Q3 Sensitivity to quality variations

The sensitivity of hierarchical control to quality variations is reflected in two aspects:

1. The reconstruction quality in stage 1 caps the controllability. If the proposed LOE cannot learn hierarchically structured latents in Stage 1, then the hierarchical control facilitated by generation in Stage 2 is inhibited. We find that when the PSNR on CelebA-HQ reaches above 25, there won't be visually apparent distortions, the hierarchical control works well.

2. The quality of ground-truth data for stage 1 caps the controllability. That is because the larger noise level makes learning conditional dependency harder. In an extreme case where the data consists solely of noise, the learned latents will lose any meaningful dependencies. To demonstrate the analysis, we experiment by perturbing ground-truth data with various noise levels as shown in the table below. As the noise level in ground-truth data increases, the quality of both reconstruction and generation deteriorates. The range of RGB values for ground-truth data is $[0,1]$.

Table: Sensitivity of conditionally sampled data to noise in ground-truth data, evaluated on CelebA-HQ.

Therefore, in real-world applications, it should be ensured the data has no apparent distortions.

Q2 CHINR for unstructured data

Our method can represent any structured signal as a function, i.e., a mapping from input coordinates to output values. To represent unstructured data such as texts, we can first project unstructured texts into structured embedding space with dimension $L$, so a phrase with $N$ words can be represented as a function: $f:x\rightarrow T: \in \mathbb{R}^{N}\rightarrow\mathbb{R}^{N\times L}$. Then the CHINR can be trained to represent any phrase with positions $x \in [1\cdots N]$ as input, and $T$ as output.

However, the control mechanism will not work on text embeddings, because the textual embedding space lacks an inherent hierarchical structure that aligns with spatial frequency, i.e. the low-frequency components represent the overall structure while high-frequency components represent finer details.

Based on the analysis, a promising future for hierarchical control of textual data is to figure out an embedding space that has a hierarchical structure aligned with spatial frequency, then the CHINR will work.

We appreciate valuable comments from Reviewer ww4M.

Q1 Optimize binary conditions

1. Yes, a promising approach for interpretability enhancement is to establish correspondences between binary conditions and ground-truth class labels. Incorporating these class labels during training enables control based on class labels or textual inputs.

2. The binary condition has no apparent effect on training efficiency.

Thanks for your insightful questions. We hope our answers address your concerns. Please let us know if you have any further questions.

Q3 Markov in eq.3

Thanks for your question. In Equation 3 the expression is $p(\mathbf h^l|\mathbf h^{<l})$ rather than $p(\mathbf h^l|\mathbf h^{l-1})$ (as in Markov chain). It is noted that $p(\mathbf h^l|\mathbf h^{<l})$ represents the dependency in layers of latents, which is not the Markov in backward diffusion steps.

Mathematically, $p(\mathbf h^l|\mathbf h^{<l})$ arises from the factorization of joint distribution into marginalized and conditional distributions. For example, for $\mathbf h=[\mathbf h^1, \mathbf h^2, \mathbf h^3]$, the joint distribution $p(\mathbf h)$ can be written by the chain rule as:

$

p(\mathbf h) &= p(\mathbf h^1)p(\mathbf h^2|\mathbf h^1)p(\mathbf h^3|\mathbf h^2,\mathbf h^1)\\ &= p(\mathbf h^1)\prod^{3}_{l=2}p(\mathbf h^l|\mathbf h^{<l}).

$

Intuitively, this reflects the hierarchical nature of the layer-wise latents. Consider a hierarchy of facial attributes (orientation--expressions--shape of the eyes), the fine details of the eyes depend on both the orientation and expressions, not just on the immediate predecessor. This fundamentally differs from a Markov chain, where each state depends only on the previous state.

