Spiral: Semantic-Aware Progressive LiDAR Scene Generation and Understanding

We sincerely thank Reviewer ZJCH for the thoughtful and constructive feedback. We are encouraged to see that you found our pipeline “innovative” and recognized our evaluation metrics (S-FPD and S-MMD) as “novel”. Below, we respond to your comments in detail.

Q1: "The pipeline is limited to generating single-frame LiDAR data, which is restrictive compared to recent advances in temporal LiDAR sequence generation. Extending the model to handle sequential data would strengthen its applicability."

A: Thank you for raising this point. While we acknowledge recent efforts on temporal LiDAR sequence generation, we respectfully note that single-frame generation remains an essential and active research direction, just as image generation continues to be studied alongside video generation. In our case, Spiral is the first to jointly generate range-view LiDAR and semantic maps in a unified framework — a task not addressed in prior works on single or sequential LiDAR generation. We agree that extending Spiral to handle temporal consistency is a promising future direction, and we plan to explore this in follow-up work. Based on your suggestion, we have revised our manuscript to better reflect this aspect.

Q2: "The method only supports unconditional generation of LiDAR and semantic maps, limiting its use primarily to data augmentation for segmentation tasks. Moreover, the segmentation performance improvement is marginal compared to traditional two-stage methods, suggesting limited contributions to the field."

A: Thanks a lot for your comment. We would like to respectfully clarify a few points:

• (1) Spiral supports both unconditional and conditional generation.

  • The unconditional generation task is performed via two options:
    1. open-loop inference that stays unconditional throughout, and;
    2. closed-loop inference that switches to conditional mode when semantic confidence is high (please see Fig. 2(b)).
  • The conditional generation task is also supported: Spiral generates a LiDAR scene given an externally provided semantic map, using only the conditional mode during inference.

• (2) Unconditional generation is crucial for data augmentation.

  • Requiring ground-truth semantic maps as input (conditional mode) would defeat the purpose of reducing annotation costs. The ability of Spiral to perform semantic-aware unconditional generation is a key contribution that enables synthetic data augmentation in practical scenarios.

• (3) The improvement from Spiral is significant, not marginal.

  • For example, on the Robo3D-fog set [R1] (see the table below) with 1% real data, Spiral-based generative data augmentation boosts the segmentation mIoU from 32.07% (w/o GDA) and 36.82% (two-step R2DM & RangeNet++) to 44.06% — a gain of +11.99% over the no-GDA baseline and +7.24% over the prior two-step method.

Table. Experiments under Robo3D-fog | Ratio of Real Data | w/o GDA | R2DM & RangeNet++ | Spiral (Ours) | |:---:|:---:|:---:|:---:| | 1% | 32.07 | 36.82 | 44.06 | | 10% | 53.93 | 54.20 | 58.61 | | 20% | 55.89 | 56.07 | 61.24 |

We hope the above clarifies and addresses your concern. We have refined our manuscript accordingly for better clarity.

Q3: It is unclear which task—conditional or unconditional generation—is evaluated in SPIRAL’s experiments. If unconditional generation is evaluated, what is the ground truth? If conditional generation is evaluated, what are the model’s inputs, given that the only condition mentioned is the semantic map, which cannot be provided during generation?"

A: Thanks for asking. We clarify that the task evaluated in this work is unconditional LiDAR scene generation with both geometric and semantic information.

Spiral is trained to jointly generate the LiDAR scene (geometry + reflectance) and its semantic map from Gaussian noise. During evaluation, we compare the generated scenes to the corresponding ground-truth LiDAR frames and their annotated semantic maps from datasets such as SemanticKITTI and nuScenes. These provide a reliable benchmark with per-point labels for both geometry and semantics.

Unlike prior methods (e.g., R2DM, LiDM) that require post-hoc segmentation using external models, Spiral eliminates this need via end-to-end semantic-aware generation, which we believe marks an important step forward in the field.

Q4: "In Table 3, the 'w/o' model’s performance still increases with a higher ratio of synthetic data, which seems contradictory. Could the authors clarify the experimental setup and explain these results?"

A: Thank you for pointing out this potential confusion. We clarify as follows:

• In Table 3, the “w/o” (or “None”) column refers to training without any generative data augmentation (GDA) — i.e., using only real, labeled LiDAR scenes.
• The “Ratio” column refers to the amount of real labeled data used in training (1%, 10%, 20%, etc.), as stated in (1) the third row of the Table 3 caption, and (2) line 265 of the main paper.

