Title: AnyMo: Scaling Any-Modality Conditional Motion Generation with Masked Modeling

Abstract:

Generating human motion conditioned on specific inputs remains a core difficulty within the fields of robotics and computer vision. While the domain has seen considerable advancements, existing approaches are frequently hampered by rigid modality setups and specialized architectures. Consequently, the dynamics of cross-modal interactions and the scaling laws governing multimodal-conditioned synthesis have remained largely uninvestigated. A primary obstacle to progress is the lack of extensive, modality-aligned motion datasets, which restricts the ability of models to generalize across varied control signals. To address these issues, this study presents OmniHuMo, a comprehensive, high-quality dataset featuring over 5,000 hours of motion data and 3.2 million sequences, all equipped with precise multimodal annotations including text, speech, music, and trajectory data. Building on OmniHuMo, we introduce AnyMo, a cohesive multimodal framework that integrates a scalable masked modeling transformer with a Residual FSQ-based motion tokenizer. This architecture facilitates the generation of high-quality motion under diverse combinations of modalities. Our extensive experimental results demonstrate that AnyMo delivers high-fidelity synthesis while providing adaptable control over both stylistic elements and spatial characteristics.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC