acceptodds
Under review as a conference paper at ICLR 2027

Hexelate: Compact Surface Primitives for Efficient Scene Rendering and Reconstruction

10 views

Abstract

Novel view synthesis and 3D reconstruction have advanced rapidly with radiance fields and splatting, yet their memory and computational demands limit their use in resource-constrained settings. We introduce Hexels, planar hexagonal surface elements with a first-order spatial color field for efficient 3D scene and 2D image representation. Hexels exploit locally planar regions with smoothly varying appearance using fewer, larger primitives. Their hexagonal support provides an area and boundary-efficient domain, while the local color field captures appearance variation within each primitive.Across 34 scenes, compact fixed-population Hexels improve PSNR over memory-matched 2DGS by 4.53–7.38 dB across dataset subsets. With densification and view-dependent appearance, Hexels improve PSNR over unconstrained 2DGS by 0.89 dB on DTU and 0.88 dB on Tanks and Temples. On Mip-NeRF360, they are 0.33 dB below 2DGS while rendering at its speed with 83.9% less deployment storage. For all-view DTU surface reconstruction, fixed-population Hexels achieve 0.978 mm Chamfer distance, reducing error by 28.9% relative to memory-matched 2DGS, although unconstrained 2DGS remains more accurate. The same primitive also improves 2D image representation over Image-GS.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Efficiency, Edge Devices, Differentiable rendering, neural rendering, image fitting, novel-view synthesis, Gaussian splatting, surface splatting, bounded primitives, local radiance fields, compact scene representations, rate–distortion, efficient rendering.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_acqEOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary
  • Scenes and images are commonly represented by Gaussian or triangle primitives. These primitives are independent, require many parameters per primitive, and can be inefficiently arranged with overlap, making them poorly suited to memory-constrained settings.
  • The paper proposes hexels that use fewer parameters per primitive and can fully tile a plane without overlap. Motivated by the smooth variation of most textures, each hexel represents appearance as a linear variation within its hexagonal region.
  • The method achieves high image fidelity under limited memory and primitive counts, and demonstrates efficiency in FPS, storage, and training VRAM.
Critical strengths
  • The paper proves that, within the same local color-error ellipse, affine hexagons cover more area than other tileable primitives, mathematically demonstrating their efficiency under this constraint.
  • When fewer primitives are used, each primitive needs to represent texture over a larger region. Linearly varying appearance within each hexel helps reduce the error over these larger regions.
  • The paper proposes variants with different levels of appearance modeling, including spherical harmonics, showing that memory efficiency can be combined with high appearance fidelity.
Critical weaknesses
  1. The paper relies heavily on empirical comparisons throughout. Even the abstract largely lists numerical improvements and differences. This makes it difficult to read and does not adequately explain why the method is superior. A paper should explain its objective, how the method is constructed, which problems in existing methods it addresses, and why its design produces the reported improvements. Simply claiming that it is better because its numbers are higher is not sufficiently convincing. Referring to the models as V1, V2, and V3 makes them difficult to distinguish. The settings should have descriptive labels, be evaluated separately, or be presented as one main model with the others treated as ablations. Section 3.1 has no corresponding Section 3.2 and discusses experimental results despite not being part of the experiments section. T&T should be introduced explicitly as Tanks and Temples, even if researchers in the area recognize the abbreviation. Some equations also lack numbers. I would normally regard writing issues as minor camera ready corrections that do not affect the score, but in this paper I think most sections need to be rewritten.

  2. Hexels are interesting, but I question why hexagons are the right choice. The paper and appendix establish that hexagons are efficient planar patches within a common local color-error ellipse. In Gaussian and triangle splatting, primitive sizes vary substantially. Although this variation can require overlap, using larger primitives that partially overlap may be more efficient. With that freedom, other primitives such as triangles may represent scenes more efficiently, while their compatibility with triangle meshes offers additional advantages, as in Triangle Splatting and MeshSplatting. Squares or rectangles may also be more natural for the many rectangular and right-angled structures in real scenes. An argument about filling a plane is not convincing enough when actual representations involve varying primitive sizes and scene boundaries.

Questions for authors
  1. Apart from the use of hexagons, I do not see the fundamental difference from existing methods, particularly Triangle Splatting. Smooth representations for differentiable rendering in Triangle Splatting and local 2D coordinate systems, as in 2DGS, have already been used. The weighted-sum rendering used in the paper's 2D image branch also has a precedent in Image-GS. With the reason for using hexagons itself insufficiently convincing, the technical novelty is unclear to me.

  2. Line 58 treats overlap between other primitives as a disadvantage, but it is unclear why overlap is inherently undesirable. As discussed above, overlap can be more efficient. More importantly, this method also uses a smooth representation for differentiable rendering and combines overlapping primitives. This seems contradictory. Although Gaussian splats cannot tile a plane, Triangle Splatting also uses a smooth representation that produces overlap, and I do not see why Hexels would be fundamentally different in this respect.

  3. The tables use too many colors, making them difficult to distinguish in some viewing conditions or in grayscale. Highlighting all first-, second-, and third-place results, including ties, within each model group can leave almost every cell in a column colored. Beyond readability, the methods address the same task, and Hexels do not appear to offer a distinctive post-training advantage unavailable to the others, for example, an exclusive ability to use the learned representation as a mesh (just an example). I therefore do not see why the groups need separate state-of-the-art highlighting.

