We greatly appreciate the time and effort you invested in providing these detailed observations, questions, and comments. We have carefully considered your comments and have outlined our responses and proposed changes below. We hope these adjustments and explanations address your concerns and further enhance the manuscript.

Before reading our feedback, we introduce a visualization website to validate the quality of our dataset at http://www.motionbase3d.com. (Please note: this link may be unstable, and refreshing the page 10~20 times may be required to view the content). In case the reviewer is unable to access the site, we have also provided an anonymous cloud storage link containing all the examples featured on the website: click this link. The platform includes visualized samples from our dataset, physically grounded examples of refined motions, rendered results generated by different motion models, and live demonstrations in both simulated environments and real-world settings (using H1 and G1 UNITREE humanoid robots). We hope this platform addresses the data quality concern reviewers may have. If the reviewer requires additional information or examples, we are more than happy to upload further relevant results upon request. Additionally, our dataset is highly scalable due to the robust data collection pipeline. As of the start of the rebuttal period, we have collected over 1.5 million motion trajectories and 4 million motion-text pairs, which are 1.5X compared to our first submission. These datasets have undergone three rounds of rigorous data filtering to ensure their quality.

W1. [The motion collection process] How do you evaluate and make sure the quality of collected motion data, avoid issues like jittering and foot sliding

We have implemented several strategies to evaluate and refine the motions generated in MotionBase. First, thanks to our constructed visualized platform, we can efficiently sample and assess a large amount of motions from MotionBase. Then, to ensure motion quality, we adopt the following steps:

1. We first train an RL-based policy $\pi_{\text{refine}}$ to elaborate and refine raw motions, ensuring they adhere to physical laws (e.g., maintaining balance) and appear more realistic. Specifically, this policy takes raw motion sequences as input, treating them as target poses, and generates new motion sequences that satisfy physical laws in a simulated environment, eliminating issues like jittering and foot-sliding.
2. While effective, the RL-based policy $\pi_{\text{refine}}$ may struggle with drastic movements in target poses, leading to slipping within the simulation. For such cases, we adopt the following considerations: If a refined motion maintains balance in the simulated environment for a specific duration, it can be regarded as high-quality with no severe issues like jittering or sliding. If a motion fails to maintain balance from the start, it is considered a bad sample and discarded.
3. For motions that pass STEP 2, we utilize a pretrained motion model like RoHM[1] for further refinement. Additionally, we experiment with more powerful motion reconstruction models such as WHAM[2]. Based on our experience, existing motion models have been able to effectively enhance the quality of raw motions.

[1] RoHM: Robust Human Motion Reconstruction via Diffusion

[2] WHAM: Reconstructing World-grounded Humans with Accurate 3D Motion

W1. [The motion collection process] If the quality of the ground truth is not good enough, how can you generate good motion? Therefore, the result in L395-402 is not solid and convincing.

We understand your concerns regarding the impact of motion quality and the results discussed in lines L395–402. First, we believe the poor FID results in L395–402 primarily stem from the limitations of the retrieval model used in the ICLR version. The model [1], initially trained on HumanML3D/MotionX, was applied to evaluate MotionBase, which introduced a mismatch. To address this, we have adopted a more robust retrieval model specifically trained on MotionBase. With this improved model, we observed significant performance gains, particularly in the FID metric. Second, as previously detailed, we have developed a functional toolbox to refine both motion and text data through various strategies. MotionBase has undergone three rounds of rigorous data cleaning and verification, and we plan to implement additional filtering processes to further enhance its quality. For better visualization, we also showcase our data on a website and hope these enhancements address your concerns and bolster confidence in our results.

[1] Generating Diverse and Natural 3D Human Motions From Text

W1. [The motion collection process] The limited contribution of the dataset. The video data comes from InternViD and WebVid and the data collection process is from motion-x and other methods. The dataset contribution is limited.

It is important to note that we do not directly use videos on the Internet, but have undertaken extensive efforts and developed a strategic combination to collect useful motion data from the Internet. Without these efforts, raw video data would remain a noisy, unusable collection. In addition to the strategies we introduced to improve the quality of motion and text descriptions, we provide more details of the data construction process.

1. Video Collection and Selection: It is worth noting that MotionX data constitutes only 5% of the entire dataset. Meanwhile, rather than relying solely on open-source datasets like InternVid, we also extract motion data from self-collected videos sourced from the Internet. For these videos, we use a pretrained 2D human keypoint detection model to filter out those without visible human activity. Additionally, rule-based methods are applied to ensure that the human bounding box occupies a significant portion of the frame, making human movement clearly visible. Videos with only partially visible humans are removed to maintain the quality of the potential motion data. Through these methods, we ensure the extracted motion data is of high quality.

2. Short Boundary Detection: Web videos are generally lengthy and feature varied camera shots. To address this challenge, we adopt the following steps:

   (1) First, for videos shorter than 30 seconds or those with explicit temporal boundaries, we directly use the video clip or the provided boundaries to segment the video into shorter clips.

   (2) For videos longer than 30 seconds, we employ a scene detection model to roughly divide the video into smaller segments.

   (3) For each segment, we further slice it into shorter clips using the following process:

   • At the beginning, the human with the largest bounding box is selected as the anchor, and their trajectory is tracked throughout the segment.
   • When the trajectory is interrupted, the start and end times of the interruption are marked as the boundaries of a new clip.
   • The process repeats by identifying the next largest visible human in subsequent frames and tracking their trajectory.
   • This process continues until no humans are visible in the video.
   • Clips without visible humans are filtered out.

   (4) After these steps, if a clip is still longer than 60 seconds, we randomly slice it into several sub-clips, ensuring that each sub-clip is shorter than one minute.

3. Removing Occlusion and blur: Occlusion and motion blur are common issues in human-related videos. To address these problems, we adopt the following steps:

   (1) First, we sample key frames from each video and use a pretrained 2D keypoint detector to extract skeleton keypoints for each human in the key frames. If a significant portion of the keypoints has predicted confidence scores below a specific threshold, we consider the human motion to be occluded and exclude it from further processing.

   (2) We then use a visual foundation model, such as Segment Anything, to generate segmentation masks for each frame. If a large object is detected in front of the human, indicating occlusion, we filter out the corresponding motion data.

   (3) To address motion blur, we track the trajectory of each human whose motion data needs to be extracted. For timestamps with low-confidence keypoint scores, we smooth the trajectory using adjacent detection results to ensure continuity and accuracy.

