FlashMo: Geometric Interpolants and Frequency-Aware Sparsity for Scalable Efficient Motion Generation

We appreciate your valuable feedback and support. Our responses to the questions are as follows.

Q1. Manifold geometry for latent representation.

A1: Thank you for bringing up this question. Below is a clear explanation.

In order to preserve latent representations on the $\mathrm{SO}(3)$ manifold, our paper simply adapts a VAE by replacing the standard Gaussian latent space with an $\mathrm{SO}(3)$-valued latent space and applying a generalized reparameterization trick.

Specifically, in a vanilla VAE, the latent variable

$

z \sim \mathcal{N}(\mu, \sigma^2 I)

$

lies in Euclidean space and is sampled using the standard reparameterization trick:

$

z = \mu + \sigma \cdot \epsilon, \quad \epsilon \sim \mathcal{N}(0, I).

$

In contrast, in an $\mathrm{SO}(3)$-VAE, the latent variable

$

R_z \in \mathrm{SO}(3)

$

is sampled using the following reparameterization trick:

$

v \sim r(v \mid \sigma)

$

$

R = \exp(v^\times) \quad \text{(mapped to } \mathrm{SO}(3) \text{ via exponential map)}

$

$

R_z = R_\mu R \quad \text{(left multiplication to center at the mean rotation } R_\mu)

$

where $v^\times$ is the skew-symmetric matrix of $v$.

This ensures all latent variables $R_z$ lie exactly on the SO(3) manifold without changing the architecture of the VAE. Hence, it’s also necessary to perform lie group geodesics interpolant in the latent space.

Similar VAE methods have been introduced in previous works such as [1]. Since it is not the main focus of our contribution, the details have not been explicitly included in the current version of the paper. We will include it in the next version of our paper.

Q2: Motion representation.

A2: HumanML3D introduces a preprocessing step that converts motion into velocity and joint position representations, along with the corresponding text descriptions. However, the original SMPL sequences can be retrieved by locating the exact index in AMASS [2], where the SMPL representation preserves the manifold geometry of joint rotations and is used for training in FlashMo. Moreover, as mentioned in the paper, the output motion sequence needs to be converted to the pose representation defined in HumanML3D to ensure consistency across evaluations and enable fair comparisons with other methods.

Once again, thank you for the valuable feedback. Your support means a lot and has helped us improve our work. We hope you will consider raising your score if you feel our response has addressed your concerns.

Reference:

[1] Explorations in Homeomorphic Variational Auto-Encoding

[2] AMASS: Archive of Motion Capture as Surface Shapes (ICCV 2019)

Dear AC,

Thank you for bringing out this important question.

There is a theoretical guarantee that the exponential map will always ensure the result lies on the SO(3) manifold if the input to the exponential map lies in Lie algebra so(3) [1]. This is the nature of the exponential map, which maps the Lie algebra of 3×3 skew-symmetric matrices (or equivalently, the angle-axis representation) to a rotation [2]. That is, as you said, with the exponential map as preprocessing to the data, you will always get inputs and outputs of a standard VAE that lie on SO(3).

However, simply converting rotation data into exponential map form does not guarantee manifold consistency in the latent space of a (latent) generative model [3]. By manifold consistency, we mean that all operations (including latent encoding, sampling, interpolation, and decoding) must preserve or respect the geometry of the SO(3) manifold. The standard VAE will treat the exp-map rotations as just a collection of real numbers ($\mathbb{R}^3$) in latent space and if we embed a non-Euclidean manifold into a Euclidean latent space, we encounter a manifold mismatch problem [4]. There is a diagram in Falorsi's paper [4] that explains this: A nontrivial manifold (top) with holes cannot be continuously mapped to a topologically trivial blob (bottom), illustrating the concept of “manifold mismatch” between data and a Gaussian latent space.