  4. Spherical harmonics account for much of the storage and per-primitive parameter count in methods such as 3DGS and 2DGS. Comparing V1, which represents only view-independent color, against the original 3DGS under the same memory limit is therefore unfair (Table 2, Figure 1). Although the comparison is intended to address resource-constrained settings, the baselines are not configured for that purpose. If the comparison uses ordinary NVS models, their memory limits should be relaxed; alternatively, it should include methods designed to reduce memory, such as NeuralGS [1].

The following are minor editorial suggestions for the camera-ready version and do not affect my review score:

  • The paper title and model name differ; this should be clarified.
  • The citation for 3D Convex Splatting needs to be corrected.
  • Correct missing spaces such as “primitive.Across” at line 36.

[1] Tang et al. NeuralGS: Bridging Neural Fields and 3D Gaussian Splatting for Compact 3D Representations

18 comments

W1. it is better because its numbers are higher

The paper already explains its objective and design rationale; we have further clarified these in the revised abstract, introduction, and related work. Our objective is to design a surface aligned general (2D and 3D) primitive which is efficient across the board (memory, VRAM, rendering speed) without sacrificing quality (NVS and surface reconstruction) especially when under constraints which come with edge device deployment. Under matched budgets, Hexels substantially outperforms the compared native surface-aligned primitives, as detailed below.

In Section 1 itself, we introduce our motivation (mentioned later in the paper as well) - We are motivated by the idea that natural images and scenes are highly structured: large regions are often locally planar with smoothly varying appearance. An efficient surface primitive should therefore be able to exploit this underlying structure, wherever possible, and do so in a compact and self-contained manner.(§1, lines 70–87)

We then explain how we go about solving this problem by mathematical analysis. To efficiently cover a plane, we need a tile-able shape which covers the plane with least overdraw and something which can represent the locally varying colors. None of these problems are trivial to solve. Appendix I.1 derives local approximation-error bounds for affine color. Appendix I.2 shows that, within the same local color-error ellipse, the largest affine hexagon covers twice the area of the largest affine triangle. Both are discussed and referenced in the main draft. In fact, both the coverage argument and the benefit of spatial color over larger regions are recognized in your Critical Strengths. We provide explicit local color-approximation and patch-area bounds supporting the joint design; the comparison with triangles is therefore mathematical as well as empirical.

Hexels are designed using such ideas - Section 3 gives the construction in Eqs. (1)-(4). Section 2 and Appendix J distinguish this objective from moving appearance into separate textures or neural fields, or relying on post-training compression. Although the matched-primitive experiments ablates the Hexel design as a whole per primitives compared to other surface aligned primitives, we have added Section 4.4 as a controlled planar shape/color ablation for further support our methodology. In fact, we even have a supplementary video, which explains the ideas briefly and even visualizes Hexel optimization.

The empirical comparisons test whether this theoretically motivated design delivers the intended quality–efficiency benefits. We evaluate 34 real-world NVS scenes-to our knowledge, the largest scene-count evaluation among all the compared methods (§4). At matched model memory, Hexels improves PSNR by 4.53-7.38 dB over 2DGS and 1.70-2.90 dB over Triangle Splatting, with lower training VRAM than both across all three benchmarks (Table 2). Together with the matched-primitive results (Table 3), these substantial margins support our state-of-the-art claim among the compared native surface-aligned primitives under matched budgets. These quality and efficiency margins support our hypothesis that exploiting local geometry and appearance improves primitive efficiency. Our justification combines design intuition, mathematical bounds, controlled ablations, and broad empirical evaluation.

The contributions also extend beyond benchmark scores. Hexels directly expose differentiable surface cells retaining geometry and spatial appearance, with O(1) work per primitive (§4.3; Appendix H.1). We will release a fully open-source iPhone/iPad app for the complete on-device workflow from images/videos through training to surface reconstruction and NVS; to our knowledge, this would be the first such release for a native surface-aligned primitive (Appendix H.2).

V1-V3 are already defined and evaluated separately: V1: spatial RGB; V2: spatial RGB + linear view dependence; V3: spatial RGB + SH (§3, lines 190–198; Appendix A.1; Tables 3 and 14–15). Descriptive labels can be added alongside this shorthand, if it helps reviewer keep better track. We chose these shorthands to reduce cognitive load.

Section 3.1 deliberately keeps the compact 2D formulation and evaluation together. Tanks and Temples is already spelled out in the abstract and introduction; we can explicitly pair it with T&T. Core Eqs. (1)-(4) are numbered; short expressions not subsequently cross-referenced were left unnumbered. These presentation details can be clarified without rewriting most sections.

Reviewer_acqEOpenReview ↗

Official Comment for W1

To summarize my response: W1 is a writing issue. My main concern is the insufficient explanation of the background and limitations of existing work, together with the lack of in-depth analysis of the results. The addition of the ablation in Figure 6 offers some improvement, but I still believe that the writing and structure of much of the manuscript need improvement.

The response appears to misunderstand the point of W1: this weakness concerns how the paper is written. For example, the abstract is the first part readers encounter. It should both persuade them to read further and make the overall work easy to understand. It should explain the research background, what problem remains, why it is a problem and needs to be addressed, what idea is proposed to address it, what results follow, and why that idea leads to improvements. The method and its results are reasonably apparent in the current abstract, but the other parts of this explanation are insufficient.