4. Single-frame Motion Processing: A substantial portion of our data consists of single-frame motions, many of which can be transformed into multi-frame sequences to enhance data diversity. To achieve this, we train a RL-based policy $\pi_{\rm multi\\_frame}$ using the AMASS dataset. This policy generates physically plausible motion sequences within a simulation environment, using the single-frame motion as the target pose. However, due to potential instability caused by drastic lower-body movements, some generated motions may fail to maintain balance. For single-frame motions that cannot be successfully converted, we use another pretrained, target-conditioned motion generator based on existing high-quality motion data. This generator uses the single-frame motion as the target pose and generates the preceding motion, effectively producing a complete sequence. While these generated motions are not fully constrained by physical laws, resulting in less consistent quality compared to those generated by the RL-based policy, they still provide an effective solution for motion conversion.

W2. [The text annotation] The annotation quality of the text by Gemini-1.5-pro is not well evaluated. In my practice, it always contains some answers like "sorry...". The results should be corrected by researchers one by one. Has the >1M data been checked?

The quality of generated texts is highly related to the prompt design. We introduce two approaches used in our work to evaluate the text description quality.

1. Firstly, we sample 10,000 motion descriptions from our MotionBase generated by Gemini, along with 10,000 descriptions each from MotionX and HumanML3D. These descriptions are scored using GPT-4o, which evaluates each description on a scale of 1 to 5 based on predefined criteria focused on clarity, detail, and accuracy. The scoring criteria are as follows.

   • Score 1: The description is vague, lacks specific details about body movement, and contains confusing or unclear expressions.
   • Score 2: The description covers some movement and posture content but lacks sufficient detail or accuracy and includes unclear expressions.
   • Score 3: The description clearly outlines movement and posture, providing basic details but lacking in-depth analysis.
   • Score 4: The description is accurate and detailed, effectively conveying the movement process and changes in body posture, with some analytical depth.
   • Score 5: The description is precise, comprehensive, and fluent, offering in-depth analysis of every detail of the movement and posture, demonstrating a high level of professionalism and clarity.

We then calculate the average scores for each dataset: MotionBase (3.837) and MotionX (1.386) and HumanML3D (1.703). These scores suggest that MotionBase descriptions are generally more detailed and accurate compared to MotionX and HumanML3D.

2. To further evaluate the quality of the generated texts for vision-based motions, we prompt Gemini-pro with text descriptions and corresponding rendered motions. Our primary focus is on the accuracy with which the text descriptions reflect the content of the visual cues. To assess this, we present 500 rendered samples with their corresponding text descriptions from each dataset to Gemini, requesting a score based on the criteria we established earlier. The evaluation results provide valuable insights. The texts of MotionX and HumanML3D receives an average score of 2.25 and 3.08, respectively. Notably, MotionBase achieves a significantly higher average score of 3.82, outperforming the other two datasets.

W2. [The text annotation] The proposed contribution of hierarchical text is not discussed well. Has it been used in the model training? If I miss, please point it out. If this annotation is not used, what is the motivation for this hierarchical text contribution? Will it make the result more fine-grained? It is quite unclear.

YES, the hierarchical text is used during pretraining to improve the diversity of text corpus and make the description finer-grained. The table compares experimental results using hierarchical text versus using only basic text. The results show that hierarchical text can effectively enhance the model's semantic understanding and thereby improve the semantic matching of generated motions.

In addition, the hierarchical texts also ensure the quality of motion description generated by Gemini-1.5 or GPT-4o. Specifically, our latest motion descriptions are structured into three hierarchical and complementary levels: an overall description of the entire body, detailed part-level description for each body part, rule-based description of each joint's relative movement derived from joint positions (e.g., "The left hand is wide apart from the right hand. It is on the ground, the left upper arm, both legs and the left forearm are aligned horizontally......"). To enhance the reliability and quality of the motion descriptions, we condition GPT-4o with two levels of description while using the remaining level as the evaluation target. GPT-4o then refines the textual content. By doing this, each level of description can provide complementary details and correction for the other two levels, enabling the generation of more precise and reliable motion descriptions.

W3. The H2VQ proposed in Humantomato (ICML-24) is missing for discussion or comparison. And the technical and evaluation contribution of motion VQ is limited.

To the best of our knowledge, Humantomato has not released their code, and no successful re-implementation has been found. As a result, it is currently impossible to conduct a quantitative comparison with the H2VQ method proposed in Humantomato. Additionally, we re-implemented RVQ and achieved better reconstruction performance (56.9 MPJPE) compared to H2VQ (62.34 MPJPE reported) on MotionX. Given this, we have chosen not to include a comparison with H2VQ in Table 6. We would be glad to perform such a comparison once Humantomato releases their code. However, to improve the reference discussion, we also include this reference to our related work.

The primary contribution of our motion tokenization technique lies in the exploration of tokenization methods specifically designed for motion generation. While previous works have used 1D discrete vectors to represent the body, we highlight the information loss that this approach may cause. In this paper, we propose a novel direction for encoding motions by treating the motion sequence as a 2D image. It's important to note that this approach has not been extensively discussed in prior work, and we validate its potential and effectiveness through our experiments. Therefore, the inclusion of LFQ in our study serves as a preliminary exploration to assess the potential of alternative tokenization methods for motion representation. We intend to conduct a more thorough investigation of this approach in future work.

W4. This work does not include any demo video, which is unacceptable in the animation community.

Thank you for your suggestion! To address your concern about the visualization of demo videos, we have created a website showcasing a variety of examples, including samples from our datasets and generated results from our models. Additionally, we present physically grounded examples, as well as examples deployed in both simulated and real-world environments ---- an aspect that has not been highlighted in previous works!

W4. The FID in Table 5 is extremely large, which strengthens my concerns about the motion quality.

We figure out that the high FID values reported in Table 5 are primarily due to the limited encoding capability of the retrieval model used for evaluation.

The model originally used in our ICLR submission was trained on HumanML3D and was unable to fully capture the characteristics of the motion data, leading to abnormally high FID scores. To address this, we train a more powerful and robust encoding model on our MotionBase dataset. After re-evaluating with this improved model, we observe a significant reduction in FID values, bringing them into a reasonable and acceptable range. We include these updated evaluation results in the revised version to ensure clarity. We also plan to use more recently proposed text-motion retrieval models like TMR[1] to obtain a more accurate evaluation of model performance.

[1] TMR: Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis.

