Research map / Video / Video generation
SimMotionEdit: Text-Based Human Motion Editing with Motion Similarity Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SimMotionEdit, a Motion Diffusion Transformer trained jointly on diffusion-based text-conditioned motion editing and an auxiliary motion similarity prediction task, evaluated on MotionFix.
Problem
Text-based 3D human motion editing, where edited motions must align with both the source motion and the text instruction while keeping realistic fidelity; existing methods struggle with precise control and semantic alignment.
Approach
Adds an auxiliary task predicting frame-wise similarity between source and edited motion (quantized into classes, with sliding-window and min-max normalization and MotionSNR filtering) and uses a separate condition transformer to fuse source and text features before a diffusion transformer generates the edit.
Key results
- On the MotionFix test set, the method reaches 25.49% R@1 generated-to-target retrieval versus 14.51% for TMED, with M-score -3.210 vs -3.512.
- Batch-level (size 32) R@1 is 70.62% for the method vs 62.90% for TMED.
- Three-class classification auxiliary loss outperforms regression (batch R@1 70.62 vs 68.96) and more classes (5, 9) degrade results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SimMotionEdit instead of TMED | MotionFix text-based motion editing · MotionFix test | R@1 (batch): 62.90 -> 70.62 | Helped |
| Use SimMotionEdit instead of TMED | MotionFix text-based motion editing · MotionFix test | M-score: −3.512 -> −3.210 | Helped |
| Add the auxiliary motion similarity loss (ablation) | SimMotionEdit · MotionFix test | R@1 (batch): 67.21 -> 70.62 | Helped |
| Filter noisy motion similarity curves during training (ablation) | SimMotionEdit · MotionFix test | R@1 (batch): 66.88 -> 70.62 | Helped |
| Use enhanced text and motion features rather than raw features (ablation) | SimMotionEdit · MotionFix test | M-score: −3.422 -> −3.210 | Helped |
| Use enhanced motion features rather than no motion features (ablation) | SimMotionEdit with enhanced text features · MotionFix test | R@1 (batch): 66.46 -> 70.62 | Helped |
| Remove enhanced text features while retaining enhanced motion features (ablation) | SimMotionEdit · MotionFix test | R@1 (batch): 70.62 -> 71.04 | Mixed |
| Replace classification with regression for motion similarity prediction (ablation) | SimMotionEdit auxiliary task · MotionFix test | M-score: −3.210 -> −3.750 | Hurt |
| Increase the number of similarity classes from 3 to 5 (ablation) | SimMotionEdit auxiliary task · MotionFix test | R@1 (batch): 70.62 -> 67.08 | Hurt |
| Increase the number of similarity classes from 3 to 9 (ablation) | SimMotionEdit auxiliary task · MotionFix test | R@1 (batch): 70.62 -> 66.88 | Hurt |
Surprising findings
- Regression-based auxiliary loss on motion similarity reduces fidelity (M-score -3.750) compared to classification, suggesting overfitting on precise similarity curves.
Limitations
- Relies on the single MotionFix dataset of 6,730 triplets for evaluation and training.
- Requires filtering of low-MotionSNR samples with an empirically chosen threshold.
- Evaluation alignment relies on TMR retrieval and fidelity on MotionCritic, which are proxy metrics.
Experimental setup
Datasets: MotionFix
Benchmarks: MotionFix
Metrics: R@1 R@2 R@3 AvgR M-score
Baselines: MDM MDM-BP TMED Ours (w/o aux loss) Ours (w/o filtering)
Builds on: MotionFix MDM DDPM AdaLN-Zero TMR
Closest papers by meaning
- Dynamic Motion Blending for Versatile Motion Editing (CVPR 2025 · similarity 0.92)
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.90)
- SALAD: Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing (CVPR 2025 · similarity 0.89)
- Towards Detailed Text-to-Motion Synthesis via Basic-to-Advanced Hierarchical Diffusion Model (AAAI 2024 · similarity 0.88)
- Enhanced Fine-Grained Motion Diffusion for Text-Driven Human Motion Synthesis (AAAI 2024 · similarity 0.88)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.