acceptodds
Under review as a conference paper at ICLR 2027

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

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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
text-driven motion editing, human motion generation, skeletal topology, representation alignment

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qCPPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work presents TAME, a flow-matching transformer for text-driven motion editing across humanoid skeletons with different topologies. The method propose both topology-constrained skeletal attention (TCSP) module which restricts skeletal attention to one-hop neighbors; and edit-focused representation alignment (EFRA) which is a self-distilled representation alignment strategy. The authors further build TopoMotionFix, which extends MotionFix to multiple skeleton topologies, and report state-of-the-art results on both seen and unseen topologies benchmarks.

Critical strengths
  1. The paper is the first to propose text-conditioned motion editing across heterogeneous humanoid skeleton topologies.
  2. The proposed method achieves state-of-the-art performance on both single toplogy or heterogeneous topologies motion editing benchmarks.
  3. The ablation studies are comprehensive, and the proposed components show clear improvements.
  4. An extended heterogeneous humanoid skeleton topologies benchmark/dataset on motionfix is proposed.
Critical weaknesses
  1. Dataset/Benchmark quality lack evaluation. In the third unseen-topology example in the supplementary video, both the ground-truth motion and TAME's output show noticeable self-penetration, raising concerns about the quality of the retargeted dataset.

  2. Lack of user study. The evaluation mainly relies on retrieval metrics, without human evaluation to assess editing quality, instruction faithfulness, and source motion preservation for comparison with baselines.

  3. Unclear unseen-topology difference. It is unclear how different these 9 held-out rigs are from the training skeletons in terms of topology and if they can really reflect generalization to new topology.

  4. Figure1 visualization unclear. Some editing effects in Figure 1 are difficult to observe, such as "turn more to the right"(the turn of TAME is unclear and the trajectory shown is a staight line) and "shake both hands"(too clustered to see).

Questions for authors

See Weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer qCPP (1/3)

We sincerely thank the reviewer for the thoughtful and constructive review and for recognising TAME as the first work on text-driven motion editing across heterogeneous humanoid skeletons, together with the ablation studies and the TopoMotionFix benchmark. We answer each point below.

W1: Self-penetration in the third unseen-topology example

We appreciate the reviewer's careful look at the supplementary video. We would like to clarify that the self-penetration comes from the heavy armour of the character that we use for display, rather than from the quality of the TopoMotionFix dataset. We examined the third unseen-topology example frame by frame and detected the frames in which the skin surfaces of two body parts intersect. The self-penetration that the reviewer notes does exist in both the ground-truth motion and the output of TAME. It occurs where the forearm guard cuts into the chest plate and where the shoulder plate reaches the helmet. This armour is specific to the character. The character is a heavily armoured soldier whose armour is large and thick. Its bone lengths are close to those of a slim character, but its surface lies 1.4 to 2.3 times further from the bones. A character of this kind is therefore prone to self-penetration once it is rendered.

TopoMotionFix is built by retargeting the motions of MotionFix onto other skeletons, and the retargeting uses the joints only. The vertices of a mesh, such as those of clothes, hats, or armour, take no part in it, and the dataset contains skeletons and joint motions only (Sec. 4.1, Appendix A). The motions were captured on human bodies, and the heavy armour in effect enlarges the body of this character, which leads to the penetration.

In Table 3 (Appendix D), the row GT (retargeted) reports the effect of the retargeting on the retrieval metrics. Table 3 follows the MotionFix evaluation on the full test set, where each motion is ranked among all test motions rather than within a batch of 32 as in Table 1. The absolute values are therefore lower than in Table 1, and what matters here is the difference between the two rows. We copy this row and the GT row below.

Generated-to-Target R@1R@2R@3AvgRGenerated-to-Source R@1R@2R@3AvgR
GT64.3688.7595.561.7420.8333.6640.4733.13
GT (retargeted)64.3488.7395.551.7320.8233.6540.4733.12

Retargeting the MotionFix test targets onto the held-out rigs and back to SMPL has a negligible effect on the retrieval metrics.

W2: User study

We thank the reviewer for this suggestion, and we agree that human evaluation is necessary for motion editing. The submission contained a user study in Appendix F, titled "Perceptual Study", but the experiments section did not refer to it. Following the reviewer's suggestion, we have conducted the study again at a larger scale, with more candidates, and with source motion preservation as a third criterion. We rename the appendix to "User Study" and refer to it in Sec. 4.2. The setting changes as follows.

