1. $**Motivation**$

1.1 One concern is to what extend we need to redistribute points in generated point clouds? Holes and clusters indeed exist in the generated point clouds, but does this problem is severe enough? Some down-streaming or related tasks or applications which are severe affected by the inequal distribution of points are needed to strength the motivation of the paper.

1.2 If we generate more points for shapes and then uniformly downsample again, will the issue of holes and clusters be alleviated? This is related with the significance of the paper.

2. $**Method**$

2.1 According to Figure 3, it seems that the hyperparameters $\epsilon$ and $\sigma$ have significant influence on the redistribution result, so it may need case-by-case tuning of $\epsilon$ and $\sigma$ to achieve a good performance. Correct me if I'm wrong.

2.2 From Figure 5 and Figure 7, incorporating LJL could lead to over-smooth shape boundaries (wings and tailplanes in Figure 5) and slightly distortion of shape details (nose in Figure 7). This may be the drawback of the proposed approach. If such a drawback could be corrected, it will strengthen the paper.

3. $**Experiments**$

3.1 Some experimental details are missing. For example, how many points are generated per shape in experiments in Section 4.2? Did you retrain ShapeGF, and Lou & Hu's model or used their released pretrained model? Why not use the original evaluation metrics adopted in ShapeGF, and Lou & Hu's paper?

3.2 It is only evaluated on three generative approaches from two categories, e.g., audo-encoding and DDPM, more experiments and evaluations are needed.

4. $**Writting**$

4.1 The legend text in Figure 18 is too small to be seen clearly.

3.3 It is preferable to add citations in Table 1 and Table 2 to make it easier to check the reference papers.

Although some results presented in the paper look good, my main concern is that the paper fails to convincingly show the value or potential of the proposed method.

• The authors choose to use the 3D point cloud generation tasks to show the value of the proposed layer. However, I think it is still questionable whether the method can be generalized to other or more advanced point cloud generation methods beyond ShapeGF and DDPM. If the method is only compatible with these two relatively old methods, the contribution of this method might not be high.

• The irregular distributions of point clouds may carry useful information about the point clouds. Normalizing point cloud distributions may not always be helpful to improve generation results. To balance the irregular structures and normalized distributions, it may need prior knowledge to adjust the hyper-parameters of the proposed method. According to Figure 14, the method seems sensitive to these hyper-parameters. Do you have a systematic/automatic solution to determine these parameters? If it is difficult to determine these parameters or the parameters need to be determined case-by-case, the method may not be able to generalize to various problems.

• Table 1 and 2 only report the relative improvement. Can you provide detailed results? If the proposed method can directly improve the number reported in the original papers, the results will be more convincing.

1. Lack of clear guidance in choosing hyper parameters (eg. \alpha and \beta in Eq. 2). Although there’s discussion in Appendix, the authors still adopt grid search. Due to the diversity in various point clouds with different local/global density, it is better to provide a more clear guidance for choosing hyper parameters.
2. The benefits of uniform distribution on downstream perception tasks (eg classification, segmentation, detection) is not verified, which is important since the application of point clouds mainly lies in perception tasks. This is not trivial since sometimes ununiform sampling is more effective (eg. Edge sampling in [1])
3. All of these experiments are based on object-level point cloud. However, scene-level point cloud is more important in real-world applications. Is the proposed method still performs well when it comes to scene-level point clouds?
4. The efficiency of this algorithm when processing large scale point clouds (~1 million points, which is common in some real-world datasets) is not mentioned. [1]. Wu, Chengzhi, et al. "Attention-based Point Cloud Edge Sampling." In CVPR, 2023.

• The proposed method might only be useful for certain classes of point cloud generative models. The paper only demonstrates improved point cloud uniformity when used in conjunction with ShapeGF (Cai et al.) and DDPM (Luo et al.). However, these two models are inherently flawed -- they formulate point cloud generation as independent points uniformly distributed on the surface, thus unable to capture the global uniformity. Methods that models joint distribution of points, such as LION (Zheng et al.) and Point-E (Nichol et al.) might already produce uniform point clouds without the proposed method.
• Evaluation is lacking -- it will be more solid if the proposed method can be benchmarked against state-of-the-art models using standard point cloud generation and reconstruction evaluation metrics on standard datasets, instead of just presenting the percentage increase over a simple baseline.
• The exposition is sometimes comfusing. For example, it is not clear how the optimal LJL parameters are found using Algorithm 2 -- it shows merely a diffusion autoencoder for point clouds.
• Minor writting issue: In Algorithm 2, "autoencoder" usually refers to an encoder-decoder that is trained to reconstruct the input. In this case, it is better to call "E_\theta" as "encoder" instead.