acceptodds
Under review as a conference paper at ICLR 2027

PhysIKDance: Contact- and Slide-Aware Dual-Space Music-to-Dance Generation via Probabilistic Inverse Kinematics

Abstract

Audio-driven 3D dance generation is essential for virtual avatars and immersive digital experiences. Despite recent progress, methods that directly predict joint rotations often accumulate errors along the kinematic chain, leading to foot skating, ground penetration and inconsistent support. To tackle the issue, we propose PhysIKDance, a simple yet effective dual-space framework decouples physical reasoning from rotation reconstruction, integrating physically supervised geometric generation with music-conditioned rotation recovery for high-fidelity dance motion generation. In Joint Coordinate Space, a diffusion model jointly predicts root translation, joint coordinates, axial twist, and foot-interaction states. A state-aware objective couples root and local-foot motion to suppress drift during planted support while preserving intentional sliding, complemented by constraints on support height and ground penetration. In Joint Rotation Space, a music-conditioned, two-stage invertible inverse-kinematics module recovers joint swing and combines it with axial twist to produce SMPL-compatible rotations. Cross-space consistency supervision further connects geometric plausibility with accurate rotation reconstruction. Extensive experiments on AIST++ and FineDance demonstrate that PhysIKDance substantially improves foot–ground consistency and overall physical plausibility while achieving competitive motion quality, musical alignment, and motion diversity. Code will be released upon acceptance.

Primary area:
generative models
Keywords:
human motion generation, human dance generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kJDNOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes PhysIKDance, a music-driven dance generation framework designed to mitigate error accumulation along the kinematic chain. Instead of directly generating joint rotations, the method first generates motion in Joint Coordinate Space, including root translation, joint coordinates, axial twist, and foot-interaction states, which enables direct geometric supervision of foot-ground interactions. A music-conditioned two-stage invertible inverse-kinematics module then reconstructs SMPL-compatible rotations in Joint Rotation Space, with cross-space consistency supervision connecting the two representations. Experiments on AIST++ and FineDance demonstrate improvements in motion quality and foot-ground consistency.

Critical strengths

Strengths

  1. The paper identifies an important limitation of directly predicting continuous 6D joint rotations: local rotational errors can accumulate along the kinematic chain and cause substantial positional errors at distal joints such as the feet. The proposed dual-space representation directly targets this issue by separating geometric generation from rotation reconstruction.

  2. The dual-space design is intuitive and effective. Direct supervision in Joint Coordinate Space provides a natural way to regulate foot-ground interactions, while the inverse-kinematics stage restores animation-ready SMPL rotations. The effectiveness of this design is supported by improvements on most core quality and physical-consistency metrics.

  3. The paper is motivated by a practical problem and is clearly illustrated. Figures 1 and 2 are particularly helpful in explaining why rotation-space errors propagate and how the proposed coordinate-space generation and rotation-space recovery interact.

Critical weaknesses

Weaknesses

  1. Although the demo generally shows better foot-ground consistency for PhysIKDance, some examples appear to exhibit unnatural upper-arm deformation or squeezing. Since axial twist is generated explicitly and then used during rotation reconstruction, it is unclear whether these artifacts are related to inaccurate twist estimation or to the inverse-kinematics recovery stage. This raises the question of whether improved foot-ground consistency may occasionally come at the cost of upper-body rotational realism.

  2. The video does not provide qualitative ablations for the core variants in Table 3. In particular, visual comparisons with direct/global rotation prediction, removing IK, or removing twist would help clarify what types of artifacts are specifically addressed by the dual-space design. Table 3 quantitatively shows large differences among these variants, but the perceptual effects are difficult to infer from metrics alone.

