acceptodds
Under review as a conference paper at ICLR 2027

Low-Redundancy Modeling for Latent-Free Human Motion Generation

Abstract

Many recent approaches to human motion generation operate in latent spaces, which require two-stage training and inherit the reconstruction errors of their autoencoders. We revisit generation in raw motion space and identify two key bottlenecks: feature redundancy, which arises from over-parameterized representations, and state redundancy, which arises from invalid off-manifold states and unnatural motions. Building on this observation, we instantiate DirectMotion, a minimalist baseline that combines three simple designs: a compact representation consisting of root features and joint rotations, probability paths constrained to the state space of human motion, and clean-sequence prediction. With only a lightweight DiT network, DirectMotion performs on par with larger latent-space models on the HumanML3D and SnapMoGen datasets. Ablation studies confirm the contribution of removing each form of redundancy. We hope these results encourage the community to reconsider latent-free generation as an inherently simpler paradigm.

Primary area:
generative models
Keywords:
Human Motion Generation, Text-to-Motion, Diffusion Model

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_RbrQOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper revisits text-driven human motion generation directly in raw motion space rather than relying on some latent representations. The authors argue motion-space diffusion is bottlenecked by two redundancies: feature redundancy from over-parameterized representations, and state redundancy from invalid off-manifold states. They instantiate this in DirectMotion, a minimalist flow-matching baseline with a compact representation (root features plus joint rotations), probability paths constrained to the state manifold via SLERP on SO(3), and direct prediction of clean motion sequences. With a 9.2M-parameter DiT, DirectMotion reports state-of-the-art FID and R-precision on HumanML3D and competitive results on SnapMoGen, with ablations isolating each design choice.

Critical strengths

[Originality]

  • The feature-redundancy vs. state-redundancy framing is a fresh lens on why latent compression has worked so well and where motion-space generation has left performance on the table. Casting the design choices as testable hypotheses makes the logic crisp.

[Quality]

  • The results are striking: state-of-the-art FID on HumanML3D with a 9.2M-parameter backbone against baselines up to hundreds of millions of parameters, leading on all R-precision metrics too. The ablations are well constructed and the physical-plausibility evaluation goes beyond perceptual metrics. The results look plausible to me.
Critical weaknesses

n/a

Questions for authors

I don't have major questions. Overall this is a surprising and well-executed manuscript, so I lean towards accept.

1 comment

We sincerely appreciate your recognition of our work. It is encouraging to know that you see value in our contribution. Thank you again for your positive feedback.

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

The paper proposes DirectMotion, which generates root features and joint rotations directly in motion space using SO(3)-constrained probability paths and clean-sequence prediction, avoiding an autoencoder latent space. It evaluates the method on HumanML3D and SnapMoGen with comparisons and ablations.

Critical strengths
  1. The method is simple, with only 9.2M parameters, yet achieves competitive results on multiple datasets.

  2. SO(3)-constrained generation and direct motion-space modeling may be useful for physical control and editing.

Critical weaknesses
  1. Limited methodological novelty. Direct motion-space generation, rotation-manifold modeling, and clean-sample prediction have all been explored previously. The paper mainly combines these ideas into a lightweight model without introducing a clearly new mechanism or theory.

  2. The central low-redundancy story is not sufficiently validated. Tables 3 and 4 show correlations but do not establish that the proposed components remove the claimed forms of redundancy. Some comparisons also change multiple factors simultaneously, creating confounding effects.

Questions for authors
  1. Does the improvement remain against parameter-matched raw motion baselines?

  2. Can the authors perform controlled ablations that change only one of the representation, probability path, and prediction target at a time?

  3. Is there direct evidence, such as manifold distance, validity, or physical-constraint metrics, supporting the claim that clean-sequence prediction better follows the data manifold?

  4. What is the single most fundamental technical contribution beyond combining existing raw motion modeling ideas?

5 comments

We sincerely appreciate your review of our manuscript. Our responses are presented below, and we hope they address your questions and concerns.


Q1: Comparison with parameter-matched raw motion baselines.

"Does the improvement remain against parameter-matched raw motion baselines?"

During the rebuttal, we selected three raw motion-space baselines for comparison: MDM-50 step [1], ACMDM-Raw [2], and Light-T2M [3]. MDM-50 step is an officially optimized variant of MDM; ACMDM-Raw is the official raw motion-space variant of ACMDM; and Light-T2M is a raw motion-space method designed for lightweight settings. We conducted experiments on HumanML3D.

Experiment 1: Comparison with parameter-matched baselines.