SubmissionRevision
Participants per study2540
Sample sets per participant1030
Candidate motions per sample set35
Seen-topology candidates(1) GT, (2) TAME, (3) SimMotionEdit(1) GT, (2) TAME, (3) TAME w/o EFRA, (4) SimMotionEdit, (5) OmniME
Unseen-topology candidates(1) GT, (2) TAME, (3) AnyTop(1) GT, (2) TAME, (3) TAME w/o EFRA, (4) TAME w/o TCSP, w/o EFRA, (5) AnyTop
Rating criteriaalignment, plausibilityalignment, plausibility, source preservation
AuthorsOpenReview ↗

Response to Reviewer qCPP (2/3)

W2: User study (continued)

The 30 sample sets of the unseen-topology study cover 15 skeletons with 2 sample sets each: the SMPL skeleton, the 9 held-out rigs, and 5 Mixamo characters. The 5 Mixamo characters are randomly selected and belong to neither the training rigs nor the held-out rigs, and their results are reported separately. Participants rate each candidate on a 3-point scale, and the candidates are shown in random order. The new criterion, source preservation, asks how well the edited motion keeps the parts of the source motion that the instruction does not ask to change (1 = unedited parts are largely changed, 2 = slightly changed, 3 = almost unchanged). Appendix F also explains how this criterion differs from the generated-to-source retrieval in Table 1. Each cell below is the mean over 1200 ratings (Seen), 800 ratings (Unseen), or 400 ratings (Mixamo), with the standard deviation after the ± sign.

ProtocolMethodAlignment ↑Plausibility ↑Source Preservation ↑
SeenGT2.65 ±0.522.61 ±0.572.93 ±0.27
SimMotionEdit2.28 ±0.752.18 ±0.622.19 ±0.39
OmniME2.40 ±0.742.29 ±0.722.26 ±0.51
TAME w/o EFRA2.36 ±0.752.24 ±0.692.25 ±0.50
TAME (ours)2.53 ±0.642.33 ±0.672.29 ±0.54
UnseenGT2.62 ±0.552.64 ±0.512.95 ±0.23
AnyTop2.20 ±0.722.12 ±0.692.16 ±0.47
TAME w/o TCSP, w/o EFRA2.12 ±0.701.91 ±0.701.90 ±0.43
TAME w/o EFRA2.28 ±0.592.14 ±0.702.19 ±0.56
TAME (ours)2.31 ±0.592.19 ±0.692.21 ±0.60
MixamoGT2.60 ±0.532.65 ±0.462.94 ±0.24
AnyTop2.23 ±0.652.08 ±0.662.13 ±0.57
TAME w/o TCSP, w/o EFRA2.09 ±0.741.96 ±0.651.90 ±0.67
TAME w/o EFRA2.29 ±0.682.10 ±0.692.16 ±0.58
TAME (ours)2.30 ±0.692.14 ±0.702.20 ±0.63

TAME receives the highest ratings among the methods on all three criteria in all three blocks. Under the seen-topology protocol, it is rated above SimMotionEdit (alignment 2.53 vs 2.28, plausibility 2.33 vs 2.18, source preservation 2.29 vs 2.19) and above OmniME (2.53 vs 2.40, 2.33 vs 2.29, 2.29 vs 2.26). Under the unseen-topology protocol, it is rated above AnyTop (2.31 vs 2.20, 2.19 vs 2.12, 2.21 vs 2.16), and the same ordering holds on the Mixamo characters (2.30 vs 2.23, 2.14 vs 2.08, 2.20 vs 2.13). The two ablated variants show that participants also perceive the effect of each component. Removing EFRA lowers alignment from 2.53 to 2.36 under the seen-topology protocol, and further removing TCSP lowers plausibility from 2.14 to 1.91 and source preservation from 2.19 to 1.90 on the unseen skeletons.

The three criteria correspond to the three aspects that the reviewer lists. Alignment measures instruction faithfulness, plausibility measures editing quality, and source preservation measures how well the source motion is preserved.

AuthorsOpenReview ↗

Response to Reviewer qCPP (3/3)

W3: Difference between the held-out rigs and the training skeletons

We appreciate this insightful point. The 9 held-out rigs are motion-capture rigs that appear only in the test set, with their original topology and bone lengths. To enlarge the difference between the held-out rigs and the training skeletons, we apply skeleton augmentation to the 83 training rigs (Algorithm 1, Appendix A). The augmentation randomly adds or removes joints along the spine and the neck, inserts dummy joints at the hip, the lower back, and the shoulder, and rescales the bone lengths and the height. The training skeletons are therefore randomised variants whose joint counts and topologies are changed, and the held-out rigs are excluded from the character pool that the augmentation samples from (Appendix B). By design, the held-out rigs thus differ from the training skeletons. To show how the held-out rigs look, we have added Figure 5 in Appendix B, which visualises the SMPL skeleton and the 9 held-out rigs in their rest poses at a common scale. We also acknowledge that this difference is not as large as the difference across species. Our work targets humanoid skeletons, and the unseen-topology protocol evaluates generalisation within this scope.

To further examine generalisation beyond the held-out rigs, the user study in Appendix F (W2 above) includes 5 randomly selected Mixamo characters. These characters belong to neither the training rigs nor the held-out rigs. The ratings of TAME on these characters are close to its ratings on the SMPL skeleton and the held-out rigs (alignment 2.30 vs 2.31, plausibility 2.14 vs 2.19, source preservation 2.20 vs 2.21), and TAME is rated above AnyTop on all three criteria. Figure 4 also shows qualitative results on Mixamo skeletons.

