Möbius moves denoising inside the language model: no motion VAE, and it leads HumanML3D on generation and captioning

Unified motion-language models had been only partially unified: understanding was handled by a language model, while generation was delegated to a separate diffusion module operating over learned motion latents. In earlier unified designs, the language model first produced a shared multimodal representation and passed it as conditioning to an external diffusion model, so the denoising process itself stayed separate from the language model. Möbius puts clean motion, noisy motion and text into the same shared attention; on HumanML3D, the 0.6B variant outperforms diffusion-based, language-model-based and unified motion-language baselines on both text-to-motion generation and motion captioning.

Generation is no longer performed by an external diffusion module; noisy motion is progressively denoised directly inside the unified multimodal space, where it can keep interacting with clean motion and language semantics. Clean and noisy motion are represented as motion tokens, routed to role-specific normalization and linear projection parameters, and then processed by shared attention. Ablations on Qwen3-0.6B with HumanML3D and only the text-to-motion and motion-to-text objectives compare dense routing, modality routing and role routing; role routing keeps strong performance on both understanding and generation.

The model operates directly on frame-level raw motion, requiring neither a discrete tokenizer nor a pretrained motion VAE. Discrete tokenizers introduce quantization error, while VAE-based approaches compress motion to temporally downsampled latents and lose frame-level information. In the ablations, VQ encoding substantially weakens generation, VAE encoding stays competitive but yields weaker motion understanding than frame-level encoding, and frame-level modeling gives the strongest balance between the two capabilities.

Training runs in two stages: motion understanding first, then generation parameters initialized from the understanding parameters while the understanding parameters are frozen. Jointly optimizing understanding and generation is unstable, because the randomly initialized generation pathway can disrupt pretrained language representations and degrade understanding. In the ablations, joint optimization notably degrades understanding, restricting denoising to final-layer context representations weakens generation, and randomly initializing denoising parameters performs worse than initializing from the understanding parameters.

Experiments focus on controlled single-person motion settings, covering text-to-motion generation and motion captioning on HumanML3D plus editing and temporal completion under the MotionLab protocol. Multi-person interaction, more detailed hand and finger motion, retargeting across heterogeneous body structures, and deployment in embodied systems such as real-world robots are listed as future work.

Observed-frame preservation in temporal completion is task-dependent: HY-Motion has the lowest [P]-MPJPE for prediction and Kimodo for in-betweening, while Möbius leads only for backcasting and keyframe infilling. The two model sizes use different training data, so their comparison does not isolate model scale. Adding multi-task training lowers the 0.6B variant's text-to-motion R@1 by 1.53 percentage points and raises FID from 6.1291 to 6.6746.