To Reviewer 76BV

We sincerely appreciate Reviewer 76BV for the thoughtful review. Below, we provide detailed responses to each comment.

Q1：What are the limitations of the proposed approach, and what could be future improvements over WaveAR?

Thanks to the reviewer for the reminder. While WaveAR achieves strong average accuracy (ADE&FDE), it does exhibit two main limitations: Our masked autoregressive diffusion module focuses on minimizing reconstruction error over the average trajectory, which can lead to overly “safe” predictions, with diversity confined to a relatively small range of motion variations; Besides,current approach relies on standard joint‑coordinate datasets containing only about 20 keypoints, limiting its ability to capture finer movements such as hand motions .Enabling fine‑grained motion details is thus an exciting direction for future research.

Q2:Ablations on the DCT vs. Wavelets

From the table, we see that incorporating spectral information improves motion prediction accuracy, and that DWT—by capturing high-frequency details in the input beyond what DCT can—yields even more precise predictions.

Q3:Differences between training and inference process

Thank you for your careful review and valuable suggestions. As stated in paper L167-L168 "To inject these spectral cues into our autoregressive diffusion backbone, we apply vanilla DWT operation to the raw 3D–joint history sequence along both temporal and channel axes." The wavelet features are computed only on the observed history frames of length H at both training and inference time.The upper limit H+F in Equation (3) merely denotes the length of the input sequence for the wavelet transform computation. This may cause ambiguity and will revise it in the next version for clarity. For the masking strategy, the masking applied to the latent sequence during training is performed randomly across the entire sequence according to a predefined probability, not limited to future prediction frames. This flexible masking strategy enables our model not only to perform standard autoregressive prediction of future frames from history but also to support more complex tasks such as motion switching. In summary, the wavelet features provide crucial spectral cues that enhance temporal understanding, and the autoregressive denoising diffusion model delivers strong generative power to ensure accurate and robust sequence reconstruction.

Q4:Minor typos

We thank the reviewer for careful attention in identifying these typos and ambiguities; we will correct them in the final version.

To Reviewer p7mw

We sincerely appreciate the valuable feedback from Reviewer p7mw. Our responses to each comment are provided below.

Q1: Results on AMASS dataset

Thank you for your suggestions. We conducted experiments on the AMASS dataset. As shown below, our model still achieves the best accuracy metrics (ADE&FDE), demonstrating strong generalization capability.

Q2：Trade-off between accuracy and diversity

Thank you very much for your valuable suggestion regarding the analysis of accuracy-diversity trade-off. We acknowledge that our current model prioritizes prediction accuracy, resulting in relatively limited diversity. Previous methods have achieved higher diversity often at the expense of generating unrealistic or physically implausible motions such as discontinuities in movement [1,2,3], an issue that does not arise in our proposed approach. To demonstrate the robustness and realism of our model predictions, we provide detailed results in the supplementary materials showing consistently accurate performance across various fine-grained action categories in the Human3.6M dataset. This further supports the validity and robustness of our method. Nevertheless, we recognize that the limited diversity remains a current limitation of our model. Improving diversity while preserving realism is an important future research direction. We will add this part in the Limitation section. In the future, we will pay more attention to this topic.

Reference:

[1]Mao, W., Liu, M., Salzmann, M.: Generating smooth pose sequences for diverse human motion prediction. In: ICCV. pp. 13289–13298 (2021)

[2]Dang, L., Nie, Y., Long, C., Zhang, Q., Li, G.: Diverse human motion prediction via gumbel-softmax sampling from an auxiliary space. In: MM. pp. 5162–5171 (2022)

[3]Wei, D., Sun, H., Li, B., Lu, J., Li, W., Sun, X., Hu, S.: Human joint kinematics diffusion-refinement for stochastic motion prediction. In: AAAI. pp. 6110–6118 (2023)

Q3:Inference time and model size

Thank you for your suggestion. We evaluated inference time and parameter count on NVIDIA V100 GPU by comparing three model variants of different sizes.

Q4:Lack of Limmitations:

Thanks to the reviewer for the reminder. While WaveAR achieves strong average accuracy, it does exhibit two main limitations: Our masked autoregressive diffusion module focuses on minimizing reconstruction error over the average trajectory, which can lead to overly “safe” predictions, with diversity confined to a relatively small range of motion variations; Besides,current approach relies on standard joint‑coordinate datasets containing only about 17–32 keypoints, limiting its ability to capture finer movements .Enabling fine‑grained motion details is thus an exciting direction for future research.

Q5:Minor typos

Thank you for the reminder of minor typos, we will correct them in the final version.

Dear Reviewer p7mw,

The deadline for author-reviewer discussion period is approaching. We kindly ask you to review the authors' rebuttal. Please provide your feedback soon. Thank you.

Best,

AC

To Reviewer fPmk

We sincerely appreciate Reviewer fPmk's thoughtful comment and insightful feedback. Please see our detailed response below.

