Cross-Axis Feature Fusion with Joint-Wise Motion Difference Prediction for Text-Based 3D Human Motion Editing

Researchers have proposed novel architectures for text-based 3D human motion editing and generation, improving semantic alignment and overall fidelity of generated motion.

A new architecture for text-based 3D human motion editing integrates joint and time dimensions through cross-axis feature fusion, consisting of two axis-anchored transformers and a cross-axis fusion block^[1]. This approach aims to preserve the style and structure of a source motion while applying edits described in natural language. An auxiliary task is introduced to train the joint-anchored transformer to regress the Soft-DTW distance between source and target joint rotations. Meanwhile, researchers have also introduced T2LM, a continuous long-term generation framework for 3D human motion generation from multiple sentences, which can be trained without sequential data and outperforms prior long-term generation models^[2]. T2LM is competitive with state-of-the-art single-action generation models. The proposed architectures demonstrate significant improvements in semantic alignment with both text instructions and source motions, as well as overall fidelity of generated motion.