This paper presents a novel approach for generating synthetic license plate (LP) images using Latent Diffusion Models (LDMs). The generation of synthetic data is crucial in the domain of LP recognition (LPR) due to strict privacy regulations limiting access to real-world datasets. The study leverages a dataset of 250,000 Ukrainian LP images to train an LDM conditioned on LP text. To that end, multiple Variational Autoencoders (VAEs) are compared in terms of their performance, and the resulting LDM is evaluated based on visual realism and control accuracy. Results demonstrate that the LDM outperforms traditional generative models, such as Generative Adversarial Networks (GANs), in producing realistic and diverse LP images while maintaining high control over the generation process. More precisely, our model yields a Fréchet Inception Distance of 16.15, compared to 31.73 obtained by the baseline. The proposed method shows promise in enhancing the quality of LPR systems while mitigating data scarcity issues.