Therefore, performance in the “w/o” column improves not due to synthetic data, but due to the increased availability of real data. For instance, in Table 3, with 10% real data, SPVCNN++ achieves 59.07% mIoU; with 20%, it improves to 61.16%. We have revised the caption of Table 3 to make this setup and terminology clearer.

References:

• [R1] Kong, Lingdong, et al. "Robo3D: Towards Robust and Reliable 3D Perception against Corruptions." Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023.

Last but not least, we would like to sincerely thank Reviewer ZJCH again for the valuable time and constructive feedback provided during this review.

Thanks for the authors' response, which solves some of my concerns. However, I still hold the following concerns.

1. The conditional generation I mentioned refers to generating LiDAR data conditioned on high-level and sparse information such as bounding boxes and key points, which would enable applications across a wider range of driving scenarios. Since several works [1,2,3] have already achieved this capability, I would encourage the authors to consider including this task. While I am not requesting direct comparisons with those works due to different experimental settings, incorporating conditional generation can demonstrate the broader potential of this approach and enhance the contribution of this work.

2. I am still confused about the data augmentation experimental setup. Does the 1% setting mean that only 1% of the real data is used during training? This appears to be an unusual setting, as it significantly reduces the available training data. More common experimental settings utilize all available real data while incorporating synthetic data as augmentation. So I said the improvements shown in the last column of Table 3 appear marginal.

[1] Wu, Z., Ni, J., Wang, X., Guo, Y., Chen, R., Lu, L., ... & Xiong, Y. (2024). Holodrive: Holistic 2d-3d multi-modal street scene generation for autonomous driving. arXiv preprint arXiv:2412.01407.

[2] Ran, H., Guizilini, V., & Wang, Y. (2024). Towards realistic scene generation with lidar diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 14738-14748).

[3] Ren, X., Lu, Y., Cao, T., Gao, R., Huang, S., Sabour, A., ... & Ling, H. (2025). Cosmos-Drive-Dreams: Scalable Synthetic Driving Data Generation with World Foundation Models. arXiv preprint arXiv:2506.09042.

Clarification on the 1% Data Setting

Thank you for seeking clarity regarding the data augmentation protocol. We confirm that:

"The 1% setting refers to using only 1% of the real SemanticKITTI training set, with the remaining training data composed of Spiral-generated LiDAR scenes."

This setup is not intended to simulate a realistic deployment scenario, but rather to rigorously evaluate whether synthetic data alone can provide meaningful learning signals. The motivations are:

1. Fidelity under low-label regimes: If Spiral-generated samples were low-quality or repetitive, training with only 1% real data would severely underperform. Observing strong gains in such a setting validates their informativeness.
2. Non-redundancy with real data: When more real data is available (e.g., 20%), improvements from synthetic data imply that Spiral contributes new, non-trivial samples beyond what the real dataset already provides.

This evaluation strategy — varying the ratio of real-to-synthetic data — follows established protocols in generative data augmentation [R5–R8], though to our knowledge, this is the first such analysis conducted for LiDAR-based semantic segmentation.

To further address your concern, we also evaluate Spiral-based GDA when using the entire real dataset (100%). Below are the results comparing no augmentation, traditional two-stage GDA (R2DM + RangeNet++), and our Spiral-based GDA:

Table. Experiment on entire real dataset (100%) setting. | Datasets | w/o GDA | R2DM & RangeNet++ | Spiral (Ours) | |-|:-:|:-:|:-:| | SemanticKITTI | 65.31 | 65.33 | 66.25 | | Robo3D-fog | 56.47 | 57.64 | 62.18 |

Even in this high-data regime, Spiral continues to deliver consistent gains, while traditional pipelines yield minimal improvement.

Furthermore, we highlight results from the Waymo generalization test. Spiral, trained on SemanticKITTI, is used to augment a 10%-subset of Waymo data. The segmentation model trained with Spiral-generated scenes achieves an mIoU of 57.52%, outperforming both the no-GDA baseline (55.89%) and the two-step baseline (55.93%), despite never seeing Waymo data during training of Spiral. This confirms Spiral’s ability to generalize across domains, reinforcing its practical utility for diverse driving scenarios.

