ANT: Adaptive Neural Temporal-Aware Text-to-Motion Model

2025-06-03

AI Summary

This work addresses the temporal misalignment between static semantic conditioning and dynamic denoising in text-to-motion generation: coarse-grained structural semantics are required early for skeletal motion modeling, while fine-grained local semantics are needed later for precise textual detail alignment. To resolve this, we propose a biologically inspired, epigenetic-regulation–motivated *stage-wise semantic adaptation mechanism*. Specifically, we introduce (i) a novel *temporal-aware semantic granularity partitioning module* based on spectral analysis; (ii) a *dynamic Classifier-Free Guidance scheduling strategy*; and (iii) a *spatiotemporal-coordinated quantitative semantic reweighting method*. To our knowledge, this is the first framework to systematically integrate developmental biology principles—particularly epigenetic regulation—into diffusion-based motion generation. Evaluated on benchmarks including StableMoFusion, our method significantly improves text-motion semantic alignment accuracy and achieves state-of-the-art performance.

Technology Category

• Natural Language Processing: Generation
• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Agentic search
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

While diffusion models advance text-to-motion generation, their static semantic conditioning ignores temporal-frequency demands: early denoising requires structural semantics for motion foundations while later stages need localized details for text alignment. This mismatch mirrors biological morphogenesis where developmental phases demand distinct genetic programs. Inspired by epigenetic regulation governing morphological specialization, we propose **(ANT)**, an **A**daptive **N**eural **T**emporal-Aware architecture. ANT orchestrates semantic granularity through: **(i) Semantic Temporally Adaptive (STA) Module:** Automatically partitions denoising into low-frequency structural planning and high-frequency refinement via spectral analysis. **(ii) Dynamic Classifier-Free Guidance scheduling (DCFG):** Adaptively adjusts conditional to unconditional ratio enhancing efficiency while maintaining fidelity. **(iii) Temporal-semantic reweighting:** Quantitatively aligns text influence with phase requirements. Extensive experiments show that ANT can be applied to various baselines, significantly improving model performance, and achieving state-of-the-art semantic alignment on StableMoFusion.

Problem

• Mismatch between static semantic conditioning and temporal-frequency demands in text-to-motion generation
• Need for adaptive semantic granularity during different denoising stages
• Improving semantic alignment and efficiency in diffusion-based motion generation

Innovation

• Semantic Temporally Adaptive Module for denoising
• Dynamic Classifier-Free Guidance scheduling
• Temporal-semantic reweighting aligns text influence