The generation of images and videos using generative models—especially diffusion- based models—is currently a very active research area, particularly following the rise of large language models (LLMs). When generation focuses on an avatar, we refer to digital twins. These generative digital twins make it possible to simulate, anticipate and op- timise movements, postures and physical efforts, while integrating the physiological, geometric and environmental constraints specific to hu- man beings and the space in which they evolve. However, these models still struggle to preserve anatomical coherence when applied to images depicting human bodies in complex motion – such as in dance, gymnastics, or other unusual contexts. This internship has two main objectives: 1. Generate anatomically complex yet realistic human motions, including atypical ones; 2. Enable automatic editing of videos involving moving human bodies while preserving their anatomical coherence. In our setting, we aim to integrate LDDMM (Large Deformation Diffeomorphic Metric Mapping), into the loss function of the diffusion network involved in image modification, in the form of explicit constraints. In particular, recent deep learning work has demonstrated the effectiveness of such metrics for constraining the training of convolutional networks through graph structures. In parallel, we plan to use reinforcement learning to dynamically correct anatomically inconsistent generations.