We introduce MixerMDM, the first learnable approach for combining multiple pre-trained text-conditioned motion diffusion models. Unlike previous static methods, MixerMDM learns a dynamic mixing strategy to adaptively combine models based on their generated motions and input conditions. This enables fine-grained and consistent individual controllability over human-human interaction motion generation. We also propose a novel evaluation technique that measures mixing quality by assessing individual and interaction alignment.
We present in2IN, a diffusion model for human-human motion generation that uses both overall and individual action descriptions to improve the realism and diversity of interactions. By augmenting the InterHuman dataset with individual textual annotations using a large language model, in2IN achieves state-of-the-art results. To further boost intra-person diversity, we propose DualMDM, a model composition method that combines motions from in2IN with a single-person model trained on HumanML3D. Together, they enhance control over individual behavior while preserving coherent interaction dynamics.
In this thesis, we introduce a novel Diffusion Model incorporating a Transformer-based architecture. This model is conditioned using textual descriptions of both the motion interactions and the individual motions within these interactions. By focusing on the individual components of the interaction, our method achieves more precise conditioning in the generation of these specific motions. Concurrently, the textual descriptions of the overall interaction enable our model to effectively capture the interplay between individual motions.