HyPlaneHead: Rethinking Tri-plane-like Representations in Full-Head Image Synthesis

Thanks for your insightful review! Here are our responses to your concerns. We will include the following content in the final version of the paper and the supplementary material.

W1: Inference Speed and VRAM Using

Due to space limitations, please refer to our reply to Reviewer Xp1z, Q1 & Q2

W2: Additional Quantitative Metrics

Following PanoHead, we compute the Inception Score (IS) for our HyPlaneHead model. The IS results of Table 1 (from No. 1 to No. 16) are as follows: 3.72, 3.86, 3.99, 4.06, 3.96, 3.61, 4.01, 3.76, 4.03, 3.97, 4.10, 4.17, 4.22, 4.18, 4.21, 4.21

W3: Diversity in Qualitative Results

Similar to previous works (e.g., EG3D, PanoHead, SphereHead), our training data includes widely used portrait datasets such as FFHQ, CelebA. These datasets are typically collected from public sources, where Caucasian individuals constitute the majority, while East Asian and South Asian individuals are present in smaller proportions. Other ethnic groups are underrepresented. As a result, the data distribution directly influences the randomly sampling results of the model. Therefore our method tends to generate predominantly Caucasian faces, with occasional East or South Asian characteristics, as seen in Supplementary Figure 3 (images 30 and 32).

However, this does not necessarily mean that our model is unable to handle underrepresented ethnicities or skin tones. Instead, it reflects the fact that the sampling distribution aligns with the distribution of the training data. By increasing the representation of diverse ethnicities and skin tones in the training set, we expect the model to generate a broader range of identities accordingly.

Moreover, when using GAN inversion techniques (e.g., PTI) on images from underrepresented groups, such as those of African descent, we consistently observe reasonable reconstructions of full-head 3D models. This is because our model has learned a rich and accurate distribution of head shapes and appearances through exposure to large-scale data, which provides a certain level of generalization. For out-of-distribution samples, targeted adaptation via inversion can further yield reasonable results.

Since we are unable to include these results here via figures or external links, we will add them to the final version of the manuscript.

W4: Discussion with Related Work

Dual Encoder [7] introduces a dual-encoder GAN inversion approach for single-view 3D full-head reconstruction. It uses one encoder for the visible front and another for the occluded back region, addressing issues like mirroring artifacts in PanoHead’s W space.

In contrast, our HyPlaneHead improves the representation structure to reduce such artifacts and enhance W space quality. As a result, standard inversion methods like PTI, which is a general-purpose technique for common GANs, already yield better performance than previous approaches. We believe that combining our method with specialized inversion strategies like Dual Encoder, which are specifically tailored for full-head generation, could further improve results. We will explore this possibility in the final version of the paper.

Tri2-plane [8] introduces a cascaded triplane representation across multiple scales of facial features, similar to a feature pyramid. This hierarchical design allows the model to generate both global and fine-grained details, leading to richer and more detailed head reconstructions. We believe that the idea of multi-scale feature pyramids could be beneficial to integrate with our hy-plane representation in future work.

Q1: Compatibility with Triplane-Localization for Editing

We have carefully read the paper, and find the proposed triplane-localization method has been successfully applied to both tri-plane (EG3D) and tri-grid (PanoHead) representations.

Given that our hy-plane shares a similar structure, we believe the method should be directly applicable to our representation as well. Moreover, we do not expect the combination of planar and spherical planes or the use of the unify-split strategy to interfere with the application of this localization approach.

Q2: Interpretability of Individual Planes

Since we are unable to include figures, we provide a qualitative explanation instead.

Firstly, when rendering using only a single planar or spherical plane, the resulting output is barely informative. This is because, for a planar plane, all rays passing through the same point on that plane share the same feature value, leading to visual artifacts such as parallel lines that reflect the shape of the feature map itself. Similarly, when visualizing only the spherical plane, the output consists of many radial lines extending from the center of the sphere, which do not convey meaningful 3D structure.

Instead, we can infer the contribution of each plane by removing one feature map at a time and observing the impact on the final rendering.

When the spherical plane is removed among the four feature planes, we observe significant missing regions in the facial area and back-of-the-head hair in the rendered result. This indicates that the spherical plane has learned to represent asymmetric features such as the front face and rear hair. However, structures on both sides, such as the shoulders and ears, remain mostly present, though with some degree of distortion.