W4: Visualisation in Figure 1

We thank the reviewer for pointing this out. We have re-rendered both examples in Figure 1. For "shake both hands", the frames were drawn on top of each other. They are now spread out along the trajectory, so that the hands can be seen in every frame. For "turn more to the right", we render the motion from a different viewpoint and space the frames apart, and the trajectory arrow now follows the path of the character instead of a straight line, so that the turn of the target motion can be seen.

Reviewer_DUGEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a novel task, text-driven motion editing on arbitrary human rigs, and a model, TAME, a diffusion based text-driven motion editing model. TAME outperforms existing methods on the existing motion editing benchmark MotionFix and outperforms the AnyTop baseline on its newly introduce TopoMotionFix evaluation.

Critical strengths

Text-driven motion editing on arbitrary skeletons is a novel and interesting problem. The paper introduces a new dataset to evaluate this new task.

Critical weaknesses

There are several gaps in terms of implementation:

How is the rest pose handled? This is crucial for any topology-agnostic approach, i.e. AnyTop has rest pose as explicit input, Skeleton-aware networks (Aberman et al.) tile static offsets along time and concat them to the rotations, and Motion2Motion aligns rest poses of source and target as an explicit step before transfer. SAME re-expresses every skeleton so that “zero rotation” means the rigs particular T-pose - this is a complex and expensive pre-processing step, and it weakens the “skeleton-agnostic” claim, as every new topology added requires this pre-processing step. This also does not exactly align characters across different topologies, as “T-Pose” can be expressed differently, depending on the rig, artist, etc. The paper claims they use the retargeting protocol from SAME (L729) but even in SAME the network gets the T-pose (rest pose) as explicit input. This is a crucial method gap which has to be addressed.

How is positional embedding utilized in TAME? I assume for the joints the T5 embeddings are used. What happens if the names are missing, or follow an unusual naming convention, i.e. “larm” for “left arm”. Another concern with this would be: naming is not necessary consistent across topologies, i.e. SMPL-X shoulder would equal Mixamo’s “Arm” while “shoulder” would instead be the clavicle.

What positional encoding is used for temporal frames?

The claim that EFRA makes “edits faithful to the instruction” (L26) is too strong, I would recommend to reword it as “EFRA improves MotionFix target retrieval”.

I recommend reporting mean and std for the experiments, as TAME is a generative model, to provide a better understanding of the spread of the method. For Table 1 - MotionFix I recommend only re-reporting the mean + std for TAME but for the methods own evaluation (TopoMotionFix) I would recommend to run multiple samples for both TAME and its ablation as well as for AnyTop.

Questions for authors

How was AnyTop adapted for this work? What adaptations were necessary? How was the rest pose handled?

Can the authors comment on Table 1 - Generated-to-Source - in the original paper of MotionFix the authors explicitly warn that that “metric should not be too high - otherwise there would be no edit”.

7 comments
AuthorsOpenReview ↗

Response to Reviewer DUGE (1/5)

We thank the reviewer for the insightful review and for recognising text-driven motion editing on arbitrary skeletons as a novel and interesting problem. We answer each point below.

W1: Gaps in terms of implementation

1. How the rest pose is handled in TAME

We thank the reviewer for raising this valuable question. The rest pose is indeed essential for the topology-agnostic approaches that the reviewer cites (AnyTop, Skeleton-Aware Networks, SAME, and Motion2Motion), whereas our paper hardly discusses it. This naturally raises the question of why TAME, which also works across heterogeneous humanoid skeletons, does not take a rest pose as input. In fact, the absence of a rest-pose input in TAME is not a gap in the method but a deliberate design choice. We analyse this in detail below, and we believe that the analysis resolves the concern of the reviewer.

Why these topology-agnostic approaches need the rest pose. These approaches have one thing in common: they are expected to generate a motion for a skeleton whose details (e.g., bone lengths and joint offsets) are unclear to the network. Put simply, the network needs something that tells it what the output character looks like, and the rest pose serves this purpose. We explain below why each of these approaches needs the rest pose, according to its task. AnyTop is a motion generation model whose only conditions are the skeleton configurations, which consist of the rest pose, the joint names, and two graphs computed from the joint hierarchy. The rest pose is prepended to the noised motion as an extra frame. Its ablation shows that removing the rest pose degrades the results, because the rest pose "encodes vital information about joint offsets and bone lengths", which the graphs and the joint names do not contain. Skeleton-Aware Networks retarget a motion from a source skeleton to a different target skeleton. The source motion does not describe the target skeleton, and the joint offsets and bone lengths of the target are supplied through its static offsets, which are tiled along time and concatenated to the rotations. SAME encodes a motion into a skeleton-agnostic embedding, which no longer contains skeleton-specific information, so its decoder "takes a target skeleton S_tgt and a skeleton-agnostic pose embedding z_t as input". Motion2Motion transfers a motion from a source skeleton to a target skeleton. Its motion features are joint rotations, which are comparable across the two skeletons only if their rest poses follow the same convention. Since "the definitions of rest poses in different files are not always the same", it aligns the rest poses of the source and the target before the transfer.