Q2 The equation for the LoE

Thank you for your suggestion. We agree that providing an equation for the Layers-of-Experts (LoE) structure would improve clarity. Each expert in the LoE is a fully connected layer, not a SIREN model. So the parameters of expert $k$ at layer $l$, denoted by $\boldsymbol\theta^l_k$ as appeared in Sec. 3.1, contain a weight matrix and a bias vector. The experts are controlled by a layer-specific gating function, which determines their contributions at each layer. Nonlinear activation (sine) is applied after the experts are combined. The equation for the LoE INR at layer $l$ is

$

\begin{split} \boldsymbol y^{l+1} = \sin(\omega_0\cdot(\bar{\boldsymbol\theta}^l\cdot\boldsymbol y^l)),~~ \bar{\boldsymbol\theta}^l&=\sum_{k=1}^K\boldsymbol\theta_k^l\cdot \alpha_k^l,\\ [\alpha_1^l,\alpha_2^l,\cdots,\alpha_K^l]^\top=\boldsymbol\alpha^l&=g_{\boldsymbol\phi}(\mathbf h^l), \end{split}

$

where $\boldsymbol y$ represents output of each layer, $\omega_0$ is a constant factor, and $\bar{\boldsymbol\theta}^l$ denotes parameters at layer $l$, modulated by a gating vector $\boldsymbol\alpha^l$, which is computed by the gating module $g_{\boldsymbol\phi}(\cdot)$. $\mathbf h^l$ denotes the $l_{th}$ layer of latent $\mathbf h$.

We have included this equation in the revised paper.

W2 Discrepancy in the conditional chain affects quality

Thank you for your insightful feedback. Controllability is a valid and valuable bonus ensured by sampling in a conditional chain. However, Figure 6 does not show the conditional chain. It is a latent replacement experiment to reveal that the layer-wise latents embed distinct and compositional semantics. As expected, the replacement disrupts the conditional chain, affecting image quality. In contrast, our controllable generation follows the conditional chain to ensure compatible latents for better quality.

We acknowledge that the conditional chain can pose challenges for generation quality due to potential error accumulation. For instance, if a generated latent lies in a low-density region of the learned distribution, subsequent layers may struggle to produce meaningful and compatible outputs. However, iterating multiple times, being time-consuming, does not essentially guarantee sampling in high-density regions. Instead, we address this in both training and inference of the conditional diffusion models.

1. During training, we first train the model to learn inter-layer dependency, using ground truth layer-wise latents (from Stage 1) to form conditions. Then, we fine-tune the model to learn a reasonable condition chain by gradually using generated latents to form conditions.

2. During inference at each layer, we further mitigate issues by adjusting the standard deviation of input noise for the diffusion process, reducing the likelihood of generating outlier latents that degrade quality.

We appreciate the insightful feedback from Reviewer LwTN.

W1 Table 1 needs uncertainties

Uncertainty

Thanks for pointing out this. Table 1 presents results from published papers to ensure fair comparisons, though none of the methods report uncertainty metrics. We understand your concern and retrain these methods four times under identical settings but with different random seeds. The uncertainties, i.e. standard deviation, are reported in tables below.

The results for CelebA-HQ are reported in:

The results for ShapeNet are reported in:

The results for SRN-Cars are reported in:

Strengths

While Table 1 demonstrates competitive generation performance, our primary focus is on controllable generation through hierarchical modeling—key aspects not addressed by prior work. Overall, our contributions are highlighted in four aspects: controllability, generalizability, versatility, and quality.

• Controllability: Our work distinguishes from previous methods in the controllability of generating fine-grain details, which has never been explored in the implicit neural representation (INR) community.

• Generalizability: Our work can avoid trivial memorization thus generating more diverse data compared with the SOTA methods, such as mNIF. Please find the discussion and visual samples from Line 357-363 in the main paper and Figure 9 in the supplementary material, respectively.

• Versatility: The CHINR's performance is validated across four different data domains, i.e. facial images, point clouds, NeRFs, and motions. This demonstrates the proposed hierarchical control and conditional dependency modeling is broadly applicable to data semantics with an inherent hierarchy structure.

• Quality: Our work outperforms the SOTA methods on both the reconstruction and generation metrics, which can be seen from Table 1 in the main paper.

[1] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one, ICML 2022.

[2] Yilun Du, Katie Collins, Josh Tenenbaum, and Vincent Sitzmann. Learning signal-agnostic manifolds of neural fields, NeurIPS 2021.

[3] Emilien Dupont, Yee Whye Teh, and Arnaud Doucet. Generative models as distributions of functions, AISTATS 2022.

[4] Tackgeun You, Mijeong Kim, Jungtaek Kim, and Bohyung Han. Generative neural fields by mixtures of neural implicit functions, NeurIPS 2024.