MethodParamsFID ↓Top 3 ↑MM Dist ↓
MDM-50 step17.88 M0.3360.7413.602
MDM-50 step*9.201 M0.3170.7523.568
ACMDM-Raw‌39.07 M0.0930.8033.127
ACMDM-Raw*‌9.226 M0.1700.7913.208
Light-T2M‌4.478 M0.0760.7813.297
Light-T2M*‌9.205 M0.0510.7843.318
Ours9.198 M0.0460.8302.963

Here, * denotes parameter-matched variants that we trained ourselves. To ensure reliable architectural modifications, we first conducted short training runs to search over architectural configurations (e.g., model dimensions and number of layers). We then selected the best configuration based on validation performance for full training. All other hyperparameters followed the official settings.

The results show that our model retains its advantages over raw motion-space baselines under matched parameter budgets.

We hope these additional experimental results address your concerns.

Revision Commitment: We will include these results in the comparison experiments and provide detailed descriptions of our baseline modifications in the appendix.


[1] Tevet, Guy, et al. "Human motion diffusion model." arXiv preprint arXiv:2209.14916 (2022).

[2] Meng, Zichong, et al. "Absolute coordinates make motion generation easy." arXiv preprint arXiv:2505.19377 (2025).

[3] Zeng, Ling-An, et al. "Light-t2m: A lightweight and fast model for text-to-motion generation." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 9. 2025.

Q2/W2: Can the authors perform controlled ablations?

"In Table 3 and 4, some comparisons also change multiple factors simultaneously, creating confounding effects. Can the authors perform controlled ablations that change only one of the representation, probability path, and prediction target at a time?"

Certainly.

Table 4 in our paper provides a complete 2×2 ablation of probability paths and prediction targets. However, in Table 3, changing the representation also changes the probability path (SO(3) to Linear). To disentangle these factors, we conducted additional experiments to complete a 2×2×2 ablation.

Experiment 2: Full 2×2×2 Ablation.

RepresentationProbability PathPredictionFID ↓CLIP-SCORE ↑
RedundantLinearVelocity26.2830.620
RedundantLinearSample23.9290.670
Redundant‌SO(3)Velocity24.2260.610
Redundant‌SO(3)Sample18.8930.684
Compact‌LinearVelocity23.4270.671
Compact‌LinearSample19.4580.686
CompactSO(3)Velocity19.9690.681
CompactSO(3)Sample17.4790.697

The bolded entries indicate newly added results; the remaining results were already reported in the paper. This complete 2×2×2 ablation allows controlled comparisons from any configuration by changing only one factor at a time: representation, probability path, or prediction target.

Furthermore, changing the probability path also changes the initial noise distribution, as the SO(3) path for rotational features is tied to a distribution on SO(3). To ensure that our conclusions are attributable to the probability path itself rather than the noise distribution.

Experiment 3: Ablation for Noise Space of Rotational Features.

RepresentationPredictionProbability PathNoise SpaceFID ↓CLIP-SCORE ↑
CompactSampleLinearEuclidean19.4580.686
CompactSampleLinearSO(3)28.2220.637
CompactSample‌SO(3)SO(3)17.4790.697

The results show that the performance improvement primarily comes from transitioning the probability path from Euclidean space to the SO(3), rather than from changing the noise distribution itself.

Finally, changing the prediction target also changes the supervision signal. We therefore conducted an additional ablation on the loss space: we reparameterized the predicted velocity field into clean samples and applied the corresponding clean-sample loss.

Experiment 4: Ablation for Supervision Signal.

RepresentationProbability PathPredictionLoss SpaceFID ↓CLIP-SCORE ↑
CompactSO(3)VelocityVelocity19.9690.681
CompactSO(3)VelocitySample26.4440.645
CompactSO(3)SampleSample17.4790.697

The results suggest that the performance gains from switching to clean-sequence prediction are not attributable to changes in the supervision signal.

All additional experiments consistently support our original conclusions and provide a more comprehensive, controlled ablation without confounding these factors.

We hope these results address your concerns.

Revision Commitment: We will include these results in the revised manuscript and update the ablation study section accordingly.

Q3/W2: Is there direct evidence supporting the clean-sequence prediction better follows the data manifold?

"Is there direct evidence, such as manifold distance, validity, or physical-constraint metrics, supporting the claim that clean-sequence prediction better follows the data manifold?"

We agree that more direct empirical evidence would strengthen our conclusions. To this end, we conducted additional experiments on SnapMoGen. We use the training dataset as an empirical approximation of the natural motion manifold and evaluate how closely generated samples align with it.