W5. Motivation. The motivation for introducing LLM is not clear. The method misses a basic baseline of a transformer (like in T2M-GPT, CVPR-23) for comparison. Besides, it is also not clear whether the usage of pre-trained parameters of LLMs or not. Whether the fine-tuning method is LoRA or not is also not well discussed. Therefore, it is not technically sound.

We incorporate LLM to enhance language understanding, a critical aspect of text-to-motion generation, which relies on strong comprehension of textual nuances. LLMs, with their proven capabilities in natural language processing, enable our model to better capture the context and subtleties of text descriptions. As a result, this leads to the generation of more accurate and contextually appropriate motion sequences. The table compares our method to baseline T2M-GPT (CVPR 2023) on HumanML3D. Our results show significant improvements. These results highlight the advantages of leveraging LLMs. We use the pre-trained LLM parameters, which offer greater capacity and superior language understanding so that our model is able to learn more complex mappings between language and motion. In Table 9 of Appendix C.3, we compare experiments with and without pre-trained parameters, showing that fine-tuned models using pre-trained parameters consistently outperform models trained from scratch. During training, we utilize full fine-tuning. While we also experimented with LoRA, it struggled to achieve competitive results. We attribute this limitation to the introduction of new motion tokens, which demand substantial parameter adjustments. LoRA, with its constrained fine-tuning approach, appears less equipped to handle these requirements effectively.

Q1. The vocabulary of the LLM and motion codebooks are different. How do authors handle this issue? What is the efficiency of the LLM-based motion generation method? Please compare with the fastest motion generation method, MotionLCM

To address varying vocabularies, we extend the LLM's vocabulary by incorporating motion codebook tokens as additional entries. The LLM is then trained directly on text-motion paired data, enabling it to effectively learn the associations between textual descriptions and motion tokens.

We test our 7M LLM-based model on RTX-4090 GPU for inference, the token generation speed is 23 token/sec. Since each token represents 4-frame motion, the generation speed of our model is around 90 FPS, denoting a real-time inference spped which can be deployed in reality, although still can not beat the speed of fastest methods like MotionLCM. However, our focus is exploring the capabilities of large models for text-to-motion generation. We are not prioritizing speed optimization in this work. Speed improvements, such as using Flash Attention, are orthogonal to our current research goals and represent an important area for future work.

Q2. Why authors should cite [1]?

The references cited here primarily aim to demonstrate that the development of multimodal large language models depends heavily on the availability of extensive data across various domains, including digital environments like games, egocentric scenarios, and others.

We hope our clarifications address your questions, and we kindly ask if there are additional concerns we can address to further improve your support for the paper. Thank you again for your valuable suggestions.

We greatly appreciate the time and effort you invested in providing these detailed observations, questions, and comments. We have carefully considered your comments and have outlined our responses and proposed changes below. We hope these adjustments and explanations address your concerns and further enhance the manuscript.

Before reading our feedback, we introduce a visualization website to validate the quality of our dataset at http://www.motionbase3d.com. (Please note: this link may be unstable, and refreshing the page 10~20 times may be required to view the content). In case the reviewer is unable to access the site, we have also provided an anonymous cloud storage link containing all the examples featured on the website: click this link. The platform includes visualized samples from our dataset, physically grounded examples of refined motions, rendered results generated by different motion models, and live demonstrations in both simulated environments and real-world settings (using H1 and G1 UNITREE humanoid robots). We hope this visualization platform addresses the data quality concern reviewers may have. If the reviewer requires additional information or examples, we are more than happy to upload further relevant results upon request. Additionally, our dataset is highly scalable due to the robust data collection pipeline. As of the start of the rebuttal period, we have collected over 1.5 million motion trajectories and 4 million motion-text pairs, which are 1.5 times compared to our first submission. These datasets have undergone three rounds of rigorous data filtering to ensure their quality.

Here are our responses to the weaknesses.

W1. Does the large proportion of single-frame motion data in the dataset contribute to static motion generation?

No.

1. First, our dataset is designed to be highly scalable, with the latest version now including a reduced share (44%) of one-frame motion data and over 1.5 million motions ---- a significant expansion over the ICLR version. Notably, this update incorporates more than 50% additional multi-frame motions extracted from open-source human behavior datasets (e.g., NTU120, Kinetics-700) and publicly available web videos. We plan to further increase the proportion of multi-frame motions in future versions.
2. Second, a substantial portion of one-frame motions can be transformed into multi-frame sequences. To achieve this, we train a RL-based policy $\pi_{\rm multi-frame}$ using the AMASS dataset. This policy generates physically plausible motion sequences within a simulation environment by using the single-frame motion as the target pose. Due to potential instability caused by drastic lower-body movements, some generated motions may fail to maintain balance.
3. For single-frame motions that fail conversion in STEP 2, we employ a pretrained, target-conditioned motion generator based on existing high-quality motion data. This generator uses the single-frame motion as the target pose and generates its preceding motion, effectively producing the entire motion sequences. Compared to motions converted through STEP 2, these generated motions may not fully adhere to physical laws.
4. To further avoid static motion generation, we provide distinct prompts for single-frame and multi-frame motions during the LLM's fine-tuning, ensuring that the model is capable of generating dynamic motion in response to specific commands.

Thanks for the rebuttal. However, I still have the remaining questions.

1. The content of dataset collections and processes, including the training of policies and motion generators, are not included in the main scripts and appendix. In addition, they lack the quantitative experiments to evaluate the policy.

2. Furthermore, the motion qualities have not been verified through quantitative experiments. I suggest the author carefully compare their methods on some benchmarks, like RICH, EMDB2, etc. The Gemini can somehow evaluate the alignment between the motion and the text. However, it is improper to evaluate the motion qualities.

3. LFQ is one of the main contributions of the paper, however, the authors scale the model with VQ instead of LFQ. I think this should not be regarded as a good thing in ICLR.

Overall, I decided to keep my score now.

Q1. What is the ratio between synthetic, static, and real data? in Table 4

In the latest version, synthetic data accounts for approximately 28% of the total dataset, while real data makes up about 72%. Additionally, static data constitutes about 44% of all data. The proportion of synthetic and static data will continue to decrease as MotionBase expands. We have included these proportion numbers in the revised version of Table 4.

Q2. The quality of occlusion cases or blurred images. How do the authors recognize that the motion is blurred or occluded? In multi-person settings, occlusion is expected to be very common.

