Cross-Axis Feature Fusion with Joint-Wise Motion Difference Prediction for Text-Based 3D Human Motion Editing
Researchers have developed a new method for editing 3D human motion based on text instructions, aiming to preserve the original motion's style and structure while incorporating the specified changes. Their approach utilizes a novel architecture with two transformers that analyze joint and time dimensions separately, integrating these features through a cross-axis fusion block. An auxiliary training task further enhances the model's ability to identify which joints require modification and which should remain unchanged, leading to state-of-the-art results in semantic alignment and motion fidelity on the MotionFix dataset. AI
IMPACT This research advances AI capabilities in motion editing, potentially impacting animation, gaming, and virtual reality by enabling more intuitive and precise control over character movements.