Optimal Transport (OT) provides a principled way to impose geometric structure on the latent spaces of deep generative models. By aligning probability distributions through optimal couplings, it enhances interpolation quality and promotes smoother, more organized latent representations. When integrated with generative architectures, OT enables the learning of interpretable and structured spaces for analyzing complex data. In this work, we present a framework that integrates Conditional Variational Autoencoders (CVAEs) with OT for modeling and reconstructing 3D human body shapes. Our model learns a continuous latent space that captures morphological variability in a compact and geometrically structured form. The use of OT encourages smoother interpolations and better organization within the latent space and the enhancing the generative capabilities of the model. The dataset comprises STAR-based parametric models obtained from multiple sessions throughout nutritional treatments. By applying Wasserstein metrics to the latent space, we improve the structure and interpretability of the learned body representations, obtaining a disentangled representation. Code is available at: https://github.com/Tech4DLab/Disentangled-Latent-Space-for-3D-Human-Body-Models-with-CVAE-and-OT .