When any one of the three planar planes (e.g., Pxy, Pyz, Pxz) is removed, significant distortions appear in the reconstructed head shape. This is because, in addition to carrying meaningful feature information, these planes also serve a role similar to "positional encodings" during training. If one of them is absent during inference, the positional representation becomes inconsistent with that used during training, resulting in structural degradation.

Among the three, the Pyz plane contributes the most to 3D structure, as its removal leads to a noticeable loss of symmetric features such as ears and shoulders. Removing the other two planar planes also introduces distortion, though to a lesser extent.

Q3: Discussion with Other Triplane-Based Methods (LRM, InstantMesh)

We believe that replacing the standard triplane representation in methods like LRM [5] and InstantMesh [6] with our hy-plane has the potential to improve their generation quality.

On one hand, our unify-split strategy eliminates inter-plane feature penetration, allowing each plane to express its features more clearly and effectively. On the other hand, by incorporating a spherical plane, we enhance the model’s ability to represent asymmetric regions, such as facial details and hair on the back of the head.

Therefore, we expect that integrating hy-plane into these models would lead to more accurate and artifact-free 3D reconstructions. Due to time constraints during the rebuttal phase and the lack of prepared data for such integration experiments, we have not yet conducted these evaluations. We will explore this direction experimentally in future work and include the results in the final version of the paper.

Q4: Single-View GAN Inversion Examples

Due to space limitations, please refer to our reply to Reviewer Xp1z, Q3.

Q5: Qualitative Comparison between 2+2 and 3+1 Configurations & FID vs. FID-random

Since we are unable to include additional figures in this response, we would like to describe our findings qualitatively here and will incorporate detailed visual comparisons into the final version of the manuscript.

No. 11, which does not use the unify-split strategy, still suffers from inter-channel feature penetration. This is visible as irregular horizontal patterns in the hair or color artifacts on clothing, which is actually because of features that leak across different channels. Nevertheless, the overall quality is already better than the baseline methods.

No. 11 and 12 also lack the near-equal-area warping, resulting in slightly blurred fine details, especially in the texture of the back-of-the-head hair.

The results of No. 13–16 show very small differences, which can only be observed through careful and extensive comparison. The main distinction lies in the orientation of the spherical plane. In No. 15 and 16, one of the spherical planes has its north pole aligned with the front of the face. Since the LAEA projection minimizes distortion around the pole, these configurations capture more facial details, leading to a slightly sharper facial appearance. Meanwhile, No. 14 and 16, which use an area-biased split, enhance the expressive power of the spherical plane, allowing for better reconstruction of asymmetric features, which is consistent with the quantitative results in Table 1.

Regarding FID vs. FID-random, as discussed in the paper, we observe that 3D-aware GANs tend to generate higher-quality results in regions aligned with the conditional camera view. For example, when the camera is conditioned to the left side of the head, the left side is well-generated, while the right side tends to be blurry or distorted. This is why we introduced the FID-random metric to evaluate how well the model generalizes across arbitrary viewpoints.

Across all 3D-aware GANs, it is common for FID-random to be higher than FID, and they reflect the same trend. However, we want to highlight an interesting anomaly in No. 2: although adding the unify-split strategy to tri-plane reduces FID, it actually increases FID-random. This occurs because the strategy eliminates inter-channel feature penetration and allows each plane to express its directional features more fully. However, since the tri-plane does not explicitly separate directional information, this enhanced expression leads to stronger mirroring artifacts on the backside of the head, thereby worsening the FID-random score.

Thanks for your insightful review! Here are our responses to your concerns.

W1 & Q2: Discussion and Comparison with Prior Works

Thank you for pointing out these related works. We agree that discussing them is essential to the completeness of our paper, and we will incorporate the following analysis into the final version.

Comparison with OrthoPlanes:

The main difference between our method and OrthoPlanes lies in their design choices. OrthoPlanes introduces additional parallel planes within tri-plane to enhance expressiveness, which is similar to the tri-grid. In contrast, our method achieves disentanglement and effective learning of symmetric and asymmetric features by combining planar planes with spherical planes.

As a result, OrthoPlanes still suffers from the following limitations, which are effectively addressed in our approach:

1. Incomplete resolution of mirroring artifacts: The root cause of mirroring artifacts lies in the use of Cartesian coordinate projection when querying features. As a result, both OrthoPlanes and tri-grid still suffer from mirroring-face artifacts and excessive left-right symmetry in certain regions.
2. Increased storage requirements: When saving a sample (e.g., for downstream tasks such as 3D model reconstruction or initialization for methods like Portrait3D [1], AnimPortrait3D [2], and ID-Sculpt [3]), they must store K times more feature planes (where K is the number of added parallel planes) leading to a significant increase in memory usage.
3. Incompatibility with unify-split strategy: Due to the increased number of planes, it becomes difficult to integrate these approaches with our unify-split strategy. As a result, inter-channel feature penetration remains an issue in these methods.