More specifically, the first sentence (lines 30–32) concludes the background discussion without adequately explaining what has prevented existing methods from addressing the problem or why resolving it matters. Lines 36–46 then largely describe results as better than the baselines or state numerical differences directly. They provide little deeper explanation of why these results were obtained or where the improvements would be particularly useful.

Listing all these numerical differences individually may be meaningful to the authors, but to readers, each becomes just another positive number in a long list. This takes up space and makes the paper harder to read. Readers could obtain these differences simply by looking at the tables without reading the surrounding text; the prose adds little information beyond what the tables already provide. What readers need is a deeper analysis that explains the results and supports their interpretation: where existing models fail, why they fail, how the proposed model improves on them, and which particular improvements lead to the reported outcomes. Reporting higher numbers demonstrates a performance difference, but does not by itself establish that explanation.

The evaluation on 34 real-world scenes can certainly be a strength. However, presenting a broad set of experimental results without sufficiently deep analysis is not enough. For example, the paper could explain that a particular dataset contains many surfaces with smoothly varying colors, examine whether an existing method struggles to represent those variations efficiently because each primitive has its own spatially constant color, and show how that limitation manifests in the reconstruction. It should then explain which aspect of Hexels addresses that limitation and provide targeted evidence that this accounts for the observed difference. This is the kind of analysis I am asking for, rather than simply stating that one method achieves a higher score than another.

I have seen the newly added ablation in Figure 6, and it is an improvement in this direction. Nevertheless, the ablation should compare against a broader range of methods and provide a more fine-grained explanation of the differences among them, with those observations connected to the analysis of the preceding experiments.

W1 did not identify the mathematical proof as a weakness. I included it in the Critical Strengths because the paper makes an effort to provide a mathematical basis for its design. That does not remove the concern raised in W2 about the scope of that justification. The paper uses tiling as a motivation and treats overlap in existing methods as a source of inefficiency. Yet, as also illustrated in the newly added Figure 5, the Hexels vary substantially in size and overlap with one another.

The result under a common local color-error ellipse addresses a restricted comparison of individual patches. It does not establish that hexagons are more efficient than other shapes in the actual representation, where primitive sizes, positions, and overlaps can vary and be jointly optimized. Thus, while I recognize the mathematical contribution under its stated conditions, I remain strongly unconvinced that it sufficiently justifies the advantage of hexagons over other structures in this setting.

AuthorsOpenReview ↗

Response to Review (Part 2)

W2. Why hexagons?

Our objective is to exploit the structure of natural scenes to improve efficiency at the primitive itself: locally planar geometry and smoothly varying appearance (§1, lines 70–87). Hexagonal support and first-order color jointly let one primitive represent a larger coherent region, while smaller Hexels capture boundaries and detail. This is the motivation explained in the introduction’s “Why a hexagon?” paragraph and developed mathematically in Appendix I.

The size and overlap freedom described in the review is already present in our comparisons. Each Hexel independently learns its position, orientation, and two tangent scales (§3, Eq. (1)); we impose neither equal-sized cells nor a fixed tessellation. Gaussian and triangle baselines also retain their native freedom to resize and overlap, and their growth/pruning policies are retained in the main matched-primitive experiment subject to the final population budget (Appendices C and E.3). Hexels nevertheless achieve better rendering quality and lower training VRAM among the compared native surface-aligned primitives under these budgets (Tables 2-3).

The mathematical argument also goes beyond plane filling. The honeycomb theorem motivates perimeter efficiency for equal-area planar partitions. Our own analysis connects appearance error to the region represented by one primitive: within the same local color-error ellipse, the largest affine hexagon covers twice the area of the largest affine triangle and about 1.299× that of the largest affine square (Appendix I.1–I.2). Translation, rotation, and anisotropic scaling are allowed and the affine-square comparison includes rectangles. The presence of rectangular structures alone therefore does not establish a better overall quality-efficiency trade-off for rectangular primitives.

We further ablate Hexels in two additional experiments. Section 4.4 crosses Gaussian/hexagonal support with constant/spatial RGB. Fixing normals and depth order approximates a local planar surface patch without coupled changes in reconstructed geometry and visibility. On the shaded target, gradient Hexels reach 37.07 dB versus 35.13 dB for gradient Gaussians, with 2.48 versus 10.47 blends per pixel. This causally ablates Hexels.

For 3D, we equip 2DGS and Triangle Splatting with the same V1 spatial color model and evaluate all nine Mip-NeRF360 scenes with exactly the same primitive count, no densification or pruning, and the same training-step budget and main hyperparameters. All three models use approximately 9 MB storage memory, retaining their native geometric parameterizations and renderers.

MethodPeak VRAM (GiB) ↓PSNR ↑SSIM ↑LPIPS ↓
2DGS + V1 color4.9523.730.66850.4521
Triangle Splatting + V1 color0.918.700.45960.6253
Hexels V10.924.840.70530.4104

Hexels achieve the best result on all three quality metrics, improving PSNR by 1.11 dB over 2DGS and 6.14 dB over Triangle Splatting. Their advantage therefore persists when both spatial appearance and primitive count are held fixed.

Triangle Splatting produces independent triangles; MeshSplatting adds connectivity, and both are already included in our comparisons (Appendix J). Hexels also directly export independent, differentiable surface cells retaining geometry and appearance (§4.3; Appendix H.1). Our practical objective is demonstrated by the complete on-device training and reconstruction/NVS workflow, which we will release as an open-source iPhone/iPad app (Appendix H.2). The case for Hexels therefore combines a structural motivation, mathematical bounds, controlled experiments, strong budgeted results, and a practical deployment-not simply an argument about tiling a plane.


