DistMovGen: A Knowledge Distillation-Enhanced Framework for Real-Time Virtual Character Motion Generation
Authors: Shi, X., Chang, J., Charles, F., Guo, S.
Conference: ICXR 2025
Dates: 01/11/2025
Journal: Extended Reality: ICXR 2025
Publication Date: 01/06/2026
Volume: 16428
Pages: 56-71
Publisher: Springer
Place of Publication: Singapore
eISSN: 1611-3349
ISBN: 978-981-95-7195-6
ISSN: 1611-3349
DOI: 10.1007/978-981-95-7195-6_4
Abstract:In natural dialogues and interactive scenarios, real-time motion generation for virtual characters faces challenges: traditional systems relying on precomputed motion libraries exhibit latency, rigid body language, and inadequate multimodal synchronization, limiting emotional engagement and immersion. We propose a hierarchical acceleration framework integrating multimodal neural networks to address these issues. The framework employs knowledge distillation to transfer motion generation expertise from a teacher model into a lightweight diffusion backbone. Our approach enables 8-step full-body motion inference and 1-step facial synthesis via diffusion-GAN fine-tuning. The distillation loss combines kinematic feature matching and output distribution alignment. Building upon this acceleration framework, we implement a dual-stream network architecture that decouples facial expression and motion generation for cross-modal coupling, directly addressing the multimodal synchronization challenges identified earlier. Experimental results demonstrate that our framework achieves real-time performance with high naturalness and performs better than existing emotional expressiveness and multimodal synchronization methods. The system supports applications including virtual social interactions, gamified entertainment, and educational training, contributing to advancing virtual character technology for enhanced human-computer interactions.
https://link.springer.com/chapter/10.1007/978-981-95-7195-6_4
Source: Manual