To Reviewer oZAK:

      Thank you for your detailed feedback and your recognition of our main motivation, ablation study designs, and the practical benefits of absolute coordinate representation for motion control and direct SMPL-H mesh generation. Below, we provide detailed responses to the weaknesses and questions you raised.

Q1: Regarding Strange Video Format.

      We thank the reviewer for the feedback and the concern regarding the video format. We would like to clarify that our qualitative results are presented through a locally browsable HTML interface, as mentioned in the Appendix Section J. In this way, readers can simply double-click the index.html file for full visualizations. The separate data folders are used solely to organize the supporting files for this interface. This design enables readers to interactively browse extensive results by task (e.g., text-to-motion, motion control, direct mesh generation) without the need to sit through a long video. We chose this format to allow faster navigation, side-by-side comparisons, and broader qualitative coverage, which are often not feasible with traditional one linear video.

Q2: Response to Joint v.s. Mesh Video Results for Motion Control Comparisons.

      We thank the reviewer for the observation. The reason we visualize joint-level visualization (rather than full meshes) in our motion control results is because we follow the experimental setup defined by prior work OmniControl $102$ , which specifically formulates the control task at the joint level (targeting key joints such as the pelvis, feet, head, and wrists). All baseline methods are not designed to operate on mesh-level control, and we adhere to this protocol to ensure a fair and consistent comparison.
      Additionally, we intentionally use joint-space visualizations to better showcase the fine-grained controllability of our model. Mesh-based visualizations can obscure the accuracy of control due to the surrounding body volume around the target joint, making it harder to judge whether the control targets are precisely achieved. In contrast, joint-level visualization offers a clearer and more accurate visualization of whether the model successfully reaches the intended control targets, especially when comparing across methods.

Q3: Response to Insufficient Visualization.

      We thank the reviewer for the comment. We would like to clarify that motion data is inherently 4D where each frame contains 3D spatial coordinates, and the temporal dimension is equally critical to understanding motion quality. Static figures in the paper cannot effectively convey this temporal structure, especially for evaluating smoothness, realism, and controllability.
      Nevertheless, we include two key visualizations in the main paper (a teaser and a side-by-side comparison). To complement this, we also provide an extensive set of results in the supplementary HTML webpage, covering over 45 motion clips across three distinct tasks. This choice reflects our belief that video results are essential for evaluating the temporal dynamics and overall quality of generated motion. We hope the reviewer will consider the full scope of visualizations included in the video results as part of the overall evaluation.

Q4: Regarding MOJO, DartControl, and Their Representations.

      We thank the reviewer for highlighting these important prior works. We will include MOJO in the revised version of the paper. However, DARTControl is in fact cited as $125$ in our current submission.
      We would like to clarify how our approach differs from both works. First, MOJO (CVPR 2021) predates HumanML3D $25$ (CVPR 2022), and primarily focuses on a different task: human motion prediction. They adopt surface keypoints as their representation, rather than joint coordinates or full mesh vertices. This also aligns with our statement in the introduction section that the field largely converged on root-relative joint representations following the introduction of HumanML3D. Our framing of absolute coordinates as “simple and long-abandoned” reflects this trajectory. We do not claim the representation is novel, but rather that it has been overlooked for text-to-motion in recent years and deserves rethinking and empirical re-examination. Second, DARTControl uses a relative representation in the form of SMPL-X $59$ parameters, which differ fundamentally from our formulation of absolute coordinates.

Q5: Response to Questionable Choice of Evaluation Metrics.

      We fully agree with the reviewer that the evaluation framework proposed in MARDM $66$ provides a more robust and fair assessment by focusing on essential, animatable dimensions and excluding redundant features. We would like to point out that our results were, in fact, obtained using the same evaluation method and protocol as MARDM, and results exactly following MARDM.

Q6: Regarding Insufficient Analysis of Kinematic Plausibility and Failure Modes.

