We request 5,000 GPU-hours per month on Arrhenius for six months to develop models that recover, interpret and generate human and animal behaviour from visual data. Our central question is how learned representations can retain geometry, motion and nonverbal cues while generalising across subjects, datasets and body structures. The project combines human motion generation, affect-aware face reconstruction, animal pose and shape estimation, generative evaluation, orchestra conducting analysis, and initial development of VITAL for video-based physiological inference. Building on the group’s work on Berzelius, we will conduct focused cross-dataset experiments and controlled comparisons. VITAL extends the group’s facial modelling methods towards physiological signals. This six-month allocation will support selected experiments in the established directions and the first stage of VITAL model development.