The source motion takes over the role of the rest pose. In text-driven motion editing, the source motion is provided as an input, and the source motion and the target motion share the same skeleton configurations. TAME can therefore learn the skeleton geometry of the target motion from the source motion. Beyond the skeleton geometry (bone lengths and joint offsets), the source motion also provides a reference for the velocity and the range of rotation of each joint, since it shows the skeleton moving over T frames rather than in the static T-pose. In motion editing, we therefore choose to use the source motion to take over the role that the rest pose plays in these topology-agnostic approaches. We compare below how the rest pose enters AnyTop and how the source motion enters TAME. AnyTop concatenates the rest pose with the noised motion along the temporal axis as an extra frame. TAME concatenates the source motion with the noised target motion along the time axis, adds learnable source/target segment embeddings to distinguish the two, and inserts a separator token to mark their boundary, which yields a sequence of 2T+1 frames (Sec. 3.2). Therefore, since the source motion and the edited motion share the same skeleton configurations (editing changes the content of a motion, not its skeleton configurations), TAME does not need to align the T-poses of different characters as Motion2Motion does. It is also unnecessary to provide the T-pose as an explicit input to the model as in SAME, because the source motion is a better choice than the T-pose for this purpose. Finally, adding a new topology at inference time does not require the T-pose pre-processing step of SAME. The retargeting pipeline of SAME is used only to construct TopoMotionFix (Sec. 4.1, Appendix A). Editing a motion sequence on a new skeleton with TAME requires only four inputs: the motion sequence to be edited, the edit instruction, the joint names normalised with the rules in Appendix G, and the adjacency matrix that contains the topology of the skeleton. We have added this discussion to Sec. 2.2.

References. AnyTop: Gat et al., SIGGRAPH 2025. Skeleton-Aware Networks: Aberman et al., TOG 2020. SAME: Lee et al., SIGGRAPH Asia 2023. Motion2Motion: Chen et al., SIGGRAPH Asia 2025.

AuthorsOpenReview ↗

Response to Reviewer DUGE (2/5)

2. Positional embedding along joints and along time

Along the joint axis, the positional embedding of a joint is the T5 embedding of its joint name. Along time, frames carry a sinusoidal positional encoding over the frame index, which is computed separately for the source motion and for the noised target motion. The source and the target are distinguished by learnable segment embeddings and a separator token (Sec. 3.2). We have added this detail to Sec. 3.2.

3. Unusual or missing joint names, and naming conventions that differ across rigs

We apologize for not describing this in the submission. In the revised version, we add Appendix G (Joint Name Normalisation) to detail how joint names are normalised into a common form before the T5 encoding. Normalisation is needed for two reasons. First, the rigs of TopoMotionFix are sourced from different motion-capture datasets, and the same joint may be written differently across them. Second, a rig unseen during training may name its joints differently from the training rigs. Normalisation will rewrite the joint names of every rig, so that the name of the same anatomical joint carries the same joint semantics on a training rig and on an unseen rig. The appendix consists of three parts: three normalisation steps for names that are written differently, a correspondence table for naming conventions that use different words for the same joint, and three rules for joints without a usable name. For the three cases that the reviewer mentions (unusual names, naming conventions that differ across rigs, and missing names), we explain below how joint name normalisation handles each of them.

Unusual names such as "larm". We adopt three steps to normalise joint names that are written differently across rigs. (i) Cleaning removes a prefix of the rig or the authoring software, rewrites the name in camel case, and separates words that are joined without spaces, so that mixamorig:Hips becomes "Hips", leftshoulder becomes LeftShoulder, and LeftForeArm becomes "Left Fore Arm". (ii) Side expansion writes an abbreviated body side in full, so that LHipJoint becomes "Left Hip Joint". (iii) Index removal replaces the numeric index of a spine or neck segment with the word Segment, so that Spine1 to Spine4 all become "Spine Segment". The number of spine and neck segments varies across rigs, so the same index may refer to different anatomical positions on different rigs, and the relations among these segments are left to the skeleton topology. With these steps, the reviewer's example, larm, is first rewritten as LArm and then becomes "Left Arm".

Conventions that use different words for the same anatomical joint. For a rig whose naming convention differs from that of our normalised names, we use a correspondence table to map its joint names to the normalised names. The lower part of Table 5 in Appendix G gives the mapping from the joint names of SMPL and SMPL-X to the normalised names, and the reviewer's example corresponds to two of its rows: left_shoulder is mapped to "Left Arm", and left_collar to "Left Shoulder". A correspondence table is suitable here because the two conventions differ in a systematic way. SMPL and SMPL-X name each joint after its anatomical location, so the joint named left_shoulder is located at the anatomical left shoulder joint. Mixamo instead gives each joint the name of the body part that it moves. For instance, the joint at the left shoulder moves the left upper arm, so Mixamo names it LeftArm. Similarly, the joint at the left clavicle moves the left shoulder, so Mixamo names it LeftShoulder, whereas SMPL-X names it left_collar. Each convention is consistent in itself, so the same anatomical joint has one name in each convention, and a table that pairs these names resolves the difference. Our normalised names are closer to the Mixamo convention, because an editing instruction often refers to the body part that should move. For instance, "raise the left arm" is carried out by rotating the joint at the shoulder, and this rotation changes the positions of its child joints along the arm. With our normalised names, the words of the instruction match the name of the joint whose rotation is to be edited, which is "Left Arm" in this example.