      We argue that since FID compares the distribution of generated motions against real human motion, samples with highly unnatural or kinematically impossible artifacts would significantly increase the distance between these distributions. Therefore, the consistently low FID scores achieved by our model suggest that such failure cases are rare.
      We would also like to clarify a common misconception: relative joint positions do not inherently guarantee constant bone lengths. It is possible to predict arbitrarily small or large relative offsets that violate bone length constraints. Only rotational representations (e.g., joint rotations) inherently preserve bone lengths with a fixed skeleton template. However, the rotation data in HumanML3D is known to be redundant and inaccurate (as discussed in our paper Line 28, and noted in HumanML3D’s GitHub issue). Consequently, no baseline method using HumanML3D’s relative joint positions can strictly preserve bone length.
      To evaluate ACMDM’s ability to maintain bone length consistency, we conducted two quantitative tests on the HumanML3D testset:
      1) Intra-sequence bone length stability – measuring changes in bone lengths within a generated sequence.
      (2) Bone length consistency with ground truth – comparing generated bone lengths against those from the ground truth.
      We also compared ACMDM to SOTA relative representation baselines with results presented in Table A4. Results indicate that ACMDM not only maintains stable bone lengths but also outperforms some relative baselines such as MotionLCM V2 $14$ and MARDM $14$ .
      Due to the NeurIPS rebuttal policy, we cannot include additional videos, but we observed no noticeable bone length artifacts in the HumanML3D testset. To further probe robustness, we tested ACMDM on 10 out-of-distribution prompts involving “flying” (e.g., “a person is flying in the air”), which are impossible motions. Despite the unusual prompts, the generated motions mostly showed realistic behavior (e.g., standing, waving) rather than attempting unrealistic "flying" motion, and intra-sequence bone length variation remained very low (0.0071 m), confirming the reliability of our approach. We will provide extended qualitative results with failure case visualizations in the revised version.

      Table A4: Bone length consistency on the humanml3d testset.

I thank the authors for putting in the effort towards this rebuttal. Thanks for the pointer towards the html file. It did not work with safari on my laptop and hence my comments about video format. However I tried it again with chrome this time and it worked. The video files indeed look much better when visualized using the html file. It would make sense in the future to provide an mp4 file for the whole video- just to make sure these troubles dont happen. I acknowledge the authors responses 1,5,6 addresses the queries I raised. However I think 2,3,4 remain valid concerns. MOJO should be cited properly in the paper and acknowledgement given to papers that use pointcloud representation for motion synthesis. Regarding 3, I think that the paper falls slightly short of the presentation standards expected at Neurips. For 2, the authors are right to point out that a mesh obscures the fine grained control that can be observed with Joint Positions however the authors claim text-to-motion - not spatially controlled motion - is their main task which does not require fine-grained details and I find it very surprising that more mesh visualizations are not provided as they claim that direct prediction of mesh vertices is a benefit of this paper. I am still inclined towards a rejection but I will downweight my vote from a strong reject to a boderline reject - the upgrade is primarily because the html file started working.

To Reviewer BjnS:

      Thank you for your thoughtful feedback. We appreciate your recognition of the clarity with which we present the problem and solution, as well as the overall readability of the paper. Below, we provide detailed responses to the weaknesses and questions you raised.

Q1: Regarding Incomplete Discussion on Underperforming Metrics in Table 3 and Strategies for Improvement.

      We appreciate the reviewer’s attention to detail and the opportunity to clarify. We would like to clarify that ACMDM demonstrates SOTA performance across nearly all metrics in Table 3, including FID, R-Precision Top-3, diversity, trajectory error, location error, and average error—significantly outperforming baseline methods. The only metric where ACMDM shows a slight shortfall is the foot skating ratio in the pelvis-only spatial control task, where it exceeds OmniControl $102$ by just 0.0044 (0.44%). We note that this margin is minimal and within a very narrow threshold, particularly considering that the metric reflects the ratio of frames with foot movement beyond 2.5 cm during contact. Furthermore, ACMDM still outperforms the next best baseline (GMD $42$ ) by a substantial margin of 0.0306 (3.06%), reinforcing its overall robustness.

