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

Fa-Ting Hong; Peter Wonka

cs.CVOct 1, 2026

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

Authors: Fa-Ting Hong, Peter Wonka

Organizations: King Abdullah University of Science and Technology

Abstract

Text-driven human motion editing aims to realize a requested change while preserving compatible source content. Existing diffusion editors rely largely on learned conditioning for preservation of the unedited part, yet their outputs can lose temporal detail as denoising proceeds. We propose the \textbf{Source-Preserving Denoising framework (SuperMotion)}, which explicitly reuses the source at each reverse step for source preservation. We first align the source motion to the output timeline and predict a preservation gate that controls reuse across frames and feature dimensions. A clean-space source anchor then utilizes the learned preservation gate to blend the predicted clean motion with the aligned source and passes the corrected estimate directly to the sampling posterior. Because the aligned source is a realized motion rather than a regression output, the anchor injects sample-level temporal detail that a reconstruction-trained denoiser tends to smooth away. To learn effective source reuse, we supervise the anchored estimate against the editing target and match its second temporal differences through a temporal high-frequency loss. These objectives require no explicit edit masks. Extensive experiments show that SuperMotion improves editing accuracy, reaching 33.20% full-pool R@1 on MotionFix, while reducing temporal-detail attenuation and preserving motion dynamics as it realizes the requested changes. Ablations confirm that the learned preservation gate is responsible for the gain and that it reuses the source to retain the unedited content properly.

Figures & tables

Figure 1: Motion editing requires both change and preservation. Given a source motion (gray), the edit should realize the requested modification while retaining compatible source content. Our outputs (salmon) and reference targets (turquoise) illustrate three instructions, “bob up and down,” “straighten knees,” and “take a long sidestep to the right.” Each strip shows successive poses.

Figure 2: Detail attenuation during denoising. Median generated-to-target second-order difference magnitude at selected timesteps, with denoising proceeding from left to right. The plotted baselines finish below the target reference of one, while the shown SuperMotion run ends at

1.035\times

. Shading marks a fixed reference band from

0.9

1.1

. Agreement in this statistic alone does not establish a correct edit.

Figure 3: Overall architecture. A Skip DiT takes the noised motion $x_{t}$ , a condition encoder over source and text, and the diffusion timestep. Shared target tokens feed a pose head that predicts velocity $v$ and a preservation head that predicts a per-frame, per-dimension gate $g$ . After converting $v$ to $\hat{x}_{0}$ , blending pulls the estimate toward the time-aligned source where $g$ is high and leaves it to the pose head where $g$ is low.

Method	B R@1 $\uparrow$	B R@2 $\uparrow$	B R@3 $\uparrow$	B AvgR $\downarrow$	FR@1 $\uparrow$	FR@2 $\uparrow$	FR@3 $\uparrow$	F AvgR $\downarrow$
Copy source	74.79	85.42	91.25	1.88	30.24	44.47	54.35	15.53
MDM ( Tevet et al., 2022b )	4.03	7.56	10.48	15.55	0.10	0.10	0.10	-
MDM-BP ( Tevet et al., 2022b )	39.10	50.09	54.84	6.46	8.69	14.71	18.36	180.99
MotionLab ( Guo et al., 2025 )	54.53	70.02	77.38	3.42	-	-	-	-
TMED ( Athanasiou et al., 2024 )	62.90	76.51	83.06	2.71	14.51	21.72	28.73	56.63
MotionReFit ( Jiang et al., 2025 )	66.63	80.05	84.98	2.64	14.13	23.52	30.53	54.06

Table 1: Comparison on MotionFix.

Method	B R@1 $\uparrow$	B R@2 $\uparrow$	B R@3 $\uparrow$	B AvgR $\downarrow$	FR@1 $\uparrow$	FR@2 $\uparrow$	FR@3 $\uparrow$	F AvgR $\downarrow$
Copy source	40.10	58.33	64.58	5.13	17.35	22.96	30.10	27.23
MotionReFit ( Jiang et al., 2025 )	25.52	34.90	45.83	8.27	9.18	14.29	16.84	48.74
TMED ( Athanasiou et al., 2024 )	39.58	49.48	60.94	6.26	13.78	22.96	30.10	34.38
SimMotionEdit ( Li et al., 2025 )	43.23	59.38	68.23	4.67	20.92	31.63	37.24	24.02
OmniME ( Shi et al., 2026 )	43.75	56.25	66.15	4.66	22.45	31.63	36.22	22.77
Ours	46.88	63.02	71.88	3.72	18.37	31.12	38.27	19.21

Table 2: Comparison on STANCE Adjustment.

Figure 4: Qualitative comparison on MotionFix dataset. Relative to prior methods, ours follows the instruction on the first two examples and keeps the unedited motion on the third.

