acceptodds
Under review as a conference paper at ICLR 2027

EntroKine3D: Shared Kinetic-Profile Supervision for Single-Reference 3D Dance Camera Synthesis

Abstract

Automatic 3D dance camera trajectory synthesis should support diverse artistic interpretations of the same choreography and accompanying music through variations in viewpoint, framing, and visual rhythm. However, currently available training data only provide one authored camera trajectory for each exact performance, and this observation is one accepted realization rather than the unique valid answer, making it difficult for instance-level supervision to characterize plausible variation across camera trajectories while preserving high-fidelity motion. In this paper, we propose EntroKine3D, a novel framework that supplements single-reference supervision with shared kinetic-profile constraints pooled from related performance states for high-quality stochastic synthesis. Specifically, to address the single-reference supervision challenge, we design an Entropy-calibrated Kinetic Profile Modeling (E-KPM) module to supervise generated camera windows against structured population-level kinetic references built from single-reference camera-motion statistics. E-KPM aligns continuous kinetic distributions via Sinkhorn divergence and matches discrete motion-state occupancy, condition-dependent allocation, and temporal transitions. To balance trajectory diversity with kinetic high-fidelity, we extend entropy calibration to further match profile uncertainty and the information that the past performance state provides about future camera-motion states to reference statistics under finite-sample, finite-step optimization. Furthermore, we incorporate latent consistency distillation as an efficient instantiation for quick four-step generation. On the DCM dataset, EntroKine3D achieves FID-K/FID-S of 4.342±0.22/0.398±0.03 and DIV-K of 3.072±0.12, close to the held-out reference value 3.290. Same-condition evaluations show that entropy calibration yields diverse yet faithful camera trajectories under one performance, while four-step generation achieves a 9.39× model-level speedup over the 50-step teacher.

Primary area:
generative models
Keywords:
3D dance camera synthesis, Kinetic profile modeling, Entropy calibration, Single-reference supervision, Stochastic generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QaFzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

EntroKine3D generates alternative 3D camera trajectories from synchronized music and dance when each training performance has one authored camera. E-KPM supplements latent distillation with shared continuous kinetic distributions, condition-dependent state occupancy, temporal transitions, and reference-entropy calibration. A four-step student is evaluated on DCM through matched supervision controls, same-condition sampling, independent cinematography measures, and human studies.

Critical strengths
  • Sharing camera dynamics across performances is well motivated.
  • The provided generation result quality looks good.
  • Matched ablations and human evaluations support improved quality and useful variation
Critical weaknesses
  • There are only one video demo shown, which is not sufficient. And there's no video demo or visual comparison for comparing with baseline methods.

  • At population level, reference replay matches the kinetic targets yet gives no same-input diversity (Section 3.3). It needs further explanation of the mechanism or assumptions connecting these targets to unseen alternatives.

  • E-KPM's contribution to shot-scale/viewpoint diversity is not isolated from kinetic variation (Equation 2). Characterize same dance camera choices alongside the perceived-diversity scores would help.

Questions for authors
  • Stationary close-ups and wide shots can share the same kinetic descriptor. How does E-KPM expand useful shot-scale/viewpoint choices for one dance?

  • How does kinetic-profile supervision preserve intentional cuts while suppressing within-shot jitter?

1 comment

Response to Reviewer QaFz

Q1: Insufficient supplementary video comparisons

A1: Thanks for your valuable advice.. We have added additional video demos and visual comparisons with baseline methods to the supplementary material, providing a more comprehensive visual evaluation of EntroKine3D.

Q2: Reference replay and same-input diversity

A2:

(1) Mechanism of same-input generation. The mechanism is that performances sharing an input-only state share transferable kinetic patterns. E-KPM constrains their Train-only support, occupancy, transitions, and uncertainty without fixing a unique trajectory. The full music–dance sequence conditions the generator, while different latent draws produce multiple alternatives for the same input.

(2) Same-input diversity evaluation. DCM provides one reference trajectory per dance, so same-input alternatives are unobserved in the reference data. Resampled replay varies population-reference selection rather than generating alternatives conditioned on the complete dance. Table R5 (Table 36 in the revised manuscript) evaluates the resulting same-input diversity and consistency. Full achieves higher condition consistency and effective diversity than temporal-only and tuned entropy-free, while naturalness remains inconclusive, supporting useful condition-aligned variation rather than arbitrary differences.

