Spatial Temporal Synergy: Balancing Change and Invariance in Text Driven 3D Human Motion Editing

Shaohui Lin; Zhenwu Shi; Jingyu Gong; Jiao Xie; Yu Zhou; Baochang Zhang; Lizhuang Ma; Chia-Wen Lin

cs.CVAug 17, 2026

Spatial Temporal Synergy: Balancing Change and Invariance in Text Driven 3D Human Motion Editing

Authors: Shaohui Lin, Zhenwu Shi, Jingyu Gong, Jiao Xie, Yu Zhou, Baochang Zhang, Lizhuang Ma, Chia-Wen Lin

Organizations: School of Computer Science and Technology, East China Normal University, Shanghai 200062, China · Shanghai Institute of Artificial Intelligence for Education, East China Normal University, Shanghai 200062, China · School of Statistics, East China Normal University, Shanghai 200062, China · School of Artificial Intelligence, Beihang University, Beijing 100191, China ·

Abstract

Text-driven human motion editing aims to modify existing motion sequences according to natural language instructions while maintaining the structural consistency of the original motion. Existing diffusion-based approaches struggle to balance text-responsive "change" and inertial "invariance". They often rely on coarse spatial constraints and rigid uniform time assumptions, leading to spatial motion distortions and the destruction of intrinsic physical rhythms during variable-length editing. To handle these challenges, we propose Change and Invariance Motion Editing (CIME), a unified framework that comprehensively decouples change and invariance into spatial pose and temporal rhythm dimensions. For spatial poses, our method integrates an omni-supervised positive-negative learning mechanism comprising hierarchical retrospective feature supervision, subtle motion preservation, and triplet-based semantic alignment. For temporal rhythms, we introduce the Riemannian Non-uniform Integral Manifold Mapping (RNIMM) module, which achieves high-fidelity reproduction of physical beats in the edited text via kinematics-aware non-uniform timestamps. Extensive experiments on the MotionFix and STANCE Adjustment datasets demonstrate that CIME achieves state-of-the-art performance in editing alignment and structural fidelity, validating the effectiveness of our unified architecture. Our source codes and models have been released at: github.com/ZhenwuShi/CIME.git

Figures & tables

Fig. 1: Display information on fast motion and slow motion.

Fig. 2: CIME is a unified framework for text-driven 3D human motion editing. Given a source motion and a natural language instruction, CIME edits the source motion to produce the desired target motion while balancing change and invariance across both spatial pose and temporal rhythm dimensions.

Fig. 3: Overview of CIME : First, the input text and source motion are projected into a shared semantic space via a CLIP encoder and aggregated by a Fusion Transformer. Next, the RNIMM module aligns these text-infused features with the target temporal dimension and injects them into the noise target. Finally, the composite features are processed by a Diffusion Transformer. This backbone is co-optimized by three auxiliary supervision mechanisms to achieve an optimal balance between precise editing and structure preservation, ultimately yielding the target motion.

Fig. 4: Illustration of arc length on Riemannian manifolds and encoding on non-uniform Riemannian manifolds. The orange boxes indicate regions of intense motion, and the red bars reflect higher instantaneous kinematic intensity.

Method	Venue	Generated-to-Target (Batch)				Generated-to-Target (Test Set)
Method	Venue	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$
Ground Truth	–	100	100	100	1.00	64.36	88.75	95.56	1.74
MDM [ 7 ]	ICLR’23	4.03	7.56	10.48	15.55	0.10	0.10	0.10	–
MDM-BP [ 8 ]	SIGGRAPH Asia’24	39.10	50.09	54.84	6.46	8.69	14.71	18.36	180.99
TMED [ 8 ]	SIGGRAPH Asia’24	62.90	76.51	83.06	2.71	14.51	21.72	28.73	56.63
MotionReFit [ 66 ]	CVPR’25	66.33	80.05	84.98	2.64	–	–	–	–

TABLE I: Comparison of retrieval performance (Generated-to-Target) on MotionFix dataset [ 8 ] . In the following tables and figures,

R@k

indicates recall at top-

k

(

\uparrow

higher is better,

\downarrow

lower is better). * : no explicit text conditioning in the DiT stage.

Method	Venue	Generated-to-Target (Batch)				Generated-to-Target (Test Set)
Method	Venue	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$
TMED [ 8 ]	SIGGRAPH Asia’24	29.69	44.01	52.08	6.97	11.22	17.86	25.51	35.56
TMED w/ MCM [ 8 , 66 ]	CVPR’25	32.22	45.03	54.83	6.56	-	-	-	-
SimMotionEdit * [ 9 ]	CVPR’25	36.46	48.96	57.81	5.71	12.76	23.98	29.59	29.05
MotionReFit [ 66 ]	CVPR’25	42.45	56.25	62.76	5.12	-	-	-	-
OmniME [ 14 ]	CVPR’26	43.75	56.25	66.15	4.66	22.45	31.63	36.22	22.77

TABLE II: Comparison of retrieval performance (Generated-to-Target) on STANCE Adjustment dataset [ 66 ] .