We deeply value your suggestions and will incorporate these clarifications and insights into the revision to improve both the clarity and scope of the manuscript.

Best regards,

The Authors of Submission 1304

References:

• [R1] Wu, Zehuan, et al. "HoloDrive: Holistic 2D-3D Multi-Modal Street Scene Generation for Autonomous Driving." arXiv preprint arXiv:2412.01407, 2024.
• [R2] Ran, Haoxi, Vitor Guizilini, and Yue Wang. "Towards Realistic Scene Generation with LiDAR Diffusion Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.
• [R3] Ren, Xuanchi, et al. "Cosmos-Drive-Dreams: Scalable Synthetic Driving Data Generation with World Foundation Models." arXiv preprint arXiv:2506.09042, 2025.
• [R4] Sun, Pei, et al. "Scalability in Perception for Autonomous Driving: Waymo Open Dataset." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020.
• [R5] Tran, Ngoc-Trung, et al. "On Data Augmentation for GAN Training." IEEE Transactions on Image Processing (30) 1882-1897, 2021.
• [R6] Feng, Chun-Mei, et al. "Diverse Data Augmentation with Diffusions for Effective Test-Time Prompt Tuning." Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023.
• [R7] Xiao, Changrong, Sean Xin Xu, and Kunpeng Zhang. "Multimodal Data Augmentation for Image Captioning using Diffusion Models." Proceedings of the 1st Workshop on Large Generative Models Meet Multimodal Applications, 2023.
• [R8] Zheng, Chenyu, Guoqiang Wu, and Chongxuan Li. "Toward Understanding Generative Data Augmentation." Advances in Neural Information Processing Systems (36) 54046-54060, 2023.

We sincerely thank Reviewer oNZH for the thoughtful and constructive feedback. We are encouraged that you recognize the state-of-the-art performance, the practical utility of our data augmentation strategy, and the novel closed-loop inference mechanism introduced in our work. We address your concerns in detail below:

Q1: "The paper introduces a closed-loop inference mechanism, but does not provide quantitative statistics on when the mode switching occurs. Including this would help clarify the behavior and stability of the strategy."

A: Thank you for raising this insightful point. We conducted a statistical analysis over 1,000 samples each from SemanticKITTI and nuScenes. Under the default 256 denoising steps, the closed-loop inference is triggered (i.e., when ≥80% of semantic predictions surpass the confidence threshold) at an average step of:

• 154 ± 34 for SemanticKITTI
• 161 ± 37 for nuScenes

This implies that the switching typically occurs in the last ~40% of the generation process, when semantic predictions are sufficiently stable. This delayed activation helps avoid early-stage noise and ensures meaningful semantic–geometric alignment. As suggested, these statistics have been added to the revised supplementary material.

Q2: "The paper does not report a comparison of inference time or time consumption for labeled LiDAR scene generation. Since SPIRAL is positioned as an efficient generative model, empirical results on generation speed would further support its suitability for large-scale data generation tasks."

A: Thanks for your comment. We fully agree with your point. As reported in Table B and Section F of the supplementary material, Spiral achieves an average generation time of 5.7 seconds per scene on an NVIDIA A6000 GPU. Compared to two-stage baselines (e.g., LiDM + segmentor), Spiral offers superior efficiency due to its end-to-end architecture, though it is slightly slower than the fastest baseline (R2DM + RangeNet++) by ~2.05 seconds. These results support the practical viability of our approach for scalable labeled LiDAR scene generation.

Q3: "Do you have empirical data on when closed-loop switching occurs?"

• "If switching occurs too early, it might resemble inefficient two-step pipelines."
• "If switching occurs too late, the semantic conditioning might be ineffective."

A: Excellent observations. Our empirical analysis (please see the response to Q1) confirms that closed-loop inference is activated neither too early nor too late, typically during the middle-to-late denoising phase. This timing ensures that the semantic map is reliable enough for conditioning, while also leaving sufficient steps for influence.

Moreover, since both geometric and semantic predictions are generated simultaneously at each step via our multi-head design, early switching (if it occurs) does not incur additional computation as would a separate two-step pipeline. Your concern regarding very early or very late switching degrading quality aligns with our findings in Table 4, where extreme settings of the threshold $\delta$ lead to suboptimal alignment. This further supports the effectiveness of our dynamic switching scheme.