Q1. Technical differences from Triangle Splatting

Hexels have a different primitive parameterization, support function, and spatial appearance. The contribution is a compact native surface primitive that jointly exploits local geometry and appearance to improve the quality-efficiency trade-off (§3, Eqs. (1)- (4); Appendix I).

Compared with the original Triangle Splatting formulation:

  • Geometry: Hexels optimize a center, orientation, and two tangent scales defining an affine canonical hexagon, rather than three independent triangle vertices.
  • Support and sharpness: Hexels evaluate a sigmoid boundary in perspective-correct local plane coordinates. Triangle Splatting uses a ReLU/power window constructed from projected triangle-edge distances. Hexel β controls the sigmoid boundary transition; Triangle’s learned exponent controls its triangular falloff (§3, Eqs. (2)–(3); Appendix B).
  • Spatial appearance: Every Hexel variant contains the field c₀ + u gᵤ + v gᵥ, optionally augmented with view dependence. Original Triangle Splatting uses per-triangle SH appearance without this intrinsic spatial field (§3, Eq. (4); Appendix A.1).
Reviewer_qw55OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper introduces Hexelate, a compact surface representation for efficient novel-view synthesis and 3D reconstruction. The proposed Hexel primitive combines a planar hexagonal support with a first-order spatial color field, allowing each primitive to represent smooth appearance variations over a larger surface region. The method extends 2DGS-style rendering with differentiable hexagonal boundaries and spatially varying colors, while also supporting view-dependent appearance. Experiments on Mip-NeRF360, DTU, and Tanks and Temples demonstrate competitive rendering quality and efficiency, particularly under limited memory and primitive budgets. The authors also evaluate the representation for surface reconstruction and 2D image fitting.

Critical strengths
  1. The method is interesting and reasonably novel. Combining hexagonal primitives with a first-order spatial color field is a simple and elegant design that achieves competitive rendering quality and efficiency without introducing additional textures or neural decoders.
  2. The demo video is well made, with clear explanations and intuitive visualizations that effectively demonstrate the method and its results. The experiments also cover multiple datasets and provide extensive comparisons in terms of rendering quality and efficiency.
Critical weaknesses
  1. The contribution of the hexagonal primitive itself is not sufficiently isolated. The paper introduces both hexagonal support and a first-order spatial color field, but lacks a controlled 3D ablation that changes only the primitive shape. It is therefore difficult to determine which component contributes most to the observed improvements.
  2. There are several inconsistencies between the reported results and the corresponding claims. Table 3 claims that Hexels outperform all baselines in rendering quality, although BBSplat achieves better results. Similarly, the claim about having the lowest training VRAM in Table 2 is not fully supported by the data. The limitations section also states that the effects of densification remain to be evaluated, despite relevant experiments already being reported in the paper and appendix.
  3. The surface reconstruction results and evaluation protocol need clarification. The DTU Chamfer Distance reported in the main text differs from Table 5, while the appendix presents additional results under different configurations and scene subsets. Furthermore, although the paper emphasizes direct polygonal surface extraction, the main geometric evaluation relies on TSDF-based extraction rather than evaluating the directly exported Hexel surfaces.
  4. Some efficiency comparisons use different evaluation conditions. The paper emphasizes that the representation does not require quantization or compression, yet the reported V3 deployment sizes use Product Quantization, while the baselines retain their respective storage formats. In addition, the matched-primitive experiments use a shared upper bound on primitive counts without clearly reporting the actual final counts for each method.
Questions for authors
  1. Could the authors provide a controlled comparison of Gaussian, triangular, and hexagonal primitives using the same first-order spatial color field? This would help clarify the contribution of the hexagonal shape itself.
  2. Could the authors clarify the differences in DTU Chamfer Distance between the main text, Table 5, and the appendix? It would also be helpful to report the geometric accuracy of the directly exported Hexel surfaces.
  3. Could the authors report the uncompressed model sizes and actual final primitive counts for each method, along with a more consistent comparison of deployment storage under comparable compression settings?
5 comments
AuthorsOpenReview ↗

Response to review (part 1)

W1/Q1. Separating hexagonal support from spatial color

While the matched primitive experiment does ablate Hexels design as a whole per primitive compared to other surfels, we add two more ablative experiments to further ablate shape and appearance. First, we causally ablate both the hexagon shape and linear color in §4.4 (lines 468–483) where Gaussian/hexagonal support is crossed with constant/first-order RGB. The planar setting allows for confounder-free experiment: and approximates a planar patch of a 3D scene without dealing with confounding effects of 3D (geometry, visibility, surface normals, and depth ordering). This lets us isolate the effects of support shape and color representation in this controlled setting.

We find that Spatial color improves both Gaussians and Hexels, and Hexels outperform Gaussians with either constant or spatial color. On the shaded target, gradient Hexels reach 37.07 dB versus 35.13 dB for gradient Gaussians, with 2.48 versus 10.47 blends per pixel (§4.4, Fig. 6). This causally ablates our primitive design - both shape and representation.