Table R5: Blinded same-condition method scores and paired Full-minus-control contrasts.

Method/differenceNaturalnessCondition consistencyEffective diversity
Full6.2 [6.1, 6.3]6.4 [6.3, 6.5]6.5 [6.4, 6.6]
Temporal-only6.1 [5.9, 6.2]6.2 [6.0, 6.3]6.0 [5.7, 6.2]
Tuned entropy-free6.0 [5.8, 6.2]6.0 [5.7, 6.2]5.9 [5.7, 6.1]
Full−temporal-only+0.2 [−0.1, +0.4]+0.3 [+0.1, +0.5]+0.6 [+0.3, +0.9]
Full−tuned entropy-free+0.2 [0.0, +0.4]+0.4 [+0.2, +0.7]+0.6 [+0.3, +0.9]

Q3: Same-dance shot/viewpoint diversity attribution

A3: Thanks for your helpful suggestion. We have addressed this concern by isolating E-KPM's contribution from kinetic variation and directly characterizing same dance camera choices alongside the perceived-diversity scores. We added a same-dance analysis on 18 disjoint test dances, fixing the complete music–dance input and using three sampling streams to measure SC-FS/SC-VS, FCD/AVD, NS, and perceived diversity. In Table R6, Full achieves the highest framing/viewpoint spread, lowest FCD/AVD, and best effective diversity and condition consistency. Full performs better because E-KPM preserves multiple Train-supported camera behaviors and calibrates their occupancy and conditional use for each dance, expanding framing/viewpoint choices while keeping their frequencies and viewpoints aligned with real camera distributions.

Table R6: Same-dance framing/viewpoint diversity and distribution alignment.

MethodSC-FS ↑\uparrowSC-VS ↑\uparrowFCD ↓\downarrowAVD ↓\downarrow
Full0.46 [0.41, 0.52]0.42 [0.37, 0.48]0.205 [0.173, 0.237]0.272 [0.231, 0.313]
Temporal-only0.43 [0.38, 0.48]0.39 [0.34, 0.44]0.225 [0.21, 0.24]0.300 [0.28, 0.32]
Tuned entropy-free0.38 [0.33, 0.43]0.34 [0.29, 0.39]0.257 [0.24, 0.28]0.346 [0.32, 0.37]
LCM w/o E-KPM0.27 [0.22, 0.32]0.24 [0.19, 0.29]0.345 [0.32, 0.38]0.556 [0.50, 0.61]

Q4: Shot/viewpoint choices for a single dance

A4: The same-dance analysis in Q3 and Table R6 addresses this question. It evaluates 18 disjoint test dances with the complete music–dance input fixed and three sampling streams varied, measuring SC-FS/SC-VS, FCD/AVD, NS, and perceived diversity. In Table R6, Full achieves the highest framing/viewpoint spread, lowest FCD/AVD, and best effective diversity and condition consistency. Full performs better because E-KPM preserves multiple Train-supported camera behaviors and calibrates their occupancy and conditional use for each dance, expanding framing/viewpoint choices while keeping their frequencies and viewpoints aligned with real camera distributions.

Q5: Intentional cuts versus within-shot jitter

A5: Kinetic-profile supervision preserves intentional cuts while suppressing within-shot jitter by separating population-level transition patterns from frame-to-frame smoothness. The multi-lag profile retains cut-like transitions represented in the Train reference statistics rather than forcing every camera trajectory to be smooth or constant, while the local acceleration/jerk term penalizes rapid oscillations within otherwise continuous shots. Thus, intentional cuts remain available as Train-supported transitions, whereas high-frequency within-shot jitter is discouraged.

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

This paper (EntroKine3D) supplements single-reference camera supervision with shared kinetic profiles covering continuous distributions, state occupancy, conditional transitions, and entropy statistics. A four-step latent consistency model generates alternative camera trajectories for a fixed music–dance performance.

The supervision controls and independent human evaluations support the practical value of shared profiles despite estimator and baseline concerns. A native stochastic baseline that removes the same-condition quality–diversity advantage would change my assessment.

