Research map / Vision-language and multimodal / Human and embodied modeling
MotionGPT3: Human Motion as a Second Modality
Bingfan Zhu, Biao Jiang, Sunyi Wang, SHIXIANG TANG, Tao Chen, Linjie Luo, Youyi Zheng, Xin Chen
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 |
|---|---|---|---|
| use continuous VAE latents instead of VQ tokens (ablation) | MotionGPT3 dual-stream · HumanML3D T2M (199 epochs) | R-Precision Top-1: 0.525 vs. 0.258 on R@1 | Improved |
| use continuous VAE latents instead of VQ tokens (ablation) | MotionGPT3 dual-stream · HumanML3D T2M | MMDist: 2.667 vs. 5.364 on MMDist | Improved |
| use continuous VAE latents instead of VQ tokens (ablation) | MotionGPT3 dual-stream · HumanML3D M2T (99 epochs) | R-Precision Top-3: 0.859 vs. 0.765 on R@3 | Improved |
| train VQ variant longer (399 epochs) (ablation) | Bimodal+VQ · HumanML3D T2M | FID: FID to 0.454, R@1 0.300 and MMDist 4.937 | Mixed |
| train VQ variant longer (199 epochs) (ablation) | Bimodal+VQ · HumanML3D M2T | R-Precision Top-3: R@3 dropping by 0.13 and BLEU@1 by 5.464 | Hurt |
| use dual-stream (bimodal) backbone instead of single-stream (ablation) | MotionGPT3 with VAE · HumanML3D M2T | R-Precision Top-3: 0.426 -> 0.853 | Improved |
| use dual-stream (bimodal) backbone instead of single-stream (ablation) | MotionGPT3 with VQ · HumanML3D T2M | R-Precision Top-3: 0.435 -> 0.532 | Improved |
| use dual-stream (bimodal) backbone instead of single-stream (ablation) | MotionGPT3 with VAE · HumanML3D T2M | FID: 0.489 -> 0.239 | Improved |
| add Stage II cross-modal alignment after Stage I (ablation) | MotionGPT3 · HumanML3D T2M | FID: 0.239 -> 0.215 | Improved |
| add Stage III joint fine-tuning (ablation) | MotionGPT3 · HumanML3D M2T | Bleu@4: 18.328 -> 19.412 | Improved |
| add Stage I text-to-motion pretraining (ablation) | MotionGPT3 · HumanML3D T2M | FID: 0.325 -> 0.208 | Improved |
| increase text backbone from GPT-2 small to medium (ablation) | MotionGPT3 · HumanML3D T2M (CFG 1.0) | R-Precision Top-1: 0.534 -> 0.558 R@1, FID 0.222 -> 0.235 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MotionGPT — MotionGPT: Human Motion as a Foreign Language (NeurIPS 2023) · better on 1 of 1 shared result
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
- M$^3$GPT: An Advanced Multimodal, Multitask Framework for Motion Comprehension and Generation (NeurIPS 2024 · similarity 0.92)
- HSI-GPT: A General-Purpose Large Scene-Motion-Language Model for Human Scene Interaction (CVPR 2025 · similarity 0.90)
- The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion (CVPR 2025 · similarity 0.90)
- ReMoGPT: Part-Level Retrieval-Augmented Motion-Language Models (AAAI 2025 · similarity 0.89)
- Motion-Agent: A Conversational Framework for Human Motion Generation with LLMs (ICLR 2025 · similarity 0.89)
- MG-MotionLLM: A Unified Framework for Motion Comprehension and Generation across Multiple Granularities (CVPR 2025 · similarity 0.89)
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.