  3. The terminology around “physical reasoning” appears somewhat overstated. The proposed physical supervision is primarily kinematic: it constrains foot positions, displacements, support states, and penetration, but does not model contact forces, torques, friction, dynamic balance, or physical executability. The paper itself acknowledges this limitation in the appendix. I therefore suggest using more precise terms such as kinematic physical supervision, contact-aware geometric supervision, or kinematic physical plausibility when describing this component, while reserving stronger “physical reasoning” claims for methods that explicitly model dynamic interactions or physical feasibility.

Questions for authors

Questions

  1. Some demo examples appear to show unnatural upper-arm deformation or squeezing. Are these artifacts related to inaccurate axial-twist prediction, the IK recovery stage, or inconsistency between generated joint coordinates and recovered rotations? Have the authors evaluated upper-body rotation or mesh deformation quality separately from foot-ground metrics?

  2. Could the authors provide qualitative ablations for the main variants in Table 3, especially direct/global rotation prediction, w/o IK, and w/o twist? These comparisons would help illustrate how the dual-space design changes both foot-ground behavior and whole-body motion quality.

  3. Since the method explicitly predicts axial twist as a separate component, how sensitive is final motion quality to twist errors? A quantitative or qualitative analysis of twist prediction error versus downstream mesh or rotation artifacts would help clarify the robustness of the reconstruction stage.

3 comments

We thank the reviewer for recognizing the motivation and effectiveness of the dual-space design. We supplement the manuscript’s ablations and limitations with upper-arm diagnostics, controlled twist comparisons, and synchronized visualizations.

W1 and Q1: What causes upper-arm deformation, and has upper-body quality been evaluated separately from foot-ground metrics?

Response: We added upper-arm intersection, mesh-deformation, and rotation-recovery evaluations independent of foot-ground metrics. Using the same SMPL model and body shape, we evaluate the same 100 FineDance clips from 18 sequences:

MethodIntersection frames (%)Intersecting faces (%)Upper-arm edge strain
GT61.26672.62800.12415
Direct rotation + FK67.60673.15810.12910
Contact consistency58.16002.01130.12428
Without sliding-specific losses57.44671.99730.12347
PhysIKDance56.31332.04560.12351

Intersection detection excludes shared-vertex face pairs and a fixed shoulder-connection region. Edge strain measures local edge-length changes relative to the same-shaped rest mesh.

Relative to direct rotation, PhysIKDance reduces the intersection-frame percentage by 11.29 percentage points and the intersecting-face percentage by 35.2%, with mean edge strain also closer to GT. After excluding boundary-only contact, the intersection advantage persists across all tested geometric crossing margins from 0 to 2 mm.

We additionally hold GT geometry, music, and IK weights fixed and change only the input twist:

IK inputUpper-arm PVE ↓ (mm)Upper-arm rotation error ↓ (°)
GT geometry + GT twist19.488.96
GT geometry + zero twist42.0942.03

Correct twist reduces upper-arm PVE by 53.7%, demonstrating its importance for local reconstruction. Together, these evaluations support improved upper-arm metrics relative to direct rotation and the value of twist information for recovery; they do not rule out twist-prediction errors as a contributor to residual squeezing. As §3.3 and Appendix A.10 explain, inconsistency between generated geometry and the fixed skeleton, together with IK recovery, may also contribute.

W2 and Q2: Qualitative comparisons of the core components.

Response: We supplement Table 3, Fig. 3, and Appendix A.9 with synchronized component-level videos. The videos use shared cameras and body shape, with full-body and enlarged upper-body views. They compare PhysIKDance with direct rotation + FK, Stage-1 skeletons with final meshes, and predicted with zero twist using identical generated joints and IK weights. The direct-generation comparison differs from Table 3’s global-versus-local rotation-recovery ablation. The latter two comparisons visualize the pre-IK representation and a controlled twist-input intervention, rather than independently retrained w/o IK and w/o twist models.

Video: https://openreview.net/attachment?id=bSMYqzbdST&name=supplementary_material W3: Precision of the physical terminology.