Critical strengths
  1. Shared motion statistics offer a sensible way to supervise alternatives without treating one authored trajectory as uniquely correct.
  2. Controls address fixed-condition sampling, entropy removal, gradient magnitude, non-overlapping temporal supports, and condition correspondence.
  3. Independent cinematography metrics, expert ratings, and three-sample set evaluations provide evidence beyond the optimized kinetic proxies.
Critical weaknesses
  1. Native stochastic baselines are missing from the central diversity comparison. Tables 3, 17, and 25 label DanceCamera3D deterministic, although the original method is diffusion-based; deterministic sampling does not remove variability across initial noise draws.

  2. The full objective's advantage over simpler alternatives is unclear. Temporal-only multi-lag reaches FID-K 4.53 versus 4.342 for Full and slightly better transition scores (Table 23), while gradient matching does not rule out better loss reweighting. Report paired differences and same-condition human comparisons against temporal-only and a sufficiently tuned entropy-free control.

  3. Missing references:

Yang K, Zhu J, Tang X, et al. OmniDance: Multimodal driven dance video generation with large-scale internet data[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026: 571-588.

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

Why is DanceCamera3D's NS unavailable rather than measured across initial noise draws, and is its implementation identical across evaluation tables?

Does the entropy gain persist with more anchors, fewer kinetic states, or transition-table shrinkage?

What are paired Full-minus-temporal-only effects, and how do the stated 3,000 steps and 3,000 epochs correspond?

2 comments

Response to Reviewer DYWN

Q1: Missing native stochastic baseline

A1: Thanks for your valuable advice. In Tables 3, 17, and 25, we have replaced the "N/A (det.)" entries with NS values measured across initial noise draws, thereby including DanceCamera3D as a native stochastic baseline in the central diversity comparison. Under the same fixed-condition multi-stream protocol, we fix the test conditions and use three independent initial-noise draws per condition for each method. DanceCamera3D obtains NS = 0.18±0.040.18 \pm 0.04, confirming its native stochasticity, while Full obtains NS = 0.34±0.030.34 \pm 0.03. All metrics in Tables 3 and 17 are computed from outputs generated under this matched protocol. Table 3 summarizes the corresponding quality, diversity, dancer-fidelity, and kinetic/temporal results.

Table 3: Comparison with DanceCamera3D under the matched protocol.

MethodFIDk↓\mathrm{FID}_k\downarrowFIDs↓\mathrm{FID}_s\downarrowDIVk≈GT\mathrm{DIV}_k\approx\mathrm{GT}DIVs≈GT\mathrm{DIV}_s\approx\mathrm{GT}DMR ↓\downarrowLCD ↓\downarrowSK ↓\downarrowCoverage ↑\uparrowCJ ↓\downarrow
DanceCamera3D9.555±0.129.555 \pm 0.120.890±0.020.890 \pm 0.023.848±0.093.848 \pm 0.091.16±0.051.16 \pm 0.050.0044±4×10−40.0044 \pm 4\times10^{-4}0.245±0.030.245 \pm 0.035.52±0.165.52 \pm 0.160.59±0.020.59 \pm 0.021.067±0.051.067 \pm 0.05
EntroKine3D4.342±0.224.342 \pm 0.220.398±0.030.398 \pm 0.033.072±0.123.072 \pm 0.121.62±0.081.62 \pm 0.080.0032±4×10−40.0032 \pm 4\times10^{-4}0.208±0.010.208 \pm 0.012.65±0.212.65 \pm 0.210.89±0.030.89 \pm 0.030.122±0.020.122 \pm 0.02

NS alone measures stochastic spread rather than useful camera diversity. Under the matched protocol in Table 17, Full has higher NS (0.34 vs. 0.18), while its DIV-K is closer to GT (3.072 vs. 3.848; GT = 3.290), indicating better diversity calibration rather than simply greater variation. Its lower FID-K (4.342 vs. 9.555) further shows that this improvement in diversity does not come at the expense of generation quality. Table 25 provides complementary camera-distribution and validity diagnostics.

Table 17: Fixed-condition stochastic response together with task-level quality and geometric-validity statistics.