Second, we provide the requested three-primitive 3D comparison by equipping 2DGS and Triangle Splatting with the same V1 spatial color model used by Hexels (§3, Eq. (4)). We evaluate all nine Mip-NeRF360 scenes with exactly the same number of primitives, no densification or pruning, and the same training-step budget and main training hyperparameters. All three models use approximately 9 MB. This controls appearance and population while retaining each method’s geometric parameterization and renderer.

MethodModel MB ↓Peak VRAM GiB ↓PSNR ↑SSIM ↑LPIPS ↓
2DGS + V1 color~94.9523.730.66850.4521
Triangle Splatting + V1 color~90.918.700.45960.6253
Hexels V1~90.924.840.70530.4104

Hexels achieve the best result on all three quality metrics, improving PSNR by 1.11 dB over 2DGS and 6.14 dB over Triangle Splatting, therefor ablating the hexagonal shape.


W2. Performance claims

Our state-of-the-art claim concerns the quality–efficiency trade-off among the compared native, surface-aligned primitives under matched-memory and matched-primitive budgets. This comparison class is discussed repeatedly in §2, §4, and Appendix J:

  • §4 identifies 2DGS, Triangle Splatting, and MeshSplatting as the closest comparisons because “these methods use explicit surface-aligned primitives without separate texture maps or neural appearance fields” (lines 304–307).
  • Appendix J explains that “3D Gaussians are not naturally surface aligned” (lines 1738-1744), distinguishing volumetric primitives from our comparison class.
  • For BBSplat, Appendix J describes its separate RGB and alpha textures and states: “These hybrid setups, however, are not the objective of our work” (lines 1788–1811). Section 2 also discusses this distinction (lines 142–145). Hybrid methods separate geometry and appearance (with significant efficiency trade-offs) and can utilize different surface primitives under the same setup. They form a different area of research and are not the focus of this work.

Volumetric and hybrid methods are included as additional cross-family comparisons. The tables already separated these families using distinct ranking colors; we have now added explicit family labels and clarified the captions.

Within our primary comparison class of surface aligned primitives, Hexels achieve the best PSNR, SSIM, and LPIPS across all three benchmarks under matched primitive budgets (Table 3), and the lowest training VRAM at matched model memory (Table 2). At matched model memory, V1 also achieves the highest PSNR and SSIM across all reported representation families.

The cross-family results further support the efficiency claim. In Table 3, every fixed Hexel variant has lower training time, lower peak training VRAM, smaller model storage, and higher FPS than both BBSplat and Nexels on all three benchmarks. For example, compared with BBSplat, V2 uses 3.36-19.53× less training VRAM and 6.08-6.71× less model storage, while rendering 6.74-11.00× faster. BBSplat’s higher quality in that setting therefore comes with substantially greater resource costs; it does not establish a better overall quality-efficiency trade-off, which is the aim of our work. We have revised the relevant captions and claims to make their scope explicit.

We apologize for the incomplete densification sentence; it has been corrected. Fixed and densified configurations are already evaluated in Tables 14-15 (Appendix K.2), with additional surface-reconstruction controls in Table 8(b) (Appendix G).

AuthorsOpenReview ↗

Response to review (part 2)

W3/Q2. DTU results and direct surface-cell evaluation

We have corrected the main-text CD to match Table 5: 0.978 mm over all 15 DTU scans (§4.2, lines 417-431). The appendix previously reported ablations on six scans (vs 15 in main text) and therefore was different; it now reports all 15 (Appendix G, Table 8).

We use TSDF for the main comparison to keep the surface-extraction protocol consistent with other methods for a direct comparison (Appendix E.5). Direct-cell export provides a complementary representation of the learned (likely unique to) Hexels: independent, differentiable surface elements retaining geometry and appearance for downstream optimization, rather than a connected mesh (§4.3; Appendix H.1).

We also evaluated the directly exported Hexel cells without TSDF fusion on the same 15 DTU scenes.

MethodMean CD (mm) ↓Training VRAM (GiB) ↓
2DGS matched memory1.3751.02
2DGS matched primitives1.4490.98
Triangle Splatting1.1242.29
MeshSplatting1.6622.65
Hexels0.9780.61
Hexels, direct cells1.1380.61

Hexels achieves the lowest CD and training VRAM among these surface baselines, supporting our state-of-the-art claim for efficient surface reconstruction. Without TSDF fusion, direct cells still outperform both budget-matched 2DGS settings and MeshSplatting in CD while using lower training VRAM.


W4/Q3. Uncompressed storage and primitive budgets

We have corrected the earlier storage reporting. All main Hexel NVS comparisons use uncompressed FP32 inference-model sizes, without quantization or compression (Appendix E.4, lines 1230-1236). Uncompressed references were already provided in the appendix earlier; optional lossless compression and PQ1024, including their quality changes, are reported separately in Appendix K.3, Table 16. Reported model memory includes geometry and all learned appearance components, including textures, feature grids, and decoders (Appendix E.3).

The matched-primitive budget equals each scene’s SfM point count (§4.1, lines 365–377). Fixed Hexels retain this population without densification or pruning (§4, lines 262–265). Baselines retain their original densification/growth and pruning and may temporarily exceed the budget during training, but their final populations must remain within it (Appendix C, lines 1010-1019; Appendix E.3, lines 1201-1206). We allow this flexibility because strictly enforcing the budgets throughout training can substantially degrade baseline methods performance (Appendix C). Even without densification, Hexels achieve the best PSNR, SSIM, and LPIPS among the compared native surface-aligned primitives under matched-primitive budgets (Table 3).