Q1 Hierarchy alignment

Thank you for your insightful question. The alignment of hierarchies in generations with human abstractions, as shown in Figure 1, is supported by: 1. the hierarchical representation ability of INRs (Sec. 2.2), embedding coarse-to-fine semantics across layers, and 2. the conditional chain in HCDM, ensuring their coherent composition. Specifically,

1. INRs naturally exhibit a spectral bias, where early layers capture coarse, low-frequency components (e.g., object categories like chair, jet, table), and later layers refine these with higher-frequency details (e.g., types of chairs). This inherent characteristic of INRs underlies the observed alignment with human abstractions.

2. In our method, the conditional chain ensures that coarse structures (e.g., object categories) are determined first, and the later layers are constrained to finer details (e.g., subcategories or textures). This controlled generation process reflects the hierarchical nature of the data.

Therefore, by embedding then conditional sampling, these two factors collaboratively generate human abstraction-aligned data as shown in Figure 1.

Thanks for your constructive comments and valuable questions. We are happy to engage in further discussions to address any additional concerns.

Q4 The design of the condition seems very ad-hoc

Thank you for your feedback. The quantization in the condition design is indeed chosen intuitively for two reasons:

1. By quantizing the condition, we introduce an information bottleneck that prevents the model from memorizing the training data, encouraging it to capture generalizable patterns.

2. Quantization helps the model discover and encode inherent clusters of the data semantics, avoiding encoding noise or overly fine details irrelevant to the hierarchy.

Q5 Citation needed in L32

Thanks for pointing out this, we have added the citation for INRs.

Q2 Minor comments

• "the high dimensionality of raw weights is intractable" -- This statement in Section 3 refers to the challenges of directly learning the distribution of high-dimensional raw INR weights for generative modeling, not just generating weights efficiently as in [1, 2, 3]. However, we understand that this phrasing may lead to misinterpretation. We have revised it to: "the high dimensionality of raw weights makes distribution modeling highly challenging"

• Typo on Line 465 -- Thank you for pointing out this. We have revised it in our new submission.

[1] Emilien Dupont, Yee Whye Teh, and Arnaud Doucet. Generative Models as Distributions of Functions, AISTATS 2022.

[2] Batuhan Koyuncu, Pablo Sanchez-Martin, Ignacio Peis, Pablo M Olmos, and Isabel Valera. Variational Mixture of HyperGenerators for Learning Distributions over Functions, ICML 2023.

[3] Siyuan Xu, Yucheng Wang, Mingzhou Fan, Byung-Jun Yoon, and Xiaoning Qian. Uncertainty-aware Continuous Implicit Neural Representations for Remote Sensing Object Counting, AISTATS 2024.

We appreciate Reviewer wk7N for their valuable comments.

Q1 Background missing reference to recent works

Thanks for this constructive suggestion. We have included these references in our background section for a more comprehensive discussion. While [1, 2, 3, 4] are included as methods that focus on latents, we do not restrict the learning methods to meta-learning or two-stage training. However, we recognize the need for a more detailed discussion of their techniques and contributions. Specifically:

• Dupont et al. [1] propose an end-to-end method using a GAN to train a hypernetwork that generates INR parameters for synthesizing new data.

• Koyuncu et al. [5] extend GASP’s capabilities for efficient inference tasks (eg. in-painting) via VAE. This framework incorporates a learned prior modeled by Normalized Flow.

• Du et al. [2] learn a signal-agnostic data manifold, where latents are mapped to INRs across different data modalities via different hyper-networks.

• Dupont et al. [3] adopt a two-stage framework, employing meta-learning to jointly optimize a shared base INR and data-specific modulation latents.

• Bauer et al. [4] extend Functa to scale to ImageNet by incorporating spatial information into the latents.

• Park et al. [7] also employ a two-stage framework, where learned latents are mapped to positional embeddings (rather than neural weights) to modulate INRs and generate new data.

• Xu et al. [6] focus on adapting INR parameters for object counting, where a hypernetwork uniquely tailors parameters for each input image to improve training convergence. This work emphasizes generalizing INRs to different inputs rather than generating new data.

[1] Emilien Dupont, Yee Whye Teh, and Arnaud Doucet. Generative Models as Distributions of Functions, AISTATS 2022.

[2] Yilun Du, Katie Collins, Josh Tenenbaum, and Vincent Sitzmann. Learning signal-agnostic manifolds of neural fields, NeurIPS 2021.

[3] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one, ICML 2022.