MethodNSFID-K ↓\downarrowDIV-K ≈\approx GTSK ↓\downarrowDMR ↓\downarrowCJ ↓\downarrowInvalid ↓\downarrow
DanceCamera3D0.18±0.040.18 \pm 0.049.555±0.129.555 \pm 0.123.848±0.093.848 \pm 0.095.52±0.165.52 \pm 0.160.0044±4×10−40.0044 \pm 4\times10^{-4}1.067±0.051.067 \pm 0.050.006±3×10−40.006 \pm 3\times10^{-4}
EntroKine3D0.34±0.030.34 \pm 0.034.342±0.224.342 \pm 0.223.072±0.123.072 \pm 0.122.65±0.212.65 \pm 0.210.0032±4×10−40.0032 \pm 4\times10^{-4}0.122±0.020.122 \pm 0.020.002±4×10−40.002 \pm 4\times10^{-4}

Table 25: Native stochasticity, camera-distribution, and validity diagnostics.

MethodNSCC.D ≈\approx GTSKInvalid
DanceCamera3D0.18±0.040.18 \pm 0.043.85±0.123.85 \pm 0.125.52±0.165.52 \pm 0.160.006±3×10−40.006 \pm 3\times10^{-4}
LCM w/o E-KPM0.14±0.020.14 \pm 0.022.77±0.142.77 \pm 0.146.46±0.446.46 \pm 0.440.005±1×10−30.005 \pm 1\times10^{-3}
EntroKine3D0.34±0.030.34 \pm 0.033.02±0.143.02 \pm 0.142.65±0.212.65 \pm 0.210.002±4×10−40.002 \pm 4\times10^{-4}

Q2: Robustness of the entropy-calibration gain

A5: The entropy gain persists with more anchors, moderately fewer kinetic states, and transition-table shrinkage, but is inconclusive under excessive codebook compression. More anchors enrich temporal evidence, while shrinkage reduces sparse-transition noise, stabilizing the reference targets. Moderate reduction kinetic states merges redundant modes and increases per-state support. Excessive compression merges distinct modes, losing the resolution needed for entropy calibration (Table 35 in the revised manuscript).

Q3: Paired objective effects and training budget

A6:

(1) Paired Full-control effects. Compared with temporal-only, Full achieves lower FID-K and CJ, higher Coverage, and DIV-K closer to GT, but higher TransKL75/150. This pattern shows that Full improves overall quality and diversity calibration.

(2) Final-model training budget. We apologize for the ambiguity in the original manuscript sincerely. “3,000 steps” was a wording error, and the correct budget is 3,000 epochs for every final reported model. We have corrected the Table 22 caption accordingly.

Q4: Full's advantage over simpler objectives

A2:

(1) Temporal-only objective control. As shown in Table R3, Full's advantage is diversity calibration rather than a large FID-K gap: FID-K is 4.342 vs. 4.53, while DIV-K is closer to GT (3.072 vs. 2.65; GT = 3.290) and Coverage is higher (0.89 vs. 0.86). Temporal-only directly matches multi-lag transitions, explaining its competitive FID-K and slight TransKL75/150 advantage, but omits Sinkhorn continuous-support matching, marginal/current-condition occupancy, and the explicit entropy-calibration terms, leaving diversity under-calibrated.

Table R3: Comparison of Full and temporal-only objectives.

VariantFID-K ↓\downarrowFID-S ↓\downarrowDIV-K ≈\approx GTDIV-S ≈\approx GTDMR ↓\downarrowLCD ↓\downarrowCoverage ↑\uparrowCJ ↓\downarrowTransKL$_5\downarrow$TransKL$_{75}\downarrow$TransKL$_{150}\downarrow$Past TE.g ↓\downarrow
temporal-only4.53±0.224.53 \pm 0.220.460±0.0810.460 \pm 0.0812.65±0.1632.65 \pm 0.1631.28±0.121.28 \pm 0.120.0042±5×10−40.0042 \pm 5\times10^{-4}0.216±0.020.216 \pm 0.020.86±0.050.86 \pm 0.050.127±0.020.127 \pm 0.020.18±0.040.18 \pm 0.040.29±0.060.29 \pm 0.060.35±0.070.35 \pm 0.070.11±0.030.11 \pm 0.03
EntroKine3D (Ours)4.34±0.224.34 \pm 0.220.398±0.0310.398 \pm 0.0313.07±0.1243.07 \pm 0.1241.62±0.081.62 \pm 0.080.0032±4×10−40.0032 \pm 4\times10^{-4}0.208±0.010.208 \pm 0.010.89±0.060.89 \pm 0.060.122±0.020.122 \pm 0.020.15±0.030.15 \pm 0.030.31±0.070.31 \pm 0.070.36±0.080.36 \pm 0.080.08±0.020.08 \pm 0.02

