acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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 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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text-driven Motion Editing, Source Preservation Denoising, Preservation head

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VDwwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates text-driven human motion editing. Given 3D human motion and a text prompt, the model produces a final motion that preserves the original motion as much as possible while satisfying the text prompt. The paper identifies detail attenuation that happens during denoising as part of the problem that washes away motion details in the unedited part. It proposes SuperMotion to temporally align the source motion at each denoising step using a learned blending gate. Experiments on MotionFix and STANCE Adjustment show improvement over the baselines.

Critical strengths
  • The paper nicely illustrates why existing diffusion editors look overly smooth by showing the second-order temporal difference ratio across denoising steps, which provides a clear motivation for this work.
  • The methods demonstrate better performance compared to the baseline on both datasets.
  • The paper also provides a clear motivation and ablation for each design, e.g., why the blending should be done in x_0 space rather than the velocity space.
Critical weaknesses
  • At its core, the model is a minor modification of SimMotionEdit, using the pre-trained SimMotionEdit weights with an additional head to predict soft gating.
  • The paper is framed around the fact that blending between the source and target motion is necessary at every step. However, from the Preservation strength ablation, setting the blending strength to 0 already produces 33.0% FR@1, almost similar to the full model with blending, which achieves 33.20%. Combined with Table 3, the model fine-tuned to predict v already achieves 31.03 FR@1. This suggests that most of the gain comes from fine-tuning the model to predict v and using the additional losses, rather than the blending itself.
  • Overall, the papers report marginal gains over SOTA OmniME. Comparing the qualitative results, they are almost indistinguishable from OmniME, with the proposed method sometimes showing more skating, possibly due to the blending of the source motion when the global positions are different.
Questions for authors
  • What happens to the detail ratio when /gamma = 0? Since /gamma=0 in Figure 5 already achieves 33.0% FR@1 (vs. 33.20% at /gamma=1), what is the median detail ratio at /gamma=0, and what is the detail ratio of the "w/o gate" ablation in Table 3? Also, how does the "w/o gate" baseline perform if you apply L_hf directly to x0 during training?

  • Why does FR@1 drop so much on STANCE Adjustment? Can you explain why SuperMotion falls behind both OmniME (18.37% vs. 22.45%) and SimMotionEdit (20.92%) on STANCE FR@1 and FR@2 in Table 2?

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

This paper presents SuperMotion, a source-preserving denoising framework for text-driven 3D human motion editing. The authors identify temporal-detail attenuation during diffusion denoising as a limitation of existing motion editors, which can lose motion dynamics and fail to preserve unedited content. SuperMotion addresses this issue by temporally aligning the source motion and introducing a learned, per-frame and per-feature preservation gate. At each reverse diffusion step, a clean-space source anchor blends the denoiser's clean-motion estimate with the aligned source according to the predicted gate. The training objective combines the standard velocity prediction loss with a source-anchor loss and a temporal high-frequency loss based on second-order temporal differences, without requiring explicit edit masks. The method is evaluated on MotionFix and STANCE Adjustment using motion retrieval metrics and qualitative comparisons. It achieves 33.20% full-pool R@1 on MotionFix and 46.88% batch R@1 on STANCE Adjustment. Ablation studies examine the contributions of the preservation gate, training objectives, and architectural components.

Critical strengths

The paper identifies temporal-detail attenuation as a practical problem in diffusion-based motion editing and addresses it by blending the predicted clean motion with a temporally aligned source at each reverse diffusion step. The learned per-frame, per-feature preservation gate allows the model to balance editing and preservation without requiring explicit edit masks. This is a useful design that directly addresses the paper's motivating problem.

The experiments on MotionFix and STANCE Adjustment provide evidence that the proposed source anchor and its associated training objectives can improve motion-editing performance. The ablations show measurable degradation when the preservation gate or either auxiliary loss is removed, while the copy-source baseline and preservation-strength analysis help distinguish genuine editing improvements from simply retaining the input motion. These analyses make the empirical contribution more informative than a comparison based on retrieval scores alone.

Critical weaknesses

The main quantitative results rely on motion retrieval metrics, while the proposed detail ratio measures second-order temporal variation rather than whether the requested edit is correct. The paper itself notes that this ratio can be close to one simply by copying the source motion, as illustrated by MDM-BP. Although the authors include source-distance analyses and qualitative examples, there is no human evaluation or independent assessment that directly separates instruction compliance from source preservation. This makes it difficult to determine how reliably SuperMotion achieves both goals.

The evaluation mixes results reported by previous papers with results obtained by running official implementations, which raises questions about the consistency of the comparison protocols. Moreover, SuperMotion does not lead on every metric: on STANCE Adjustment, its full-pool R@1 is 18.37%, compared with 22.45% for OmniME. The ablation studies also use a single seed, with no reported variation across independent runs. These limitations make the broad claim of consistent superiority over existing methods less convincing.