Response: We agree that “contact-aware geometric supervision” is more precise and will use it consistently. Appendix A.10 already identifies the supervision as kinematic and excludes forces, torques, friction cones, dynamic balance, and biomechanical feasibility; §2 likewise distinguishes kinematic constraints from simulation-based methods. Equations (5) and (41) constrain geometric quantities, not dynamic executability. This terminology aligns the broader wording with the method’s stated scope: improved foot-ground kinematic consistency and motion quality, with residual upper-body artifacts acknowledged.

Dear Authors,

Thank you for your detailed response and clarifications.

The experiments on upper-arm deformation and qualitative ablations have addressed my concerns. I maintain my original score and lean towards acceptance.

Thank you for your time and thoughtful feedback. We are pleased that the upper-arm deformation experiments and qualitative ablations have addressed your concerns. We sincerely appreciate your positive assessment and support for our work.

Reviewer_xskaOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces PhysIKDance, a two-stage music-to-dance generation framework aiming to address unrealistic foot–ground interactions (e.g., foot skating and ground penetration) caused by error propagation in joint rotation-based motion generation. The key idea is to decouple motion generation in joint coordinate space from rotation reconstruction. Accordingly, a music-conditioned diffusion model first generates joint coordinates and axial twists, with explicit foot-contact supervision to improve planted or supported sliding foot movements. A subsequent two-stage invertible inverse kinematics (IK) module converts the generated joint coordinates and twists into valid joint rotations. Experiments on FineDance and AIST++ demonstrate improvements in physical and foot plausibility over existing baselines.

Critical strengths
  • The paper addresses an important challenge in music-to-dance generation, where inaccurate foot–ground interactions often lead to unrealistic movements such as foot skating and ground penetration. The proposed separation between joint coordinate generation and rotation reconstruction is meaningful and intuitive, since it allows direct supervision of foot movements while maintaining valid rotations to alleviate error accumulation of forward kinematics from rotation to joint positions.
  • The experiments on FineDance and AIST++ datasets show good improvements over existing baselines, particularly in terms of foot quality. These results demonstrate the practical potential of the proposed framework for improving the quality of generated dance motions.
Critical weaknesses
  • The technical contribution and specific motivation for the music-to-dance generation are not sufficiently clear and convincing. The problem of error propagation in joint rotation-based representations has been studied in prior motion literature (for example, QuaterNet, BMVC 2018), while the swing-twist decomposition and two-stage invertible IK module are adapted from NIKI.
  • The main physical plausibility claim seems somewhat overstated, while the framework is mainly designed around foot supervision and the evaluation mainly focuses on foot movement metrics without considering other important aspects such as motion smoothness or temporal consistency, or plausible pose angle. It remains unclear whether the reported improvements in foot quality metrics could indicate better overall physical plausibility of the generated dance motions.
Questions for authors
  • I think the proposed designs can generally be applied to other motion generation tasks, such as text-to-motion. Could the authors clarify what challenges specific to music-to-dance generation motivate this framework, and what distinct advantages or insights the paper offers over existing approaches?
  • Since the IK module already receives joint coordinates and axial twist information (which is sufficient to reconstruct rotations), could the authors further justify the necessity of the music conditioning during rotation reconstruction?
4 comments

We thank the reviewer for recognizing the value of separating geometric generation from rotation reconstruction. We clarify the contribution and the role of music using the manuscript’s formulation and ablations.

W1 and Q1: What motivates the framework specifically for music-to-dance generation, and what does it contribute beyond existing approaches?

Response: The framework first addresses error propagation in rotation-based generation. Local rotations are recursively composed through FK, so small rotational errors can produce substantial positional deviations at distal joints such as the feet. The second paragraph of the Introduction and Fig. 1 explain this challenge. Controlling foot-ground relationships through rotation variables is particularly relevant to dances involving body translation and changes in support state, motivating our dual-space design.