Occlusion and motion blur are indeed very common in human-related videos. To avoid these issues, we adopt the following steps:

• 2D Keypoint Detection Fittering. We first sample key frames from each video. A pretrained 2D keypoint detector is then used to extract skeleton keypoints for each human in the key frames. If a significant portion of keypoints has predicted confidence scores below a specific threshold, we consider the human motion to be occluded and exclude it from further processing.

• Segment Filtering. We utilize a visual foundation model, such as Segment Anything, to generate segmentation masks for each frame. If a large object is detected in front of the human, indicating occlusion, we filter out the corresponding motion data.

• Adjacent Frame Smoothing. To handle motion blur, we track the trajectory of each human whose motion needs to be extracted from the video. For timestamps with low-confidence keypoint scores, we smooth the trajectory using adjacent detection results to ensure continuity and accuracy.

In addition to occlusion, we apply many additional post-processing techniques to enhance the quality of the dataset. If the reviewer has any further questions about the process, please feel free to ask. We are willing to provide more details.

We hope our clarifications address your questions, and we kindly ask if there are additional concerns we can address to further improve your support for the paper. Thank you again for your valuable suggestions.

We appreciate the reviewer’s careful assessment and acknowledgment of our paper's clarity, motivation, and contributions.

Before reading our feedback, we introduce a visualization website to validate the quality of our dataset at http://www.motionbase3d.com. (Please note: this link may be unstable, and refreshing the page 10~20 times may be required to view the content). In case the reviewer is unable to access the site, we have also provided an anonymous cloud storage link containing all the examples featured on the website: click this link. The platform includes visualized samples from our dataset, physically grounded examples of refined motions, rendered results generated by different motion models, and live demonstrations in both simulated environments and real-world settings (using H1 and G1 UNITREE humanoid robots). We hope this platform addresses the data quality concern reviewers may have. If the reviewer requires additional information or examples, we are more than happy to upload further relevant results upon request. Additionally, our dataset is highly scalable due to the robust data collection pipeline. As of the start of the rebuttal period, we have collected over 1.5 million motion trajectories and 4 million motion-text pairs, which are 1.5X compared to our first submission. These datasets have undergone three rounds of rigorous data filtering to ensure their quality.

Here are our responses to weaknesses and questions.

W1. The layout of the paper is somewhat challenging for readers. It contains numerous messages and analyses, requiring readers to scroll up and down frequently to locate referenced tables. Additionally, due to page limitations, many explanations are placed in the Appendix. Tables and figures are positioned mid-page without aligning well with the paragraph height, disrupting the flow.

Thank you for your comments on the paper's layout. We want to present as much content as possible therefore the layout could be not well-aligned. In the revised manuscript, we will: (1) Optimize Figure/Table Placement: Align figures and tables with the paragraph height and place them as close as possible to the relevant text to enhance readability. (2) Improve Referencing: Ensure clear and unambiguous references to figures and tables, reducing the need for excessive scrolling. (3) Streamline Appendix: Integrate critical explanations and analyses into the main text where appropriate, while reorganizing the remaining appendix content for better structure and clarity.

W2 & W3. Important reference and minor issues and typos. Thank you for your notice. We have cited this important reference and corrected the issues and typos you pointed out in the revised paper.

Dear reviewer pT95.

Sincerely thank you for your reviewering time. We would like to kindly remind you that the rebuttal period has been extended to December 2nd, giving us an additional week to address your concerns. Please feel free to reach out if you have any further questions, and we guarantee a prompt response.

Considering your mentioned the concerns about static data, we attach this part in the bottom. If you have any additional question, please let us know. We have provided over 20 pages of rebuttal material, which has been a time-intensive effort. We greatly value this opportunity to discuss and refine our work, and we sincerely hope we are moving in the right direction.

Best regards, Authors

Responses to static data concern:

• Most importantly, our dataset comprises over 800K dynamic motion sequences, making it at least 10 times larger than existing benchmarks. This scale ensures that even if all static poses were removed from MotionBase, the dataset's overall contribution would remain largely unaffected.
• We argue that static poses provide valuable additional knowledge about human activity, similar to how static images contribute to video understanding or single-frame object bounding boxes aid video tracking. The effectiveness of static data is validated in the ablation studies presented in Table 4 and Table 8 in the appendix, where we test on datasets without static data. In fact, we suppose static poses are particularly beneficial for our LLM-based decoder, as they enhance its understanding of positional and rotational relationships among joints. This is especially important when motion vocabulary is learned from scratch during pretraining.
• The static data will not result in "over smoothing" because:
  • During fine-tuning, we design specific prompts to guide the model's generation. For static poses, we include additional language prompts, such as 'keep the action still,' to help the model distinguish between generating a static pose and a dynamic motion.This approach leverages the well-known ability of large generation models based on LLMs [1,2,3,4,5] to produce diverse outputs through carefully crafted prompts. In addition, no negative effects are observed when combining images and videos, or bounding boxes and segmentation masks, or 3D skeleton poses and sequences, or vision and audios during training. We believe this principle should extend to motion as well.
  • Visualization. To better prove our conclusion, in the visualization appendix, we present generation results from both models trained with and without additional language prompts. The clear differences demonstrate that with additional language prompts, the model can distinguish between different motion distributions, thus not affecting the generation of dynamic motions.
  • Many common training strategies, like weighted sampling, weighted gradient, progressively increasing the ratio of dynamic motion, all these strategies can improve the stability of training, ensuring the model to avoid "over smoothing".

Dear reviewer pT95.

Thank you again for your valuable review. We have responded to your concern about static data quality. We hope to hear back from you. If you have any unresolved concerns, please feel free to let us know! Or if our responses have addressed your questions, we would be grateful if you could consider adjusting your score accordingly.

Best regards,

Authors

Thank you for your appreciation of our work, and proposing thoughtful feedback which helps us further improve our work.

Before reading our feedback, we introduce a visualization website to validate the quality of our dataset at http://www.motionbase3d.com. (Please note: this link may be unstable, and refreshing the page 10~20 times may be required to view the content). In case the reviewer is unable to access the site, we have also provided an anonymous cloud storage link containing all the examples featured on the website: click this link. The platform includes visualized samples from our dataset, physically grounded examples of refined motions, rendered results generated by different motion models, and live demonstrations in both simulated environments and real-world settings (using H1 and G1 UNITREE humanoid robots). We hope this visualization platform addresses the data quality concern reviewers may have. If the reviewer requires additional information or examples, we are more than happy to upload further relevant results upon request. Additionally, our dataset is highly scalable due to the robust data collection pipeline. As of the start of the rebuttal period, we have collected over 1.5 million motion trajectories and 4 million motion-text pairs, which are 1.5 times compared to our first submission. These datasets have undergone three rounds of rigorous data filtering to ensure their quality.

