TM-Edit: Text Guided Diffusion-Transformer Based Motion Editing With Temporal Soft Masks Learning

Abstract

Text-driven 3D human motion editing aims to modify an existing motion sequence following natural language instructions, which is a crucial task for character animation, virtual agents, and motion authoring. Recent diffusion-based methods have shown remarkable success in text-to-motion generation. Editing existing motions requires precise spatiotemporal control to localize modifications while preserving context. Current diffusion-based motion editing methods lack explicit fine-grained control over when and how strongly to edit. To address this, we propose TM-Edit, a text guided Diffusion-Transformer based motion editing framework which introduces learned temporal soft masks to provide explicit frame-wise editing guidance. The proposed model predicts an editing intensity mask to encode high-level intent from both the source motion and the text instruction. This mask is then used to modulate source motion features within a conditional diffusion process via an uncertainty-aware gating mechanism, ensuring robust training and inference. Additionally, a feature semantic alignment loss is employed by using a pre-trained motion retrieval model to enhance cross-modal consistency. Extensive experiments on the MotionFix benchmark dataset demonstrate that our approach achieves state-of-the-art performance. Code will be made publicly available.