We therefore generate joint coordinates and axial twist first, then recover rotations through IK. Section 3.1, Eq. (1), defines the intermediate representation comprising contact/slide states, root translation, joint coordinates, and axial twist; §3.3, Eqs. (7)–(10), describes rotation recovery. This places geometric supervision directly in coordinate space, reducing reliance on recursive rotation composition during generation, while FK consistency constrains the recovered motion (Appendix A.4.5, Eq. (41)). We agree that this formulation could extend to text-to-motion and other generation tasks, which we regard as a promising direction for future work.

Dance also requires distinguishing stationary support from supported sliding rather than penalizing all foot movement. Section 3.2, Eqs. (2) and (5), expresses world-space foot displacement as the sum of root and local-foot displacement. These contributions should cancel during planted support and match reference displacement during supported sliding. The music and motion-state condition in §3.3, Eq. (6), further provides temporal context for recovery.

The manuscript and additional experiments support this design. Table 1 reports FineDance FIDₖ/FIDɡ of 13.162/5.239. In the additional configuration comparisons, removing sliding-specific losses gives 21.21/5.95, and contact-consistency supervision gives 20.08/6.64, with PFC values of 0.2521 and 0.2526, respectively, versus 0.2390 for the full model. On 100 shared clips, mean endpoint drift within GT-planted intervals decreases from 7.03 cm for direct rotation + FK to 6.07 cm for PhysIKDance. These results support the benefits of the proposed representation and supervision for motion quality and foot stability.

Q2: Why condition IK on music when coordinates and twist already provide rotation information?

Response: Music supplies context for learned recovery from imperfect geometry, rather than information required by exact analytical IK. Section 3.3 explains that generated coordinates and twist may not form a configuration realizable by the prescribed skeleton. Equation (6) combines temporal music features with contact/slide states, root motion, and root displacement as a shared recovery condition. The error-augmented mappings in Eqs. (7)–(10) accommodate geometric inconsistency, while Appendix A.5 progressively increases the contribution of generated inputs during training. This is the intended role of music beyond the geometric variables themselves.

The relevant comparisons are already reported in Table 3:

ConfigurationFIDₖ ↓FIDɡ ↓PFC ↓
Without direct IK music conditioning13.42529.4480.5939
Without rotation-to-joints training15.69020.5770.3676
PhysIKDance13.1625.2390.2390

The no-music configuration retains Stage-1 music conditioning and removes only the direct audio input to recovery. The no-reverse configuration removes rotation-to-joints supervision while retaining the invertible recovery path. The full configuration achieves better FIDₖ, FIDɡ, and PFC. These are configuration-level comparisons: their evaluated Stage-1 outputs differ, so they do not isolate either component’s effect on fixed inputs.

W2: Scope of physical plausibility and evaluation beyond foot metrics.

W2: Scope of physical plausibility and evaluation beyond foot metrics.

Response: Our evaluation extends beyond foot metrics. Section 4.1 and Table 1 report FID, motion diversity, and activity statistics. On FineDance, PhysIKDance achieves Divₖ/Divɡ of 8.829/6.894, compared with GT’s 9.709/7.179, and an active-joint ratio of 0.833, compared with 0.850 for GT. These results provide evidence of retained motion diversity and activity alongside improved foot metrics. The user study in §4.4 and Appendix A.8 further evaluates expressiveness and perceived physical plausibility.

For jitter and smoothness, joint-velocity and temporal-twist supervision in Appendices A.4.1 and A.4.3 is designed to constrain unnatural frame-to-frame changes. However, these training objectives and FID/Div are not direct jitter measurements, so we do not claim comprehensive validation of temporal smoothness. Consistent with the scope already stated in Appendix A.10, we will use contact-aware geometric supervision and describe the demonstrated benefits as improvements in motion quality and foot-ground kinematic consistency.