Here are our responses to weaknesses.

W1 & Q1. The paper lacks a thorough comparative analysis across varied methods to demonstrate MotionBase's influence on model efficacy. Additional baselines and a broader selection of models trained on MotionBase would more robustly substantiate its claimed advantages.

In this work, our primary focus was on constructing the high-quality MotionBase dataset and investigating the scaling law of model size and dataset scale. Due to time limitation, we were unable to perform a comprehensive comparison among all existing methods. We acknowledge this as an important area for improvement and plan to include additional baseline models and experimental results to provide a more thorough analysis of our findings.

W2. The paper does not include visual comparisons of motions generated by models trained on the baseline Motion-X dataset versus those trained on the proposed MotionBase dataset.

We showcase the generated results from different models on the model generation result page of our visualization platform, which allows you to directly compare the motions generated by models trained on Motion-X and MotionBase. We believe this visual comparison will provide a clearer and more intuitive demonstration of the improvements MotionBase brings to model training.

W3. Including the ground truth R-Precision and FID scores in relevant tables would strengthen the presentation and transparency of the results.

We have added the ground truth R-Precision and FID scores to Table.2 and Table.3 in our revised version. Here are the results:

W4. The paper would benefit from dynamic visualizations within the qualitative analysis of the motions in the proposed datasets, which could provide a clearer and more engaging illustration of the dataset's scope and quality.

We fully agree that dynamic visualizations are essential for effectively demonstrating the value of our dataset. This concern has also been raised by other reviewers. To address this, we build a visualization platform that allows all reviewers to directly and clearly explore a wide range of motion examples, providing a more intuitive and comprehensive understanding of the richness and high quality of our dataset. We sincerely invite you to visit the platform and experience these dynamic examples. Your comment is important.

We appreciate the reviewer’s careful assessment and acknowledgment of our paper's clarity, motivation, and contributions.

Before reading our feedback, we introduce a visualization website to validate the quality of our dataset at http://www.motionbase3d.com. (Please note: this link may be unstable, and refreshing the page 10~20 times may be required to view the content). In case the reviewer is unable to access the site, we have also provided an anonymous cloud storage link containing all the examples featured on the website: click this link. The platform includes visualized samples from our dataset, physically grounded examples of refined motions, rendered results generated by different motion models, and live demonstrations in both simulated environments and real-world settings (using H1 and G1 UNITREE humanoid robots). We hope this visualization platform addresses the data quality concern reviewers may have. If the reviewer requires additional information or examples, we are more than happy to upload further relevant results upon request. Additionally, our dataset is highly scalable due to the robust data collection pipeline. As of the start of the rebuttal period, we have collected over 1.5 million motion trajectories and 4 million motion-text pairs, which are 1.5 times compared to our first submission. These datasets have undergone three rounds of rigorous data filtering to ensure their quality.

Here are our responses to the questions.

Q1: How did the author verify the correctness/accuracy of the pose estimation?

We acknowledge that this is a key concern for all reviewers so we provide a visualization platform for all reviewers to easily examine our data. In addition, we outline the strategies used to verify the accuracy of the estimated human poses in our dataset:

• Physical Grounding: We train an RL-based policy $\pi_{\rm refine}$, which takes raw poses as input (target poses) and generates refined pose sequences. If the policy successfully tracks the raw poses and produces smooth, balanced motions, we assess the data as high-quality. This approach not only allows the policy $\pi_{\rm refine}$ to serve as a discriminator for verifying pose accuracy and avoiding issues like jittering or sliding, but also acts as an effective method to refine the raw motions.
• Rule-based Methods: In the latest version of MotionBase, we incorporate an additional level of text: rule-based descriptions derived from joint positions and angles. These precise descriptions enable us to assign rule-based scores to each motion. Motions with low scores are filtered out to improve overall quality.
• Manual Review via Visualization Platform: We conduct random sampling for manual check and leverage a visualization platform to facilitate efficient assessment of motion quality.

Q2. What do authors think about properties of a new metric?

Regarding this, we believe that an ideal text-to-motion metric should exhibit the following properties:

• Fine-grained Quality Assessment: The new metric should capture detailed motion features, such as local hand and leg movements in addition to global postures. The metric should evaluate multiple aspects of quality, including naturalness, smoothness, fidelity, and style. In fact, the fine-grained assessment has long been ignored by previous works, partially because the lack of motions required to provide subtle differences. Our MotionBase provides an opportunity to achieve this.
• Human-like Perception: The metric should strongly align with human evaluation, considering factors like human kinematics and biomechanics. It should also account for perceptual similarities, even in the absence of reference motions.
• Physical Plausibility: The metric should evaluate adherence to physical laws, such as gravity and inertia, and analyze dynamic properties like joint torques and ground reaction forces. This is particularly important for applications in robotics.
• Strong Generalization: Ideal metrics should be capable of handling the one-to-many nature of text-to-motion mapping. They should generalize well across datasets and various motion types, such as walking, running, and jumping. Due to time and space constraints, we could not comprehensively address or explore all these aspects. We plan to delve deeper into these directions in future work.

We hope our clarifications address your questions, and we kindly ask if there are additional concerns we can address to further improve your support for the paper. Thank you again for your valuable suggestions.

After checking everything here and the discussion I’m keeping my score as is. The works itself is valuable with some minor concerns that can be addressed or progressed in a later work.

Thank you for your prompt response! We sincerely appreciate the opportunity to address your concerns.

To begin, note the corresponding revisions in our latest manuscript have been masked in blue, and we would like to highlight two key points:

1. Revised manuscript: We have not made significant modifications to our revised manuscript. All updates reflect responses to the reviewers' comments and questions, which align with our initial conclusions and do not contradict them.

2. External links: The primary reason for not using the OpenReview Supp to provide our demos is due to OpenReivew's file size limitation of 100MB, which is insufficient for our larger demo files. Since the anonymous website may be unstable, we provide an anonymous cloud drive link to show the demos, which is a widely accepted practice in current conference submissions. Additionally, we include a smaller zip file in OpenReview Supp, containing a subset of examples that fit with the 100MB restriction. After reviewing ICLR's guidelines (https://iclr.cc/Conferences/2025/CallForPapers), we confirm that anonymous links are not forbidden during both the paper submission and discussion phase.

