preserving, and pose regularization losses. Extensive experiments on MultiPIE, AFW, and LFW datasets demonstrate that the method improves both visual fidelity and face recognition, particularly at extreme yaw angles: on MultiPIE we reduce FID to 15.90 (from 18.32 with CycleGAN) and achieve 98.9% rank-1 accuracy at ± 90^∘; on LFW we obtain 90.20% accuracy with LPIPS=0.3052. Quantitative evaluations further validate the contribution of deformable convolutions and pose supervision. Our work presents a robust solution for pose-invariant face recognition and establishes a strong benchmark for identity-preserving face frontalization. Model implementation is available on the author's GitHub page https://github.com/Shak97/PoseDefCycleGAN.