Motion AI models typically handle understanding or generation separately, using a language model for one and a separate diffusion module for the other. This disconnection limits performance. Möbius integrates both by having clean motion for understanding and noisy motion under generation attend to text in a shared representation. Special parameters manage each role. It works directly on raw motion frames, needing no separate encoder. On HumanML3D, Möbius achieves top results for both text-to-motion and motion captioning. Scaling to 1.7B parameters and 1,700 hours of data shows the framework is effective and scalable. This matters because it proves that understanding and generation can be unified simply, improving efficiency for applications like animation and robotics.
Motion model understands and generates in one shared space
Möbius operates directly on raw motion without discrete tokenization, outperforming prior unified models on both text-to-motion generation and motion captioning.
Chinese Tech
Jiazhi Shu · Chengfeng Zhao · Jinyi Wang · Yubo Zhao · Zhiyang Dou · Ronghui Li · +4 more
The Hong Kong University of Science and Technology · Tencent Games Motus · Massachusetts Institute of Technology
Research Digest··1 min read
In one line
Möbius unifies motion understanding and generation by diffusing noisy motion directly in the same multimodal space as text.
Before you read this
Three ideas this paper builds onunified multimodal understanding and generationA single end-to-end model processes and generates multiple data types (e.g., images, text) with equal competence.27 papers →role decouplingSeparating distinct functions such as planning, reasoning, and execution into separate components improves efficiency, quality, and reliability.45 papers →diffusion modelA type of generative model that learns to reverse a gradual noising process to create data.170 papers →
Read it as
The authors introduce Möbius, a unified motion-language model that processes clean motion, noisy motion, and text in shared attention with role-specific parameters.
The authors propose Möbius, which brings the entire diffusion process into the same multimodal representation space used for motion understanding.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.