Experiment 5: Manifold Distance and Joint Rotation Validity

PredictionEMD ↓PMD ↓Joint Outlier ↓
Velocity0.3340.0880.681
Sample0.2980.0570.645

The metrics are defined as follows:

  1. Empirical Manifold Distance (EMD) measures the distance between generated samples and the training set, based on the nearest-neighbor distance from each generated sample to the training samples.

  2. Perceptual Manifold Distance (PMD) measures the distance between generated samples' perceptual features (extracted by the evaluator) and those of the training set, based on nearest-neighbor distances in the perceptual feature space.

  3. Joint Outlier measures the validity of generated joint rotations, based on their nearest-neighbor angular distances to training samples. The 99th percentile of the corresponding nearest-neighbor distances in the validation set is used as the outlier threshold.

In addition, we extend the physical plausibility evaluation in the appendix to compare how well velocity prediction and clean-sequence prediction satisfy physical constraints.

Experiment 6: Physical Plausibility Evaluation

PredictionJitter Degree ↓Foot Floating ↓Foot Sliding ↓Body Penetration ↓
Velocity0.01700.1650.00891.57
Sample0.01010.1160.00731.47

These results consistently support our hypothesis. We hope these additional results provide useful insights into the question.

Revision Commitment: We will include these results in the appendix and provide detailed descriptions of the evaluation metrics and their computation.

Reviewer_9zhiOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DirectMotion for generating 3D human motion sequences. The authors motivate the method based on the following observations: first, generation methods based on latent encoded representations of 3D human motion sequences may compound errors from both the encoding and generation stages, which is detrimental to improving the quality of generated sequences; second, raw 3D motion data representations used by existing motion generation methods, such as pose coordinates plus motion velocity representation, have considerable data redundancy, which likewise affects the performance of generative models. DirectMotion mainly consists of three aspects: a compact representation of motion data, including a 4-dimensional root-joint description of the human skeleton and rotation representations of other joints; a probabilistic path constrained by the human motion state space; and a clean-sequence prediction method.

Critical strengths
  1. Proposes a direct data representation of human poses that effectively eliminates data redundancy; applying this representation to a lightweight motion generation model also achieves SOTA generation results.
  2. The discussion and assumptions about human motion pose representation are informative and thought-provoking.
Critical weaknesses
  1. The proposed direct motion data representation method also leads to error accumulation: any joint rotation error along the path from the root to a leaf joint will amplify the position error of the leaf joint;
  2. Although experiments show that the proposed raw data representations is better than others, it is still insufficient to prove that, for motion generation, "the less redundant the representation, the better."
Questions for authors
  1. From the perspective of motion sequence generation, is a redundancy-free representation necessarily optimal? For example, as mentioned earlier, joint rotation-based representation may cause cumulative errors. If positional redundancy is introduced, could it be used to constrain cumulative errors? Although Table 3 supports the hypothesis that "redundancy-free representation is better," the authors may do not fully explore the benefits of redundancy in this setting.
  2. There are many assumptions about human motion representations in the discussion of the paper, and the reliability of some assumptions is not intuitive. For example, in the first paragraph of Section 3.3, does natural motion really constitute a manifold? Does state manifold refer to the manifold of natural motion or to the state space? Some experimental studies do not sufficiently support the relevant assumptions in the paper. For example, although Table 3 supports the hypothesis that "redundancy-free representation is better," the authors do not detail the processing of the redundant data representation method. Have they fully exploited the potential benefits that redundant representations may bring?
6 comments

We sincerely thank you for reviewing our manuscript. Below are our responses, which we hope will address your questions and concerns.


Q1.1/W1: Error accumulation.

"The direct motion data representation leads to error accumulation. If positional redundancy is introduced, could it be used to constrain cumulative errors?"

We agree that representing motion using rotations introduces the issue of error accumulation. Moreover, incorporating redundant positional features may enable the model to learn the consistency between rotations and positions, thereby mitigating error accumulation.

This question is closely related to the seminal motion-space diffusion model MDM [1], which introduces several auxiliary losses. In particular, the forward kinematics (FK) position loss may help alleviate error accumulation. Interestingly, based on our understanding of the paper and its implementation, MDM does not apply these losses to redundant representations, arguing that such representations already explicitly encode positional information. This suggests two key perspectives:

1. Redundant features can help mitigate error accumulation.

2. Without redundant features, FK position loss can serve the same purpose.

The first addresses Q1.1, suggesting that redundant positional features may alleviate error accumulation. The second addresses W1, showing that FK position loss may offer an alternative without introducing redundancy. To investigate them, we conducted additional experiments on HumanML3D during the rebuttal.