Comparison with SYM3D:

1. Goal Difference

   • SYM3D enhances the symmetry of tri-plane representations through symmetric regularization, which is beneficial for generating fully symmetric artificial objects.
   • Our hy-plane is designed to support a broader range of real-world scenarios where both symmetric and asymmetric structures coexist, such as in full-head portraits.

   Nevertheless, we believe that the symmetric regularization used in SYM3D could be beneficial if selectively applied to the planar planes in our method, especially the Pyz plane, which is responsible for learning symmetric facial features such as ears. We will explore this possibility in future work.

2. Different Strategies to Address Feature Penetration

   • SYM3D employs an attention-based scheme (View-wise Spatial Attention) to learn how to alleviate feature penetration across channels.
   • Our hy-plane, utilizes the unify-split strategy to geometrically and fundamentally prevent feature penetration at its source.

[1] Portrait3D: Text-Guided High-Quality 3D Portrait Generation Using Pyramid Representation and GANs Prior

[2] Text-based Animatable 3D Avatars with Morphable Model Alignment

[3] ID-Sculpt: ID-aware 3D Head Generation from Single In-the-wild Portrait Image

W2: Discussion and Comparison with SYM3D

Thank you for introducing this related work. This is a very insightful and meaningful comparison, as it supports our findings from a different perspective.

In fact, the correlation across xy, yz, and xz planes discussed in SYM3D essentially corresponds to the inter-channel feature penetration problem we identify in our paper, though they observe it from a different angle. We detect this issue through visual inspection of feature maps (as shown in Figure 1(a,b)), while SYM3D quantifies it using correlation metrics.

Inter-channel feature penetration causes similar values at the same UV positions across different feature maps, resulting in visually similar patterns, exactly what can be seen in our Figure 1(a,b). This phenomenon is also visible in the top-right panel of Figure 7 in SYM3D, where GET3D’s feature maps exhibit repetitive vertical lines at the same spatial locations when zoomed in. Numerically, this leads to high correlations between different feature maps.

The difference lies in terminology: SYM3D does not explicitly refer to this as inter-channel feature penetration, but rather describes it as correlation.

SYM3D attributes this issue to the use of single-view images during training, as opposed to multi-view data. While this is certainly true—more views provide richer geometric cues and reduce ambiguity—we also discuss this point in our paper (lines 65–71 and 224–227). From a more fundamental perspective, however, the issue arises due to the structural nature of convolutional networks, where different channels are prone to interfere with each other, especially in the absence of direct supervision. Our solution addresses this root cause directly by modifying the architecture via the unify-split strategy, which effectively eliminates inter-channel feature penetration at its source.

In contrast, SYM3D uses view-wise spatial attention to alleviate the issue. As shown in the bottom-right panel of Figure 7 in SYM3D, this approach does reduce correlation to some extent. However, as illustrated in the middle-bottom panel, strong correlations still exist between certain planes, particularly Pyz and Pxz. We have also computed and visualized similarity matrices similar to those in SYM3D's Figure 7, and our results show significantly lower feature correlations across different planes. Please refer to our response to W4 & Q1 below for specific quantitative results.

Moreover, in terms of implementation complexity, our unify-split strategy is simpler and more intuitive compared to SYM3D’s view-wise attention mechanism, while achieving a more complete resolution of the problem.

We will include this detailed comparison in the final version of the manuscript to further clarify how our method improves upon previous approaches.

W3 & Q1: View-Specific Disentanglement Visualization

Thank you for pointing out the need for clearer visualization of view-specific disentanglement. Due to the NeurIPS’25 rebuttal policy, we are unable to include additional figures or external links in this response. However, we would like to provide a qualitative explanation based on Figure 4 in the paper.

As designed in our hy-plane representation, the planar planes learn highly symmetric features, such as ears and shoulders, which are captured by the Pyz plane. On the other hand, the spherical plane captures asymmetric information such as the front face and back hair. This clear separation between symmetric and asymmetric features is a key aspect of the disentanglement achieved by our method.

We will add more detailed visualizations and explanations of view-specific feature learning in the final version of the manuscript to further clarify this point.