Thank you for the response and clarifications. Overall I think the paper presents a valuable empirical contribution toward generating high-quality dance motion from music. However, I still have some reservations regarding the limited technical contribution and the broader physical plausibility claims. For example, the main evaluated metrics such as FID, div_k, div_g are statistical metrics computed from motion features, while the active-joint ratio mainly reflects motion activity. These do not directly evaluate the physical properties of the motion. I acknowledge the merits of the proposed framework, so I would like to increase my rating, but lower confidence given the remaining concerns.

Thank you for your positive assessment of our work and your helpful suggestions. We appreciate your concern regarding the evaluation of physical plausibility. Our intention is to use foot-skating rate and support ratio to assess specific aspects of foot–ground interaction, while FID and DIV measure motion quality and diversity in feature space, and the active-joint ratio reflects motion activity. Although complementary metrics of this kind are commonly used in dance generation, we agree that they provide only a partial assessment of physical plausibility. The supplementary videos are intended to offer additional qualitative evidence by allowing visual inspection of foot–ground interactions and overall motion. Your feedback also highlights a valuable direction for future work: exploring simulation-based evaluation, potentially incorporating rewards related to balance, contact stability, and motion tracking to obtain a more comprehensive quantitative assessment.

Reviewer_bKZcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PhysIKDance, a dual-space framework for music-to-dance generation. The method incorporates contact- and slide-aware modeling into motion generation and employs a music-conditioned invertible IK module.

Critical strengths
  1. Clear presentation. The paper is well organized and generally easy to follow, with helpful illustrations of the overall framework.

  2. Simple and reasonable design. The separation of joint-coordinate generation and rotation recovery is intuitive and well motivated by the challenges of foot-ground interaction.

  3. Relevant empirical evaluation. The experiments cover two established benchmarks and include both motion-quality and physical-consistency metrics, providing a useful assessment of the method's performance.

Critical weaknesses
  1. Insufficient validation of the core design. Although Table 3 provides several ablations, the experiments do not sufficiently isolate the contributions of the proposed dual-space formulation and slide-aware supervision. In particular, a controlled comparison against direct rotation generation with equivalent FK-based physical constraints, as well as an ablation of the sliding-specific objective, would better substantiate the proposed design.

  2. The slide-aware claim appears somewhat overstated. EDGE has already considered preserving intentional sliding while reducing foot-skating artifacts through contact-consistency supervision. In this paper, sliding states are determined by geometric thresholds on foot height and velocity, rather than choreographic intent. Therefore, the proposed formulation distinguishes different foot-motion patterns but does not necessarily identify intentional sliding. Its additional benefits over existing contact-aware constraints remain unclear.

  3. Limited differentiation from closely related work. Lodge and PAMD have already addressed foot-ground constraints and the mismatch between joint-position objectives and rotation-based representations. Moreover, swing–twist decomposition and invertible IK are established techniques in HybrIK and NIKI. While their integration into a music-to-dance generation framework is reasonable, the paper could more clearly identify its distinctive methodological contributions and demonstrate their advantages over these prior approaches.

  4. Remaining motion artifacts and limited scope of physical plausibility. Despite the reported improvements, some qualitative examples and demo sequences still appear to exhibit inconsistent foot-ground contact (still floating) and noticeable arm–body interpenetration. The proposed constraints primarily target foot-ground interactions, while other forms of body interpenetration remain unaddressed. Although the paper acknowledges limitations in physical modeling, its broader claims regarding physical plausibility should be interpreted more cautiously.

Questions for authors
  1. Can the authors compare coordinate-first generation with direct rotation prediction under equivalent FK-based physical constraints?

  2. What additional benefit does the sliding-specific objective provide over conventional contact-consistency supervision? Could the authors separately evaluate planted-foot stability and sliding-motion fidelity?

  3. Given that sliding labels are derived from velocity thresholds, how does the method distinguish intentional choreographic sliding from unintended foot skating?

  4. What are the key methodological contributions relative to Lodge, PAMD, and existing IK-based methods?

  5. Could the authors further discuss the remaining floating artifacts and body-interpenetration issues observed in the qualitative results? How broadly should the reported improvements in physical plausibility be interpreted?

