One of the biggest challenge to train robust deep learning models in medical imaging is the acquisition of high-quality 3D cardiac segmentation labels. Full volumetric annotations are costly, time consuming, and often limited across cardiac pathologies, blocking the development of generalizable solutions. To address this issue, we propose a novel text-conditioned framework for generating realistic 3D multi-label cardiac segmentation masks directly from clinical descriptions. Our approach adapts CogVideoX, a state-of-the-art text-to-video diffusion model with an expert transformer to operate in a volumetric segmentation domain, enabling the synthesis of anatomically coherent 3D masks from natural language prompts such as “dilated left ventricle” or “right ventricular hypertrophy.” By treating 3D segmentation volumes as pseudo video sequences, the model learns to translate semantic indicators of pathology into plausible geometric representations. This work opens new perspectives in the generation of text-based, anatomy-aware cardiac data and establishes the feasibility of bridging clinical language and 3D morphology.