[4] Matthias Bauer, Emilien Dupont, Andy Brock, Dan Rosenbaum, Jonathan Richard Schwarz, and Hyunjik Kim. Spatial functa: Scaling functa to imagenet classification and generation, arXiv preprint arXiv:2302.03130, 2023.

[5] Batuhan Koyuncu, Pablo Sanchez-Martin, Ignacio Peis, Pablo M Olmos, and Isabel Valera. Variational Mixture of HyperGenerators for Learning Distributions over Functions, ICML 2023.

[6] Siyuan Xu, Yucheng Wang, Mingzhou Fan, Byung-Jun Yoon, and Xiaoning Qian. Uncertainty-aware Continuous Implicit Neural Representations for Remote Sensing Object Counting, AISTATS 2024.

[7] Dogyun Park, Sihyeon Kim, Sojin Lee, and Hyunwoo J Kim. DDMI: Domain-agnostic Latent Diffusion Models for Synthesizing High-Quality Implicit Neural Representations, ICLR 2024.

Thanks for your valuable feedback on improving the background and presentation. We are happy to engage in further discussions. Please let us know if you have any additional concerns.

I thank the authors for their thorough rebuttal. All of my concerns have now been addressed, and I continue to recommend acceptance.

Q1 Model size & inference cost

We focus on improving the controllability of data generation in INR, where model efficiency is not our primary focus. Nevertheless, we achieve better quality with a smaller number of parameters and comparable inference costs.

Reviwer concerns

but I noticed that mNIF reports the number of parameters and inference cost across different generation models

Answer: It is noted that mNIF evaluates the number of parameters and inference cost purely on the INR, not the generator (diffusion model).

Would it be important to consider the size of the data generator as a factor here?

Answer: For the generator, the backbone of HCDM has the same size as mNIF. Our HCDM employs an extra $128\times12$ fully-connected layer to form the conditions and a $12\times4096$ fully-connected layer to embed binary conditions, which brings negligible model size increment.

How does CHINR's inference cost and model size compare to its competitors?

Answer: We report the number of parameters, inference cost, and size of the generator in table as below.

Table: Quantitative results retrained on CelebA-HQ.

W2 Quantitative comparison of conditional dependency

Thanks for your suggestion, we evaluate the importance of conditional dependency by progressive ablation, as shown in the table below. The experiments in the table reveal that enhancing the modeling of conditional dependency (from bottom to top) helps generate higher-quality data.

The experiments start with all five layers (the top row) in the conditional chain and then progressively exclude layers from the chain. In each setting, the excluded layers are trained using unconditional diffusion models. TThe table shows that the conditional chain is essential in producing high-quality generations from hierarchical sampling, as it ensures latent compatibility across layers and alignment with the data's semantic structure.

Table: Ablation on the number of layers in the conditional chain, evaluated on CelebA-HQ.

We appreciate constructive comments from Reviewer 1yWi.

W1 Incremental contribution over mNIF

Our proposed CHINR targets different goals from previous works, e.g. mNIF, Functa. We aim to achieve data generation controllability by modeling the conditional dependencies of model parameters, while others, such as mNIF, focus on improving generation quality. We are the first in INR's community to explore controllability by leveraging layer-wise hierarchical structure in parameter space. Moreover, our experimental analysis reveals that semantics information can be disentangled in parameter space, which hasn't ever been observed. Additionally, our work avoids trivial memorization and generates more diverse data than mNIF as presented in L357-L363 of the main paper and Figure 9 of the supplementary materials.

Overall, the main contributions of our work are demonstrated across four aspects: controllability, generalizability, versatility, and quality. Therefore, the quality improvement is only one of our contributions. Specifically, our main contributions are:

• Controllability: Our work distinguishes from previous methods in the controllability of generating fine-grain details, which has never been explored in the implicit neural representation (INR) community.

• Generalizability: Our work can avoid trivial memorization thus generating more diverse data compared with the SOTA methods, such as mNIF. Please find the discussion and visual samples from Line 357-363 in the main paper and Figure 9 in the supplementary material, respectively.

• Versatility: The CHINR's performance is validated across four different data domains, i.e. facial images, point clouds, NeRFs, and motions. This demonstrates the proposed hierarchical control and conditional dependency modeling is broadly applicable to data semantics with an inherent hierarchy structure.

• Quality: Our work outperforms the SOTA methods on both the reconstruction and generation metrics, which can be seen from Table 1 in the main paper.

Thanks for your constructive comments and insightful questions. We are happy to engage in further discussions. Please let us know if your have any other concerns.