Experiment 1: Generate from t=0t=0 to t=1t=1

RepresentationAdditional LossFID ↓FK-GT err. ↓FK-Pred. err. ↓
CompactN/A0.046161.52 mmN/A
CompactFK loss0.042162.78 mmN/A
Compact + PositionN/A0.082173.01 mm13.38 mm

Experiment 2: Denoise from t=0.5t=0.5 to t=1t=1

RepresentationAdditional LossFID ↓FK-GT err. ↓FK-Pred. err. ↓
CompactN/A0.04058.20 mmN/A
CompactFK loss0.03959.01 mmN/A
Compact + PositionN/A0.01830.03 mm11.06 mm

In these experiments, we evaluate positional errors by applying FK to the generated rotations and comparing the resulting positions against the ground-truth (GT) positions or the model-predicted positions. Notably, the evaluator we use takes only positional features as input, so FID can also reflect error accumulation. For models without redundant representations, we also include a baseline trained with the FK loss.

We consider two settings: generation from pure noise (t=0t=0 to t=1t=1) and denoising from noisy GT samples (t=0.5t=0.5 to t=1t=1). The former represents a generation task, while the latter is closer to a reconstruction task.

The results are insightful: redundant features help reduce error accumulation, but only in the denoising setting. In the generation setting, where sampling starts from pure noise, the results do not support this hypothesis. We offer the following explanation:

1. The model can learn consistency between positional and rotational features.

2. In denoising, this consistency allows the model to leverage partially noisy GT positions to better predict rotations, thereby reducing FK errors.

3. In generation, however, positional features are also initialized from pure noise, which may limit the effectiveness of the consistency in reducing FK errors.

In other words, the extent to which position features can mitigate error accumulation may depend on how informative the positional features themselves are.

On the other hand, adding the FK loss yields slightly lower FID in these runs, but does not reduce the reported FK-to-GT discrepancy. Thus, the extent to which this auxiliary loss benefits positional accuracy remains unclear. We plan to investigate in future work.

We hope these results address your concerns regarding error accumulation.

Revision Commitment: We will add a Limitations section to discuss the potential error accumulation associated with rotation representations. We will also include the above experimental results in the appendix for readers' reference.


[1] Tevet, Guy, et al. "Human motion diffusion model." arXiv preprint arXiv:2209.14916 (2022).

Q1.2/W2: Is a redundancy-free representation necessarily optimal?

"Although experiments show that the proposed raw data representations is better than others, it is still insufficient to prove that, for motion generation, "the less redundant the representation, the better."

No, non-redundant representations are not universally optimal.

Our hypothesis is that non-redundant representations may be easier for generative models to predict directly. This does not rule out the benefits of redundant features in other settings. For example, as noted in the Introduction (lines 39–42), redundant features can serve as inputs to autoencoders for learning expressive latent representations. We conducted a simple experiment on SnapMoGen during the rebuttal to test this.

Experiment 3: Reduntant vs. Compact for VQ-VAE

RepresentationDim.rFID ↓MPJPE ↓
Redundant2964.5828.984 mm
Compact1484.9769.419 mm

Experiment 4: Reduntant vs. Compact for Latent Generative Model

RepresentationDim.gFID ↓CLIP-Score ↑
Redundant29615.060.685
Compact14815.380.683

Here, rFID measures the FID of VQ-VAE reconstructions, MPJPE measures the mean per-joint position error of the reconstructions, and gFID measures the FID of samples generated by the downstream generative model. We use the VQ-VAE and latent generative model from MoMask++ [1], a latent-space method that performs denoising in a discrete latent space. The results show that redundant features improve both VQ-VAE reconstruction and downstream generation, demonstrating their benefits in latent-space models.

We do not claim that non-redundant representations are necessary for high-quality generation or optimal in all settings. Our hypothesis is more precisely stated as follows:

Predicting redundant features may burden motion-space models; non-redundant representations may therefore be easier to predict.

The experiments in Table 3 consistent with this hypothesis. The above experiment 1 provide further support: adding positional features worsens generation FID.

We hope these results help clarify the scope of our experimental findings.

Revision Commitment: We will revise the presentation of our hypothesis and clarify the scope and applicability of redundancy-free representations.


