Research map / Vision-language and multimodal / Human and embodied modeling
MotionChain: Conversational Motion Controllers via Multimodal Prompts
Biao Jiang, Xin Chen, Chi Zhang, Fukun Yin, Zhuoyuan Li, Gang Yu, Jiayuan Fan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decode concatenated motion tokens jointly instead of independent per-turn decoding (ablation) | MotionChain multi-turn composition · HumanML3D motion composition | MPJPE: 350.79 -> 108.77 | Improved |
| use joint token decoding instead of past-condition decoding (ablation) | MotionChain · HumanML3D motion composition | MPJPE: 232.46 -> 108.77 | Improved |
| joint token decoding versus independent decoding on diversity (ablation) | MotionChain · HumanML3D motion composition | Diversity: 6.47 -> 5.56 | Hurt |
| use linear projection instead of Q-former or perceiver vision tokenizer (ablation) | MotionChain · BEDLAM first-frame | MPJPE: 195.49 / 185.61 -> 144.37 | Improved |
| linear versus Q-former on last-frame condition (ablation) | MotionChain · BEDLAM last-frame | MPJPE: 134.73 -> 133.73 | No clear effect |
| do not share motion codes across quantizer layers (separate tokens per layer) (ablation) | MotionChain · HumanML3D text-to-motion | FID: 0.291 -> 0.248 | Improved |
| increase codebook dimension d from 128 to 1024 (ablation) | motion VQ-VAE · HumanML3D | FID: 1.698 -> 0.014 | Improved |
| increase residual quantizer layers from 2 to 8 or 16 (ablation) | motion VQ-VAE · HumanML3D | MPJPE: 79.7 -> 49.7 | Improved |
| use multimodal motion-aware model instead of text-only LLMs for motion reasoning | MotionChain · Motion reasoning | Bleu@4: 7.39 -> 19.19 | Improved |
| use MotionChain versus MotionGPT for motion prediction | MotionChain · AMASS motion prediction | FID: FID 0.905 -> 1.053; ADE 4.745 -> 4.388 | Mixed |
| use MotionChain versus T2M-GPT for text-to-motion FID | MotionChain · HumanML3D text-to-motion | FID: 0.116 -> 0.248 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- MotionGPT — MotionGPT: Human Motion as a Foreign Language (NeurIPS 2023) · better on 2 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Motion-Agent: A Conversational Framework for Human Motion Generation with LLMs (ICLR 2025 · similarity 0.92)
- The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion (CVPR 2025 · similarity 0.91)
- MotionGPT: Finetuned LLMs Are General-Purpose Motion Generators (AAAI 2024 · similarity 0.89)
- AvatarGPT: All-in-One Framework for Motion Understanding Planning Generation and Beyond (CVPR 2024 · similarity 0.88)
- ReMoGPT: Part-Level Retrieval-Augmented Motion-Language Models (AAAI 2025 · similarity 0.88)
- HSI-GPT: A General-Purpose Large Scene-Motion-Language Model for Human Scene Interaction (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