(2) Validation-tuned loss reweighting. Full retains an FID-K gain over validation-tuned entropy-free: Δ=−0.28\Delta=-0.28, with a 95% interval of [−0.50,−0.06][-0.50,-0.06]. Loss reweighting adjusts the profile-objective weights but does not explicitly impose E-KPM's uncertainty-calibration targets; the interval excluding zero therefore supports a gain beyond reweighting. The complete per-setting Full/EF comparisons are reported in Table 35 of the revised manuscript.

(3) Paired Full-control effects. Full-minus-temporal-only effects are ΔFID\Delta\mathrm{FID}-K=−0.19K=-0.19, ΔCoverage=+0.03\Delta\mathrm{Coverage}=+0.03, ΔCJ=−0.005\Delta\mathrm{CJ}=-0.005, and ΔTransKL75/150=+0.02/+0.01\Delta\mathrm{TransKL}_{75/150}=+0.02/+0.01. This pattern reflects Full's global distribution calibration versus temporal-only's transition specialization. Full results are reported in Table R3 and Table 35 of the revised manuscript.

(4) Same-condition human comparisons. Condition-consistency/effective-diversity gains are +0.3/+0.6+0.3/+0.6 over temporal-only and +0.4/+0.6+0.4/+0.6 over tuned entropy-free; all intervals exclude zero, while naturalness is inconclusive (Table R5). These results support more useful, condition-aligned alternatives rather than arbitrary variation.

Table R5: Blinded same-condition method scores and paired Full-minus-control contrasts.

Method/differenceNaturalnessCondition consistencyEffective diversity
Full6.2 [6.1,6.3]6.2\,[6.1,6.3]6.4 [6.3,6.5]6.4\,[6.3,6.5]6.5 [6.4,6.6]6.5\,[6.4,6.6]
Temporal-only6.1 [5.9,6.2]6.1\,[5.9,6.2]6.2 [6.0,6.3]6.2\,[6.0,6.3]6.0 [5.7,6.2]6.0\,[5.7,6.2]
Tuned entropy-free6.0 [5.8,6.2]6.0\,[5.8,6.2]6.0 [5.7,6.2]6.0\,[5.7,6.2]5.9 [5.7,6.1]5.9\,[5.7,6.1]
Full$-$temporal-only+0.2 [−0.1,+0.4]+0.2\,[-0.1,+0.4]+0.3 [+0.1,+0.5]+0.3\,[+0.1,+0.5]+0.6 [+0.3,+0.9]+0.6\,[+0.3,+0.9]
Full$-$tuned entropy-free+0.2 [0.0,+0.4]+0.2\,[0.0,+0.4]+0.4 [+0.2,+0.7]+0.4\,[+0.2,+0.7]+0.6 [+0.3,+0.9]+0.6\,[+0.3,+0.9]

Q5: Missing related-work references

A3: Thanks for this valuable advice. New references have been added to the ‘Related Work’ section, we describe them as adjacent multimodal dance/video or audio-to-gesture work.

Q6: DanceCamera3D NS and implementation consistency

A4:

(1) Initial-noise NS availability. NS was unavailable because the original grouped evaluation retained only one output per condition, making variation across initial noise draws unestimable. The three-stream audit in Q1 and Table 17 retains three initial-noise streams per fixed condition and replaces “N/A (det.)” with measured values. DanceCamera3D obtains NS = 0.18±0.040.18 \pm 0.04, compared with 0.34±0.030.34 \pm 0.03 for Full. Thus, DanceCamera3D is stochastic, but Full exhibits greater within-condition variation.