Q4: "In Figure 4, what is the source of the ground-truth (GT) semantic maps?"

A: Thanks for your question. The ground-truth (GT) semantic maps shown in Figure 4 are from the official SemanticKITTI annotations. In contrast:

• Spiral generates semantic maps jointly with the LiDAR scene in a unified diffusion process;
• The baselines (e.g., R2DM, LiDM) generate semantic maps via post-hoc segmentation using a pretrained RangeNet++ model.

This illustrates a key distinction: Spiral performs end-to-end semantic-aware generation, while the others follow a two-step pipeline.

Q5: "Could the authors provide a comparison of inference time for labeled LiDAR scene generation?"

A: Thanks a lot for asking. Please refer to the details in our response to Q2.

Q6: "[Suggestion] Consider classifier-free guidance (CFG) with semantic maps during inference."

A: Thank you for this thoughtful suggestion. We conducted a comprehensive study by applying CFG in our conditional steps with varying guidance weights $w \in \{ 1, 2, 4, 8 \}$, using the formula:

where:

• $\hat{\epsilon}$: final noise prediction,
• $\epsilon_{\text{cond}}$: noise predicted with conditional input,
• $\epsilon_{\text{uncond}}$: noise predicted with unconditional input,
• $w$: guidance strength controlling the influence of the condition.

In the conditional steps of Spiral, we set the guidance weight $w = 0$, which makes the inference to standard conditional generation.

The results (see table below) show minimal improvements when $w = 1$, particularly in Cartesian metrics, and degradation for larger weights. We attribute this to intermediate semantic predictions still containing artifacts, which can mislead the generator when overly trusted. Thus, while promising in principle, CFG offers limited gains in our current setup. As suggested, we have included these results and the discussion in the revised supplementary material.

Note: All metrics are lower-is-better.

Last but not least, we would like to sincerely thank Reviewer oNZH again for the valuable time and constructive feedback provided during this review.

Dear Reviewer oNZH,

Thank you once again for your thoughtful and constructive feedback!

In response to your comments, we have added quantitative statistics on the closed-loop switching step, showing that it typically occurs in the final 40% of the generation process, balancing stability and influence. We also clarified the generation time of Spiral (5.7s/scene), which remains efficient for large-scale data generation. Your concern about early or late switching is well-taken and reflected in our experiments (Table 4), supporting our dynamic threshold design. As requested, we clarified the source of ground-truth semantic maps in Fig. 4 and included empirical results on applying classifier-free guidance (CFG), which yielded limited gains due to intermediate prediction noise. These updates are now included in the revised manuscript and supplementary material.

We are actively participating in the Author-Reviewer Discussion phase and would be happy to provide further clarification if needed!

Best regards,

The Authors of Submission 1304

Thank you for the detailed rebuttal. Regarding the answer to Q4, what confused me is how the generation model could have ground-truth segmentation annotations when the generated data is inherently unpredictable. Please correct me if I’ve misunderstood anything.

Dear Reviewer oNZH,

We sincerely thank you for your continued engagement!

In our previous reply to Q4, we misunderstood your original question. To clarify:

1. The segmentation maps attached to the generated LiDAR samples are generated jointly with the point clouds during generation.
2. However, when we report evaluation results, the ground truth always refers to all real LiDAR scenes and their corresponding human-annotated segmentation labels.

Below, we would like to explain this in more detail:

• We would like to emphasize that, in generation tasks, the evaluation does not compare each generated sample against a specific "ground truth" counterpart. Instead, it evaluates the statistical distributional similarity between the generated set and the real set, as is commonly done with metrics like FID [R1] in the 2D image generation community.