Specifically, even if we map the rotations $\mathbf{M}$ into exp-map form $\mathbf{M}^* \in \text{SO}(3)$ via exponential maps and feed $\mathbf{M}^*$ into a standard VAE, this alone does not guarantee that the learned latent representations will remain on the SO(3) manifold. This is because a standard VAE maps $\mathbf{M}^*$ into an unconstrained latent space (typically latent vectors $\mathbf{x} \in \mathbb{R}^d$), optimized to match an isotropic Gaussian prior, such that:

which implies $\mathbf{x} \notin \text{SO}(3)$ in general. As a result, the manifold structure of the input rotations is not preserved — the latent variables $\mathbf{x}$ no longer lie on SO(3). That is, the latent distribution in a standard VAE is trained to match a unit Gaussian, not a curved manifold, and thus fails to preserve the rotational geometry of $\mathbf{M}^*$.

Furthermore, standard stochastic interpolants used in SiT operate in Euclidean latent space:

which linearly blends the latent code $\mathbf{x}^*$ with Guassian noise $\boldsymbol{\epsilon}$, without respecting any underlying geometry (manifold structure) present in the source data. Therefore, even if the input is on SO(3), the diffusion interpolant operates off-manifold, potentially producing invalid rotations or semantically inconsistent transitions.

It is worth noting that, as you mentioned, Grassia's [5] paper provides a solid theoretical foundation for handling rotations in computation for our work. However, since it (1998) predates modern deep generative models, it does not discuss VAEs or latent diffusion.

We hope the above clarification addresses your question, and we are happy to respond to any further questions.

Best regards,

Authors #6198

Reference:

[1] PA Absil, R Mahony, R Sepulchre, Optimization algorithms on matrix manifolds. (2008)

[2] Richard Hartley, et al. Rotation averaging. (IJCV 2013)

[3] Falorsi et al. Reparameterizing Distributions on Lie Groups (AISTATS 2019)

[4] Falorsi et al. Explorations in Homeomorphic Variational Auto-Encoding (2018)

[5] Grassia, F. S. Practical parameterization of rotations using the exponential map.Journal of Graphics Tools. (1998)

Dear AC,

We now see that one of your suggested potential justifications is for the frequency filtering. However, as you mentioned, there is already existing work that applies FFT to abstract latent representations.

We agree with the AC that this is a good point, and the suggested vanilla VAE setup with the tangent space helps make our work more comprehensive.

Meanwhile, please forgive us for the delayed response, as we needed some time to run experiments to verify our justification, and some of the responses arrived around midnight in our local time. But we will do our best to address any questions!

Best regards,

Authors #6198

Dear AC,

Thank you for your further clarification and patience. Now we understand that what you’re suggesting is that there seems to be a simpler approach: applying the logarithmic map to convert the rotations into the tangent space first, then training a standard VAE and SiT in the tangent space (valid rotation), and finally applying the exponential map on decoding.

One of the reasons we designed a more complex approach (adopting SO(3)-VAE and a geometric interpolant) is that the simpler method you suggested may encounter a non-unique mapping problem.

The exponential map from the Lie algebra $\mathfrak{so}(3)$ (tangent space) to the group $SO(3)$ is surjective but not injective [1]. In other words, many different tangent vectors can correspond to the same rotation. For example, a rotation by 180° about some axis is represented in the tangent space by a vector of length $\pi$ in that direction, but also by a vector of length $\pi$ in the opposite direction – two distinct vectors map to one rotation. Hence, mapping a Euclidean latent to $SO(3)$ inherently involves such non-unique correspondences. A vanilla VAE in $\mathbb{R}^3$ would have to arbitrarily choose or learn a single representative for each rotation, or risk outputting inconsistent representations (e.g. spontaneously jumping between equivalent rotations). This ambiguity can complicate learning and sampling in vanilla VAE. By constraining the SO(3)-VAE in our method, it will not face this issue (homeomorphic).

To verify our justification, we conducted a fair experiment on HumanML3D with a consistent training setup (6k epochs). The results in the following table indicate that this learning approach introduces more difficulty in VAE training and decreases performance.

I hope this additional explanation answers your question. We are happy to address any further questions.