Thank you for the detailed comments! Even if the quality improvement is incremental, the strength of the paper lies in controllability and generalizability. I think authors clarification during rebuttal makes this aspect way more clear! I am not an expert in this area and nor have I worked in this domain or even a related one. Therefore, I do not feel too qualified to provide any insights on the quality of this paper. I have decreased my confidence score to reflect this, and have notified ACs to seek an opinion from different reviewers. Best of luck!

We thank Reveiwer gfwF for the constructive comments.

W1.1 Blank cells in Table 1

In Table 1 of the main paper, for fair comparison, we adopt only the reported results of the existing methods. Moreover, the missing results (blank cells) come from:

1. Some existing methods cannot deal with all three datasets. For example, GASP cannot work on the NeRF data, such as SRN-Cars, because GASP requires the ground-truth radiance fields for training, which are unavailable.

2. Some existing methods are not evaluated on all the three datasets. For example, GEM is not evaluated on the NeRF data.

3. Some existing methods report only the qualitative results without the quantitative ones. For example, Functa only provides visual samples on Shapenet generation without quantitative results evaluated by the metrics.

Nevertheless, we understand your concern and address it by retraining these methods under identical settings. To measure the standard deviation, each model is retrained four times with different random seeds. The results for CelebA-HQ are reported in:

The results for ShapeNet are reported in:

The results for SRN-Cars are reported in:

W1.2 Marginal quality improvement

The main contributions of our work are demonstrated across four aspects: controllability, generalizability, versatility, and quality. Therefore, the quality improvement is only one of our contributions. Specifically, our main contributions are:

• Controllability: Our work distinguishes from previous methods in the controllability of generating fine-grain details, which has never been explored in the implicit neural representation (INR) community.

• Generalizability: Our work can avoid trivial memorization thus generating more diverse data compared with the SOTA methods, such as mNIF. Please find the discussion and visual samples from Line 357-363 in the main paper and Figure 9 in the supplementary material, respectively.

• Versatility: The CHINR's performance is validated across four different data domains, i.e. facial images, point clouds, NeRFs, and motions. This demonstrates the proposed hierarchical control and conditional dependency modeling is broadly applicable to data semantics with an inherent hierarchy structure.

• Quality: Our work outperforms the SOTA methods on both the reconstruction and generation metrics, which can be seen from Table 1 in the main paper.

Moreover, the analysis presented in Section 4.3 of the main paper further reveals how each layer of CHINR governs disentangled data semantics, which mathematically applies to other sinusoid-activated INR models. This property sheds new lights on the inductive bias of INR architecture.

[1] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one, ICML 2022.

[2] Yilun Du, Katie Collins, Josh Tenenbaum, and Vincent Sitzmann. Learning signal-agnostic manifolds of neural fields, NeurIPS 2021.

[3] Emilien Dupont, Yee Whye Teh, and Arnaud Doucet. Generative models as distributions of functions, AISTATS 2022.

[4] Tackgeun You, Mijeong Kim, Jungtaek Kim, and Bohyung Han. Generative neural fields by mixtures of neural implicit functions, NeurIPS 2024.

Thanks for your constructive comments and insightful questions. We are happy to engage in further discussions. Please let us know if your have any other concerns.

Thank you for the additional experiments and clarifications.

The additional benchmarks add some robustness to the method's performance relative to other models, and the table showing the hierarchical analysis adds strength to the claims of controllability.

Overall, the performance of the method is still only comparable to mNIF. Figure 9 does indeed show better generalization, but I assume it is curated without further explication in the caption or text below the caption.

Moving past the benchmarks, the approach to controllable generation and modeling hierarchical features does seem to be somewhat novel. While modeling hierarchical features itself is not new to the field [1] (thanks to reviewer wk7N for pointing out this paper) or [2] for example, the usage of Layer-of-Experts seems to be new. However, the diffusion-generated latents is not novel [3]. Therefore, we are left with looking to the Layer-of-Experts in combination with its effect on hierarchical generation.

In my opinion, the novelty and contribution does not outweigh the lack of clear improvement and/or benefit from CHINR's hierarchical approach. I will raise my score to a 5 for the authors' additional work in addressing my concerns.

References

[1] Park, Dogyun, et al. "DDMI: Domain-agnostic Latent Diffusion Models for Synthesizing High-Quality Implicit Neural Representations." The Twelfth International Conference on Learning Representations.

