Hong Kong University of Science and Technology, Tencent Games Motus and MIT (Jiazhi Shu, Chengfeng Zhao et al.) · Motion-language model

Möbius

Möbius is a unified motion-language model from researchers at HKUST, Tencent Games Motus and MIT. It handles text-to-motion generation, motion-to-text captioning, motion completion (with or without a text prompt) and motion editing in one model. A single backbone connects text and motion through shared multimodal attention. On the HumanML3D test set with the MotionStreamer evaluator, the 0.6B model trained on HumanML3D reports R@1 of 0.8209 and FID of 6.1291, against 0.6303 and 12.2110 for MotionStreamer. The 1.7B model trained on a large-scale dataset reports R@1 of 0.7968 and FID of 10.6642, compared with 0.7899 and 13.5194 for HY-Motion (1B). On motion captioning under the TM2T protocol, the 1.7B model scores a CIDEr of 0.5836 and the 0.6B model 0.5754, against 0.4077 for MotionGPT3. The paper is arXiv:2611.05279 and a GitHub link is listed on the project page; the text does not state a licence or release date.