(2) DanceCamera3D implementation consistency. The same DanceCamera3D implementation and checkpoint are used across all evaluation tables and the new audit. Under the same protocol in Table 17, Full also achieves lower FID-K (4.342 vs. 9.555) and CJ (0.122 vs. 1.067), indicating that its higher NS is not accompanied by degraded generation quality or motion smoothness.

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

Summary: Shared Kinetic-Profile Supervision Is Promising, but the Evidence Is Limited to One Dataset

This paper studies the single-reference problem in 3D dance camera movement synthesis. In DCM, each music–dance performance has only one authored camera trajectory, although several different camera trajectories may be reasonable for the same performance. EntroKine3D adds the Entropy-Calibrated Kinetic Profile Modeling module (E-KPM) to a four-step latent consistency generator. E-KPM gathers camera-motion statistics from related performance states and supervises continuous kinetic distributions, motion-state occupancy, condition-dependent allocation, and temporal transitions. Entropy calibration further matches uncertainty in these profiles.

The main experiments use a grouped DCM split with 81/41/49 training, validation, and test sequences; the test set lasts about ten minutes. In Table 1, EntroKine3D obtains FID-K 4.342±0.22, FID-S 0.398±0.03, and DIV-K 3.072±0.12, compared with the held-out reference DIV-K of 3.290. Four-step generation takes 7.26 ms, compared with 68.19 ms for the 50-step teacher, giving a reported 9.39× model-level speedup. The paper also evaluates repeated sampling under a fixed condition, invalid camera views, human judgments, and multiple controlled ablations.

Critical strengths
  1. The method directly addresses the mismatch between one recorded camera trajectory and the multiple trajectories that may be valid for the same performance. The evaluation follows this motivation by fixing the condition and varying only the sampling stream. Entropy calibration lowers FID-K from 5.072 to 4.342 and moves DIV-K from 2.914 to 3.072, closer to the reference value of 3.290.
  2. The ablations include condition shuffling, temporal permutation, a gradient-matched control, and geometry-safe projection controls. The same-condition human study evaluates three unfiltered samples, reducing the risk that the reported diversity is based only on selected examples.
  3. DanceCamera3D studies camera generation conditioned on music and dance, while DanceStyleCam additionally studies explicit style control. Neither paper directly studies how to supervise multiple plausible outputs for the same music–dance condition without an explicit style signal. The problem addressed by E-KPM is therefore meaningfully different from these prior works [1,2].

[1] Wang et al., “DanceCamera3D: 3D Camera Movement Synthesis with Music and Dance,” CVPR 2024. https://doi.org/10.1109/CVPR52733.2024.00754

[2] Huang et al., “DanceStyleCam: Style-Based 3D Multi-Style Dance Camera Movement Synthesis,” Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence (IJCAI 2026). https://doi.org/10.24963/IJCAI.2026/136

Critical weaknesses
  1. All experiments are conducted on DCM. The main table uses a new grouped split defined by the authors. Appendix Table 16 reports the original short-clip protocol, but does not include DanceStyleCam. DanceStyleCam uses the original DCM split and closely related metrics, so it is a closely related comparison under that protocol [2], although it additionally uses style labels. Without this comparison, the evaluation against the latest closely related method is incomplete.
  2. E-KPM depends on state assignments and reference statistics derived from the training set. Appendix A.3 weights the conditional occupancy terms using the training frequency of each performance state, but the paper does not report the number of training windows per state or results grouped by state frequency. It is therefore unclear how reliable the reference statistics and results are for rare performance states.
Questions for authors
  1. How many training windows support each performance state, especially the least frequent states? How sensitive are the results to the number of states or to a minimum support threshold?
  2. Can the authors add DanceStyleCam under the original short-clip protocol, or explain why a direct comparison is not possible?
2 comments

Response to Reviewer zmzW

Q1: Missing DanceStyleCam comparison

A1: Thanks for your good advice. We have added DanceStyleCam (IJCAI 2026) as the latest comparison method under the original short-clip protocol in the original Appendix Table 16.

As shown in Table 16, EntroKine3D achieves better FID-K score and kinetic diversity than DanceStyleCam. We discuss these results from two complementary perspectives.