• Specifically:

  • Given a generated set consisting of $M$ LiDAR scenes {$(x_1^g, y_1^g), (x_2^g, y_2^g), ..., (x_M^g, y_M^g)$}, where $x^g$ denotes the generated LiDAR point coordinates and $y^g$ denotes the generated semantic maps, and a real set consisting of $N$ LiDAR scenes and their human-annotated segmentation labels: {$(x_1^r, y_1^r), (x_2^r, y_2^r), ..., (x_N^r, y_N^r)$}. For a generated LiDAR scene $x_i^g$, there is no corresponding human-annotated segmentation label, so the evaluation cannot rely on per-sample comparisons.
  • Instead, we transform $(x_i^g, y_i^g)$ into a semantic-aware feature $f^s_g$ by concatenating the encoded features from the RangeNet++ encoder and the LiDM semantic conditional module, as illustrated in Eq. 9 and Fig. 3 of the manuscript.
  • This gives us a feature set for the generated samples: $\mathcal{F}^s_g =$ {$f^s_{g1}, f^s_{g2}, ..., f^s_{gM}$}.
  • Similarly, we obtain the feature set for real samples: $\mathcal{F}^s_r =$ {$f^s_{r1}, f^s_{r2}, ..., f^s_{rN}$}.
  • We then compute the statistical similarity between $\mathcal{F}^s_g$ and $\mathcal{F}^s_r$ using our proposed metrics, such as S-FRD and S-MMD, as defined in Eqs. 2-5 in the supplementary material.

• Only when the generated LiDAR scenes exhibit both high geometric quality and strong geometric-semantic alignment (i.e., well-aligned semantic maps), will the features $f^s_g$ fall into a reasonable domain and $\mathcal{F}^s_g$ share statistical properties with $\mathcal{F}^s_r$.

• Thus, by comparing the semantic-aware feature distributions, we bypass the need for ground-truth semantic labels for the generated samples.

We appreciate your thoughtful question and hope the above explanation resolves the ambiguity. We are happy to elaborate further if any details are still unclear.

Yours sincerely,

The Authors of Submission 1304

Reference:

• [R1] Heusel, Martin, et al. "GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium." Advances in Neural Information Processing Systems 30, 2017.

We sincerely thank Reviewer gApr for the thoughtful and constructive feedback. We are encouraged that you found our closed-loop inference design both “novel” and “reasonable,” and appreciated our improvements in semantic-scene alignment. Below, we address each of your concerns in detail.

Q1: "What is the memory and runtime compared to your baselines?"

A: Thank you for this practical question. The parameter sizes of Spiral and baselines are reported in the “Param” column of Table 1 and Table 2 in the manuscript. The average inference time per LiDAR scene is presented in Table B and Section F of the supplementary material. Specifically, Spiral takes 5.7 seconds per scene on an NVIDIA A6000 GPU, which is comparable to or faster than two-stage baselines, thanks to its streamlined end-to-end generation pipeline.

Q2: "The proposed S-x metrics compare the semantic map with the ones in the dataset, but it does not account for the alignment between the generated scene and the map. For example, a garbage scene with perfect segmentation would still pass the metric. This might over-weight scene generation in the evaluation."

A: We appreciate this insightful concern. In practice, we find that the scenario of a “garbage scene” with strong semantic segmentation performance rarely occurs. From both empirical observation and theoretical reasoning:

• In our two-stage baselines (e.g., LiDARGen + RangeNet++), moderately corrupted LiDAR scenes already lead to significant segmentation failures, even when the segmentation network is extensively trained with noise augmentations. This is expected, since 3D segmentation is fundamentally geometry-dependent — poorly generated structures yield unreliable spatial features and semantic predictions.
• Our semantic-aware evaluation metrics are explicitly designed to measure cross-modal alignment, not scene quality in isolation. As illustrated in Fig. 3(a), we encode the depth image and semantic map independently using two modality-specific encoders (RangeNet++ and LiDM-Seg), and fuse the extracted features into a unified semantic-aware feature $f_s$. Therefore, even in the rare cases where a scene may have poor geometric quality but good semantic map, the geometric quality is not "overly out-weighted".
• At the BEV level (please see Fig. 3(b)), we compute semantic histograms across classes and spatial cells. These are highly sensitive to both the correct geometry (e.g., point distribution) and the correct semantic allocation (e.g., class labels). If either modality fails, due to noisy generation or poor segmentation, the BEV histogram will diverge from the real data, and the model will be penalized accordingly.

In summary, geometry and semantics contribute equally in our metrics by design, and misalignment across modalities is indeed captured. As suggested, we have further clarified this detail in the revised manuscript.

Q3: "Regarding closed/open-loop inference: I understand that open-loop directly diffuses both modalities, which might lead to discrepancies. However, the metrics may not be able to reflect misalignment between the generated scene and the map."