Questions for authors

1. How do you distinguish successful editing from simply preserving the source motion?

The detail ratio measures temporal variation but does not establish that the requested edit has been correctly performed. The paper also shows that increasing the preservation strength can improve retrieval against the source more than against the target. Do you have a direct evaluation of instruction compliance and source preservation, such as a human study or separate quantitative metrics? Such evidence would help establish whether the proposed method improves editing quality rather than primarily retaining the input.

2. Are the comparisons with prior methods conducted under consistent evaluation protocols?

The appendix states that some MotionFix results are taken from the original papers, while the STANCE results combine reported numbers with evaluations using official implementations. Were all methods evaluated using the same test samples, retrieval gallery, motion representation, and evaluation settings? In addition, can the authors clarify the claim of consistent superiority given that SuperMotion has lower full-pool R@1 than OmniME on STANCE Adjustment? Clarifying

Reviewer_63qaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SuperMotion edits an input motion using a textual instruction while retaining compatible source content. A learned gate blends each clean-motion prediction with a linearly time-aligned source before the reverse posterior update. The gate varies by frame and feature, is trained without explicit edit masks, and is supervised through anchor reconstruction and second-temporal-difference losses. The implementation also uses velocity prediction and a Skip DiT backbone. Experiments on MotionFix and STANCE Adjustment report improvements in several retrieval measures and temporal-detail statistics. The design is coherent, but the small leading margin and the limited effect of disabling the inference anchor leave uncertainty about the strength and mechanism of the contribution.

Critical strengths

S1. Explicit source reuse addresses a concrete editing requirement

Equations 5–8 (pp. 5–6) define a bounded blend between the clean prediction and an aligned source, applied after guidance. This gives a clear interpretation to the preservation head and requires no annotated edit mask. The same-schedule fine-tuning control without the gate reaches 29.45% full-pool R@1 versus 33.20% for the complete method (Table 3 and discussion, p. 8), supporting the value of the training setup beyond simply running more optimization.

S2. The paper includes useful controls for copying and temporal detail

The copy-source baseline is strong and is reported alongside learned editors (Table 1, p. 6). Table 4 (p. 13) reports detail ratio, distance to target, distance to source and the fraction closer to source. SuperMotion has target distance 1.586 versus 1.642 for OmniME and detail ratio 1.035. The authors correctly state that matching this statistic does not establish correct editing, and Figure 5 (p. 9) exposes how increasing preservation can improve retrieval through copying.

Critical weaknesses

W1. The leading editing margin is small and its stability is not established

MotionFix full-pool R@1 is 33.20% versus 32.02% for OmniME (Table 1, p. 6). Appendix A (p. 12) specifies 506 full-pool queries, so this 1.18-point margin corresponds to roughly six queries. The main comparison lacks uncertainty across training seeds, and the component ablations use the same seed (Section 4.3, p. 8). On STANCE, full-pool R@1 is 18.37%, below OmniME's 22.45%, despite better average rank and several other metrics (Table 2, p. 7). This supports a tradeoff, but not a uniform leading result on both benchmarks.

W2. The necessity of the proposed inference operator is unclear

Figure 5 and its discussion (p. 9) report approximately 33.0% target R@1 when the learned model's anchor is removed at inference, almost unchanged from 33.20% with the anchor. This suggests that much of the gain may be a training effect. It does not make the training contribution invalid, but it weakens the explanation that explicit source injection at every reverse step is essential. Velocity prediction, anchor supervision and the high-frequency objective also change together relative to the parent method (pp. 5–6); the current controls do not fully determine which training change explains the improvement.

Overall assessment

2 / 4. Weak rejection. This is close to the acceptance boundary. The method is sensible and the copy-versus-edit analysis is useful, but the leading margin lacks a stability assessment and the main inference mechanism can be disabled with little change in the reported retrieval score. Evidence resolving those two points could move my recommendation to weak acceptance.

Questions for authors

Q1. When the inference anchor is removed as in Figure 5 (p. 9), what happens to the detail ratio, target distance and preservation metrics in Table 4 (p. 13)? If these also remain similar, should the contribution be framed primarily as an anchor-based training method rather than a necessary sampling-time correction?

Q2. How stable is the 33.20% versus 32.02% MotionFix result (Table 1, p. 6) across training seeds and paired test queries? Please clarify whether any repeated-run or paired uncertainty estimates are available. The mixed STANCE result (Table 2, p. 7) also needs a clear account of which aspect of editing improves.

Q3. Can the training effect be isolated with the same backbone, velocity parameterization, guidance and high-frequency objective, varying only anchor supervision and the inference blend? Please also separate test-only temporal diagnostics from the 836-file aggregate that includes validation outputs (Appendices A and C, p. 12), so that these diagnostics support the same held-out editing claim as the retrieval results.

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

Related papers

Discussion (0)

Sign in to comment.