Best regards,

Authors #6198

Reference:

[1] Luca Falorsi et al. Reparameterizing Distributions on Lie Groups (AISTATS 2019)

We appreciate your valuable feedback and support. Our responses to the questions are as follows.

Q1: High-low frequencies in Figure 1.

A1: In Figure 1, the frequency of the motion features is mapped onto the motion sequence to better illustrate motion frequency. This visualization shows that motion frequency varies across both the temporal (frame-wise) and spatial (body-part-wise) dimensions. Dynamic frames appear in lighter colors compared to others, and similarly, dynamic body parts are shown in lighter colors compared to static ones. In our paper, frequency-aware sparsification is applied to the temporal attention, i.e., token pruning is performed along the temporal dimension.

To extend sparsification to both temporal (frame-wise) and spatial (body-parts) dimensions, one could simply adopt a standard temporal-spatial attention structure, where temporal and spatial attention layers are alternately stacked. To verify this, we conducted an ablation study on HumanML3D, which shows slight improvements in both generation quality and efficiency. This is because applying sparsification to both temporal and spatial dimensions provides finer-grained sparsification granularity. These results are not included in the current version of the paper, as the spatial-temporal architecture has been widely discussed in prior works [1,2] and is not the main focus of our contribution. However, we will include them in the next version of our paper.

Once again, thank you for the valuable feedback. Your support means a lot and has helped us improve our work. We hope you will consider raising your score if you feel our response has addressed your concerns.

Reference:

[1] MoGenTS: Motion Generation based on Spatial-Temporal Joint Modeling (NeurIPS 2024)

[2] Motion Mamba: Efficient and Long Sequence Motion Generation (ECCV 2024)

Dear Reviewer,

Thank you for your valuable feedback and support. If we have addressed your concerns, we kindly ask you to consider raising your score. We are happy to address any further questions.

Best regards,

Authors #6198

Thanks for your prompt response and postive feedback. We would like to provide further clarification based on your comment.

Q2. Efficiency improvement comes with a trade-off in quality.

A2. First, it is important to clarify that our method achieves an extraordinary balance between generation quality and efficiency, both of which are critical and essential to the overall performance of a model. This has already been acknowledged in your original review: “Frequency-Aware Sparsification… This enables faster computation while balancing with quality.”

The trade-off between quality and efficiency is not a problem unique to sparsification; it is a general challenge faced by efficiency AI. Token sparsification prunes redundant and less important tokens from the original set to improve efficiency, but this inherently introduces information loss—a limitation not exclusive to our method. Similarly, step distillation methods, such as MotionLCM and MotionPCM, accelerate inference by relying on consistency models but require fewer inference steps. Linear architectures, such as Motion Mamba, reduce computational complexity but often introduce more complex scanning mechanisms to maintain model capacity.

Hence, the goal of our efficient method is not simply to trade off quality for efficiency, but to achieve an extraordinary balance — which is exactly what our method accomplishes.

Q3. Falls short of the performance achieved by MoGenTS and MotionPCM, and the results on the MotionHub V2 pretraining may not be a fair basis for comparison.

A3. This question has been addressed in 8uxk A2.1. As mentioned in the rebuttal, MotionPCM and MoGenTS do not outperform FlashMo (without pretraining). Therefore, the comparison is fair, as FlashMo is also a train-from-scratch model in the first line of Table 1.

Below is a sub-table extracted from the original Table 1, which presents the comparison on HumanML3D.

It is evident that FlashMo (without pretraining) achieves SOTA performance on 6 out of 8 metrics on HumanML3D, while MotionPCM achieves state-of-the-art results on at most two metrics (across two model variants), and MoGenTS does not achieve state-of-the-art on any metric.

Moreover, in terms of FID, the only method with close efficiency to FlashMo (0.027 AITs) is MotionPCM (1-step) at 0.031 AITs. However, its FID score is 0.044, which does not outperform FlashMo’s 0.041.

This performance is unrelated to the MotionHubV2 dataset, as FlashMo in this table does not use any pretraining.

We hope this clarification addresses your concerns, and we are happy to answer any further questions. Once again, thank you for your positive feedback and support.

