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

TL;DR

ANT introduces an adaptive neural architecture for text-to-motion generation that dynamically adjusts semantic granularity and guidance to improve temporal coherence and semantic alignment, inspired by biological morphogenesis.

Contribution

The paper proposes ANT, a novel adaptive neural model with spectral analysis-based semantic partitioning and dynamic guidance scheduling for improved text-to-motion synthesis.

Findings

Achieves state-of-the-art semantic alignment on StableMoFusion.

Significantly improves performance across various baseline models.

Demonstrates effective temporal and semantic control in motion generation.

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. 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.