AuthorsOpenReview ↗

Response to Reviewer DUGE (3/5)

3. Unusual or missing joint names, and naming conventions that differ across rigs (continued)

Missing names. A joint may have an empty name, or a name without anatomical meaning, such as Joint012. An editing instruction refers to body parts by name, so every joint needs a name that tells the model which body part it belongs to. We infer the missing name from the named joints around it in the skeleton hierarchy, using three rules. (i) A joint that lies between two "Spine/Neck Segment" joints is named "Spine/Neck Segment". (ii) On a limb, the name is inferred from the names of the neighbouring joints and the skeleton topology, so that a joint between "Left Shoulder" and "Left Fore Arm" is named "Left Arm". (iii) If the joint at the same position on the opposite side of the body has a name, the joint receives this name with Left and Right swapped. For instance, if "Left Arm" exists and the joint at the same position on the right side has no usable name, this joint is named "Right Arm". If none of the three rules applies, the joint is reported to the user, who names it by hand.

W2: The EFRA claim

The claim at L26. We agree that "faithful to the instruction" is too strong. We have replaced it with a statement that our evaluation measures directly, and the abstract now reads "improving the alignment of the edited motion with the instruction". The contribution list is reworded in the same way. This statement is supported by two kinds of evidence, and both compare TAME with TAME w/o EFRA. First, on the retrieval metrics, EFRA raises the generated-to-target R@1 by 4.05 on MotionFix (Table 2) and by 1.42 on the unseen-topology protocol of TopoMotionFix (72.84 ± 0.81 vs 71.42 ± 0.39 over 20 seeds, Table 1). Second, on human judgment, the user study in Appendix F now includes TAME w/o EFRA as a candidate, and participants rate the alignment with the instruction of TAME above that of TAME w/o EFRA (2.53 vs 2.36 under the seen-topology protocol, and 2.31 vs 2.28 under the unseen-topology protocol).

W3: Multi-seed evaluation

We thank the reviewer for this recommendation. We have added multi-seed results to Table 1, and we copy them below.

MotionFix. As recommended, we report the mean and the standard deviation for TAME. We sample the test set 20 times with the checkpoint of Table 1, changing only the sampling seed, and report the statistics over these 20 runs.

Generated-to-Target R@1R@2R@3AvgRGenerated-to-Source R@1R@2R@3AvgR
TAME (multi-seed)79.70 ± 0.8589.81 ± 0.7793.22 ± 0.721.64 ± 0.0485.76 ± 0.4994.47 ± 0.5096.95 ± 0.391.39 ± 0.05

The upper part of Table 1 keeps the single-run value of the submission (78.75), so that TAME and the baselines are compared under the same single-run setting. This value is 0.95 below the multi-seed mean, about one standard deviation.

TopoMotionFix. We run TAME, its two ablated variants, and the adapted AnyTop 20 times each with different sampling seeds, with all other settings unchanged.

MethodGenerated-to-Target R@1R@2R@3AvgRGenerated-to-Source R@1R@2R@3AvgR
AnyTop (adapted)68.54 ± 0.5383.79 ± 0.5489.52 ± 0.402.00 ± 0.0383.93 ± 0.3092.46 ± 0.4795.22 ± 0.301.48 ± 0.02
TAME w/o TCSP, w/o EFRA68.69 ± 0.6682.33 ± 0.5288.02 ± 0.452.08 ± 0.0282.40 ± 0.4092.00 ± 0.3795.63 ± 0.241.52 ± 0.01
TAME w/o EFRA71.42 ± 0.3985.17 ± 0.4890.48 ± 0.481.96 ± 0.0286.71 ± 0.3993.77 ± 0.4096.45 ± 0.261.39 ± 0.01
TAME72.84 ± 0.8186.78 ± 0.6291.47 ± 0.361.89 ± 0.0589.95 ± 0.6695.74 ± 0.6497.33 ± 0.591.29 ± 0.03

The standard deviation of the generated-to-target R@1 is between 0.4 and 0.8 for every method. TAME w/o EFRA is 2.7 to 2.9 points above the two methods without TCSP, which is more than three times the standard deviation of any row, so the gain of TCSP is not an effect of the sampling seed.

Reviewer_yHT2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TAME, a text-driven motion editing framework for heterogeneous humanoid skeletons. It introduces Topology-Constrained Skeletal Propagation (TCSP) to model skeletal structures and Edit-Focused Representation Alignment (EFRA) to improve editing fidelity. The authors also construct TopoMotionFix, a multi-topology motion editing dataset. Experimental results demonstrate improvements in edit alignment, source preservation, and generalization to unseen skeletal topologies.

