CLDM-Palm: A controllable latent diffusion model for high-fidelity palmprint generation based on Bézier curves

Using limited real data to synthesize realistic palmprints and expand training samples for recognition models has become a promising direction in palmprint recognition. However, the pseudo-palmprints generated by existing models still exhibit significant discrepancies from real ones, particularly in crease structures and fine-grained details. In this paper, we first introduce Latent Diffusion Models (LDM) as the backbone to improve the quality of palmprint generation. Secondly, to incorporate Bézier curves as control conditions into the model, we propose the Palm-to-Bézier Module (P2BM), which maps real palmprints to their corresponding Bézier-style pseudo-Bézier curves. These curves establish the connection between real palmprints and Bézier curves, which are used as conditional inputs during diffusion model training. At inference time, Bézier curves can be provided as conditions to generate high-resolution, fine-grained, and highly realistic synthetic palmprints. Thirdly, to enable Bézier curves to better model palmprint creases, we propose 12 Bézier curves templates based on real crease distribution priors. With only 10-step Denoising Diffusion Implicit Models (DDIM) sampling, our method achieves a significantly lower Fréchet Inception Distance (FID) compared to existing palmprint generation approaches. Moreover, the recognition models trained on the synthetic palmprints generated by our model achieve new state-of-the-art results in both Fisher Discriminant Ratio (FDR) and TAR@FAR metrics. Under a 1:1 train-test fine-tuning setting, our model improves average TAR@FAR=\(10^{-6}\) performance by over \(10\%\) compared to prior methods. We name our model CLDM-Palm (Controllable Latent Diffusion Model-Palm).