Method	B R@1 $\uparrow$	B R@2 $\uparrow$	B R@3 $\uparrow$	B AvgR $\downarrow$	FR@1 $\uparrow$	FR@2 $\uparrow$	FR@3 $\uparrow$	F AvgR $\downarrow$
Ours	79.17	89.17	92.92	1.68	33.20	50.79	59.09	12.87
w/o gate	74.79	87.08	91.04	1.95	29.45	44.86	54.74	16.22
tanh gate	76.04	89.58	92.29	1.75	29.84	46.05	55.34	13.35
per-frame	74.17	89.17	93.54	1.79	29.64	46.44	57.11	14.12
w/o $\mathcal{L}_{\mathrm{anc}}$	76.25	87.92	92.71	1.89	27.27	43.48	55.14	15.66
w/o $\mathcal{L}_{\mathrm{hf}}$	76.88	88.54	91.46	1.75	30.24	44.66	55.14	13.61

Table 3: Ablation study on different components (MotionFix).

Figure 5: Preservation strength at inference. Scaling the learned gate by

\gamma

without retraining raises retrieval against the source far more than against the target.

Figure 6: Behavior of the learned gate. (a) Mean gate value $g$ against edit magnitude $d(\text{target},\text{time-aligned source})$ , with interquartile and 10–90% bands across samples in each bin. (b) Mean gate over frames and channels along the reverse trajectory, with the same bands.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Detail ratio	$d(\mathrm{gen},\mathrm{tgt})$	$d(\mathrm{gen},\mathrm{src})$	Closer to src
TMED	0.330	1.885	1.791	60.9%
SimMotionEdit	0.708	1.661	1.616	56.1%
OmniME	0.658	1.642	1.496	60.8%
SuperMotion	1.035	1.586	1.563	54.8%
Copy source	–	1.745	0	–

Appendix

Table 4: Copy-versus-edit diagnostic on MotionFix (

N{=}836

Figure 7: Qualitative comparison on STANCE Adjustment. Each instruction targets a specific limb or side. Left to right: source, TMED, SimMotionEdit, OmniME, SuperMotion, and the reference target.

Explore similar work

May 29, 2026cs.CV

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

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

Zhenwu 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

Text-To-Motion Generation Image-Text Alignment Motion Positive

Ask PDF

Aug 10, 2026cs.CV

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

Instruction-driven editing of 3D human motion requires precise spatiotemporal localization, rich semantic grounding, and strict preservation of unmodified content. Existing methods either resort to training-free adaptation of generative models or rely solely on triplet supervision; however, adaptation often yields suboptimal control, and manually curated triplet datasets remain severely limited in scale and semantic diversity. To overcome this bottleneck, we ground motion editing directly within text-to-motion generation across data, architecture, and inference. At the data level, we develop a closed-loop synthesis-and-verification pipeline that produces Omni-MoEdit, a large-scale dataset spanning body-part, amplitude, temporal, action, and style edits. At the architectural level, we introduce UniMoFlow, a unified latent flow-matching model that shares broad semantic and kinematic knowledge between generation and editing. At the inference level, SAFE (Source-Anchored Flow Editing) complements UniMoFlow with controllable, source-anchored refinement. Furthermore, we augment standard evaluations with semantics-aware metrics to account for valid edits that inherently deviate from a single ground-truth reference. Extensive experiments demonstrate improved target-text alignment, edit effectiveness, and cycle consistency, while maintaining competitive source fidelity and text-to-motion generation quality.

Yilei Hua, Beibei Jing, Ce Zheng +3

School of Computer Science, Huazhong University of Science and Technology · School of Software Technology, Zhejiang University

Human Motion Generation Text-To-Motion Generation Human Motion Contrastive Flow Matching Kinematics Process-Level Supervision+1

Ask PDF

May 31, 2026cs.CV

Cross-Axis Feature Fusion with Joint-Wise Motion Difference Prediction for Text-Based 3D Human Motion Editing

We address text-based 3D human motion editing, where the goal is to preserve the style and structure of a source motion while applying edits described in natural language. The release of the MotionFix dataset has spurred active research into training-based diffusion models that directly generate an edited motion from a source motion and a text instruction. While previous works have focused primarily on learning when an edit should occur temporally, our goal is to create a model that understands not only this temporal aspect but also which specific joints are responsible for the change. Targeting this, we propose a novel architecture and a complementary auxiliary task to aid its training. Our architecture consists of two axis-anchored transformers, which extract distinct features along the joint and time dimensions respectively, and a cross-axis fusion block that integrates these representations. We further introduce an auxiliary task that trains the joint-anchored transformer to regress the Soft-DTW distance between source and target joint rotations. This objective teaches the module to understand which joints to modify and which to preserve. Through comprehensive experiments on the MotionFix dataset, we demonstrate that our method significantly improves semantic alignment with both the text instruction and the source motion, as well as the overall fidelity of the generated motion, achieving state-of-the-art results.

Gyojin Han, Junmo Kim

School of Electrical Engineering, KAIST

3D Motion Training-Set Long-Tail Motions Human Motion Prediction Motion Diffusion Models

Ask PDF