A: Your understanding of the open-loop setting is correct — it involves jointly generating the LiDAR scene and semantic map without enforcing consistency during generation. This indeed introduces cross-modality discrepancies.

• As shown in Table 4, closed-loop inference with default hyperparameters (the highlighted row) outperforms open-loop inference (first row). Importantly, we emphasize that the semantic map in closed-loop inference is still generated by Spiral, and is not externally provided. Therefore, we believe that the proposed closed-loop inference strategy is a novel contribution of this work.
• Our proposed semantic-aware metrics are designed to effectively evaluate cross-modality alignment.
  • For range-view and Cartesian-based metrics, as illustrated in Fig. 3(a), the semantic-aware feature $f_s$ for each scene is formed by concatenating the global features extracted from the depth image and semantic map, respectively. This ensures that both geometric and semantic quality are jointly evaluated.
  • For BEV-based metrics, they assess alignment at a finer granularity. Specifically, for each generated scene, we compute semantic-aware histograms $h_s$ over the bird’s-eye-view by combining the histogram of each semantic class (please see Fig. 3(b)). As a result, a generated scene with either:
    1. good semantic segmentation but poor geometric quality, or;
    2. poor semantic segmentation but good geometric quality

will fail to produce a histogram that matches the real scene. Additionally, this also ensures that the BEV metrics penalize the misalignment in geometry or semantics, making them strong indicators of overall generation fidelity.

Furthermore, as elaborated in Q2, our semantic-aware metrics, including modality-specific encoders and BEV histograms, are designed to detect cross-modal misalignment, ensuring that semantic and structural coherence is thoroughly assessed.

Q4.1: "The data augmentation results should be more extensively compared, especially as this is a main application of the paper."

A: We appreciate your emphasis on this point. In addition to the experiments presented in the manuscript, we further conduct a comprehensive evaluation of generative data augmentation (GDA) under two challenging tests:

1. Cross-Domain Generalization (Waymo Test):

   • We use LiDAR scenes generated by Spiral, trained on SemanticKITTI, to augment the training set of Waymo [R1] (when 10% of real data is available), and evaluate the model's performance on the Waymo val set.
   • The segmentation model SPVCNN++ is trained under three settings:
     1. without GDA,
     2. with synthetic data generated by R2DM & RangeNet++,
     3. with our Spiral-generated data.
   • The segmentation performance (mIoU) under these settings is 55.89%, 55.93%, and 57.52%, respectively. These results demonstrate the strong generalization ability of our Spiral-based GDA approach.

2. Out-of-Distribution (OOD) Extreme Weather (Robo3D Test):

   • We evaluate SPVCNN++ trained under the same three settings as in the above "Cross-Domain Generalization Test", on challenging out-of-distribution subsets, fog and wet ground, from Robo3D [R2].
   • The results demonstrate that, although Spiral is not specifically fine-tuned for these OOD extreme weather conditions, the data generated by Spiral still effectively enriches the training set and improves performance.

Table. Experiments under fog scenarios: | Ratio of Real Data | w/o GDA | R2DM & RangeNet++ | Spiral (Ours) | |:---:|:---:|:---:|:---:| | 1% | 32.07 | 36.82 | 44.06 | | 10% | 53.93 | 54.20 | 58.61 | | 20% | 55.89 | 56.07 | 61.24 |

Table. Experiments under wet ground scenarios: | Ratio of Real Data | w/o GDA | R2DM & RangeNet++ | Spiral (Ours) | |:---:|:---:|:---:|:---:| | 1% | 34.26 | 37.96 | 44.19 | | 10% | 54.70 | 55.92 | 59.31 | | 20% | 56.81 | 57.53 | 62.02 |

Across all scenarios, our Spiral-based GDA consistently outperforms all baselines. These results demonstrate our effectiveness in both cross-domain and OOD generalization settings. As suggested, we have improved the presentation of these results and emphasized the comparisons further in the revised manuscript.

Q4.2: "Please show the mIoU of SPVCNN++ without any augmented data."

A: Thank you for pointing this out. The mIoU results of SPVCNN++ trained without any generative data augmentation (GDA) are already reported in the “None” column of Table C in the supplementary material (see also "w/o" column of Table 3 in the manuscript).