Critical strengths
  1. Relevant Problem Setting. The paper addresses an important limitation of existing motion editing methods by targeting heterogeneous humanoid skeletons with varying topologies.
  2. Comprehensive Framework. TAME integrates topology-constrained skeletal propagation (TCSP) and edit-focused representation alignment (EFRA) to improve topology generalization and instruction fidelity.
Critical weaknesses

1, Limited Novelty. The method mainly combines standard techniques (self-distillation and skeletal, temporal, and text cross-attention). The overall framework is relatively straightforward and lacks specialized designs to address the challenges posed by varying skeletal topologies. 2, Insufficient Experimental Validation. The experiments are not comprehensive enough to convincingly demonstrate the method's ability to handle variations in joint count and skeletal hierarchy, as claimed in the paper. 3, Robustness to Different Joint Definitions. It remains unclear whether TAME can perform accurate motion editing across datasets with different numbers of joints and skeletal configurations.

Questions for authors

see the weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer yHT2 (1/3)

We thank the reviewer for the constructive comments and respond to each point below.

W1: Limited novelty

We propose TAME to accommodate the varying skeletal structures of different characters and to improve the alignment of the edited motion with the text instruction.

Standard techniques (self-distillation). A central difficulty in text-driven motion editing is to make the edit follow the text instruction. We identify one cause of edits that deviate from the instruction: motion editing is an asymmetric learning problem (Sec. 3.3). The unedited parts of the target closely follow the source motion, which serves as a reliable reference for them, whereas the edited parts have to be inferred from the source motion and the text instruction. The edited parts are therefore harder to learn, yet they decide whether the instruction is followed, so they call for stronger guidance.

Edit-Focused Representation Alignment (EFRA) provides this guidance with representation alignment, a recent advance in image generation that has rarely been applied to human motion, and builds on its self-distilled form, SRA (Jiang et al., ICLR 2026). In image generation, REPA (Yu et al., ICLR 2025) introduced the idea: it aligns the noisy hidden states of a diffusion transformer with clean features from a pretrained external encoder, and SRA removes the external dependence by self-distilling across noise levels, matching an earlier student feature under higher noise to a later EMA-teacher feature under lower noise. In the motion domain, representation alignment is far less explored. To our knowledge, only three text-to-motion works use it (LUMA, Jia et al., 2025; CMDM, Yu et al., CVPR 2026; MoLingo, He et al., CVPR 2026), all for generation and all aligning to an external target. EFRA is the first to bring representation alignment to motion editing, and the first motion model to do so without any external component. In this domain, therefore, representation alignment is far from a standard technique. We have added a discussion of LUMA, CMDM, and MoLingo to Sec. 2.3 to make clear that representation alignment is far less explored in the motion domain.

Rather than using SRA off the shelf, we tailor it to motion editing in three respects: a DTW edit map marks the edit-relevant joint-time tokens of each source–target pair, an edit-aware dual-timestep schedule sets the noise level of the teacher token by token from this map, and the alignment loss covers only the marked tokens. Under this schedule the student is noised uniformly at the higher timestep, while the teacher receives the lower timestep only on the edit tokens, so that the teacher is ahead of the student exactly where the asymmetry above calls for it. The teacher is the network deployed at inference, so EFRA adds no inference cost. Table 2 shows that this editing-specific design, not the alignment itself, carries the gain: aligning all tokens (SRA applied uniformly) improves generated-to-target R@1 by +1.01 (74.70 → 75.71), while the edit-aware schedule and mask add a further +3.04 (75.71 → 78.75). Overall, EFRA improves both edit fidelity and source preservation (Table 2), and with it TAME outperforms all fixed-topology baselines on MotionFix (Table 1). In the user study (Appendix F), participants also rate the alignment with the instruction of TAME above that of TAME w/o EFRA (2.53 vs 2.36 under the seen-topology protocol).

Specialized designs for varying skeletal topologies. TAME has a dedicated design for each of the three ways in which skeletons vary: the number of joints, their naming, and the skeletal hierarchy. For the number of joints, motion is represented as per-joint tokens, so the feature dimension of each joint stays fixed however many joints a rig has. For naming, we normalise the joint names of every rig into a common form (new Appendix G on joint names) to address the inconsistency of joint naming across rigs. For the hierarchy, every joint is physically connected to a few neighbours on any rig, so Topology-Constrained Skeletal Propagation (TCSP) restricts joint interaction to these connections: in each layer a joint exchanges information only with its physical neighbours, and relations between distant joints build up over layers by propagating along the kinematic chain. The identity of a joint is thus established twice: once by its own name, and once by the names of the neighbours it reads through TCSP. In implementation, TCSP reads the adjacency matrix of the input skeleton (Sec. 3.1) and uses it as the mask of the joint-axis attention, so that topology enters the network as a structural communication rule, read from whatever skeleton is given, rather than as a learnable bias in full attention.

AuthorsOpenReview ↗

Response to Reviewer yHT2 (2/3)

W2: Experiments on variations in joint count and skeletal hierarchy

We test the ability of TAME to handle variations in joint count and skeletal hierarchy with the unseen-topology protocol of TopoMotionFix: the model is trained on 83 rigs with skeleton randomisation, 25,794 distinct skeletal configurations in total, and tested on 9 further rigs that are held out from training (Sec. 4.2, Appendix B). Four pieces of evidence show that TAME edits motions accurately on rigs with different joint counts and hierarchies, including rigs it has not seen in training: the TCSP ablation (Table 1), the comparison with AnyTop (Tables 1 and 3), the user study (Appendix F), and the qualitative results (Fig. 4 and the supplementary video).