Dear Reviewer uEGj,

Thank you for your consistent support for our work!

We welcome further discussion to see whether our response addresses your remaining concerns. To save your time, we provide a brief summary below:

1. FlashMo outperforms SOTA methods with comparable AIT (including FID).

The only method with comparable efficiency to FlashMo (0.027 AITs) is MotionPCM (1-step) at 0.031 AITs (even slower). FlashMo (sparse, no pretrain) achieves a better FID (0.041) than MotionPCM (1-step, FID 0.044). Moreover, FlashMo (full attention, no pretrain) achieves the lowest FID (0.029), outperforming all methods. Moreover, our method achieves significant improvements in other metrics as well. (The detailed results is are tables of 8uxk A4&A5.)

2. FlashMo's performance gains from its advanced interpolant and architectural design, not simply from data-driven training.

(1) To showcase and ablate the performance gains from our method design, we adopt full attention with temporal-spatial attention. The results of FlashMo (full attention, no pretrain) demonstrate that our method design improve not only FID (0.029) but also overall generation. This improvement is unrelated to MotionHub V2, as no pretraining is used in this additional ablation.

(2) The other key contribution of FlashMo is scalability. We conducted a fair comparison between MoGenTS and ours on HumanML3D, both of which are pretrained on MotionHub V2. FlashMo (FID 0.041→ 0.029) demonstrates significantly better scalability than MoGenTS (FID 0.033 → 0.052).

We hope these answers address your remaining concerns, and we kindly hope you will consider raising your rating if your concerns have been mainly addressed.

Once again, thank you for your time and support! We wish you a successful run at NeurIPS!

Best regards,

Authors #6198

We appreciate your valuable feedback and support. Our responses to the questions are as follows.

Q1: Figure 2(d) details.

A1: As shown in Figure 2(d) and Section 3.3, the input tokens 𝑋 are first projected into queries 𝑄 and keys 𝐾, which are used to compute attention scores. These attention scores indicate the importance of each token and guide the selection of a subset of tokens 𝑍, where low-importance tokens are pruned. This adaptive token selection reduces redundancy while preserving key information for subsequent attention computation. We will include a more detailed figure in the next version of the paper.

Q2: Method robustness to the hyperparameter γ.

A2: Thank you for bringing up this important question. The attention threshold γ is related to the number of adaptively selected tokens, as mentioned in Section 3.3 (Line 173). In token sparsification, model performance is sensitive to the number of pruned tokens, and pruning more tokens leads to greater information loss, which the similar pattern can be observed in both vision and language models [1]. Hence, sparsification methods aim to prune redundant tokens while preserving as much performance as possible.

This exactly demonstrates the benefit of our adaptive token pruning, which significantly reduces information loss compared to methods using a fixed number of pruned tokens (e.g. predefined attention mask, or hard token threshold). This advantage can be also demonstrated in the experiment on KIT-ML in the following table, where the same sparsification hyperparameters remain applicable, which verify the method robustness to our hyperparameters.

Q3: Outperform the real data in R precision.

A3: This is a classic question. R-Precision evaluates the alignment between a generated motion and its corresponding text description. For each motion sample, a pool of 32 descriptions is created, including the ground-truth text and 31 randomly selected mismatched texts from the test set. The Euclidean distances between the motion feature and the text features in the pool are computed and ranked. If the ground-truth description appears in the top-k closest matches, it is counted as a successful retrieval. The final R-Precision score is the average retrieval accuracy at Top-1, Top-2, and Top-3 positions [2].

Since the R-Precision calculation relies on the distances in the feature space, in some cases, the generated motion may align more closely with the semantics of the text in the embedding space than the ground-truth motion itself. This could happen if the generated motion is smoother, less noisy, or more "canonical" than the real motion, making it easier for the model to match it to the text. As a result, it's possible for the R-Precision of generated motion to slightly surpass that of the ground truth. You can also observe this pattern in other works, such as MoGenTS [3], MotionPCM [4], etc. Hence, in some works, such as MoMask [5], ground truth values are not mentioned in their tables.