This baseline reflects the model trained using only real data. For example, under the 10% supervision setting on the Robo3D-fog set, this baseline achieves 53.93% mIoU, while Spiral-based GDA improves the score to 58.61%. We have revised the table caption to make this clearer.

Q4.3: "The 'None' column in Table C is unclear and not explained."

A: Thanks for pointing out our oversight! The “None” column in Table C (corresponding to the "w/o" column in Table 3 in the main paper) refers to the baseline performance of SPVCNN++ trained without any synthetic data, i.e., real-only training. This column is essential for measuring the absolute contribution of generative augmentation. We have revised the table caption and corresponding text to explicitly define “None” and "w/o" to avoid confusion.

Q4.4: "What happens when the real data ratio is high (e.g., 80%)?"

A: Great question. We have conducted additional experiments under the 80% supervision setting, where the amount of real data is abundant. The resulting mIoU scores on the SemanticKITTI are:

• None (real only): 64.48%
• R2DM + RangeNet++: 64.51%
• Spiral (ours): 65.44%

Even in this data-rich regime, Spiral continues to provide non-trivial performance gains, while R2DM shows negligible benefit. This result highlights that Spiral not only excels in low-data regimes but also contributes value in high-data scenarios by generating semantically aligned and diverse LiDAR scenes.

References:

• [R1] Sun, Pei, et al. "Scalability in Perception for Autonomous Driving: Waymo Open Dataset." CVPR, 2020.
• [R2] Kong, Lingdong, et al. "Robo3D: Towards Robust and Reliable 3D Perception against Corruptions." ICCV, 2023.

Last but not least, we would like to sincerely thank Reviewer gApr again for the valuable time and constructive feedback provided during this review.

Dear Reviewer gApr,

Thank you once again for your thoughtful and constructive review!

In response to your comments, we clarified the runtime and memory usage of Spiral compared to baselines, and explained how our semantic-aware metrics jointly assess geometric and semantic fidelity to capture modality misalignment. We also emphasized the role of closed-loop inference in promoting semantic-scene consistency and provided further evaluation results under both cross-domain (Waymo) and out-of-distribution (Robo3D) settings. These results demonstrate consistent performance gains over existing augmentation methods, even under high-data regimes. We have revised the manuscript and supplementary materials accordingly to improve clarity and completeness.

We are actively participating in the Author-Reviewer Discussion phase and would be happy to provide further clarification if needed!

Best regards,

The Authors of Submission 1304

We sincerely thank Reviewer XYPz for the thoughtful and constructive feedback. We are encouraged that you found our work “useful,” “well-written”, and that our results demonstrate “good” performance. Below, we address your key concerns in detail.

Q1: "Limited semantic label set: semantic output is constrained to pre-defined categories from SemanticKITTI and nuScenes, reducing applicability to varied domains."

A: Thanks for your question. Our current design follows a widely accepted convention in 3D perception benchmarks, where tasks such as segmentation and detection rely on fixed taxonomies, as in SemanticKITTI and nuScenes. These datasets are representative and diverse, encompassing a broad range of scenes (urban, residential, highway, rural) across different geographic regions (Germany, Singapore, USA), and have become de facto standards in the field.

While constrained to pre-defined classes, these benchmarks include out-of-distribution (OOD) labels, such as the “unlabeled” category in SemanticKITTI (e.g., tunnel walls, scaffolding, bus stops). Spiral is capable of modeling such categories as well, ensuring that OOD content is not omitted in our generative process.

We believe that extending Spiral to support an unlimited number of categories is a meaningful direction. However, it faces a critical bottleneck: the lack of reliable text–LiDAR paired data, as also observed by LiDM [R1] and Text2LiDAR [R2]. Although Text2LiDAR [R2] introduces language annotations, they are typically coarse, scene-level descriptions (e.g., "cars and buses at a busy street", "surrounded by cars") and lack fine-grained, point-level semantics that are essential for point-wise labeled LiDAR scene generation. This makes open-vocabulary generation technically under-constrained. Addressing this challenge is a key goal in our ongoing research. Thanks again for your valuable suggestion.

Q2: "Long-tailed label distributions in LiDAR data (e.g., rare classes like bicyclists) can cause the model to underfit these classes. The paper doesn’t detail any strategies (e.g., re-weighting) to balance this."