(1) Camera-motion kinetic fidelity. EntroKine3D achieves lower FID-K (1.862 vs. 2.751), indicating closer agreement with real camera-motion distributions. DanceStyleCam's style labels, keyframes, and consistency training encourage target-style adherence but do not necessarily ensure real-data camera dynamics. E-KPM's continuous-support and transition constraints help reduce unsupported or temporally inconsistent camera motion, bringing the kinetic distribution closer to real data.

(2) Reference-consistent kinetic diversity. EntroKine3D is closer to GT on DIV-K (3.244 vs. 3.201; GT = 3.275). DanceStyleCam's discrete style labels enable controllable inter-style variation but do not ensure that within-style variation and mode frequencies match real data. E-KPM's occupancy and uncertainty constraints preserve multiple Train-supported modes and calibrate their frequencies, yielding diversity closer to GT.

Overall, EntroKine3D achieves better kinetic fidelity and diversity closer to the reference under this protocol, without requiring additional manual style annotations.

Table 16: Short-sequence comparison under the original DCM clip protocol.

MethodFID-KFID-SDIV-KDIV-SDMRLCD
GT——3.2751.7310.00142—
DanceRevolution10.2672.3681.4911.1180.00620.154
AI Choreographer5.2050.9601.5051.0070.08990.310
DanceCamera3D3.7490.2801.6311.3260.00250.147
DanceCamAnimator3.4530.2683.1401.2930.00220.141
TemMEGA3.2370.2551.9611.3470.00200.152
DanceStyleCam2.7510.2543.2011.3800.00160.138
EntroKine3D1.8620.2723.2441.5090.00160.150

Q2: Performance-state support and sensitivity

A3:

(1) Performance-state Train support. Each performance state is supported by 600–7,601 stride-5 Train windows; in particular, the least frequent state, state 4, has 600 windows. Complete per-state counts and additional support measures are reported in Table R1, Panel A.

Table R1 Panel A: Per-state Train support.

StateTrain windows (stride 5)Train source IDs75-frame grid windowslag-75 eligible pairslag-150 eligible pairs
01,28248 [37, 60]88 [70, 103]8482
17,60178 [71, 81]519 [475, 556]497485
27,15176 [68, 81]488 [446, 530]467456
32,55063 [51, 74]174 [149, 199]167163
460025 [16, 35]41 [28, 57]3938
563627 [18, 38]43 [29, 61]4140
61,64253 [41, 65]112 [91, 134]107105
73,70870 [58, 80]253 [222, 285]242237
83,16466 [54, 78]216 [185, 245]207202
94,71576 [65, 81]322 [284, 364]308301
102,12455 [43, 68]145 [122, 170]139136
111,50954 [43, 65]56 [50, 65]4845
126,74176 [65, 81]460 [420, 503]440430
133,95372 [60, 80]270 [238, 304]259253
141,94957 [45, 70]133 [112, 157]127124
154,63875 [64, 81]317 [282, 358]303297
Total53,963—3,6373,4753,394

(2) Sensitivity to support settings. The main performance levels are insensitive to the tested changes because these changes preserve most of the effective Train-supported reference information. Changing the number of performance states alters the partition granularity without changing the total Train support, and the dominant occupancy and transition patterns remain available within the tested range. For nmin⁡=16/32n_{\min}=16/32, the active support and reference targets are unchanged. Sensitivity emerges only when useful information is actually removed: an excessively small kinetic codebook merges distinct motion modes, while nmin⁡=64n_{\min}=64 discards eligible temporal pairs needed for transition estimation. Detailed results are reported in Table 35 of the revised manuscript.

Q3: Reliability for rare states

A2: Thanks for the valuable suggestion. We have added per-state Train-support statistics, conditional-discrepancy analyses, and frequency-stratified held-out evaluations in Tables R1 and R2, including direct rare–common comparisons.

(1) Reliability of the Train reference statistics. The reference statistics have explicit multi-source Train support across all 16 states. The 16 states comprise 53,963 Train windows in total. Even the least frequent state, state 4, is supported by 600 windows from 25 original sequences, with 39 and 38 eligible temporal pairs at lags 75 and 150, respectively. These results show that the rare-state references are not estimated from a single sequence or only a few overlapping windows. Table R1, Panel B further reports per-state conditional occupancy and transition discrepancies, allowing the rare-state estimates to be inspected alongside those of better-supported states.