Once again, thank you for the valuable feedback. Your support means a lot and has helped us improve our work. We hope you will consider raising your score if you feel our response has addressed your concerns.

Reference:

[1] HiRED: Attention-Guided Token Dropping for Efficient Inference of High-Resolution Vision-Language Models (AAAI 2026)

[2] Generating Diverse and Natural 3D Human Motions from Text (CVPR 2022)

[3] MoGenTS: Motion Generation based on Spatial-Temporal Joint Modeling (NeurIPS 2024)

[4] MotionPCM: Real-Time Motion Synthesis with Phased Consistency Model

[5] MoMask: Generative Masked Modeling of 3D Human Motions

Dear Reviewer,

Thank you for your valuable feedback and support. If we have addressed your concerns, we kindly ask you to consider raising your score. We are happy to address any further questions.

Best regards,

Authors #6198

Dear Reviewer GZdj,

Thank you for your positive feedback. We’re glad to hear that your concerns have been adequately addressed.

We kindly hope you will consider raising your rating to further champion our paper. Your support means a lot to us, and we are happy to answer any further questions.

Once again, thank you for your time and support! We wish you a successful run at NeurIPS!

Best regards,

Authors #6198

We appreciate the reviewer’s valuable feedback and reponse to the concerns as follows.

Q1.1: Motivation of sparse multi-head attention.

A1.1: Our motivation of sparse multi-head attention is recognized by Reviewer GZdj, “The frequency-aware sparse attention mechanism is well-motivated…”.

Futhermore, in our paper, the motivation of sparse multi-head attention has explicitly mentioned in lines 28-31 “Efficiency. Motion diffusion suffers from high computational cost, as well as long training and inference times. Existing efficient methods such as step distillation [1,2], step reduction [3], and linear models [4,5] either require additional training of teacher models or involve complex scanning mechanisms, resulting in overengineering and low efficiency.” and lines 37-40 “We observe that dynamic motion, which is more important, exhibits higher frequency compared to static motion in generation process, as shown in Figure 1. Hence, we design a frequency-aware sparsification mechanism that dynamically prunes tokens corresponding to low frequency motion, enhancing efficiency while preserving high frequency motion at the attention head level.”

Q1.2: Ablation study.

A1.2: The ablation study of each component including geometric factorized interpolant and frequency-aware sparsification have shown in Table 2 and 3.

(1) In Table 2, our interpolant method significantly improves generation quality compared to other methods.

(2) In Table 3, our frequency-aware sparsification significantly improves efficiency while largely preserving generation quality compared to the full attention baseline.

Both generation quality and efficiency are non-negligible perspectives of overall motion generation performance, which is recognized by Reviewer uEGj, “This enables faster computation while balancing with quality”.

Q1.3: Frequency-aware sparsification for different architecture.

A1.3: Sparsification methods are often designed for a series of models with similar or shared architectures. e.g. Sparse VideoGen [6] introduces a semi-structured sparsification method by designing a predefined attention mask based on the observed attention patterns in CogVideoX [7] and HunyuanVideo [8], both of which share a common video DiT architecture.

Moreover, our frequency-aware sparsification is a structured sparsification method with adaptive token selection, without predefined attention masks, which expands the potential of our method to adapt to other architectures.

Q1.4: Exhibit more realistic dynamics or richer frequency.

A1.4: This question relates to the fundamental mechanism of token sparsification methods, which prune tokens from the original set, reducing redundant and less important information without introducing new content [9,10], i.e., token sparsification methods aim to improve efficiency while preserving important information as much as possible (which impacts generation quality); they do not enrich motion frequency or exhibit more realistic dynamics.

Q2.1: Evaluation and concerns about w/o pretrain.

A2.1: In Table 1 and Figure 6 of our paper, the performance of FlashMo w/ pretraining is to demonstrate the outstanding scalability of our method, which does not hinder the performance gains achieved through our method’s advancements.