A: Thank you for highlighting this. We acknowledge the class imbalance present in both SemanticKITTI and nuScenes. To address this, Spiral incorporates a class-reweighting strategy, consistent with the official RangeNet++ codebase.

Specifically, the loss weight for category $i$ is defined as:

where $\text{freq}_i$ is the normalized frequency of category $i$, and $\epsilon$ is a small constant (default: $0.001$).

This strategy mitigates the dominance of frequent classes and encourages the model to attend to under-represented semantic categories during training. As suggested, we have further clarified this detail in the revised manuscript.

Q3: "Although the approach reduces collection and annotation costs, since the model is trained on a specific dataset, it is questionable how well it captures real-world variability and complexity, which may limit the practical utility of the generated data."

A: We appreciate this concern. To evaluate how well Spiral captures real-world variability and how robust it is to real-world complexity, we conduct experiments on generative data augmentation (GDA) under two challenging scenarios:

1. Cross-Domain Generalization (Waymo Test):

   • We use LiDAR scenes generated by Spiral, trained on SemanticKITTI, to augment the training set of Waymo [R4] (when 10% of real data is available), and evaluate the model's performance on the Waymo val set.
   • The segmentation model SPVCNN++ is trained under three settings:
     1. without GDA,
     2. with synthetic data generated by R2DM & RangeNet++,
     3. with our Spiral-generated data.
   • The segmentation performance (mIoU) under these settings is 55.89%, 55.93%, and 57.52%, respectively. These results demonstrate the strong generalization ability of our Spiral-based GDA approach.

2. Out-of-Distribution (OOD) Extreme Weather (Robo3D Test):

   • We evaluate SPVCNN++ trained under the same three settings as in the above "Cross-Domain Generalization Test", on challenging out-of-distribution subsets, fog and wet ground, from Robo3D [R3].
   • The results demonstrate that, although Spiral is not specifically fine-tuned for these OOD extreme weather conditions, the data generated by Spiral still effectively enriches the training set and improves performance under these challenging scenarios.

Table. Experiments under fog scenarios: | Ratio of Real Data | w/o GDA | R2DM & RangeNet++ | Spiral (Ours) | |:---:|:---:|:---:|:---:| | 1% | 32.07 | 36.82 | 44.06 | | 10% | 53.93 | 54.20 | 58.61 | | 20% | 55.89 | 56.07 | 61.24 |

Table. Experiments under wet ground scenarios: | Ratio of Real Data | w/o GDA | R2DM & RangeNet++ | Spiral (Ours) | |:---:|:---:|:---:|:---:| | 1% | 34.26 | 37.96 | 44.19 | | 10% | 54.70 | 55.92 | 59.31 | | 20% | 56.81 | 57.53 | 62.02 |

These results confirm that Spiral generates semantically and structurally diverse scenes that enhance real-world performance under both domain and weather shifts — supporting its practical utility.

References:

• [R1] Ran, Haoxi, et al. "Towards Realistic Scene Generation with LiDAR Diffusion Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.
• [R2] Wu, Yang, et al. "Text2LiDAR: Text-Guided LiDAR Point Cloud Generation via Equirectangular Transformer." European Conference on Computer Vision, 2024.
• [R3] Kong, Lingdong, et al. "Robo3D: Towards Robust and Reliable 3D Perception against Corruptions." Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023.
• [R4] Sun, Pei, et al. "Scalability in Perception for Autonomous Driving: Waymo Open Dataset." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020.

Last but not least, we would like to sincerely thank Reviewer XYPz again for the valuable time and constructive feedback provided during this review.

Dear Reviewer XYPz,

Thank you once again for your thoughtful and constructive feedback!

In response to your comments, we clarified the scope and motivation of using fixed semantic taxonomies (e.g., SemanticKITTI, nuScenes), and discussed the potential and current limitations of open-vocabulary generation due to the scarcity of fine-grained text–LiDAR pairs. We also detailed the class-reweighting strategy adopted during training to mitigate long-tailed label distributions. Finally, we added new results on cross-domain (Waymo) and OOD (Robo3D) scenarios, demonstrating that Spiral-generated data significantly improves real-world robustness and generalization.

Your suggestions have greatly helped us improve the clarity and completeness of our work.

We are actively participating in the Author-Reviewer Discussion phase and would be happy to provide further clarification if needed!

Best regards,

The Authors of Submission 1304