arXiv cs.AI· Yiheng Li, Zhuo Li, Ruibing Hou, Yingjie Chen, Hong Chang, Hao Liu, Shiguang Shan·· 14 小时前AI 评分29
AnyMo：用掩码建模扩展任意模态条件动作生成
AnyMo: Scaling Any-Modality Conditional Motion Generation with Masked Modeling
AI 导读
AnyMo 是一个统一多模态框架，结合基于 Residual FSQ 的动作 tokenizer 与可扩展掩码建模 Transformer，支持任意模态组合下的高质量动作合成。研究团队同时构建了 OmniHuMo 数据集，包含超 5,000 小时动作、320 万条序列，并精确对齐文本、语音、音乐和轨迹等多模态标注。实验显示 AnyMo 在实现高保真合成的同时，可对空间与风格属性进行灵活控制。
正文 · 原文
Abstract:Conditional human motion generation remains a fundamental challenge in computer vision and robotics. Despite significant progress, current methods are often constrained by fixed modality configurations and task-specific architectures, leaving cross-modal interactions and the scaling laws of multimodal-conditioned synthesis largely underexplored. A key bottleneck is the scarcity of large-scale modality-aligned motion data, limiting generalization across diverse control signals. In this work, we introduce OmniHuMo, a large-scale, high-quality dataset comprising over 5,000 hours of motion and 3.2 million sequences with precisely aligned multimodal annotations (e.g., text, speech, music, and trajectory). Leveraging OmniHuMo, we propose AnyMo, a unified multimodal framework combining a Residual FSQ-based motion tokenizer with a scalable masked modeling transformer, enabling high-quality motion synthesis under arbitrary modality combinations. Extensive experiments show that AnyMo achieves high-fidelity synthesis while offering flexible control over both spatial and stylistic attributes.
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2605.29488 [cs.CV] |
| (or arXiv:2605.29488v3 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2605.29488 arXiv-issued DOI via DataCite |
Submission history
From: YiHeng Li [view email]
[v1]
Thu, 28 May 2026 07:15:19 UTC (32,642 KB)
[v2]
Sun, 31 May 2026 04:55:43 UTC (32,642 KB)
[v3]
Thu, 29 Oct 2026 06:08:58 UTC (32,642 KB)
来源：arXiv cs.AI · arxiv.org