(1) TCSP ablation (Table 1). Between two models that differ only in TCSP and share the same training recipe, TCSP raises generated-to-target R@1 from 68.69 ± 0.66 to 71.42 ± 0.39 under the unseen-topology protocol (mean and standard deviation over 20 sampling seeds). On the fixed SMPL skeleton it makes no difference (74.80 vs 74.70, Table 2), so the gain comes from handling the variation in topology.

(2) Comparison with AnyTop (Tables 1 and 3). We adapt AnyTop, an arbitrary-topology motion generation method, to editing and train it on the same data. TAME reaches 72.84 ± 0.81 generated-to-target R@1 against 68.54 ± 0.53 for AnyTop over 20 sampling seeds, and the ordering holds under the test-set gallery of Table 3.

(3) User study (Appendix F, Table 4). In the study for the unseen-topology protocol, 40 participants rate TAME above AnyTop on alignment (2.31 vs 2.20), plausibility (2.19 vs 2.12), and source preservation (2.21 vs 2.16).

(4) Qualitative results (Fig. 4, supplementary video). Fig. 4 compares TAME with AnyTop on Mixamo skeletons unseen during training. TAME produces plausible local edits and preserves the source motion, whereas AnyTop more often shows unnatural limb trajectories, over-smoothed edits, or unintended changes to unedited joints. The supplementary video shows further edits on Mixamo characters.

Joint count and hierarchy vary in both training and test. In training, Algorithm 1 (Appendix A) adds or removes joints along the spine and neck chains and inserts joints next to the hip, lower back, and shoulder, which changes both the joint count and the hierarchy. In test, the 9 held-out rigs are used for evaluation with their original topology. They are removed from the training samples and from the character pool that Algorithm 1 draws from, so these rigs never appear in training.

AuthorsOpenReview ↗

Response to Reviewer yHT2 (3/3)

W3: Motion editing across datasets with different numbers of joints and skeletal configurations

TAME is designed to edit motions on skeletons with different numbers of joints and skeletal configurations, and we agree that the paper should state more clearly how this is tested and how different joint definitions are handled. We clarify both points below.

TopoMotionFix spans several datasets. It contains the SMPL skeleton of MotionFix together with rigs sourced from five motion-capture datasets (LAFAN1, ACCAD, TotalCapture, PFNN, and SFU, Sec. 4.1), and these rigs differ in the number of joints and in the skeletal hierarchy. TAME is trained on these rigs jointly and is then tested on 9 held-out rigs, where it reaches 72.84 ± 0.81 generated-to-target R@1 over 20 sampling seeds, compared with 68.54 ± 0.53 for AnyTop (Table 1). Further results under this protocol are given in our reply to W2. Beyond TopoMotionFix, we also evaluate TAME in a user study on 5 randomly selected Mixamo characters, which belong to neither the training rigs nor the held-out rigs. In this user study, the ratings that participants give to TAME on these characters are close to its ratings on the SMPL skeleton and the held-out rigs (alignment 2.30 vs 2.31, plausibility 2.14 vs 2.19, source preservation 2.20 vs 2.21), and participants rate TAME above AnyTop on all three criteria. The complete setting and results of the user study are given in Appendix F (Table 4).

Since these rigs come from different datasets, their joint names are not always consistent, and a rig that is unseen during training may name its joints in yet another way. We therefore normalise the joint names of every rig into a common form before the T5 encoding, so that the name of the same anatomical joint carries the same joint semantics on every rig. The new Appendix G (Joint Name Normalisation) describes this normalisation, which handles three cases.

First, the same joint name may be written differently. We rewrite such names into one common form in three steps, namely cleaning, side expansion, and index removal. Cleaning removes a prefix of the rig or the authoring software and separates words that are joined without spaces, so that mixamorig:Hips becomes "Hips" and LeftForeArm becomes "Left Fore Arm". Side expansion writes an abbreviated body side in full, so that LHipJoint becomes "Left Hip Joint". Index removal replaces the numeric index of a spine or neck segment with the word Segment, so that Spine1 to Spine4 all become "Spine Segment".

Second, two naming conventions may use different words for the same anatomical joint. SMPL and SMPL-X name each joint after its anatomical location, whereas Mixamo gives each joint the name of the body part that it moves. The joint at the anatomical shoulder is therefore named left_shoulder in SMPL-X and LeftArm in Mixamo. Our normalised names are closer to the Mixamo convention, because an editing instruction such as "raise the left arm" often refers to the body part that should move, and its words then match the name of the joint to be edited. The names of a rig that follows another convention are converted through a correspondence table, and Table 5 in Appendix G gives the complete table for the body joints of SMPL and SMPL-X.

Third, a joint may have no usable name. Its name can then be inferred from its neighbouring joints or its counterpart on the opposite body side, or specified by the user.