In Table 1, FlashMo (without pretrain) achieve SOTA in 5 out of 8 metrics in HumanML3D and 4 out of 7 metrics in KIT-ML, while MotionPCM achieves SOTA on at most two metrics (1-step) and MoGenTS does not achieve any SOTA on HumanML3D, similarly, MoGenTS achieves SOTA on only two metrics and MotionPCM does not achieve any SOTA on KIT-ML.

This has been widely recognized by Reviewer GZdj, 'The proposed method shows better performance and becomes more efficient,' and by Reviewer HnUJ, 'The proposed method achieves state-of-the-art performance with the lowest inference time, training time, model size, and GFLOPs.’

Q2.2: No public access for MotionHub V2.

A2.2: The MotionHub V2 dataset is publicly available on the GitHub repository of MotionLLAMA [11], provided via a BaiduPan link.

Q3: Frequency-aware head partition and figure 4.

A3: As defined in lines 156–158 of our paper, attention heads are partitioned based on their frequency into Hi-Fi and Lo-Fi groups, where β is the partition ratio, which is ablated in Table 3.

As for the procedure for computing the frequency pattern in the figure, the caption of Figure 4 in our paper mentioned: 'The frequency magnitude is computed using the Fast Fourier Transform (FFT) and averaged over 100 latent motion features’. Specifically, the frequency magnitude of each attention head is calculated with a Fast Fourier Transform (FFT). We take the output feature maps from either the Hi-Fi or Lo-Fi attention head. For each feature map, we apply FFT to convert the latent representation into the frequency domain. We then compute the magnitude of the frequency components and apply a logarithmic transformation for stability and better interpretability. To ensure a reliable estimate, we randomly select 100 latent motion features and average their frequency magnitudes. This averaged result reflects the typical frequency characteristics of each attention head and is used to analyze or visualize the frequency sensitivity of different groups. Adapting FFT to calculate frequency magnitude is a common approach, and similar methods have also been introduced in works such as [12] and [13].

Once again, we appreciate the reviewer’s valuable feedback. We kindly hope you will consider raising your score if we have adequately addressed your concerns.

Reference:

[1] MotionLCM: Real-time controllable motion generation via latent consistency model (ECCV 2024)

[2] MotionPCM: Real-time motion synthesis with phased consistency model.

[3] Emdm: Efficient motion diffusion model for fast and high-quality motion generation (ECCV 2024)

[4] Motion mamba: Efficient and long sequence motion generation. (ECCV 2024)

[5] Light-t2m: A lightweight and fast model for text-to-motion generation (AAAI 2025)

[6] Sparse VideoGen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025)

[7] CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

[8] HunyuanVideo: A Systematic Framework For Large Video Generative Models

[9] Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers (CVPR 2023)

[10] PACT: Pruning and Clustering-Based Token Reduction for Faster Visual Language Models (CVPR 2025)

[11] MotionLLaMA: A Unified Framework for Motion Synthesis and Comprehension (also known as “VersatileMotion”)

[12] Improving Vision Transformers by Revisiting High-frequency Components (ECCV 2022)

[13] Parameter-Inverted Image Pyramid Networks for Visual Perception and Multimodal Understanding (TPAMI 2025)

Dear Reviewer,

Thank you for your valuable feedback and support. If we have addressed your concerns, we kindly ask you to consider raising your score. We are happy to address any further questions.

Best regards,

Authors #6198

Thank you for your prompt response. Below is our detailed response to your comment.

Q4. Performance.

A4. Below is a sub-table extracted from the original Table 1 in our paper, which presents the comparison on HumanML3D, as reviewer emphasized.

(1) The reviewer claimed that our method fails to outperform prior state-of-the-art methods with comparable AIT in terms of FID, which he/she claimed as the most critical metric. However, this claim is not supported by the results.

In terms of FID, the only method with close efficiency to FlashMo (0.027 AITs) is MotionPCM (1-step) at 0.031 AITs (slower). However, its FID score is 0.044, which does not outperform FlashMo’s 0.041.