The following sections are our latest replies to the questions:

Q1. For your motion-capturing process, what is the policy you use? PHC? RFC? ASE? For the statement “If a refined motion maintains balance in the simulated environment for a specific duration, it can be regarded as high-quality with no severe issues …”, what is the duration, and how to compute the success rate? What is your success rate? Besides, how do you deal with videos with shot cuts?

We train the single primitive policy from PHC, which takes a raw motion sequence as input and generates refined joint positions using the simulator (IsaacGym). To ensure consistency in data format, we then convert the global joint positions into HumanML3D's representation using HumanML3D's conversion scripts. The success is determined by the duration of the simulation process. Following PHC's termination conditions, the policy is tasked with tracking the provided motion sequences. During simulation, we log the step number (corresponding to the frame index of the motion sequence) at which termination is triggered and calculate the duration for each sample. Later for each sample, we set a threshold as 50% of its original length, and if the duration of the refined motion sequence exceeds this threshold, it is considered a successful sample. The success rate is then defined as the ratio of successful samples to the total number of samples, with our conversion achieving a success rate of 51.4%. Regarding short cuts, if you mean "videos that have abrupt transitions between different scenes or camera angles", we apply a human tracking algorithm to ensure consistency in human motions across each video, as detailed in the data construction section.

Q2. I also noticed reviewer qpUP has similar concerns on static motions. I think the data is the pose, not motion or animation. Our animation community does not recognize the pose as motion. When the data distribution includes such data, it will be harmful to your generation results. I have also provided the blog by Daniel in my previous review.

• Most importantly, our dataset comprises over 800K dynamic motion sequences, making it at least 10 times larger than existing benchmarks. This scale ensures that even if all static poses were removed from MotionBase, the dataset's overall contribution would remain largely unaffected.

• We argue that static poses provide valuable additional knowledge about human activity, similar to how static images contribute to video understanding or single-frame object bounding boxes aid video tracking. The effectiveness of static data is validated in the ablation study presented in Table 4. In fact, we suppose static poses are particularly beneficial for our LLM-based decoder, as they enhance its understanding of positional and rotational relationships among joints. This is especially important when motion vocabulary is learned from scratch during pretraining. Furthermore, we ensure the reliability of our LLM-generated motion sequences by using purely dynamic motion data for instruction tuning.

• Of course, we fully agree with both you and Daniel's blog that dynamic data plays a more critical role in motion learning. This is precisely why we prioritize collecting such data from human-related videos.

Dear reviewer, we thank you once again for your valuable feedback, which has greatly helped improve the quality of our paper. As the rebuttal deadline approaches, we would like to confirm whether our responses have adequately addressed your concerns. Specifically, we have:

• (1) explained the details of the motion-capturing details.
• (2) clarified that the dataset includes massive dynamic motion sequences, which we believe is sufficient large to substantiate our data contribution, while also demonstrating the benefits of static data via our ablation in Table 4 and Appendix C.1 (Tables 7 & 8).
• (3) uploaded a supp material to OpenReview.
• (4) provided a detailed explanation of the scoring process based on the Gemini-score.
• (5) introduce the usage of hierarchical text.
• (6) address why the results of Hierarchical and Full settings differ.
• (7) clarified the purpose of citing reference [1].

All new additions and modifications are highlighted in blue in the manuscript for your easy reference.

We would greatly appreciate your confirmation on whether these responses have fully resolved your concerns. If not, we welcome any additional feedback or further questions you may have.

Best regards,

The Authors

Dear Reviewer,

Thank you for sharing your remaining questions. Before attaching our further responses, we are encouraged to see that the discussion is becoming more focused and specific with reduced questions, which suggests we are moving in the right direction!

Here is our summary to the remaining five questions: (1) details issues about PHC; (2) discussion about static data; (3) video concatenation; (4) sim-to-real demo videos; (5) text hallucinations; (6) the FID of T2M-GPT. Hoping that our understanding of the remaining questions is correct.

We will attach our responses very soon. Thanks again for your constructive feedbacks!

Best regards,

The Authors

Thank you for your thoughtful reply. While we deeply appreciate your feedback, we would like to respectfully address some points of disagreement and provide clarification.

1. We believe there may be a misunderstanding in your question and suggestion. How could we use to-be-refine data (our data) to train PHC for data to be refined (our data), akin to the idea of lifting oneself with their own hands. It is important to note that no large-scale benchmark can guarantee 100% clean data. Therefore, even with a successful conversion rate of approximately one-third, this demonstrates the effectiveness of our data processing strategy with at least 30% additional improved motion with high quality.

2. We respectfully note that citing an informal blog as valid evidence in a professional rebuttal may not be appropriate. Additionally, we respectfully disagree with the blog's conclusion as it has not been validated under the context of large-scale generative pretraining. Considering this, we believe the use of static data are still an open question instead of a conclusion. We kindly recommend that the reviewer refer to newest well-recognized generative works in the field, such as UnifiedIO-2, VideoPoet, EMU3, and multimodal tuning methods like LLaVA, Qwen-2-vl, to gain a deeper understanding of the capabilities of current generative models. If you have already read these works, you might observe that concerns about "over-smoothing" are not self-standing when considering large-scale pretraining, well-designed prompts, and effective training strategies. As we understand the reviewer might not specialize in the area of LMM, we have provided a more detailed explanation below to offer further clarity.

   1. From a statistical learning perspective, directly combining data from different distributions without incorporating conditional information may result in learning a biased marginal distribution, which could potentially degrade generation quality. However, by introducing additional conditional information—in our case, specific language prompts—we guide the model to learn the conditional distribution P(motion|text, condition) rather than the marginal distribution P(motion). This allows the model to accurately switch between different distribution modes during generation, effectively avoiding distribution confusion. Moreover, static data can be seen as a special case of dynamic data, functioning as a "snapshot" of motion at a particular moment.

   2. More importantly, static data contributes to improved performance by helping the model better understand spatial constraints between joints. Static poses provide abundant examples of valid joint positions and rotations, enabling the model to learn what constitutes physically plausible human postures. Furthermore, static data enhances the model’s ability to capture correlations between joints, such as motion range limitations and inter-joint dependencies. With the guidance of conditional prompts, the model can not only distinguish between static and dynamic data but also leverage the foundational knowledge learned from static data to enhance dynamic motion generation, thereby improving overall quality.