Appendix G gives the complete rules and examples for all three cases.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their time and their constructive comments. We are encouraged that the reviewers find text-driven motion editing across heterogeneous humanoid skeletons to be a novel and important problem (Reviewers DUGE, qCPP, and yHT2). We have uploaded a revised version of the paper, in which all changes are marked in blue. We reply to each reviewer in a separate comment. Here we summarise the main changes to the paper and the key points of our replies.

1. Larger-scale user study (Reviewer qCPP, W2). We have conducted the user study again at a larger scale (Appendix F, Table 4). Each study now has 40 participants, and source preservation is added as a third rating criterion besides alignment and plausibility. The candidates now include OmniME and the ablated variants of TAME, and the unseen-topology study additionally covers 5 randomly selected Mixamo characters. TAME receives the highest ratings among the methods on all three criteria under the seen-topology protocol, under the unseen-topology protocol, and on the Mixamo characters.

2. Generalisation to unseen skeletons (Reviewer qCPP, W3; Reviewer yHT2, W2). The 9 held-out rigs appear only in the test set and are excluded from the character pool that the skeleton augmentation samples from (Algorithm 1, Appendix A and Appendix B). We have added Figure 5 in Appendix B, which visualises the SMPL skeleton and the 9 held-out rigs in their rest poses. The user study also covers 5 Mixamo characters outside the training and held-out rigs, and the ratings of TAME on them are close to those on the SMPL skeleton and the held-out rigs.

3. Rest pose and the adaptation of AnyTop (Reviewer DUGE, W1 and Q1). In text-driven motion editing, the source motion is an input and shares its skeleton with the target motion, so the source motion takes over the role that the rest pose plays in topology-agnostic generation and retargeting approaches. Sec. 2.2 now includes this discussion. Our reply to Reviewer DUGE gives this analysis in detail, and the new Appendix H describes how AnyTop is adapted to editing.

4. Missing and inconsistent joint names (Reviewer DUGE, W1; Reviewer yHT2, W3). We have added Appendix G (Table 5), which describes how joint names are normalised before the T5 encoding. Names that are written differently across rigs are unified by cleaning, side expansion, and index removal (e.g., LHipJoint becomes "Left Hip Joint"). When two naming conventions use different words for the same anatomical joint, such as shoulder in SMPL and Arm in Mixamo, a correspondence table maps them to the normalised names. The name of a joint without a usable name can be inferred from its neighbouring joints or its counterpart on the opposite body side, or specified by the user.

5. Novelty of our method (Reviewer yHT2, W1). In our reply to Reviewer yHT2, we explain that Edit-Focused Representation Alignment (EFRA) is the first to bring representation alignment to motion editing, and that it is tailored to this task rather than applied off the shelf. We also explain that TAME has a dedicated design for each of the three ways in which humanoid skeletons vary: the number of joints, their naming, and the skeletal hierarchy.

6. Multi-seed evaluation (Reviewer DUGE, W3). On MotionFix, TAME is sampled 20 times with different seeds, and its generated-to-target R@1 is 79.70 ± 0.85. On TopoMotionFix, TAME, its two ablated variants, and the adapted AnyTop are each run 20 times. The generated-to-target R@1 is 72.84 ± 0.81 for TAME, 71.42 ± 0.39 for TAME w/o EFRA, 68.69 ± 0.66 for TAME w/o TCSP, w/o EFRA, and 68.54 ± 0.53 for AnyTop. The standard deviation is between 0.4 and 0.8 for every method, and the gain of TCSP (2.7 to 2.9 points) is more than three times the standard deviation of any row.

7. Concern about self-penetration in the supplementary video, and Figure 1 (Reviewer qCPP, W1 and W4). Our reply to Reviewer qCPP analyses this self-penetration and shows that the retargeting has a negligible effect on the retrieval metrics (Table 3). We have also updated the visualisation in Figure 1.

8. Wording of the EFRA claim (Reviewer DUGE, W2). We have replaced "faithful to the instruction" in the abstract and in the contribution list with "improving the alignment of the edited motion with the instruction", which our retrieval metrics and the user study measure directly.

9. Ethics concern on TopoMotionFix (Reviewer DUGE). Sec. 4.1 now cites each of the five motion-capture datasets that provide the rigs (LAFAN1, ACCAD, TotalCapture, PFNN, and SFU). The ethics statement now states that any work that uses TopoMotionFix should cite MotionFix, AMASS, and these five datasets in addition to our paper, and that the released dataset will list every source with its reference and licence.

We hope that the revision and our replies address the concerns of the reviewers, and we are happy to answer any further questions.

AuthorsOpenReview ↗

Follow-up on Our Author Responses

Dear Reviewers,

Thank you again for your time and your constructive reviews. We have posted a detailed reply under each review, together with a general response and a revised paper in which all changes are marked in blue.

If you have a chance to read our replies before the discussion closes on November 18, we would greatly appreciate your feedback, and we would be happy to clarify any remaining questions. If our replies address your concerns, we would be grateful if you would consider revisiting your scores.

Thank you for your time.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.