# MotionFix: Text-Driven 3D Human Motion Editing

**Authors:** Nikos Athanasiou, Alpár Ceske, Markos Diomataris, Michael J. Black, Gül Varol  
**arXiv:** [2408.00712](https://arxiv.org/abs/2408.00712)  
**Date:** 2024-08-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.00712v1/

## Key points

- Authors focus on 3D motion editing based on textual descriptions
- Challenges addressed include scarcity of training data and accurate editing of source motion
- Methodology introduced for collecting dataset consisting of triplets: source motion, target motion, and edit text
- Conditional diffusion model named TMED trained on MotionFix dataset shows superior performance over baseline models
- New retrieval-based metrics introduced for evaluating motion editing
- Code and models to be made publicly available for future research in fine-grained motion generation

## Summary

, , , , 


In their paper "MotionFix: Text-Driven 3D Human Motion Editing," authors Nikos Athanasiou, Alpár Ceske, Markos Diomataris, Michael J. Black, and Gül Varol focus on the challenging task of 3D motion editing. Their goal is to generate an edited motion based on a textual description of the desired modification provided alongside a 3D human motion. The primary challenges they address include the scarcity of training data and the development of a model capable of accurately editing the source motion in accordance with the text description. To tackle these challenges, the authors introduce a methodology for semi-automatically collecting a dataset consisting of triplets: (i) a source motion, (ii) a target motion, and (iii) an edit text. This dataset creation process results in the establishment of the MotionFix dataset. By leveraging this curated dataset, they train a conditional diffusion model named TMED that takes both the source motion and edit text as input. In their study, various baseline models trained solely on text-motion pairs datasets are compared with their proposed model trained on triplets. The results demonstrate superior performance of their TMED model over these baselines. Additionally, new retrieval-based metrics for evaluating motion editing are introduced, leading to the establishment of a new benchmark using the evaluation set from MotionFix. The promising outcomes presented in this paper pave the way for further advancements in fine-grained motion generation research. The authors plan to make their code and models publicly available for future exploration and utilization by other researchers in this field.