   3. We carry out experimental results to validate the above points: when specific conditional prompts are added, the model effectively distinguishes between data distributions and generates outputs accordingly.

3. We provided a new video with higher resolution to the supp.

4. The robot videos are intended as demos to showcase the potential applications of our work.

5. We kindly request the reviewer to revisit our response, as we have addressed your question. Key points have been clearly marked with significant symbols, such as (1.; 2.), directly following the sentence you referenced. We hope this format makes it easier to locate our responses.

6. We understand that you may have a different interpretation of the experimental results. However, we respectfully suggest that this difference in interpretation should not overshadow the significance of our contributions. We retrain a T2M-GPT model with a parameter count comparable to GPT-2 medium on MotionBase. Nevertheless, the results indicate that it struggles to match the performance of the GPT-2 architecture. We believe that large motion models based on decoder-only LLMs, which jointly train text tokens and motion tokens, achieve better text-motion semantic alignment and stronger motion generation capabilities.

6. For T2M-GPT, we follow the official training configuration, using the text embedding from the pretrained CLIP ViT-B/32 version, and for GPT-2, we similarly use the official pre-trained model [8]. The experimental results show inferior performance, which we believe is self-evident. T2M-GPT uses a frozen CLIP text encoder, resulting in fixed text representations. In contrast, GPT-2 enables joint training of text and motion tokens, allowing text representations to be optimized along with the task, thus achieving better text-motion semantic alignment.

Sincerely, we hope the discussion can be based on the officially published literature and experimental results in our paper, rather than unofficial blogs. We also provide our code and checkpoints from the beginning. We are more than happy you can run it by yourself if you are not convinced by our shown results.

If you have further questions, feel free to tell us, or if, our responses successfully address your concern, would you mind reconsidering your review based on these more reliable references?

Sincerely,

Authors

[1] MotionX: A Large-scale 3D Expressive Whole-body Human Motion Dataset, CVPR 24

[2] Holistic-Motion2D: Scalable Whole-body Human Motion Generation in 2D Space

[3] Morph: A Motion-free Physics Optimization Framework for Human Motion Generation, 2024

[4] https://github.com/IDEA-Research/DINO-X-API

[5] https://github.com/ViTAE-Transformer/ViTPose, 1.4K star

[6] https://github.com/caizhongang/SMPLer-X, 1K star

[7] Reconstructing World-grounded Humans with Accurate 3D Motion, CVPR2024

[8] https://huggingface.co/openai-community/gpt2-medium

Dear reviewer,

We have further questions: Based on your feedback, do you believe the paper would be improved by removing all of our static data and corresponding experimental results? Doesn’t this suggestion seem somewhat unreasonable? Additionally, do you recognize the contribution of the dynamic part of MotionBase? If you believe the static data is more problematic, then the dynamic data should be of higher quality to compensate, right? Otherwise, how would the results improve?

Best,

Authors

6-round questions about "LLM Motivation and T2M-GPT Comparison"

• Q1: Hierarchical text contribution isn’t well discussed. Has it been used in training? What’s the motivation?
• A1: Yes, hierarchical text is used in pretraining. It improves (1) text corpus diversity, (2) finer-grained descriptions, and (3) performance over basic text.
• Q2: The motivation for introducing LLM is unclear. Why no T2M-GPT baseline? Are pre-trained parameters and fine-tuning (e.g., LoRA) used?
• A2: LLMs enhance language-motion understanding. Training: (1) pre-trained parameters, (2) full fine-tuning (LoRA struggles with motion tokens). HumanML3D results outperform T2M-GPT baseline.
• Q3: Why differ between hierarchical and full-param results? I do not see the result of T2M-GPT as a baseline.
• A3: Full-param used only basic text; hierarchical used both basic and detailed text. T2M-GPT results on MotionBase are in Appendix C.7, showing lower performance than ours.
• Q4: T2M-GPT's FID is surprisingly high. Did you scale its size?
• A4: The gap isn’t size-related but due to T2M-GPT’s frozen CLIP encoder, which limits text understanding versus our joint training approach.
• Q5: What proves frozen CLIP is ineffective? CLIP mainly helps T-M alignment.
• A5: We retrained T2M-GPT with comparable GPT-2 medium parameters (380M vs. 355M). It still underperforms our approach.
• Q6: Why does T2M-GPT struggle against GPT-2? Pretraining or CLIP settings?
• A6: T2M-GPT’s frozen CLIP encoder provides fixed text representations, while GPT-2 enables joint optimization of text and motion, achieving better semantic alignment.
• Q7: NO FURTHER REPLY

3-round questions of "Should the reviewer base on initial paper?"

• Q1: I would like to clarify that the significant revisions related to contributions are not supported officially, according to the review guidance.
• A1:We have not made significant changes. All updates address reviewers' comments, align with our initial conclusions, and do not contradict them.
• Q2: According to the ICLR review pipeline, I have the right to ignore these major changes.
• A2: Disregarding our responses would be unfair. ICLR allows revisions during rebuttals to address reviewer concerns, unlike conferences like CVPR and NeurIPS. These updates aim to clarify misunderstandings and do not alter our original contributions.
• Q3: PHC and the dynamic data expansion are major revisions. The discussion process is for addressing misunderstandings in the reviews, not for providing major revisions.
• A3: The ICLR reviewer guideline states: Engage in discussion: The discussion phase at ICLR is different from most conferences in the AI/ML community. During this phase, reviewers, authors and area chairs engage in asynchronous discussion and authors are allowed to revise their submissions to address concerns that arise. It is crucial that you are actively engaged during this phase. Maintain a spirit of openness to changing your initial recommendation (either to a more positive or more negative) rating.
• Q4: NO FURTHER REPLY

7-Round Questions about "Static data"