Q2: Response to Questionable Novelty of ACMDM and Additional Relative Motion Representations.

      We thank the reviewer for raising this point. We would like to clarify that the goal of our work is not to “propose a global motion representation to address the lack of local relative motion representations,” nor to introduce a novel model architecture. Rather, our motivation is to critically revisit the dominant practice in text-driven motion generation—the wide-spread adoption of root-relative (local) representations proposed by HumanML3D $25$ —and demonstrate that a simple yet long-overlooked global (absolute) representation, when paired with a simple Transformer model, without task-specifc designs(e.g., UNet or altered attentions) and any additional kinematic losses, can achieve significantly improved motion fidelity, better text alignment, strong scalability, and easy adaptability to subtasks such as motion control (Table 2) and direct mesh generation (Section 3.3), without requiring task-specific reengineering.
      Notably, nearly all state-of-the-art baselines in both text-to-motion and motion control—such as MDM $91$ , MotionDiffuse $120$ , ReMoDiffuse $119$ , MLD++ $14$ , MotionLCM v2 $14$ , and MARDM $66$ —rely on HumanML3D-style root-relative representations. In Table 1, we also include comparisons with the InterGen $52$ representation (absolute + redundancy), and following the reviewer’s suggestion, we also trained ACMDM with SMPL $59$ parameters—another relative representation but rarely used in text-to-motion generation—to provide a broader empirical perspective.
      As shown in Table A2, our approach significantly outperforms the SMPL parameter representation, also highlighting why such representations are rarely adopted in text-to-motion tasks. This further indicates that, despite its simplicity, our absolute-coordinate formulation is both highly effective and generalizable, with strong potential to serve as a foundation for future research in motion generation.

      Table A2: Quantitative text-to-motion evaluation with SMPL parameter

Q3: Response to Incomplete Theoretical Derivation and Formulas in Section 3.

      We thank the reviewer for the suggestion. We would like to clarify that our work does not propose a new model architecture or motion representation that necessitates new theoretical derivations. Instead, our contribution lies in critically re-examining the dominant practice in text-driven motion generation—the widespread use of root-relative (local) representations—and showing that a simple yet long-overlooked absolute-coordinate representation can be surprisingly effective. Even combining this representation with a simple Transformer architecture and no auxiliary kinematic losses, we achieve strong empirical performance across multiple tasks. We emphasize that our goal is not to introduce complex new formulations with theoretical derivation, but rather to challenge de facto assumptions through extensive experiments. Our motivation for this shift in representation is discussed in Lines 35–43. We believe the strength of our work lies in its empirical rigor and potential to reframe future research directions in this text-driven motion generation field.

Thank you for your detailed response. I think current rebuttal has not fully addressed my concern.

For Q3, although it has been experimentally demonstrated that its method works, it does not explain why it works from a theoretical perspective? In the new data application scenario, the model will have different feature representations, which I think is also a necessary part of the explanation.

To Reviewer 4XhZ:

      Thank you for your insightful feedback and for recognizing the strength of our core idea, the quality of our visualizations, the clarity of our writing, and the design of our model scaling experiments. Below, we provide detailed responses to the weaknesses and questions you raised.

Q1: Response to Diversity Concerns.

      We thank the reviewer for raising this important point. The MultiModality metric proposed by $25$ and reported in Table 2 evaluates the diversity of generated motion embeddings conditioned on the same text prompt, as detailed in the Appendix. It is important to note that this is a secondary motion quality metric, meaning its value is only meaningful when core metrics such as FID and R-Precision are also strong.
      In our case, ACMDM achieves better FID and R-Precision compared to prior methods, while also maintaining strong MultiModality scores (2.156 for ACMDM-S-PS2), even surpassing several relative-representation baselines such as MLD++ $14$ (1.924) and MotionLCM v2 $14$ (1.780). This demonstrates that our absolute-coordinate formulation does not compromise motion diversity while maintaining better motion quality.