Thank you for the detailed rebuttal and additional experiments. The new evidence addresses several of my concerns, and I have therefore updated my Soundness score from 1 to 2.

I have two remaining questions:

W1/Q1: The new 3D ablation is helpful, but the methods still use different geometric parameterizations and renderers. Could the authors clarify how much of the improvement can be attributed specifically to the hexagonal shape?

W4/Q3: Thank you for clarifying the shared primitive budget. Could the authors also report the actual final primitive counts for each baseline, as requested in my original review?

I appreciate the authors' efforts and look forward to these clarifications.

Reviewer_Nkw8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a new primitive for novel-view synthesis, the Hexel, a planar hexagonal surface combined with first-order spatial color to increase color expressiveness. Each primitive is defined by its center (3 values), two scale values, a quaternion for its orientation (4 values), an opacity, and a sharpness parameter. Those 11 parameters form the geometric representation of the primitive. The authors define 3 appearance variants: V1 uses a base RGB value and two RGB spatial gradients; V2 adds 9 view-dependent coefficients on top of V1, and V3 replaces the view-dependent coefficients by the classical spherical harmonics.

The paper is benchmarked on classical NVS datasets.

Critical strengths

The motivation to replace current primitives by the Hexel for both memory and efficiency constraints is sound.

The hexagonal support follows a mix between Convex splatting (a sigmoid support function with a sharpness parameter) combined with the "max" evaluation of Triangle splatting (instead of the LogSumExp).

In the constrained setups, Hexels show strong results against state-of-the-art methods.

Critical weaknesses

[W1] The paper structure is difficult to follow, as some key information is present only in the Appendix. For me, parts of Appendix A (Hexel parameterization) and B (rasterization) deserve to be in the core of the paper as they present key information. On the other hand, I believe that Section 3.1 (Representing 2D Images as Hexels) is not directly relevant for the scope of this paper, which for me targets novel-view synthesis, as a 3D problem. This ``toy example" could be placed in the appendix, making some space for the parts listed above.

[W2] I do not understand what the motivation is to use Hexels (instead of 2D convexes or triangle) for native novel-view synthesis, as it shows worse performance on most benchmarks and scores. What is the "raw" advantage of the Hexel compared to the Triangle, since one of its advantages is that it can be directly optimized as a mesh (MeshSplatting). I would find it interesting to have the results for the ``mesh-based novel view synthesis" against MiLo and MeshSplatting.

[w3] There are no discussions in the main paper about the influence of the first-order spatial color compared to a unique color per primitive in the 3D setup. I believe this discussion is important, since this feature could be applied to other planar primitives.

Misc. The following remarks are not critical but worth mentioning:

  • There is a typo in your Convex splatting citation, as it appears as (?). Same issues for other references in the Appendix.
  • It is not clear to me what the “V1 memory budget" means. How does it affect the other methods.
  • In Table 4, as well as in some Figures, you used a “*" for Convex and Triangle Splatting, but its signification is not defined.
  • The notations used in Section 3.1 do not match the ones in Equation 19 in the Appendix.
Questions for authors

[Q1] Can you provide some quantitative results for a V4 version, which would use c0\mathbf{c}_0 + SH (the version with the first-order spatial color)? I believe that this is a mandatory baseline that can be used to assess the influence of a hexagonal shape compared to a triangle one.

[Q2] From a practical perspective, and without considering the different memory and efficiency constraints, what is the advantage of the Hexel compared to the Triangle? As an independent future work, what could be the advantage of a mesh-based Hexel compared to MeshSplatting?

[Q3] One key change between Convex Splatting and Triangle Splatting is the change of the support function, which changed from a sigmoid one (which therefore defines an influence outside the convex shape) towards a ReLU one (which is only defined inside the shape). Could you please elaborate why you made this choice, as it needs to define a cut-off threshold as well as a larger number of pixels to evaluate for rasterization? What is the average value of β\beta?

5 comments

W1. Main-text formulation and the role of 2D representation We agree that the core formulation and rendering pipeline should be understandable from the main text. Section 3 defines the Hexel's position, orientation, scales, hexagonal support, and spatial color field in Eqs. (1)-(4) (lines 162-198). The rendering steps are: intersect the camera ray with the Hexel plane, evaluate its support and color at the resulting local coordinates, and composite the opacity-weighted contributions front to back (§3, lines 199–204). Appendices A-B provide derivative calculations and rasterization details.

Our scope is a general, efficient primitive for novel-view synthesis, surface reconstruction, and image representation (§1, lines 80-87). Section 3.1 and Appendix D evaluate the same support and spatial color model on the Image-GS benchmark, improving reconstruction quality and converging faster than Image-GS (Table 1). Separately, the controlled planar ablation in §4.4 (lines 468-483) fixes surface normals and depth order to isolate the effects of hexagonal support and first-order spatial color in 2D. Efficient surface reconstruction is another core application, evaluated in §4.2.


W2. Design motivation The basic idea behind Hexels is to build the primitive around the structure of what it represents. Many real surfaces are locally flat, with smoothly varying color that can be approximated locally by a linear field. Our Hexel design - an affine hexagonal patch with spatially varying color - converts these regularities into efficiency within the primitive itself (this is discussed in the paper: §1, lines 70–87; §3).

