LLM-Guided Dual-Branch Diffusion Model for Fine-Grained Motion Synthesis

Abstract

Given a complex text description, existing studies have indeed made significant progress in motion synthesis, which largely overlook the fine-grained action information from expressive texts. In this paper, we introduce LGDDM, a LLM-Guided Dual-branch Diffusion Model, which relies on the LLM’s ability to decompose text and provide fine-grained and interpretable guidance to generate motions. To ensure the coordination and coherence of target motions, we then design an attention-structured dual-branch fusion block, which adaptively balances high-fidelity and diverse motion synthesis. Furthermore, we develop a large-scale decomposed human motion dataset to justify the effectiveness of our approach. Extensive experiments prove that combined with long text description, decomposed action information is rather important and essential for fine-grained human motion synthesis.