2503.06119

Total: 1

#1 Unlocking Pretrained LLMs for Motion-Related Multimodal Generation: A Fine-Tuning Approach to Unify Diffusion and Next-Token Prediction [PDF] [Copy] [Kimi] [REL]

Authors: Shinichi Tanaka, Zhao Wang, Yoichi Kato, Jun Ohya

In this paper, we propose a unified framework that leverages a single pretrained LLM for Motion-related Multimodal Generation, referred to as MoMug. MoMug integrates diffusion-based continuous motion generation with the model's inherent autoregressive discrete text prediction capabilities by fine-tuning a pretrained LLM. This enables seamless switching between continuous motion output and discrete text token prediction within a single model architecture, effectively combining the strengths of both diffusion- and LLM-based approaches. Experimental results show that, compared to the most recent LLM-based baseline, MoMug improves FID by 38% and mean accuracy across seven metrics by 16.61% on the text-to-motion task. Additionally, it improves mean accuracy across eight metrics by 8.44% on the text-to-motion task. To the best of our knowledge, this is the first approach to integrate diffusion- and LLM-based generation within a single model for motion-related multimodal tasks while maintaining low training costs. This establishes a foundation for future advancements in motion-related generation, paving the way for high-quality yet cost-efficient motion synthesis.

Subject: Machine Learning

Publish: 2025-03-08 08:16:16 UTC