FLAME | Jihoon Kim

Summary

FLAME is a diffusion-based model that synthesizes high-fidelity human motion from free-form text. It can edit selected frames or joints without fine-tuning, using a Transformer designed for variable-length motion and language conditioning. Experiments show state-of-the-art generation on HumanML3D, BABEL, and KIT, with editing also extending to motion prediction and in-betweening. (Kim et al., 2023)

Contribution

During my KakaoBrain research internship, I contributed across FLAME’s full research lifecycle and published the work as first author at AAAI 2023, where it received an oral presentation.

Resources

Project Page · HTML · PDF · arXiv · GitHub · Video

References

2023

AAAI

FLAME: Free-form Language-based Motion Synthesis & Editing

Jihoon Kim, Jiseob Kim, and Sungjoon Choi

In Proceedings of the AAAI Conference on Artificial Intelligence; oral presentation , 2023

DOI arXiv Bib HTML PDF Video Code Website

@inproceedings{kim2023flame,
  title = {{FLAME}: Free-form Language-based Motion Synthesis \& Editing},
  author = {Kim, Jihoon and Kim, Jiseob and Choi, Sungjoon},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume = {37},
  number = {7},
  pages = {8255--8263},
  year = {2023},
  doi = {10.1609/aaai.v37i7.25996},
}