Depth-Adaptive Injection for Stylized Text-to-Motion.

Abstract

Discrete text-to-motion generation models are powerful tools for synthesizing diverse human motions, yet endowing them with controllable, high-fidelity stylistic variations remains challenging. Direct retraining or fine-tuning these models often disrupts the fragile discrete token distributions learned by masked generative models, leading to optimization instability, style-content entanglement, and degraded motion quality or text alignment. To address these limitations, we propose the Modular Injection Framework (MIF), a framework for conditional adaptation of masked discrete text-to-motion models. For optimization stability, MIF employs a continuous residual injection mechanism to fuse conditioning signals within the hidden states of the pretrained backbone. Crucially, through systematic ablation, we uncover a depth-aware structural behavior: assigning injection to different transformer layers with depth-adaptive weights introduces a beneficial architectural bias. This design effectively aids to strike better trade-offs between style representation and motion quality, allowing the network to modulate high-frequency stylistic details while preserving low-frequency semantic content. Furthermore, to explore the generality of MIF, we adopt a distribution-aware adaptation strategy to extend MIF to novel motion adaptation and cross-modal music-conditioned dance synthesis. Based on target-domain reconstruction and codebook utilization statistics, the strategy determines which pretrained components remain fixed and which are trainable. Extensive experiments demonstrate our method's superior stylized motion quality and robust cross-domain adaptability. Notably, our method also supports temporal and cross-task composition.