Weakness (1):Limited Scope of Evaluation Benchmarks

Thank you for your suggestions. We conducted experiments on the AMASS dataset. As shown below, our model still achieves the best accuracy metrics (ADE&FDE), demonstrating strong generalization capability.

Weakness (2):Omission of a Key Baseline Comparison

Thank you for your suggestion. We compare MotionWavelet with our model and get the following results. FDE measures accuracy on the final frame, while ADE measures accuracy over the entire predicted sequence. Although MotionWavelet achieves a better FDE, its relatively poor ADE suggests it can produce inaccurate predictions during intermediate frames. Since ADE is a more important metric for overall trajectory accuracy, our method clearly outperforms MotionWavelet. Moreover, because MotionWavelet’s code is not publicly available, we re-implemented its algorithm based on the paper and compared inference speeds, demonstrating that our model runs significantly faster than MotionWavelet. We will add the results to the final version.

Weakness (3). Potentially Overstated Novelty Claim

We appreciate the reviewer pointing out this issue. We agree that our original statement lacked precision and could lead to misunderstandings. To clarify, our approach is the first to apply masked autoregressive modeling for human motion prediction. We will correct our statement in the final version.

Questions (1). Evidence for the statement in line3

Autoregressive methods relying on VQ have achieved excellent performance on text-to-motion [1,2,3]. However, the presence of quantization error [2] and codebook collapse [4] makes it difficult to generalize these methods to other motion generation tasks. As noted in MARRS [5], VQ-VAE-based methods achieve poor performance for human action-reaction systhesis. For the human motion prediction, the results in Table 3 of our paper also demonstrate the instability of the VQ-VAE-based methods. For ease of reading, we have put the results from Table 3 in the paper into the rebuttal. When we replace our proposed ST-VAE with a VQ-VAE, the model’s performance degrades significantly across all metrics.

References:

[1] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730–14740, 2023. 2, 3

[2] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900–1910, 2024. 1, 2, 4

[3] kkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1546–1555, 2024. 1, 2, 4, 8

[4] huanxia Zheng and Andrea Vedaldi. Online clustered codebook. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22798–22807, 2023. 1

[5] YB Wang, Shuo Wang, JN Zhang, JF Wu, QD He, CC Fu, CJ Wang, and Yong Liu. Marrs: Masked autoregressive unit-based reaction synthesis. arXiv preprint arXiv:2505.11334, 2025. 8

To Reviewer n6UW

We sincerely appreciate the valuable and insightful feedback from Reviewer n6UW. We will address all the concerns and polish our paper accordingly.

Weakness (1). advantages of continuous autoregressive generation paradigm

First, continuous generation avoids the inherent limitations of vector quantization (VQ) (e.g. the loss in mapping continuous motion sequences to discrete tokens[1], and the difficulty in optimizing the training process due to notorious codebook collapse[2]). Second, autoregressive generation has the following advantages: 1) The motion sequences are temporal, which is a natural fit with autoregressive modeling. 2) Better ability to capture long-distance dependencies and better ability to model long sequences. 3) Faster training and inference speed for autoregressive generation compared to diffusion's modeling of global distributions. Thanks for your valuable suggestion about re-summarizing the innovation point. We will change this part in the revised version.

Weakness (2). inference speed computational efficiency

Thank you for your suggestion. We evaluated inference time and parameter count on an NVIDIA V100 GPU by comparing three model variants of different sizes. Thanks to the reduced number of denoising steps, our model achieves faster inference while maintaining high prediction accuracy.

Questions (1). in-depth analysis of DWT

First, we visualize the spectrum of WaveAR w/o and w/ DWT, respectively. Since it is not possible to add images in the rebuttal, we count the percentage of high-frequency component averages. We randomly sampled 200 instances, and w/ DWT, the average proportion of the amplitude of high-frequency components is 0.2429, whereas w/o DWT, the average proportion of the amplitude of high-frequency components is 0.1624. This demonstrates that the DWT effectively captures the high-frequency component of the motion. Second, we found some “static” phenomena in the generated motion, i.e., the person is hardly moving. After applying DWT, these cases will generate motion with more details instead of "static", , demonstrating that high-frequency information leads to richer motion details. Since it is not possible to add images in the rebuttal, we will update this part of the visualization to the supplementary materials in the revised version. We also conduct an experiment on the ablation between DCT and Wavelet module. The results are shown below. We can see that incorporating spectral information improves motion prediction accuracy, and that DWT—by capturing high-frequency details in the input beyond what DCT can—yields even more precise predictions.

Questions (2). inference time and model size

Please refer to Weakness(2).

References:

[1] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900–1910, 2024. 1, 2, 4

[2] huanxia Zheng and Andrea Vedaldi. Online clustered codebook. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22798–22807, 2023. 1