acceptodds
Withdrawn from ICLR 2027

DanceFields: Learned Deformation Fields for Few-Keypose Dance Style Transfer

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Even for the same choreography, different dancers leave distinct impressions—style is not an accessory, but a core component that defines dance. While motion style transfer has predominantly focused on locomotion with consistent stylistic expressions across movements, dance fundamentally breaks this universality: the same style manifests through entirely different body parts and dynamics depending on the choreography. We formalize this choreography-dependent realization as the reading of a style and introduce the task of few-shot dance style transfer, which transfers expressive style while strictly preserving choreographic content. Because this artistic interpretation is extrinsic to visual pose features, it cannot be inferred from content alone and must instead be observed. We propose DanceFields, a representation that disentangles shared choreographic content from choreography-specific style readings. DanceFields learns a shared bank of content-conditioned deformation fields across choreographies, mapping a style to linear coordinates solved analytically in closed form via meta-learning. By observing merely three still keyposes of the target choreography in the desired style, our model resolves the style coordinates without per-style or test-time optimization. Furthermore, DanceFields explicitly models time as style, decoupling local temporal warping from global tempo variations. On a newly captured dataset of 30 choreographies across 16 styles (Laban Effort axes and character traits), DanceFields significantly outperforms per-style-trained baselines in style fidelity, content preservation, and physical plausibility. Crucially, representing style as coordinates naturally unlocks intuitive arithmetic controls—including continuous intensity grading, interpolation, and part-wise compositional editing—and demonstrates robust zero-shot generalization across unseen dance choreographies and even cross-domain locomotion benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dance Style Transfer, Motion Style Transfer, Character Animation, Laban Movement Analysis, In-Context Learning, One-Shot, Human Motion

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

This paper focuses on unseen dance restyle. It adopts concepts from Labanotation motion analysis. Given three keyframes, the method uses LMA to discretize motion into multiple LMA-based attributes and then pretrains an MLP-based field to predict the residual between the input dance and the target style dance, thereby completing the restyle task.

Critical strengths
  • The use of LMA is interesting. Many works have shown that Labanotation offers advantages for motion understanding and modeling.
  • The claim that the method can transfer an unseen style using only three keyframes is ambitious. If the paper has indeed solved this task, I would be very excited.
  • Although the writing is poor, the paper consistently tries to distinguish itself from existing work, and want to tell something to the audience. I appreciate it.
Critical weaknesses

The introduction is very poorly written. Fortunately, I am familiar with Labanotation and phase representation; otherwise, I would have had no idea what the paper was discussing. The main text tries to highlight the core differences between the proposed framework and existing style transfer methods. I appreciate this effort, and the proposed ideas are interesting and exciting. However, the English writing is poor. Sentences are over-compressed and readability is low.

The related work should at least introduce what Labanotation is. Alternatively, the appendix should describe LMA, including what it is, where it can be used, and why it is effective.

The English writing in the methodology is even more confusing. It does not read as academic English, and I found it difficult to follow. What does "no support" in Figure 2 mean? The figure gives me the impression that it was produced with an AI-assisted research tool. Overall, the writing is too poor. Even if the idea is strong, I cannot accurately judge from the methodology description whether the method works or what its defects are. I simply cannot understand it.

The main text mentions a newly collected Laban dancing dataset. I am willing to treat this as a contribution, but I searched the entire paper with CTRL+F and found no related content. How was this dataset constructed? How can the fidelity of Laban information extraction be guaranteed? These points are unclear. These critical problems lead me to a clear rejection.

From the supplementary video, my main concern is that the results are not good. I kept observing motions under the "sexy" label. The hands move abruptly, which makes the overall motion poorly coordinated. Visually, the restyle results are not intuitive, and the method does not achieve the stylization seen in methods on the XIA dataset. The sentences in the appendix are not grammatical at all. Overall, the paper falls below the acceptance bar.

Questions for authors

Please answer the questions I mentioned in Critical Weaknesses section.

