TAME:Topology-Aware Text-Driven Motion Editing across Heterogeneous Humanoid Skeletons

Qichen Zheng; Siyuan Yang; Chong Wang; Jun Liu; Shijian Lu; Alex Kot; Kwok-Yan Lam

cs.CVOct 3, 2026

TAME:Topology-Aware Text-Driven Motion Editing across Heterogeneous Humanoid Skeletons

Authors: Qichen Zheng, Siyuan Yang, Chong Wang, Jun Liu, Shijian Lu, Alex Kot, Kwok-Yan Lam

Organizations: Nanyang Technological University · KTH Royal Institute of Technology · Lancaster University

Abstract

Text-driven motion editing modifies an existing motion sequence according to a text instruction while preserving the content of the source motion. Existing methods are typically built for a single, fixed skeletal topology, which limits their use in animation pipelines where characters differ in joint count and skeletal hierarchy. We present Topology-Aware Motion Editor (TAME), a flow-matching transformer that edits motions on humanoid skeletons of varying topology. TAME represents motion as per-joint, per-frame tokens and models interactions among joints, across frames, and with the text instruction through skeletal, temporal, and text cross-attention layers. To make the skeletal attention follow each character's hierarchy, TAME replaces full joint attention with Topology-Constrained Skeletal Propagation (TCSP), which restricts attention to one-hop kinematic neighbors in the skeleton's adjacency matrix. We further introduce Edit-Focused Representation Alignment (EFRA), a self-distilled representation alignment strategy that aligns student features with cleaner EMA-teacher features exclusively on edit-relevant joint-time tokens, making edits faithful to the instruction. To make this setting trainable and comparable, we construct TopoMotionFix, a multi-topology extension of MotionFix with seen- and unseen-topology evaluation protocols. TAME outperforms previous methods in edit alignment and source preservation on MotionFix and reliably edits motions on unseen skeletons in TopoMotionFix.

Figures & tables

Dataset	Protocol	Method	Generated-to-Target (Batch-32)				Generated-to-Source (Batch-32)
Dataset	Protocol	Method	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$
		GT	100.00	100.00	100.00	1.00	74.01	84.52	89.91	2.03
MotionFix	Seen	MDM ( Tevet et al., 2023 )	4.03	7.56	10.48	15.55	2.62	6.15	9.38	15.88
MotionFix	Seen	MDM-BP ( Athanasiou et al., 2024 )	39.10	50.09	54.84	6.46	61.28	69.55	73.99	4.21
MotionFix	Seen	TMED ( Athanasiou et al., 2024 )	64.11	76.51	82.56	2.72	73.19	85.08	91.03	1.92
MotionFix	Seen	MotionReFit ( Jiang et al., 2025b )	66.33	80.05	84.98	2.64	83.47	90.42	92.84	1.73

Table 1: Quantitative comparison.

\uparrow

indicates higher values are better,

\downarrow

indicates lower values are better; bold indicates best; dash indicates unavailable.

				Generated-to-Target (Batch-32)				Generated-to-Source (Batch-32)
	PH	Block	Edit Mask	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$
TAME w/o TCSP , w/o EFRA	—	—	—	74.80	88.00	92.64	1.80	84.38	92.64	95.16	1.52
TAME w/o EFRA	—	—	—	74.70	87.50	90.62	1.85	84.17	92.34	95.06	1.51
(a) Projection-head dimension	256	4/7	✓	76.11	89.11	93.45	1.66	84.27	92.84	94.86	1.53
	512	4/7	✓	77.32	88.31	92.74	1.74	84.48	92.44	94.76	1.59
	1024	4/7	✓	78.75	91.67	94.79	1.55	86.04	94.58	97.50	1.36

Table 2: Ablations on MotionFix. The first two rows ablate TCSP ( TAME w/o TCSP , w/o EFRA vs. TAME w/o EFRA ); blocks (a)-(c) keep TCSP and EFRA enabled and vary one hyperparameter: PH (projection-head dimension), Block (student/teacher layer pair

(l_{s},l_{t})

), Edit Mask (✓/

\times

: with/without the DTW edit mask). Bold marks our default.