W4 & Q1: Missing Quantitative Evaluation for Feature Penetration

Thank you for this valuable suggestion. We agree that a quantitative evaluation is essential to validate the effectiveness of our method in reducing feature penetration across channels.

We recognize that the similarity matrix is a suitable metric for measuring inter-plane correlation and thus serves as an effective indicator of feature penetration. However, due to the NeurIPS’25 rebuttal policy, we are unable to include visualizations or figures in this response.

Instead, we show the similarity matrix below and we will add it in the final manuscript. The results demonstrate that hy-plane(3+1) with the unify-split strategy achieves significantly lower correlations between different planes compared to hy-plane(3+1) without the unify-split strategy, indicating that the unify-split strategy effectively reduces inter-channel feature penetration.

Similarity Matrix of hy-plane(3+1) with unify-split strategy:
[[ 1.          0.13848868  0.19367712 -0.24672825]
 [ 0.13848868  1.         -0.11731562  0.31794619]
 [ 0.19367712 -0.11731562  1.         -0.47516028]
 [-0.24672825  0.31794619 -0.47516028  1.        ]]

Similarity Matrix of hy-plane(3+1) without unify-split strategy:
[[1.         0.59808226 0.63007857 0.4739821 ]
 [0.59808226 1.         0.6070948  0.71127384]
 [0.63007857 0.6070948  1.         0.48041942]
 [0.4739821  0.71127384 0.48041942 1.        ]]

W5 & Q3: Comparison in Single-view Setting

Thank you for raising this important point.

We have conducted single-view reconstruction experiments using the same input image via PTI inversion. However, due to the NeurIPS’25 rebuttal policy, we are unable to include figures or external links in this response. Instead, we provide a qualitative discussion based on our results.

The results show that all 3D-aware GAN-based methods (EG3D, PanoHead, SphereHead, and our HyPlaneHead) are capable of reconstructing highly detailed front-view images that closely match the input, as PTI finetunes the model parameters to fit the target image. However, significant differences emerge in the quality of back-view reconstructions.

Specifically, EG3D and PanoHead often suffer from mirroring-face artifacts or lack sufficient expressiveness, leading to noticeable distortions in the back-of-the-head region. While SphereHead alleviates these issues, it tends to produce blurred and messy hair textures in the back view. In contrast, our method not only avoids such artifacts but also reconstructs clearer and more realistic details in the rear area.

We will include the full set of single-view GAN inversion results and their corresponding analysis in the final version of the manuscript.

Dear Authors,

Apologies that I could not respond earlier and thanks for the rebuttal. Yes, this rebuttal adressed most of my concerns and I am happy to keep my score.

Please include the additional results in the rebuttal, especially W4, and relation to the previous work I mentioned (W1-2) in the revised paper.

Best.

Thanks for your insightful review! Here are our responses to your concerns.

Q1 & Q2: Training/Inference Speed & VRAM Usage

To provide a comprehensive comparison of training and inference speed, as well as VRAM usage across different methods and HyPlaneHead configurations, we conducted measurements for each experiment listed in Table 1 of the paper. The results are as follows.

To provide a comprehensive and fair comparison, we clarify the definitions of the metrics used in our evaluation. Representation Parameters refer to the number of parameters in the feature maps of different tri-plane-like representations (e.g., tri-plane, spherical plane, hy-plane, etc.). Total Learnable Parameters denote the total number of trainable parameters in the entire model architecture, such as EG3D, PanoHead, SphereHead, and HyPlaneHead. Training Speed and Training VRAM Usage are measured on a single V100 GPU with a batch size of 2, representing the average time and memory consumption required to train 1,000 images. Similarly, Inference Speed and Inference VRAM Usage are evaluated under the same hardware setup but with a batch size of 1, by generating 100 images and computing the average time and memory cost per image.

From the statistics in the table, we can observe that different representations vary significantly in terms of feature plane parameter count. The tri-plane uses the fewest parameters (3 × 256 × 256 floating-point values), while the tri-grid with 512×512 resolution uses the most (9 × 512 × 512). Our proposed hy-plane uses 1 × 512 × 512 floating-point values, which is only 1.33 times the number used by the tri-plane. However, the overall difference in total learnable parameters across models is relatively small. The minor variations are mainly due to differences in the final convolutional layer configuration of StyleGAN2, which depends on how each representation is generated.