Q2: Response to Failure Cases.

      We did not observe any noticeable artifacts on the HumanML3D $25$ testset. However, similar to all baseline methods, for certain out-of-distribution prompts involving physically implausible scenarios such as "a person is flying", the model tends to produce motions like standing while waving hands, rather than attempting to achieve unrealistic flight behavior.
      Due to the NeurIPS rebuttal policy, we are unable to include additional video and image samples at this stage. However, we will provide extended qualitative results with failure case visualizations in the revised version.

Q3: Regarding Duplicate Citations.

      We thank the reviewer for pointing this out. We will carefully review and correct this in the revised version.

Q4: Training and Computational Cost.

      Thank you for the question. Our model uses a simple Transformer backbone, which is computationally similar to the baseline MDM, with the addition of AdaLN for conditioning. For the smallest variant, ACMDM-S-PS22 (8 layer 512 dim same as MDM $91$ ), training is computationally efficient, which only requires approximately 8 GB of GPU memory and completes in approximately 8 hours on RTX 4090. It is also faster to train than MDM, thanks to the temporal downsampling in our 2D-AutoEncoder. Full scaling details and configurations are provided in the Appendix. While the largest variant is more resource-intensive, it still trains within 2 days on an NVIDIA H200. Importantly, inference remains lightweight and can be performed on consumer GPUs such as the RTX 4090 or 5090. ACMDM-S-PS22 achieves an AITS (Average Inference Time per Sample) of 1.35s. Moreover, since our diffusion objective is formulated as velocity, it supports the use of even fewer-step ODE solvers (e.g., ODE-10) during inference for further efficiency.

To Reviewer suN9:

      Thank you for your insightful feedback and your recognition of our work’s clarity in writing, thoroughness in analysis, and the quality and quantity of our visual results. Below, we provide detailed responses to the weaknesses and questions you raised.

Q1: Response to Fairer Comparison with Baseline Methods.

      Following the reviewer’s suggestion, and to ensure a fairer comparison, we retrained all baseline methods that originally used the full HumanML3D representation with the reduced, essential representation and present the results in Table A1. While most baselines showed only minor performance fluctuations, MLD++ $14$ and MotionLCM v2 $14$ exhibited more noticeable changes on the FID metric. This suggests that these methods may overfit to redundant dimensions when trained on the original HumanML3D $25$ representation. Nevertheless, our method consistently achieves state-of-the-art performance across major metrics. We would also like to emphasize that, unlike baseline methods, our approach is not only effective for text-to-motion generation but also well-suited for downstream tasks and demonstrates strong scalability.
      As highlighted in MARDM $66$ , the original HumanML3D evaluator is problematic because (1) it unnecessarily emphasizes redundant dimensions, and (2) all baseline methods can directly output essential, animatable dimensions without any further conversion. We thus strongly argue against evaluating methods with redundant dimensions, as it can be both inaccurate and misleading for real animation quality. Importantly, we believe that all motion generation methods should be optimized for producing the most perceptually faithful outputs, rather than fitting to evaluator-specific artifacts. Therefore, we also include results from a user study (see Response 4) to provide a more reliable and human-aligned assessment of generated motion quality.

      Table A1: Quantitative text-to-motion evaluation with baseline methods retrained with reduced representation.

Q2: Response to Bone Length Consistency with Absolute Joint Coordinates.