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

DanceFields represents style through coefficients over shared, content-conditioned deformation fields, solved from a few target keyposes by ridge regression. Separate timing fields model local warping and global tempo, evaluated with captured dance data, user studies, and cross-dataset transfer. The representation is promising, but the practical support budget, unseen-style protocol, and evaluation ambiguities affect central claims. Successful support-only alignment, style-disjoint adaptation, and harmonized comparisons would substantially improve my assessment.

Critical strengths
  1. Separating choreography from execution quality, including timing, is useful for motion authoring.
  2. Shared fields and closed-form adaptation provide a compact, editable representation with substantial solver and representation ablations.
  3. Identity references, alignment sensitivity, and cross-domain results help expose the method's operating conditions.
Critical weaknesses
  • The three-keypose input budget excludes alignment information. Appendix A obtains keypose correspondence from a frozen DTW alignment of complete neutral–styled pairs, and a five-frame offset reduces Effort cosine from 0.607 to 0.222 (Table 7). The current evidence therefore establishes sparse adaptation with accurate phase labels; an annotation or matching protocol using only deployment-available inputs is needed.
  • Unseen-style adaptation is insufficiently specified. The six-fold split holds out choreographies, while the method uses style-specific mean fields and spatial/timing priors computed from training performances (§§3.2–3.3). MOCHA is useful external evidence, but the construction of these quantities for a genuinely new style—particularly timing from static keyposes—requires an explicit style-disjoint protocol and prior ablations.
  • Metric and pipeline ambiguities weaken the comparisons. Table 1 reports relative jerk changes of −444%/−683%, which are impossible under the usual relative-change formula for a nonnegative jerk magnitude; provide the actual formula and raw values. All baselines also lose to Identity on SC, motivating a matched postprocessing and root-handling audit before attributing the gap to the learned fields.
  • Missing references:

Zhang J, Kang Z, Liu L, et al. OpenDance: Multimodal Controllable 3D Dance Generation with Large-scale Internet Data[J]. arXiv preprint arXiv:2506.07565, 2025.

Li J, Kang D, Pei W, et al. Audio2gestures: Generating diverse gestures from audio[J]. IEEE Transactions on Visualization and Computer Graphics, 2023, 30(8): 4752-4766.

Questions for authors
  • How are phases obtained when the full target performance is unavailable and only three keyposes plus feasible user annotations remain?
  • What produces the negative jerk percentages, and do the gains survive common projection, smoothing, and root treatment?
Reviewer_E3AvOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces few-shot dance style transfer and proposes DanceFields, which disentangles choreographic content from choreography-dependent style representations. By observing only three key poses, the method analytically infers style coordinates without per-style or test-time optimization, while explicitly modeling temporal dynamics. Experiments demonstrate improved style fidelity, content preservation, and physical plausibility, alongside intuitive style editing and generalization to unseen choreographies and locomotion tasks.

Critical strengths
  1. The problem of transferring few keyposes to new subject (or body) for choreography is interesting.

  2. A unified representation that disentangles shared choreographic content from choreography-specific style readings is proposed.

Critical weaknesses
  1. What is the difference between the few keypose transfer and few-shot keypoint detection [a-c]? And also the difference to motion transfer [d,e]. Their connection and difference should be described for differentiation.

  2. What is the core contributions of DanceField?

  3. The dance content and dance style have vague definitions. The central challenges of few-shot keypose transfer for choreography should be clearly illustrated.

[a] Few-shot Keypoint Detection with Uncertainty Learning for Unseen Species @ CVPR 2022

[b] OpenKD: Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection @ ECCV 2024

[c] GKDT: General Keypoint Detection Transformer @ ECCV 2026

[d] Towards High-Quality 3D Motion Transfer with Realistic Apparel Animation @ ECCV 2024

[e] Learning High-Fidelity Garment Deformation via Skinning-Free Image Transfer @ 3DV 2026

Questions for authors

See Critical Weaknesses

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

Related papers

Discussion (0)

Sign in to comment.