4 comments

We thank the reviewer for recognizing the motivation and clarity of PhysIKDance.

W1 and Q1: What advantages does coordinate-first generation offer over direct rotation prediction with FK-based constraints?

Response: Coordinate-first generation makes foot trajectories directly accessible to geometric supervision. The abstract (lines 014–016), Introduction (lines 047–050), and Fig. 1 explain how local rotation errors can propagate through FK and produce positional deviations at distal joints. Generating joint coordinates reduces reliance on this recursive composition during geometric generation, making the relationship between root motion and foot displacement explicit (§3.2, Eqs. (2) and (5)). IK then recovers skeletal rotations, with FK consistency connecting them to the generated geometry (Appendix A.4.5). Table 2 reports improved foot-ground metrics over the evaluated baselines. The additional direct-rotation comparison below provides more targeted evidence: on the same 100 clips, PhysIKDance has lower mean endpoint drift within GT-planted intervals (6.07 versus 7.03 cm). These results support the proposed pipeline without implying that coordinate generation universally outperforms rotation-based methods.

W1–W2 and Q2–Q3: Does sliding-specific supervision provide additional benefits, and do geometric sliding labels identify choreographic intent?

Response: We agree that geometric sliding should not be equated with choreographic intent. Section 3.2 and Appendix A.2 explicitly define the labels in geometric terms. We therefore interpret support ratio, foot-skating ratio (FSR), and grounded foot speed relative to GT, rather than treating all foot motion as an error. The additional comparisons below show that the full configuration achieves lower FIDs and PFC, with aggregate foot statistics closer to GT, than either removing sliding-specific losses or using a contact-consistency control inspired by EDGE. The latter is an internal supervision variant, not a reproduction of EDGE.

ConfigurationFIDₖ ↓FIDɡ ↓PFC ↓Foot speedFSRSupport ratio
GT≈0≈00.15790.13440.34550.5768
Contact consistency20.086.640.25260.21740.85000.5300
Without sliding-specific losses21.215.950.25210.21180.85300.5200
PhysIKDance13.165.240.23900.15100.58700.5659

Foot speed, FSR, and support ratio are interpreted by proximity to GT; bold indicates the best generated result under the stated criteria.

To address the request for state-specific evaluation, we additionally analyze 100 clips from 18 sequences using fixed GT-derived intervals shared by all methods:

ConfigurationPlanted-interval speed ↓ (m/s)Planted-interval endpoint drift ↓ (cm)Sliding speed W1 to GT ↓ (m/s)Sliding displacement W1 to GT ↓ (cm)
Direct rotation + FK0.2547.030.1972.70
Contact consistency0.2776.470.2012.76
Without sliding-specific losses0.2846.930.1982.61
PhysIKDance0.2316.070.2273.10

W1 denotes Wasserstein-1 distance. This diagnostic uses a GT speed threshold of 0.05 m/s, distinct from the 0.12 m/s training-label threshold in Appendix A.2. Intervals are not selected using predicted foot speed.

The full model has lower mean movement within GT-planted intervals, but larger sliding-distribution distances. These results support improved aggregate quality and reduced drift in reference-defined intervals, not superior sliding fidelity. Since generated dances need not follow GT choreography at identical times, this interval analysis is diagnostic. Appendix A.7.4 already explains that FSR includes supported sliding and does not semantically identify unintended skating. We will align the broader “intentional sliding” wording with this definition.

W3 and Q4: What distinguishes PhysIKDance from Lodge, PAMD, and existing IK methods?