Notably, the total learnable parameters for entries No. 3 and No. 5 are identical. This is because in experiment No. 3, we did not modify the model code at all, which means only the rendering pipeline was adjusted to use one spherical tri-plane. We reported the actual parameter count to remain consistent with our experimental setup.

In terms of training and inference speed, as well as VRAM usage, there are some differences among the experiments. Overall, however, the additional computational overhead introduced by our innovations is relatively small.

Replacing the tri-plane with hy-plane introduces:

• +5.5% training time,
• +3.8% inference time,
• +8.8% training VRAM,
• −0.8% inference VRAM.

Adding the evenly split strategy further increases the cost by:

• +8.4% training time,
• +6.9% inference time,
• +12.2% training VRAM,
• +16.6% inference VRAM.

When combining with the near-equal-area projection, the overhead becomes:

• +0.5% training time,
• +3.5% inference time,
• −5% training VRAM,
• −5.7% inference VRAM.

Please note that in some cases, VRAM usage may actually decrease, likely due to memory fragmentation causing inaccuracies in the nvidia-smi measurement.

In summary, while our method does introduce some computational and memory overhead, the increase is relatively modest and justifiable given the significant improvements in disentanglement and reconstruction quality.

In addition, regarding Q2, which asks "Could this method be extended to higher-resolution image generation without a super-resolution module?" This is indeed a very insightful question. Current 3D-aware GANs typically rely on super-resolution modules to enhance the rendering resolution. However, it is possible that alternative strategies could achieve similar results. For example, Tri2-Plane [1] proposes a method of incorporating a feature pyramid into the tri-plane structure to improve fine-grained detail generation. Whether such an approach can fully replace the widely used super-resolution modules remains an open and interesting question. We consider this a promising direction for future work and will explore it in more depth in subsequent research.

[1] Tri2-Plane: Thinking Head Avatar via Feature Pyramid, 2024

Q3: Single-View GAN Inversion Results:

We have conducted single-view reconstruction experiments using the same input image via PTI inversion. However, due to the NeurIPS’25 rebuttal policy, we are unable to include figures or external links in this response. Instead, we provide a qualitative discussion based on our results.

The results show that all 3D-aware GAN-based methods (EG3D, PanoHead, SphereHead, and our HyPlaneHead) are capable of reconstructing highly detailed front-view images that closely match the input, as PTI finetunes the model parameters to fit the target image. However, significant differences emerge in the quality of back-view reconstructions.

Specifically, EG3D and PanoHead often suffer from mirroring-face artifacts or lack sufficient expressiveness, leading to noticeable distortions in the back-of-the-head region. While SphereHead alleviates these issues, it tends to produce blurred and messy hair textures in the back view. In contrast, our method not only avoids such artifacts but also reconstructs clearer and more realistic details in the rear area.

We will include the full set of single-view GAN inversion results and their corresponding analysis in the final version of the manuscript.

Thanks for your insightful review! Here are our responses to your concerns.

W1: Surface Comparisons

We apologize that, due to the NeurIPS’25 rebuttal policy, we are unable to include additional figures or external links in this response. However, we would like to provide a qualitative discussion based on the geometry visualizations presented in the paper and some results not included in the main submission.

All surface/shape/geometry comparisons mentioned below were extracted from tri-plane-like representations using the Marching Cubes algorithm and saved as .mrc files for visualization with ChimeraX software. This is consistent with the extraction methods used in EG3D, PanoHead, and SphereHead.

For the tri-plane representation, originally proposed in EG3D (as shown in Figure 5(a)), one can observe a face-like artifact appearing on the back of the head when viewing at 90°, 135°, and 180° angles. In the corresponding 3D surface, this manifests as an unnatural facial-shaped protrusion. This issue is commonly referred to as the "mirroring-face artifact," as discussed in the SphereHead paper.

For the tri-grid representation, introduced in PanoHead, the addition of parallel planes enhances representational capacity, which alleviates the mirroring problem to some extent. However, a similar face-like artifact still appears on the back of the head. From the extracted surface, this artifact is less pronounced and more flattened compared to the tri-plane case. Nevertheless, the issue is not fully resolved. Moreover, both tri-grid and tri-plane always suffer from high-frequency grid-like noise patterns, which are visible as horizontal and vertical stripes, in certain regions of the extracted surface, especially around the back of the head. This is caused by the use of Cartesian coordinate projection for feature querying.

For the single spherical tri-plane representation, as proposed in SphereHead (Figure 2(f)), the extracted surface appears smoother but lacks fine details. This is primarily due to the uneven expressiveness across different spatial locations and the underutilization of the square feature map of the spherical plane. Additionally, polar artifacts appear on left and right sides and seam artifacts are visible on the back of the head.

