COSMI: COmpositional Synthesis of Multi-object Interactions

Daniel Eskandar^*   Ilya A. Petrov^*   Gerard Pons-Moll
^*Equal contribution

COSMI composes multi-object human-object interactions from single-object captures. The COSMI dataset holds 222k sequences and 275 hours of interactions with up to five objects, and the COSMI method, a text-to-interaction diffusion transformer with weight-shared object slots, generates interactions with a variable number of objects.

The code, models, and the dataset pipeline will be released soon.

@article{eskandar2026cosmi,
  author  = {Eskandar, Daniel and Petrov, Ilya A. and Pons-Moll, Gerard},
  title   = {COSMI: Compositional Synthesis of Multi-Object Interactions},
  journal = {arXiv preprint arXiv:2610.03252},
  year    = {2026}
}