• Q1: How is motion quality ensured? I suggest authors read Daniel Holden's blog, who is a well-known graphics scientist.
• A1: We use a 3-step method to ensure motion quality.
• Q2: Our animation community considers static data harmful. Refer to Daniel's blog.
• A2: (1) Our dataset contains 400K motions in the initial version, 800K in the latest version, 5 times and 10 times larger than the previous ones, respectively. (2) Results on Table 4 show effectiveness of static data. (3) Dynamic motion tuning ensures high-quality generation.
• Q3: Once you introduce the static ones, it will make your motion over-smoothing. Could you please clarify this?
• A3: Strategies like (1) prompt design, (2) weighted sampling, and (3) progressive training help prevent over-smoothing. We provided cases to show the effectiveness of prompt design to avoid "over-smoothing" in Supp.
• Q4: I still cannot understand why static motions enhance the quality. If you read Daniel's blog, you will find that the provided result is not self-standing. These motions will introduce noise into the dataset.
• A4: Informal blogs are not valid evidence. Static data is an open question, not a conclusion. Ablations comparing training "with" and "without static data" demonstrate its effectiveness.
• Q5: I asked for ablation settings but missed the details.
• A5: Details are clearly stated in L435–L438 and L1049–L1059 under "TRAIN SET w/o static data."
• Q6: I think this is not a suggested choice to use static data in the animation community.
• A6: We’ve provided experiments, codes, and checkpoints proving effectiveness. Subjective conclusions alone are unconvincing. Besides, our datasets contain 400K motions in the original version and 800K in the latest one.
• Q7:Expanding from 400K to 800K is a major revision and is not worthy of being considered.
• A7: Based on your feedback, do you believe the paper would be improved by removing all static data, experiments and discussion? Doesn’t this suggestion seem somewhat unreasonable?
• A7: Do you recognize the contribution of our dynamic data, no matter 400K or 800K?
• A7: There is a logical contradiction in reviewer's opinion: The more harmful you believe static data is, the higher-quality dynamic data should be. Otherwise, how would our results improve?
• Q8: NO FURTHER REPLY

8-round questions about "Data Quality and PHC"

• Q1: The data contribution is limited. The videos comes from InternViD and WebVid.
• A1: We don’t use raw videos directly. Instead, we process them through a rigorous framework: (1) 2-step video selection, (2) 4-step boundary detection, (3) 3-step occlusion removal, and (4) 2-step single-frame processing. (5) ...
• Q2: How is motion quality ensured?
• A2: (1) An RL policy refines motions to be physically plausible; (2) empirical quality assessment; (3) a pretrained motion model for further refinement.
• Q3: What RL policy do you use? PHC? RFC? ASE? What is the duration and success rate?
• A3: We use PHC: 50% of the original duration, 51.4% success rate.
• Q4: Why is PHC’s success rate so low, and how do you handle cases like "sit"?
• A4: Low rates stem from (1) interaction motions like "sit" failing and (2) PHC being trained only on AMASS, limiting generalization.
• Q5: Why not train a new PHC on your dataset?
• A5: It’s illogical to use to-be-refined data to train PHC for refining that data to-be-refined.
• Q6: If PHC doesn’t generalize well, how can success be ensured?
• A6: We never claim 100% success. PHC refines 30% of the data, which is still an improvement over prior works with 0%.
• Q7: (Repeat of Q5) Why not train a new PHC on your dataset?
• A7: Same as A5.
• Q8: PHC is highlighted as a revision and should be seen as a major change.
• A8: PHC is a minor step in data refinement, itself a small part of the pipeline. It’s not a core contribution of our paper, and implementation details are not central to our work. Highlighting it for clarity doesn’t make it a major revision.
• Q9: NO FURTHER REPLY

Rebuttal Summary with WfZ5

Given the lengthy and fragmented rebuttal history with WfZ5, we provide a brief summary to facilitate reading. We hope the AC, Senior AC, and PC will carefully review this summary before making a final decision. In total, WfZ5 raised 51 questions, including 40 detailed ones and 11 of less importance.

2-round questions of "External links are not allowed"

• Q1: External links are not permitted as they cannot be tracked fairly. Why not use the OPenreview's Supp?
• A1: OpenReview's 100MB file size limit prevents us from including all materials in Supp. We also confirmed ICLR guidelines do not forbid anonymous links.
• Q2: So many motion generation papers submitted to ICLR/SIGGRAPH using Supp. I don't know what your barrier is. You can concatenate them into a video demo. Is it very challenging? External links are not appropriate.
• A2: This is the first time we’ve been informed that external links for codes and resources are not only discouraged but permitted, despite their common use in ICLR 2025 submissions.
• Q3: NO FURTHER REPLY

8-round questions about demos.

• Q1: The work should include demo videos.
• A1: We provide an anonymous website and cloud link.
• Q2: The videos on the website are not user-friendly. Please add them to the Supp.
• A2: Our videos exceed the 100MB limit, so we offer an anonymous cloud link for easy access.
• Q3: Many motion papers use the Supp. Why can’t you concatenate them into a single video?
• A3: We provided a zip file with all demos to the Supp.
• Q4: Could you concatenate the videos?
• A4: We concatenated the demos and updated the supp file.
• Q5: The resolution is too low; I can’t see fine-grained motions.
• A5: We provided higher-resolution videos.
• Q6: I do not know why it cannot be downloaded now. I will try it later.
• A6: We verified successful downloads worldwide via OpenReview.
• Q7: I am not clear what the core scientific problem you resolve for robot demos.
• A7: The robot videos demonstrate potential applications, not a core contribution.
• Q8: The demos seem straightforward with no technical innovation. If you emphasize them, reviewers might treat them as a major contribution. It’s better to include them in the appendix or show the scientific problem solved.
• A8: The robot demos are in the supp file only. We never claimed them as a core contribution. These comments contradict our statements.
• Q9: NO FURTHER REPLY

6-round Questions about "Text annotations"

• Q1: Has hierarchical text been used in model training?
• A1: Yes, hierarchical text was used during pretraining, and relevant experiments were provided.
• Q2: Has the text data been verified?
• A2: Yes, using two methods: (1) a text-only method and (2) a vision-text method.
• Q3: How can a text-only method fairly evaluate alignment with motion?
• A3: As stated in A2, we have two evaluation methods: Firstly, we independently use text. More importantly, we use both vision and text.
• Q4: LLMs hallucinate without visual input. Should the method be improved?
• A4: We don’t rely solely on text-only scoring. Both text-only and vision-text methods are used, as explained in A1 and A2.
• Q5: You stated "Firstly, we independently use...". I identify the evaluation is text-only according to "independently".
• A5: We clearly outlined two methods using "1... 2..." in A1 and "Firstly... More importantly..." in A2.
• Q6: I asked the evaluation protocol previously and got the reply "Firstly, we independently assess the quality of the text description.". According to the word "independently", I treat the evaluation as text only. However, after checking the further response, I noticed that it has visual grounds. As a result, I clarified why I treated it as text only in the last reply. The authors asked me to revisit the response, I do not know what I missed.
• A6: We don't know why the reviewer only focused on the word "independently" while overlooking the initial word, "Firstly."
• Q7: NO FURTHER REPLY