Motion AutoStyling via Seperable Modelling of Posing and Dynamic LoRAs

Motion restyling is a key challenge in animation, aiming to transform a neutral motion sequence into one that embodies a specific style while maintaining the original semantics and execution fidelity. Although recent motion synthesis models, such as motion diffusion and controllable generation, can produce stylized motions from text prompts, they are typically computationally intensive, lack fine-grained control, scale poorly to new styles, and often introduce artifacts when applied to unseen motions, thereby limiting their effectiveness for style transfer. We present MoRAL, an interactive motion restyling framework based on Separable Modeling of Posing and Tempo LoRAs (Low-Rank Adaptation). Our approach addresses the high-dimensional challenge of spatio-temporal motion style modeling by decomposing it into two lightweight and tractable submodules. The first is a spatial stylization module that captures spatial motion styles through an exchangeable posing LoRA over a base automatic posing network, enabling users to intuitively manipulate and refine key poses while preserving stylistic consistency. The second is a temporal stylization module that restores the temporal dynamics between key poses across the trajectories of major control joints, such as the feet and pelvis, using a tempo LoRA. This dual-LoRA design simplifies and strengthens the restyling process, ensuring both controllability and temporal coherence. As a plug-and-play solution, MoRAL can be integrated into existing animation pipelines, offering a practical and flexible tool for creative animation.