For the dual spherical tri-plane representation, also from SphereHead (Figure 5(e)), while polar artifacts and seam artifacts are disappeared, the extracted surface looks even smoother. This is largely because each spherical plane uses its weakly expressive equatorial region to cover the other’s strongly expressive poles, leading to suboptimal overall detail and quality.

In contrast, our hy-plane representation (Figure 1(d) and Figure 5(j)) achieves a surface that retains rich geometric detail while eliminating the grid-like noise patterns seen in tri-grid/tri-plane methods. It also completely resolves the mirroring-face artifact, just as the spherical tri-plane does. We further observe that the introduction of the unify-split strategy eliminates cross-channel feature penetration artifacts. For example, the abnormal horizontal streaks often seen in hair regions in tri-plane/tri-grid methods (caused by features from Pyz leaking into Pxy) no longer appear in our method.

Additionally, the integration of the near-equal-area warping strategy (LAEA projection + elliptical grid mapping) significantly improves the clarity of fine details in the extracted surface. This is due to the more efficient utilization of the square feature map for the spherical plane.

We will include detailed surface comparison figures in the final version of the manuscript.

W2: Lack of Clear Visual Labels in Qualitative Comparisons

We sincerely thank the reviewer for this valuable feedback. We agree that clearer visual labels would greatly improve the readability and comparison of our qualitative results. In the final version of the paper, we will revise all figures (including those in the supplementary material) to include explicit text labels next to each method’s output for better clarity and ease of comparison.

W3 & Q2: Novelty and Generalization to Beyond Full-Head Scenarios

We appreciate the reviewer’s thoughtful question regarding the novelty and general applicability of our method.

While the current work focuses on full-head generation due to data availability and practical relevance, we have also conducted preliminary experiments on other objects to evaluate the generalization potential of hy-plane.

Specifically, we collected multiple categories (e.g., cars, motorcycles, furniture, fruits, and houses) from public 3D model datasets (e.g. Objaverse, OmniObject3D and ShapeNet), and trained our model on each category. The preliminary results show that hy-plane achieves performance comparable to the standard tri-plane representation, particularly for symmetric objects such as cars and furniture. For asymmetric objects like fruits or houses, our method can produce more accurate asymmetric regions and generates fewer mirroring artifacts.

For example, when generating a house, although some houses exhibit symmetry in shape and window placement, they typically do not have two symmetric doors, i.e. one on each side. However, using a standard tri-plane tends to generate either two unrealistic "doors" or door-like artifacts on the opposite side. In contrast, our hy-plane naturally captures the asymmetry and typically generates only one realistic door, without any mirror-like artifacts on the other side.

Exploring the application of hy-plane to more general objects and scenes is an interesting and valuable direction. Due to time and data constraints during the rebuttal phase, we have not yet extended our experiments to more general scenarios such as object generation across arbitrary categories. We will include these additional experiments and discussions in the final version of the manuscript.

Q1: Resolution in Unify-Split Strategy

In our experiments, we used a larger unified feature map of size 512×512 so that after splitting, each individual feature map would have the same resolution (256×256) as those used in tri-plane, tri-grid, and spherical tri-plane methods. Although this increases the model’s parameter count to some extent, our experiments show that the performance improvement is not due to the increased number of parameters.

Due to space limitations, please refer to the Representation Parameters and Total Learnable Parameters columns in the table provided in our response to Reviewer Xp1z (Q1 & Q2) for detailed information.

As shown in Table 1 (rows 8, 9, and 10), even when the feature maps of tri-plane, tri-grid, and spherical tri-plane are also upscaled to 512×512, their performance does not improve compared to the 256×256 version. This indicates that the performance gain in our method is attributed to the novel representation design rather than an increase in model capacity. Besides, the same observation can be found in Table 2 of OrthoPlanes [1]. The first and second rows compare the same EG3D-based setup with only one difference: the former uses a 256×256 feature map, while the latter uses a 512×512 feature map. However, the FID score of the latter is actually higher than that of the former, demonstrating that simply increasing the feature map size does not necessarily lead to better performance.

Furthermore, although we use a larger unified feature map (512×512), it has little impact on inference speed or GPU memory usage. Therefore, we believe the comparison between different methods remains fair and meaningful.

[1] OrthoPlanes: A Novel Representation for Better 3D-Awareness of GANs