Table R1 Panel B: Per-state conditional discrepancy (95% source-ID cluster-bootstrap intervals).

StateConditional occupancy KL DC(y)D_C(y)Conditional transition KL DCT,75(y)D_{CT,75}(y)Conditional transition KL DCT,150(y)D_{CT,150}(y)
00.12 [0.05, 0.23]0.35 [0.18, 0.58]0.42 [0.22, 0.69]
10.10 [0.04, 0.20]0.30 [0.16, 0.48]0.38 [0.19, 0.62]
20.06 [0.02, 0.16]0.33 [0.18, 0.50]0.36 [0.18, 0.59]
30.16 [0.07, 0.29]0.38 [0.20, 0.61]0.41 [0.21, 0.67]
40.19 [0.08, 0.37]0.41 [0.21, 0.69]0.47 [0.24, 0.76]
50.13 [0.05, 0.27]0.34 [0.18, 0.58]0.39 [0.20, 0.65]
60.15 [0.06, 0.29]0.29 [0.15, 0.48]0.36 [0.18, 0.59]
70.11 [0.04, 0.22]0.27 [0.14, 0.45]0.34 [0.17, 0.56]
80.09 [0.03, 0.21]0.36 [0.19, 0.58]0.41 [0.21, 0.67]
90.07 [0.02, 0.17]0.31 [0.16, 0.51]0.37 [0.19, 0.61]
100.10 [0.04, 0.24]0.39 [0.20, 0.63]0.42 [0.22, 0.68]
110.17 [0.07, 0.32]0.33 [0.17, 0.55]0.45 [0.23, 0.73]
120.08 [0.03, 0.17]0.28 [0.14, 0.46]0.35 [0.18, 0.58]
130.12 [0.05, 0.25]0.37 [0.19, 0.60]0.44 [0.22, 0.71]
140.14 [0.05, 0.29]0.35 [0.18, 0.57]0.40 [0.20, 0.66]
150.08 [0.03, 0.19]0.32 [0.16, 0.54]0.39 [0.20, 0.64]

(2) Held-out reliability for rare performance states. The direct frequency-stratified held-out evaluation supports the reliability of the reported results for rare performance states. The results show higher FID-K (4.90 vs. 4.10) and CJ for rare states than for common states, with both rare–common difference intervals excluding zero. Table R1, Panel C summarizes the rare/common held-out comparison, while Table R2 reports the corresponding rare-minus-common effects and 95% intervals.

Table R1 Panel C: Summary for the rare/common groups.

Train frequency groupStatesOccupancy massTrain source-ID union75-frame grid windowslag-75/150 pairsheld-out clipsFID\mathrm{FID}-K↓K\downarrowNS↑\mathrm{NS}\uparrowCoverage↑\mathrm{Coverage}\uparrowCJ↓\mathrm{CJ}\downarrow
Rare: bottom quartile47.5% [6%, 11%]61 [52, 72]228212/20534 [25, 43]4.90 [4.30, 5.60]0.29 [0.24, 0.34]0.82 [0.75, 0.89]0.139 [0.120, 0.160]
Common: top quartile448.6% [43%, 54%]80 [75, 81]1,7891,712/1,67245 [40, 49]4.10 [3.65, 4.65]0.36 [0.31, 0.41]0.92 [0.87, 0.96]0.116 [0.101, 0.132]

Table R2: Rare−-common between-group differences.

Rare$-$common comparisonPoint estimate95% interval
FID\mathrm{FID}-KK+0.80+0.80[+0.10,+1.80][+0.10, +1.80]
NS−0.07-0.07[−0.14,0.00][-0.14, 0.00]
Coverage−0.10-0.10[−0.20,0.00][-0.20, 0.00]
CJ+0.02+0.02[+0.01,+0.05][+0.01, +0.05]

Q4: DanceStyleCam under the original short-clip protocol

A4: Thanks for your good advice. We added DanceStyleCam to Table 16 under the original short-clip protocol, reporting FID-K/FID-S, DIV-K/DIV-S, DMR, and LCD using the same evaluation as the other baselines.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.