[1] Guo C., Hwang, I., Wang, J., & Zhou, B. (2026). SnapMoGen: Human motion generation from expressive texts. Advances in Neural Information Processing Systems, 38, 99939-99955.

Q2.1/W2: The processing of the redundant data representation.

"Although Table 3 supports the hypothesis that "redundancy-free representation is better," the authors do not detail the processing of the redundant data representation method. Have they fully exploited the potential benefits that redundant representations may bring?"

In Section 5 (line 323), we state that we directly use the officially preprocessed version of the SnapMoGen dataset. Therefore, the preprocessing details for redundant features follow the official implementation. We agree that the manuscript should additionally specify how these features are incorporated, supervised, and used during evaluation.

Our modeling procedure in Table 3 for redundant features is as follows:

  1. Concatenate redundant features along the feature dimension.

  2. Add noise to all features jointly and feed them into the model.

  3. Supervise the model outputs using the GT of all features, including redundant ones.

We treat all features equally, regardless of whether they are redundant. During evaluation, the official SnapMoGen evaluator takes only compact features as input, while generated redundant features are discarded. Table 3 therefore presents a fully controlled experiment testing whether modeling redundant features degrades the quality of the compact features required for downstream applications.

However, we acknowledge that this approach may not fully exploit the potential benefits of redundant representations. For example, consistency supervision between positional and rotational features could further mitigate error accumulation. However, such specialized treatment would introduce additional confounding factors, making it difficult to isolate the actual causes of performance differences.

How to fully exploit redundant features is an interesting question that deserves dedicated investigation. To explore this, we conducted an additional experiment on HumanML3D during the rebuttal, introducing a consistency loss on redundant positional features to enforce agreement between the predicted rotations and positions.

Experiment 5: Generate from t=0t=0 to t=1t=1 with Consistency Loss

RepresentationAdditional LossFID ↓FK-GT err. ↓FK-Pred. err. ↓
RotationN/A0.046161.52 mmN/A
Rotation + PositionN/A0.082173.01 mm13.38 mm
Rotation + Position‌Consistency‌ loss0.059168.38 mm11.56 mm

Experiment 6: Denoise from t=0.5t=0.5 to t=1t=1 with Consistency Loss

RepresentationAdditional LossFID ↓FK-GT err. ↓FK-Pred. err. ↓
RotationN/A0.04058.20 mmN/A
Rotation + PositionN/A0.01830.03 mm11.06 mm
Rotation + Position‌Consistency‌ loss0.01828.66 mm10.14 mm

The results show that consistency loss can improve overall quality while reducing positional errors. We therefore agree that specially handling redundant features may bring additional benefits. However, the consistency loss alone is insufficient to offset the modeling burden introduced by redundant features, as the overall quality (FID) still does not surpass that of the non-redundant representation. This suggests that fully realizing the benefits of redundant features may first require overcoming their additional modeling burden.

We hope these results clarify how we handle redundant features and the rationale behind our treatment.

Revision Commitment: We will include these results in the appendix and detail our treatment of redundant features in the main paper. We will also clarify in the Limitations section that our current experimental setup may not fully explore the potential benefits of redundant features.

Comments on the paper

5 comments
~Kent_Fujiwara2OpenReview ↗

Technical Note on Missing Baselines and Quantitative Control/Editing Evaluation

Having worked closely on 3D human motion generation, I wanted to share a brief technical note, as several critical elements have been overlooked in the reviews. While I leave the assessment of conceptual novelty to the reviewers and AC, the experimental validation has two significant limitations that affect the paper’s core claims:

  • Omission of Recent Latent-Space Baselines and Overstated SOTA Claims: The manuscript omits key recent progress in latent- and token-space motion generation, for example, SALAD (CVPR2025) , MotionHiFlow (CVPR2026) and Causal Motion Diffusion Models (CVPR2026). Excluding these stronger methods gives an incomplete picture of the current landscape and leaves the paper’s performance claims unsubstantiated.

  • Lack of Quantitative Evaluation for Control and Editing: Although motion control and editing are central applications of the paper, they are supported only by qualitative visualizations. In controllable generation, recent works like OmniControl (ICLR2024), ProgMoGen (CVPR2024), and ProjFlow (CVPR2026) establish standardized quantitative protocols (evaluating both spatial/joint control accuracy and generation fidelity). Similarly, recent motion editing benchmarks and methods, such as MotionFix (SIGGRAPH Asia2024) and MotionLab (ICCV2025), demonstrate the necessity of quantitative comparisons against existing baselines rather than relying solely on visual examples.