Moreover, the advantages of our method extend beyond FID. It is evident that FlashMo (without pretraining) clearly outperforms MoGenTS and MotionPCM, achieves SOTA performance on 6 out of 8 metrics on HumanML3D, while MotionPCM achieves state-of-the-art results on at most two metrics (but across two model variants), and MoGenTS does not achieve state-of-the-art on any metric.

The results in the table speaks for itself.

(2) The reviewer claimed that FID is the most critical metric, disregarding our significant improvements in other metrics. However, this claim is one-sided and tenuous.

FID exhibits notable limitations in various aspects. For example, as ParCo [1] mentioned: 'FID does not assess the alignment between textual descriptions and generated motions.' Similarly, as mentioned in Motion-Agent [2]: 'variable length can lead to higher FID scores.” That is, FID can only assess the distance between the distributions of two sets of motions [2].

Hence, it is essential to evaluate motion generation comprehensively, considering multiple aspects such as motion distribution distance (FID), text matching (R-Precision), text-motion alignment (MM Dist), diversity (MModality), and efficiency (AITs), etc. Judging a model’s performance primarily based on FID while disregarding other metrics is misleading.

It is worth noting that many latest models published in 2025—such as Motion-Agent (ICLR 2025) [2], ReMoGPT (AAAI 2025) [3], ScaMo (CVPR 2025) [4], and others—did not achieve the FID levels required by the reviewer, as shown in the table below. However, this did not prevent them from generating high-quality motions or being accepted by top-tier conferences. (Their FIDs are highlighted in the table.)

Reference:

[1] ParCo: Part-Coordinating Text-to-Motion Synthesis (ECCV 2024)

[2] Motion-Agent: A Conversational Framework for Human Motion Generation with LLMs (ICLR 2025)

[3] ReMoGPT: Part-Level Retrieval-Augmented Motion-Language Models (AAAI 2025)

[4] ScaMo: Exploring the Scaling Law in Autoregressive Motion Generation Model (CVPR 2025)

The authors have addressed most of my concerns in their rebuttal, but I remain unconvinced about the model’s performance with and without pretraining. While the frequency-aware sparsification indeed accelerates training and inference, the non-pretrained model still fails to outperform prior state-of-the-art methods with comparable AITs (e.g., MotionPCM at 0.031–0.045 s).

Although the authors claimed in their rebuttal that they achieved SOTA results on most HumanML3D metrics, they do not lead on the most critical FID score. This raises doubts about whether the claimed key contribution truly delivers across the board. The only substantial improvement in FID comes from augmenting the training set with MotionHubV2 data (from 0.041 down to 0.029). For these reasons, I maintain my original rating.

Q5. Whether the claimed key contribution truly delivers across the board.

A5. (1) The reviewer acknowledged that 'frequency-aware sparsification indeed accelerates training and inference,' but expressed concerns about whether the remaining key contribution is truly effective. It is important to note that, beyond efficiency, another key contribution of our method is scalability. That is, both our geometric factorized interpolant design and MotionSiT architecture are aimed at improving scalability in a data-driven manner.

Since MotionPCM is not yet open-sourced, we conducted a fair comparison between MoGenTS and our method, both of which are pretrained on MotionHub V2 and evaluated on HumanML3D. As shown in the table below, our method demonstrates significantly better scalability than MoGenTS. Moreover, after pretraining, MoGenTS does not show performance improvement and even exhibits a decline in some metrics—including FID (0.033 to 0.052), which the reviewer obviously cares about most.

This verifies that our key contribution delivers not only efficiency but also scalability.

(2) The reviewer also claimed that “The only substantial improvement in FID comes from augmenting the training set with MotionHubV2 data (from 0.041 down to 0.029)”, but ignoring the performance boost by our interpolant and architecture design. However, the claim is not true.

To fully showcase and ablate the performance gains from our interpolant and architecture design, we adopt full attention with standard temporal-spatial attention, excluding the effect of sparsification. The results in the following table demonstrate that our interpolant and architecture improve not only FID (0.029 without pretrain) but also overall generation quality. This improvement is unrelated to MotionHub V2, as no pretraining is used in this setting.