[2] Hao, Zekun, et al. "Implicit Neural Representations with Levels-of-Experts" NeurIPS, 2022

[3] Erkoç, Ziya, et al. "Hyperdiffusion: Generating implicit neural fields with weight-space diffusion" Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

Dear Reviewer gfwF,

As the deadline for author-reviewer discussions approaches, please don’t hesitate to reach out if you need any further clarification or have additional questions.

Best,

Authors

Thank you for your valuable feedback! We are happy that our responses addressed your concerns regarding CHINR’s robustness and controllability. Below, we clarify your remaining points.

Point 1

Overall, the performance of the method is still only comparable to mNIF.

As replied in our general response and responses to reviewers (“Response to Reviewer gfwF (1/4): W1.2,” “Response to Reviewer 1yWi (1/3),” and “Response to Reviewer LwTN (1/6): W1-Strengths”), our work outperforms other methods across four key aspects: controllability, generalizability, versatility, and quality. In particular, our novel controllable data generation has been recognized by Reviewer 1yWi, wk7N, LWtN, and ww4M.

In particular, CHINR is designed with distinct goals compared to prior works such as mNIF. While these methods primarily focus on improving quality metrics, CHINR aims to achieve controllability by modeling conditional dependencies in the parameter space. To our knowledge, we are the first to explore controllability by leveraging a layer-wise hierarchical structure in the INR community. This hierarchical control mechanism is particularly valuable in scenarios requiring fine-grained modifications, an aspect lacking in mNIF and similar methods.

Point 2

Figure 9 does indeed show better generalization, but I assume it is curated without further explication in the caption or text below the caption.

Thank you for recognizing CHINR’s generalizability. We want to clarify that Figure 9 is not curated but rather randomly generated and retrieved from the training set. The process for retrieval is described in detail in the text below the caption.

Point 3

While modeling hierarchical features itself is not new to the field [1] (thanks to reviewer wk7N for pointing out this paper) or [2] for example, the usage of Layer-of-Experts seems to be new.

We appreciate your acknowledgment of the novelty in our Layer-of-Experts. However, we would like to emphasize the key distinctions from [1] and [2]:

1. Our work does not model the hierarchical features but instead focuses on the hierarchy of INR parameters. This distinction highlights a new direction in exploring semantic hierarchy in parameter space.
2. Our work is new to the INR community in modeling the semantics hierarchy by capturing conditional dependencies across layer-wise parameters.
3. Different Goals:The works cited ([1] and [2]) aim primarily to improve quality, while CHINR focuses on controllable data generation.
4. Architectural Differences
   • [1]: This work uses layer-wise positional embeddings without modeling their conditional dependencies. In contrast, we model conditional dependencies through layer-wise modulation latents.
   • [2]: While its levels-of-experts overfit each data instance with spatially varying neural weights, CHINR generalizes to different data instances with combinations of experts in a spatially invariant way.

Point 4

However, the diffusion-generated latents is not novel [3].

We did not claim as a novelty to use diffusion-generated latents. We have already presented in background section (L133-L134) that the referred work [3] generates latents by a diffusion model.

Our novelty about diffusion is to design an HCDM that enables control over different semantic granularities. It is also a recognized strength in your previous comments ("Official Review of Submission5442 by Reviewer gfwF") and other reviewers' comments ("Official Review of Submission5442 by Reviewer 1yWi" and "Official Review of Submission5442 by Reviewer ww4M").

Point 5

In my opinion, the novelty and contribution does not outweigh the lack of clear improvement and/or benefit from CHINR's hierarchical approach.

We respectfully disagree and reiterate that our primary objective is to improve controllability in data generation, as reflected in our title.

Thanks again for your valuable feedback. We hope these clarifications address your remaining concerns. We are happy to engage in further discussions. If you find that all your concerns have been resolved, we would sincerely appreciate your consideration in raising your score.

References

[1] Park, Dogyun, et al. "DDMI: Domain-agnostic Latent Diffusion Models for Synthesizing High-Quality Implicit Neural Representations." The Twelfth International Conference on Learning Representations.

[2] Hao, Zekun, et al. "Implicit Neural Representations with Levels-of-Experts" NeurIPS, 2022

[3] Erkoç, Ziya, et al. "Hyperdiffusion: Generating implicit neural fields with weight-space diffusion" Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023