\uparrow

\downarrow

indicates higher/lower is better.

Figure 1: Qualitative comparison on the seen-topology setting. Columns are edit prompts; rows are source, ground truth, TMED ( Athanasiou et al., 2024 ) , SimMotionEdit ( Li et al., 2025 ) , OmniME ( Shi et al., 2026 ) , and our TAME . Each cell shows 8 keyframes (light to dark in time).

Figure 2: Qualitative comparison on the unseen-topology setting, evaluating TopoMotionFix -trained models on Mixamo skeletons. Columns are edit prompts; rows are source, ground truth, AnyTop ( Gat et al., 2025 ) , and our TAME . Each cell shows 6 keyframes (light to dark in time).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Protocol	Method	Generated-to-Target (Test set)				Generated-to-Source (Test set)
Dataset	Protocol	Method	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	AvgR $\downarrow$
		GT	64.36	88.75	95.56	1.74	20.83	33.66	40.47	33.13
		GT (retargeted)	64.34	88.73	95.55	1.73	20.82	33.65	40.47	33.12
MotionFix	Seen	MDM ( Tevet et al., 2023 )	0.00	0.30	0.49	476.65	0.10	0.10	0.10	486.26
MotionFix	Seen	MDM-BP ( Athanasiou et al., 2024 )	8.69	14.71	18.36	180.99	30.21	36.82	40.47	106.05
MotionFix	Seen	TMED ( Athanasiou et al., 2024 )	14.51	21.72	28.73	56.63	22.41	34.45	40.57	31.42

Appendix

Table 3: Quantitative comparison under the test-set retrieval setting of the official SimMotionEdit evaluation implementation, with the same metrics as Table 1 . Baseline numbers are taken from the respective papers; dash indicates unavailable; bold indicates best within each protocol.

Protocol	Method	Alignment $\uparrow$	Plausibility $\uparrow$
Seen	GT	2.53 $\pm$ 0.66	2.61 $\pm$ 0.69
	SimMotionEdit ( Li et al., 2025 )	2.08 $\pm$ 0.81	2.28 $\pm$ 0.76
	TAME (ours)	2.31 $\pm$ 0.74	2.46 $\pm$ 0.71
Unseen	GT	2.53 $\pm$ 0.66	2.61 $\pm$ 0.69
	AnyTop ( Gat et al., 2025 )	1.73 $\pm$ 0.84	1.98 $\pm$ 0.81
	TAME (ours)	1.96 $\pm$ 0.79	2.14 $\pm$ 0.77

Appendix

Table 4: Perceptual evaluation. Mean

\pm

standard deviation on the

3

-point Likert scale across

25

participants and

10

sample sets per participant (

250

ratings per cell). Higher is better. Each block compares ground truth (GT), our TAME , and one baseline: SimMotionEdit for the seen-topology protocol and AnyTop for the unseen-topology protocol.

Human Motion Generation

Human motion generation creates articulated movement from language, audio, video, spatial constraints, or interaction goals.

See all 292 papers

Spatial Temporal Synergy: Balancing Change and Invariance in Text Driven 3D Human Motion Editing
Aug 17, 2026Shaohui Lin, Zhenwu Shi, Jingyu Gong +5Motion Diffusion Models
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 +5
TopX: Text-to-Motion for Unseen Skeletons from the Skeleton File Alone
Oct 25, 2026Tengjiao Sun, Pengcheng Fang, Xiaoyu Zhan +3Text-to-Motion Generation Human Motion Generation
University of Southampton · MOGO AI · Nanjing University
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

Aug 17, 2026J/K move · Enter open · S save

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

Shaohui Lin, Zhenwu Shi, Jingyu Gong, Jiao Xie+4

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+5

Motion Diffusion Models

Open paper AskPDF arXiv

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

Abstract

Human Motion Generation

Explore similar work

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

TopX: Text-to-Motion for Unseen Skeletons from the Skeleton File Alone

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