We hope these additional information has addressed your remaining concerns. Thank you.

Dear reviewers 8uxk,

We welcome further discussion to see whether our response addresses your concerns. To save your time, we provide a brief summary below:

1. FlashMo outperforms SOTA methods with comparable AIT (including FID).

The only method with comparable efficiency to FlashMo (0.027 AITs) is MotionPCM (1-step) at 0.031 AITs (even slower). FlashMo (sparse, no pretrain) achieves a better FID (0.041) than MotionPCM (1-step, FID 0.044). Moreover, FlashMo (full attention, no pretrain) achieves the lowest FID (0.029), outperforming all methods. Moreover, our method achieves significant improvements in other metrics as well.

2. FlashMo's performance gains from its advanced interpolant and architectural design, not simply from data-driven training.

(1) To showcase and ablate the performance gains from our method design, we adopt full attention with temporal-spatial attention. The results of FlashMo (full attention, no pretrain) demonstrate that our method design improve not only FID (0.029) but also overall generation. This improvement is unrelated to MotionHub V2, as no pretraining is used in this additional ablation.

(2) The other key contribution of FlashMo is scalability. We conducted a fair comparison between MoGenTS and ours on HumanML3D, both of which are pretrained on MotionHub V2. FlashMo (FID 0.041→ 0.029) demonstrates significantly better scalability than MoGenTS (FID 0.033 → 0.052).

We hope these answers address your remaining concerns, and we look forward to your positive feedback!

Once again, thank you for your time and support! We wish you a successful run at NeurIPS!

Best regards,

Authors #6198

Dear AC,

Thank you for your continued support and patience, especially your encouragement regarding our extensive experiments (the authors are truly grateful and moved).

Meanwhile, we share the same understanding with AC that FID does not always reflect the perceptual quality of motions. We hope our response addresses 8uxk’s concerns, and we are happy to answer any further questions.

As AC mentioned, unfortunately we cannot upload videos at this stage, but we will include improved visualizations with full attention and different design choices in the next version of our paper.

Best regards,

Authors #6198

Dear reviewer 8uxk,

Thank you for your further discussion and critical thinking!

We’re glad to see that you mentioned the limitation of the current FID metric in the community. The authors agree with the reviewer and share the same concern that the current motion FID in the community heavily relies on the feature extractor provided by Guo et al., which has limitations in reflecting the quality of generated motions.

This exactly reflects the importance of evaluating motion generation from a comprehensive perspective, not only using FID, which we mentioned in the previous discussion. The advantages of our method extend beyond FID, with significant improvements also observed in other metrics.

From a deeper perspective, this might require modern evaluation metrics and benchmarks that are carefully designed to better reflect motion generation quality and can be seamlessly adapted to various recent baselines. However, this cannot be easily accomplished by a single work or paper, but may require support and effort from the broader community. Otherwise, it will just be an ordinary dataset and benchmark paper.

Both the reviewers and the AC also mentioned a good point regarding visualization. I must admit that artifacts such as jitter and foot sliding are common problems in motion generation tasks, and this is exactly the motivation for us to keep exploring novel methods to address these challenges. If you look at modern models such as FlowMDM, it is an excellent work in long motion generation with high quality and ease of use, but you can still observe the foot sliding problem in their teaser video. Similarly, recent SOTA methods such as MoGenTS set a good standard in text-to-motion, but they include only four comparison demos to MoMask and also show some small artifacts.

From another perspective, as mentioned in the paper, recent VQ-VAE-based methods achieve outstanding quantitative results but often show small artifacts due to tokenization, e.g., frame-wise noise arising from directly decoding discrete tokens. Diffusion models, however, seem to perform much better in this regard, which is also the reason why we aimed to design a new method based on the diffusion architecture. We agree with the AC and reviewers that providing more ablation videos and user studies comparing different architectural designs will better showcase the effectiveness of our method. We will continue to improve these aspects to make our work more comprehensive.

Once again, thank you for the further discussion and support of our work.

Best regards,

Authors #6198