Response: Our contribution connects state-aware geometric generation with context-conditioned rotation recovery. As discussed in §2, Lodge uses coarse-to-fine dance generation and foot refinement, while PAMD combines a pose-manifold prior with foot-ground refinement. PhysIKDance generates the intermediate state [contact, slide, root, joints, twist] (§3.1, Eq. (1)), directly supervises root–foot displacement, and passes this state to music-conditioned recovery (§3.3). Bidirectional reconstruction, FK consistency, and a teacher-to-predicted curriculum expose recovery to generation errors during training (Appendices A.4–A.5). Section 2 explicitly credits HybrIK’s swing–twist decomposition and NIKI’s invertible error modeling; we do not claim either technique itself as new. The additional PAMD comparison provides complementary evidence: PhysIKDance has lower FIDₖ/FIDɡ (13.16/5.24 versus 14.38/6.17), whereas PAMD has lower PFC (0.1627 versus 0.2390).

Table 3 also reports better FIDₖ, FIDɡ, and PFC for the full configuration than for configurations without direct IK music conditioning or rotation-to-joints supervision. These are configuration-level comparisons: their evaluated Stage-1 outputs differ, so they do not isolate either component’s effect on fixed inputs.

W4 and Q5: Why do floating and body interpenetration remain, and how broadly should the physical-plausibility improvements be interpreted?

Response: Our learned geometric constraints improve motion quality but do not guarantee exact contact or collision-free motion. As Appendix A.10 explains, the supervision is kinematic, and fixed-skeleton recovery can alter generated foot positions. The objectives also do not impose whole-body collision avoidance or dynamic feasibility. These limitations allow residual floating and body interpenetration. We will align the terminology throughout the paper with this existing scope: improved foot-ground kinematic consistency and motion quality, rather than a guarantee of whole-body physical feasibility.

Reviewer_bKZcOpenReview ↗

Final assessment

Thank you for the detailed response and the additional experiments. I appreciate the authors' efforts to address the concerns, particularly the new sliding-specific ablation and direct-rotation comparison. These results strengthen the empirical support for the framework and partially address my original concerns. However, I still find the methodological novelty somewhat limited. The approach primarily integrates established ideas in contact-aware supervision, joint-coordinate modeling, and inverse kinematics. While the integration is reasonable and empirically effective, the rebuttal does not fully demonstrate a sufficiently distinctive technical contribution beyond these existing components. The additional results also suggest that the benefits are not consistent across all physical-consistency metrics, particularly sliding fidelity. Therefore, although the rebuttal has improved my assessment of the paper's empirical support, my concerns regarding its limited methodological novelty remain. I will maintain my original rating.

Thank you for your careful assessment and for recognizing the value of the additional sliding-specific ablation and direct-rotation comparison. We appreciate your remaining concerns regarding methodological novelty and sliding fidelity and would like to clarify two points.

Regarding our methodological contribution, we acknowledge the foundations provided by prior work on contact supervision, joint-coordinate modeling, and inverse kinematics. Our specific design connects state-aware foot–ground constraints with rotation recovery: Eq. (5) couples root and local-foot displacement according to the support state, enforcing cancellation during planted support and matching reference displacement during supported sliding. We also jointly generate joint coordinates and axial twists, recover skeletal rotations through music-conditioned, error-augmented two-stage invertible IK, and apply cross-space supervision to constrain the reconstructed geometry. The additional experiments support these design choices, and we will clarify their specific distinctions from prior approaches.

Regarding physical consistency and sliding fidelity, Table 2 shows that our method outperforms the compared baselines in PFC, grounded foot speed, and foot-skating ratio, while achieving a support ratio close to ground truth. We acknowledge that the full configuration does not achieve the best value on every metric across all ablations. For example, removing IK yields a lower PFC but substantially worsens geometric FID, highlighting the importance of considering motion quality and foot–ground consistency together.

Thank you also for highlighting sliding fidelity. As explained in Appendix A.7.4, the skating statistic includes supported sliding, so we interpret it alongside ground-truth statistics, support ratio, and motion activity. Together, these complementary measures provide evidence of improved foot–ground consistency. In future work, we will further discuss the distinction between geometric sliding and choreographic intent and explore more targeted evaluations of sliding fidelity.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.