Data-driven methods are increasingly explored as alternatives to classical plasticity theory for constitutive modelling of sand. This study introduces a conditional generative adversarial network (cGAN-LD) for modelling state-dependent sand behaviour by integrating the predictions from the Li-Dafalias (LD) model [1] with experimental data. The generator network learns the discrepancy between the LD model prediction and real data, producing a residual stress–strain response conditioned on the initial state. Adding this learned residual to the LD predictions yields the synthetic states. The discriminator network guides the generator training process so that the predicted states are indistinguishable from the real data. The predictive performance of the cGAN-LD model is validated using drained triaxial compression tests on Karlsruhe fine sand [2] and undrained triaxial compression tests on Toyoura sand [3]. Comparative studies with standard GAN, GAN-LD, and Wasserstein-GAN-LD architectures demonstrate that cGAN-LD provides the most accurate and physically consistent predictions among the tested approaches. Keywords: Sand behaviour, constitutive model, generative adversarial networks, residual learning