      We thank the reviewer for raising this point. We would like to clarify a common misconception: relative joint positions do not inherently guarantee constant bone lengths. It is possible to predict arbitrarily small or large relative offsets that violate bone length constraints. Only rotational representations (e.g., joint rotations) inherently preserve bone lengths by using a fixed template skeleton. However, the rotation data in HumanML3D is known to be redundant and inaccurate (as discussed in our paper Line 28, and noted in HumanML3D’s GitHub issue), thus unable to be used for animation. Consequently, no baseline method using HumanML3D’s relative joint positions can strictly preserve bone length.
      To evaluate ACMDM’s ability to maintain bone length consistency, we conducted two quantitative tests on the HumanML3D testset:
      (1) Intra-sequence bone length stability – measuring changes in bone lengths within a generated sequence.
      (2) Bone length consistency with ground truth – comparing generated bone lengths against those from the ground truth.
      We also compared ACMDM to SOTA relative representation baselines with results presented in Table B1. Results indicate that ACMDM not only maintains stable bone lengths but also outperforms some relative baselines, such as MotionLCM V2 and MARDM.
      Due to the NeurIPS rebuttal policy, we cannot include additional videos, but we observed no noticeable bone length artifacts in the HumanML3D testset. To further probe robustness, we tested ACMDM on 10 out-of-distribution prompts involving “flying” (e.g., “a person is flying in the air”), which are impossible motions. Despite the unusual prompts, the generated motions mostly showed realistic behavior (e.g., standing, waving), and intra-sequence bone length variation remained very low (0.0071 m), confirming the reliability of our approach.

      Table B1: Bone length consistency on the humanml3d testset.

Q3: Regarding Future Envision and Practical Usability of Direct Mesh Generation.

      In the paper, we demonstrate strong results with direct SMPL-H $59$ mesh generation. We view direct mesh generation as the next step toward simplifying the 3D character animation pipeline. Historically, animators had to draw each frame of a character’s motion manually. The introduction of rigging and skinning revolutionized this process by enabling skeleton-based animation to drive meshes, significantly reducing human labor. Our approach aims to take this one step further, providing an end-to-end solution where a user only needs to provide a text prompt or a static 3D character, and the generative model directly produces realistic animated motion without intermediate skeletons or rigging.
      As highlighted in our visualizations, direct mesh modeling captures finer 3D details and implicit motion dynamics that skeleton-to-mesh pipelines either fail to reproduce or require additional manual work to achieve. Moreover, by representing all characters as sets of absolute coordinates, the absolute-coordinate approach has the potential to generalize across diverse characters without the need for predefined kinematic chains for each character.
      We believe that, in the future, this direction will offer a more universal, streamlined, and automation-friendly solution for character animation, reducing manual labor while achieving higher motion fidelity.
      Regarding computational resources, as discussed in Section 3.3, our diffusion model operates in latent space by encoding SMPL-H meshes (6890×3) into compact latent representations of size (28×12). With motion diffusion patch size set to 1x28, the inference cost becomes equivalent to that of joint-level generation (for smallest variant ACMDM-S-PS28, AITS=2.37s), making our approach very efficient despite working directly with mesh data.

Q4: User Study.

      We conducted a user study with 15 participants to evaluate ACMDM against SOTA baseline methods across a total of 21 motion comparisons, including both Text-to-Motion generation (15 comparisons) and Motion Control tasks (6 comparisons), with motion captions drawn from the HumanML3D test set. For each comparison, participants were shown side-by-side outputs from ACMDM and baseline methods and asked to choose the motion that appeared most natural, text-aligned, and, on Motion Control, also accurately followed the control signal.
      In Text-to-Motion generation, ACMDM was the most preferred method, selected in 62.22% (140/total 225) of all cases, compared to 15.56% (35/225) for MLD++, 8.44% (19/225) for MotionLCM v2, and 13.78% (31/225) for MARDM. In Motion Control, ACMDM was again the most preferred, with 78.89% (71/total 90), outperforming OmniControl (21.11%; 19/90 ) and MotionLCM v2 + CtrlNet (0% 0/90). These results demonstrate that ACMDM is consistently favored by users in terms of perceptual quality and control fidelity.