Importantly, our motivation is also supported by appropriate theoretical results - Appendix I, Propositions I.1-I.2, bounds the local affine color-approximation error and supports Hexel's design. Within the same local color-error ellipse on a planar patch at a fixed view, the maximal affine hexagon covers twice the triangle’s area (the "raw" edge over triangles). The affine color field also represents linear color ramps exactly, while one spatially constant color cannot. In addition to this theory, the implementations also differ: Hexels use local-plane sigmoid support (§3, Eqs. (2)-(3)); densification uses gradient-triggered cloning and two-child splitting (Appendix C), while Triangle Splatting uses a projected triangular window and opacity/sharpness-driven sampling with midpoint subdivision (Triangle Splatting, §3).

Performance We believe there has been some confusion regarding the experiment. Please refer to the updated tables. Our main evaluations concern matched-memory and primitive-budgeted NVS (Tables 2,3, Appendix K), efficient surface reconstruction (Table 5,6), and 2D image representation (Table 1). Our primary comparison is with explicit surface primitives with local appearance like 2DGS, Triangle Splatting, and MeshSplatting (discussed in §2; §4, lines 299-309). Volumetric and hybrid parameters are provided for completeness. Please refer to the previous section for our reasoning as to why. For our main aforementioned evaluations, across 34 scenes (which is the largest amongst all the baselines) Hexels largely outperform the compared surface-aligned primitives on all benchmarks in quality and efficiency. Hexels have the highest PSNR and SSIM across all three benchmarks in Table 2; V2 improves all three quality metrics over the closest surface baselines with lower model storage and training VRAM in Table 3. DTU Chamfer distance is 28.9% lower than 2DGS at matched model memory (Table 5), Tanks and Temples mean F1 is highest among the compared methods (Table 6), and Hexels improve image quality and shorten fitting time relative to Image-GS (Table 1). We also please know that we have even more memory-matched settings in the Appendix, which we couldn't add to the main draft but are our main experiments (Tables 9-13). Even on unconstrained NVS, which is not our core experiment (as mentioned in §4.1, lines 392–394), Hexels perform competitively (achieving higher PSNR and SSIM, lower LPIPS, and higher FPS with lower deployment size than MeshSplatting across all three datasets, and TriangleSplat on DTU/Tanks and Temples higher FPS with lower memory)

Mesh-based NVS Can you please clarify what you mean by this? Also, we'd like to point out that triangle soup produced by standard Triangle Splatting cannot be turned into a clean mesh trivially.

Connected to meshes, Hexels also provide direct surface-cell export: each cell’s six vertices are obtained differentiably in O(1) work (§4.3, Fig. 5; Appendix H.1, Fig. 13). The cells retain their geometry and appearance parameters for downstream optimization.

MeshSplatting is already included in Tables 2-4. Tables 5-6 show that Hexels outperform MeshSplatting in efficient surface reconstruction as well.

AuthorsOpenReview ↗

Response to Review (part 2)

Here, we include results for MiLO for our memory/primitive matched (Table 2-3) settings. as the tables show, we outperform both of the methods significantly, both in performance and in efficiency.

Matched model memory. T&T: 10 scenes; Mip-NeRF360: 9; DTU: 15.

DatasetMethodTrain (min) ↓VRAM (GiB) ↓FPS ↑PSNR ↑SSIM ↑LPIPS ↓
T&TMILO90.08.2369.219.160.5950.564
T&TMeshSplat17.22.15119.616.290.5650.567
T&THexels13.30.56438.423.050.8110.295
Mip360MILO174.711.6141.520.840.4750.646
Mip360MeshSplat33.14.4960.819.310.4750.634
Mip360Hexels24.20.90377.824.820.7050.410
DTUMILO25.91.6072.221.170.6360.541
DTUMeshSplat13.31.92130.721.140.6660.505
DTUHexels8.40.55830.928.940.8900.303

Matched primitive budgets

MethodT&T Train min ↓T&T VRAM GiB ↓T&T FPS ↑T&T PSNR ↑T&T SSIM ↑T&T LPIPS ↓T&T Deploy MB ↓Mip360 Train min ↓Mip360 VRAM GiB ↓Mip360 FPS ↑Mip360 PSNR ↑Mip360 SSIM ↑Mip360 LPIPS ↓Mip360 Deploy MB ↓DTU Train min ↓DTU VRAM GiB ↓DTU FPS ↑DTU PSNR ↑DTU SSIM ↑DTU LPIPS ↓DTU Deploy MB ↓
MILO119.015.5668.919.790.6160.4611731.71205.514.8843.021.740.5780.4381730.4253.014.7773.021.770.7080.3741729.12
MeshSplat18.02.59116.816.420.5990.54828.3331.54.5460.918.930.4980.62019.9012.51.96135.322.960.7290.4714.58
Hexels V113.30.56438.423.050.8110.29516.2224.20.70377.824.820.7050.41010.228.40.55830.828.940.8900.3033.04

Connected to meshes, Hexels also provide direct surface-cell export: each cell’s six vertices are obtained differentiably in O(1) work (§4.3, Fig. 5; Appendix H.1, Fig. 13). The cells retain their geometry and appearance parameters for downstream optimization.


W3/Q1. First-order spatial color and hexagonal support

The motivation for first-order spatial color, compared with one color per primitive, is discussed in §1 (lines 70-87) and §3 (Eq. (4), lines 190-198). In fact, we provide a local approximation error bound in Appendix I.1 which explains the representation capacity of spatial color.