I hope these pointers are helpful both to the authors for strengthening the empirical validation and to the Area Chair during the final evaluation.

Thank you for your comments. We appreciate your interest in our work and provide our responses below.


1. On Recent Latent-space Baselines and Our Performance claims

Thank you for pointing out the missing baselines. To address your concerns and provide reviewers and the Area Chair with a more comprehensive reference, we followed your suggestion and conducted comparisons between DirectMotion and the baselines you mentioned. The results are presented below.

Experiment 1: Comparison on HumanML3D

MethodParamsFID ↓R@1 ↑R@2 ↑R@3 ↑MM Dist ↓MM ↑
SALAD9.89+0.16M0.1310.5570.7520.8422.9431.858
CMDM38.70+76.31M0.0780.5630.7590.8492.9201.827
MotionHiFlow30.77+2.37M0.0300.5430.7390.8313.0040.885
Ours9.20M0.0460.5480.7370.8302.9631.254

For latent-space methods, the model size is reported as "generator + VAE." The evaluation follows the 67D non-redundant protocol, consistent with the main paper.

The comment suggests that we have overstated our performance. We would like to clarify that our claims are relatively measured. In the Abstract, we state that

Line 19-21: "DirectMotion performs on par with larger latent-space models on the HumanML3D and SnapMoGen datasets."

In the Conclusion, we state that

Line 482-483: "We demonstrate that latent-free generation can match or exceed latent-space methods."

After incorporating the additional baselines you suggested, we believe these claims remain largely unchanged, as our model remains competitive.

However, we acknowledge that our statement in the Introduction and experimental section that "DirectMotion outperforms far more elaborate latent-space methods on HumanML3D" could imply universal superiority over all latent-space methods. We will revise this statement accordingly.

Revision Commitment: We will revise the performance statements in the Introduction and experimental sections to ensure consistency with those in the Abstract and Conclusion. We will also incorporate the above results into the revised manuscript.

2. On Our Claims on Motion Control and Editing.

We respectfully believe that characterizing motion control and editing as our core claims is not appropriate. As clarified in the Introduction, we do not present motion control or editing as contributions of this work. Nowhere in the paper do we claim to propose a novel motion control or editing method, nor do we claim state-of-the-art performance on either task.

In Section 5.3, we use a simple inpainting approach to explore whether DirectMotion learns a motion prior. We explicitly state the purpose of these experiments:

Line 426-428: "Since DirectMotion also learns in the raw motion space, it may acquire the same prior. We examine this possibility qualitatively below."

Our conclusions are:

Line 458-459: "This suggests that DirectMotion may form a motion prior through learning in the raw motion space."

Line 463-464: "Indicating that the motion prior captured by DirectMotion may be fine-grained enough to predict motion at the level of individual joints."

These findings do not imply that we propose a motion control or editing method, nor do we claim state-of-the-art performance on these tasks.

Our DirectMotion is a minimalist baseline and has not been specifically optimized for controlled generation. Nevertheless, this raises an interesting question: could DirectMotion achieve competitive results when equipped with task-specific control techniques?

To investigate this, we conducted experiment. We adapted the constrained projection method proposed by ProjFlow (really nice work) to our model, enabling zero-shot path control with zero error. The comparative results under the ProjFlow protocol are presented below.

Experiment 2: Zero-shot Path Control

MethodZero-shot?FID ↓R@3 ↑Diversity →Foot Skating ↓Traj. err. ↓Loc. err. ↓Avg. err. ↓
GT-0.0000.79510.455-0.00000.00000.0000
MDM✓1.7920.6739.1310.10190.40220.30760.5959
PriorMDM✗0.3930.7079.8470.08970.34570.21320.4417
GMD✓0.2380.76310.0110.10090.09310.03210.1439
OmniControl✗0.0810.78910.3230.05470.03870.00960.0338
MotionLCM V2+CtrlNet✗3.9780.7389.2490.09010.10800.05810.1386
MaskControl✗0.0660.79910.4740.05430.00000.00000.0093
ACMDM-S-PS22+CtrlNet✗0.0670.80510.4810.05910.00750.00100.0100
ACMDM-S-PS22+DNO✓0.1510.802-0.06100.00270.00020.0089
ACMDM-S-PS22+ProjFlow✓0.1070.78410.6440.06290.00000.00000.0000
Ours✓0.1370.81810.8360.06090.00000.00000.0000

Although our method does not achieve the best results on every metric, we believe its performance is competitive.

We hope these clarifications help convey our contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.