#	Components				Generated-to-Target (Batch)				Generated-to-Target (Test Set)
#	$\mathcal{L}_{\text{retro}}$	$\mathcal{L}_{\text{triplet}}$	$\mathcal{L}_{\text{pres}}$	$\mathcal{M}_{\text{RNIMM}}$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$
1					71.04	83.96	89.58	2.22	26.88	44.27	51.98	20.88
2	✓				72.71	85.62	88.75	2.09	30.63	45.26	54.35	18.62
3		✓			73.54	85.00	88.54	2.04	28.26	42.49	51.78	17.99
4			✓		75.62	87.50	90.42	1.88	30.24	45.65	56.92	15.58
5				✓	73.54	87.92	91.04	1.88	29.64	44.27	55.34	15.71

TABLE III: Ablation study on the MotionFix dataset [ 8 ] . Comparison of retrieval performance (Generated-to-Target).

Method	Generated-to-Target (Batch)				Generated-to-Target (Test Set)
Method	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$
Baseline	71.04	83.96	89.58	2.22	26.88	44.27	51.98	20.88
infoNCE [ 71 ]	68.33	83.12	86.04	2.38	25.10	38.14	48.62	24.14
K-means	70.21	82.50	87.92	2.35	27.47	42.29	50.79	21.06
Triplet(Ours)	73.54	85.00	88.54	2.04	28.26	42.49	51.78	17.99

TABLE IV: Analysis of Negative Text Sampling Strategies. The evaluation is conducted on the MotionFix dataset [ 8 ] .

Method	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$
Curriculum Scheduling	72.92	84.79	90.21	2.07
Ours	73.54	85.00	88.54	2.04

TABLE V: Comparison with Curriculum Scheduling based on Generated-to-Target (Batch).

Method	Venue	Generated-to-Target (Batch)				Generated-to-Target (Test Set)
Method	Venue	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$
SimMotionEdit* † [ 9 ]	CVPR’25	21.43	33.04	42.41	8.80	7.42	11.35	14.41	58.67
OmniME [ 14 ]	CVPR’26	22.40	35.94	47.40	7.44	6.63	10.71	15.82	41.23
CIME (Ours)	—	29.69	43.23	54.69	6.36	10.20	18.37	23.47	34.74

TABLE VI: Trained on MotionFix and tested on STANCE adjustment.

{{\dagger}}

indicates our re-implementation.

Fig. 7: Subjective evaluation: CIME vs SimMotionEdit [ 9 ] . The charts report average human ratings (scale 1–10) across four perceptual metrics: Semantic Alignment, Motion Preservation, Transition Smoothness, and Overall Naturalness. Results for the MotionFix [ 8 ] and STANCE Adjustment [ 66 ] datasets are displayed on the left and right , respectively. Our proposed CIME consistently outperforms the baseline across all metrics on both datasets.

Motion Diffusion Models

Motion diffusion models generate, edit, reconstruct, and refine articulated movement from text, audio, video, and other conditions.

See all 22 papers

Omni-Supervised Motion Editing: Balancing Change and Invariance through Positive-Negative Learning
May 29, 2026Zhenwu Shi, Jingyu Gong, Peiwei Wang +7
Shanghai Institute of Artificial Intelligence for Education, East China Normal University, China · School of Computer Science and Technology, East China Normal University, China · Shanghai Key Laboratory of Computer Software Evaluating and Testing, China +4
UniMoFlow: Grounding Instruction-Driven 3D Human Motion Editing in Generation
Aug 10, 2026Yilei Hua, Beibei Jing, Ce Zheng +3Flow Matching Text-to-Motion Generation
School of Computer Science, Huazhong University of Science and Technology · School of Software Technology, Zhejiang University
SuperMotion: Source-Preserving Denoising for Text-Driven Human Motion Editing
Oct 1, 2026Fa-Ting Hong, Peter WonkaHuman Motion Generation Diffusion Models for Motion Generation
King Abdullah University of Science and Technology

May 29, 2026J/K move · Enter open · S save

Omni-Supervised Motion Editing: Balancing Change and Invariance through Positive-Negative Learning

Zhenwu Shi, Jingyu Gong, Peiwei Wang, Xingzan Wang+6

Shanghai Institute of Artificial Intelligence for Education, East China Normal University, China · School of Computer Science and Technology, East China Normal University, China · Shanghai Key Laboratory of Computer Software Evaluating and Testing, China+4

Open paper AskPDF arXiv

Text-based human motion editing aims to modify existing motion sequences according to natural language instructions while maintaining the consistency of the original motion. Existing diffusion-based approaches often rely on heuristic similarity cues or coarse global conditioning, leading to motion distortion and suboptimal semantic alignment. The key challenge lies in balancing change (i.e. precisely editing target regions) and invariance (i.e. preserving unedited parts). To handle such challenge, we propose an Omni-Supervised Positive-Negative Learning framework, named OmniME. Our method integrates three complementary components: (1) retrospective feature supervision that enforces coarse-to-fine consistency across transformer layers,(2) motion preservation mechanism that focuses on subtle variations according to the source-target similarity, and (3) triplet-based semantic alignment that strengthens text-motion correspondence. Together, these components form a unified supervision paradigm that balances change and invariance. Extensive experiments on the MotionFix and STANCE Adjustment datasets demonstrate that OmniME achieves state-of-the-art performance in editing alignment, validating the effectiveness of our unified learning framework. Our source codes and models have been released at: https://github.com/rocket-ycyer/OmniME.git

Abstract

Motion Diffusion Models

Explore similar work

Omni-Supervised Motion Editing: Balancing Change and Invariance through Positive-Negative Learning

UniMoFlow: Grounding Instruction-Driven 3D Human Motion Editing in Generation

SuperMotion: Source-Preserving Denoising for Text-Driven Human Motion Editing