We added a controlled planar ablation crossing Gaussian/hexagonal support with constant/spatial RGB (§4.4, lines 468-483). The planar setting helps us clearly ablate the low overdraw and better representation hypothesis for Hexels: a full 3D reconstruction introduces coupled changes in geometry, visibility, surface normals, and depth ordering, making it difficult to isolate the effects of support shape and color representation. Instead, we approximate a local planar surface patch of a 3D scene, eliminating all of other confounders while keeping primitive count, initialization, loss, and optimization setup the same. This lets us causally test the effects of support shape and affine color in this controlled setting.

Could you clarify whether V4 means c₀ + SH without spatial gradients, or c₀ + spatial gradients + SH? The latter is already V3: V1, V2, and V3 all retain the spatial field, with V2/V3 adding view dependence (§3; Appendix A.1).

To additionally compare the primitives in 3D under a shared appearance model (which we believe you are asking with the V4 experiment), we equip 2DGS and Triangle Splatting with V1 color and evaluate all nine Mip-NeRF360 scenes. All three methods use exactly the same number of primitives, with no densification or pruning. They share the V1 spatial color representation and have approximately the same model memory. We also match the training-step budget and main training hyperparameters to minimize other confounding factors.

MethodModel MB ↓Peak VRAM GiB ↓PSNR ↑SSIM ↑LPIPS ↓
2DGS + V1 color~94.9523.730.66850.4521
Triangle Splatting + V1 color~90.918.700.45960.6253
Hexels V1~90.924.840.70530.4104

Hexels achieve the best PSNR, SSIM, and LPIPS over 2DGS and Triangle Splatting with the same spatial color model. This ablates hexagonal support of Hexels as a surface-aligned primitive over Gaussians and Triangles.

AuthorsOpenReview ↗

Response to Review (part 3)

Q2. Practical advantages and mesh-based extensions Please refer to W2 and W3/Q1 for the design motivation, theoretical analysis, and comparison with triangles. Our work fundamentally targets efficiency: exploiting locally planar geometry and smoothly varying appearance to obtain high quality at a lower cost. Beyond cost, the new comparison in W3/Q1 shows higher PSNR and SSIM, and lower LPIPS than Triangle Splatting when both use V1 color and exactly the same number of primitives.

A practical application of this efficiency is fully on-device reconstruction and rendering. As described in Appendix H.2, we will release an open-source iOS app supporting the complete workflow from images/videos to 3D reconstruction and novel-view synthesis on an iPhone or iPad. To our knowledge, this will be the first fully open-source app supporting the complete on-device workflow - from images or videos through training to 3D reconstruction and novel-view synthesis - on an iPhone or iPad, without requiring post-training quantization.

Our current Hexels are not a connected mesh. They instead support direct, differentiable surface-cell export in O(1) work per primitive (§4.3; Appendix H.1). Each cell retains its geometry and spatial appearance parameters for downstream optimization. A future mesh-based extension could investigate whether these larger, spatially varying patches reduce the number of surface elements needed.


Q3. Sigmoid support and learned sharpness

We use sigmoid support rather than a ReLU-based one because it allows retained pixels just outside the current hexagon to provide gradients during training. With ReLU, those pixels have zero contribution and no gradient through the support until the boundary reaches them. The sigmoid therefore gives existing primitives a way to adjust their position and extent using nearby reconstruction errors, which is useful when optimizing a fixed population of primitives.

It also fits our representation: at higher β, the interior has nearly uniform weight, allowing the affine color field to describe appearance across the patch, while β controls the edge transition. Changing β preserves the hexagon’s K = 1/2 contour. We evaluate support in normalized local surface coordinates, so its softness is relative to the patch rather than a fixed number of screen pixels (§3, Eq. (3); Appendix A).

This does require evaluating some extra pixels outside the hexagon. We control that work explicitly. The 3D rasterizer uses an alpha threshold of 1/255, which gives the bound d(u,v) ≤ log(254)/β. We use this bound for tile assignment, tighten it using the primitive’s opacity, and reject contributions below the threshold (Appendix B, lines 972–997). The sigmoid therefore has nonzero tails mathematically, but we only evaluate a finite region.

We initialize β at 10 and learn it separately for each Hexel, together with geometry, opacity, and appearance (Appendix E.2, lines 1166–1176). Below we report mean ± population standard deviation across all trained Hexels pooled within each benchmark, separately for each variant. Scenes containing more Hexels therefore contribute more weight.

BenchmarkV1V2V3
Tanks & Temples (10)4.42 ± 7.314.32 ± 7.254.63 ± 7.33
Mip-NeRF360 (9)4.64 ± 6.274.49 ± 6.554.76 ± 6.09
DTU (15)4.68 ± 88.844.35 ± 5.964.72 ± 14.24

The benchmark means are similar, while the spreads show variation across individual Hexels. These standard deviations describe the spread of learned values, not uncertainty in the mean. However, these aggregate statistics do not show where different sharpness values occur within a scene. Our supplementary video already shows how Hexels evolve during optimization. We will try to add selected views of reconstructed scenes, highlighting specific patches with high β values, to make this spatial variation clearer.

Comments on the paper

1 comment
Area_Chair_UHbkOpenReview ↗

Discussions

Dear Reviewers,

Could you please engage in further discussions with the authors based on their response?

Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.