acceptodds
Withdrawn from ICLR 2027

Quadrilateral Splatting for Real-Time Radiance Field Rendering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Triangle Splatting (TS) has recently emerged as an effective explicit, non-volumetric surface representation for novel-view synthesis (NVS) and real-time rendering, but each triangular primitive is defined over a three-vertex surface domain, which can limit its flexibility for modeling fine-scale local geometry and spatial appearance variation. This limited geometric flexibility can result in less clearly resolved surface-normal structure in regions containing fine geometric detail. We propose Quadrilateral Splatting (QS), which represents scenes using independently optimized planar quadrilateral elements with a continuous bilinear appearance field. The quadrilateral domain provides additional flexibility for modeling local surface geometry, while the continuous bilinear appearance field represents spatial color variation within each primitive. QS jointly optimizes quadrilateral geometry, per-primitive opacity, and view-dependent appearance through differentiable polygon rasterization and alpha compositing, with planar-convex retraction maintaining valid quadrilateral geometry during optimization. Experiments on Mip-NeRF 360 and NeRF Synthetic show improved PSNR, SSIM, and LPIPS over TS and 2DGS on indoor scenes, with clearer fine-scale geometry in rendered normal maps and comparable outdoor quality while preserving real-time rendering. These results demonstrate the effectiveness of quadrilateral primitives with continuous bilinear appearance fields for real-time radiance field rendering.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Novel View Synthesis, 3D Gaussian Splatting, Triangles, Quadrilaterals

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_waYpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes using learnable quadrilateral planes for non-volumetric primitive based 3D reconstruction improve the fine geometry and appearance reconstruction capabilities while also inheriting the efficient rasterization property of non-volumetric primitive based 3D reconstruction models. The claimed contributions are the (1) quadrilateral primitive and its respective appearance modeling, (2) training framework including planar-convex retraction, (3) and improvement of reconstruction metrics PSNR SSIM LPIPS on indoor Mip-Nerf 360 and NeRF Synthetic when comparing with Triangle Splatting and 2DGS. The authors also include remarks on rendering speed and geometric quality, but rendering speed is slower than its non-volumetric competitors and no comparison with volumetric representations and no quantitative comparison for geometric accuracy is available to back the claims. Especially, the paper repeatedly highlights the lack of flexibility of Triangle Splatting in reconstructing fine detail geometry and appearance and claims that the proposed quadrilateral primitives are more flexible in doing so, but there is no logical or empirical evidence to support this, apart from the qualitative comparisons in Figures 6,7,8 which is insufficient. Moreover, the paper fails in showing that non-volumetric representation is important at all.

Critical strengths

None

Critical weaknesses
  1. Unclear advantage of proposed method. It seems that the only advantage of this method is reconstruction quality for indoor scenes. If so, the paper should be written with the focus on indoor scenes, which it currently isn’t. If there is a tradeoff that would make the proposed method better applicable in particular situations it should be mentioned. Otherwise the quantitative comparisons show - with comparison to quite outdated methods - no improvement in overall reconstruction quality while also being large in model size and slower in rendering speed. 


  2. Lack of related work and comparisons. More comparison should be made with both volumetric and non-volumetric representations to show the advantage of non-volumetric representations when compared to volumetric methods, and the superiority - in any aspect - when compared with other recent non-volumetric representations. Related Works section and Experiments section should include recent works like Radiant Triangle Soup (Burgdorfer and Mordohai, 2025), Deformable Beta Splatting (Liu et al., 2025), Nexels (Rong et al., 2025), or explain why the comparison is limited to Triangle Splatting and 2DGS. Currently the paper seems to lack study in relevant literature, and need clear demonstration of the advantage of the proposed method in relation to recent literature.

Questions for authors
  • Inconsistent terminology or unclear expressions in the paper should be made consistent, clearly defined or revised.
What does the following mean: “local point spacing” “primitive adaptation process” “auxiliary shell points” “bilinear spatial residuals” “local child domains” “in-plane displacement” “insufficiently observed” “boundary/edge falloff”


  • What triggers the split, clone and pruning of the quadrilateral primitives? Why are these methods adequate for quadrilateral primitives? 



  • What is the model size and number of primitives for NeRF Synthetic comparison? (Table 2) Does the improvement in PSNR SSIM LPIPS accompany larger model size and slower rendering speed?

What is the model size, number of primitives for Outdoor/Indoor Mip-NeRF 360? The improvement in PSNR SSIM LPIPS for Indoor Mip-NeRF 360 in Table 1 could be due to larger model size.

Minor questions / suggestions for future work:

  • The paper could be restructured to put less important details like dataset or evaluation metrics details in the appendix to have more important details like “primitive adaptation” or ablation tables included in the main paper. 



  • The meanings of equation 2 and equation 3 should be explained. Currently it only defines the variables and there is insufficient explanation of why equation 2 is derived (equation 2). Define gamma_u and gamma_v. Elaborate on what the update to d does to the quadrilateral (equation 3).

Reviewer_jDLTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

In this work, the authors propose a Quadrilateral Splatting (QS) method for Novel View Synthesis (NVS) rendering. In contrast to the Triangle Splatting (TS) representing the scene with optimized triangle elements, QS introduces independently optimized planar quadrilateral elements as the primitive representation, where each element consists of four cyclic vertices, opacity, a positive edge-falloff exponent, spherical-harmonic (SH) coefficients, and three RGB spatial residual vectors. During rendering, the appearance field is modeled as bilinear interpolation in the elements. The results on Mip-NeRF 360, NeRF Synthetic, Tanks & Temples confirms that the proposed QS method has about 0.33 db improvements on the indoor Mip-nerf 360 data.

Critical strengths
  1. This work proposes a more informative 3D NVS representation over the triangle primitives;

  2. Though heavier, the proposed method achieves better rendering performances than triangle primitives.

Critical weaknesses
  1. My major concern about this work is the motivation to introduce such a QS representation for neural rendering. From my perspective, increasing the representation complexity may not be obviously superior to simply introduce more TS primitives, especially when TS still has higher efficiency as shown in Table 4. As shown in Table.7, QS can achieves 0.4 dB improvements by doubling the number of primitives, which is even larger than its improvements over baselines on Mip-NeRF 360. This also confirms the influence of primitive capacity.

  2. The improvements over baselines are still limited, and slightly inferior on outdoor scenes.

Questions for authors
  1. As claimed in the weaknesses section, I think a further comparison under similar rendering cost may be essential, e.g., increasing the number of TS primitives until it has similar FPS with QS, to confirm whether QS is superior under similar cost;

  2. The authors miss some related works such as [1,2] in this paper. [1] also introduces a 3D rectangle primitive representation, which maybe similar as the QS in this work, while the bilinear interpolation of appearance field modeling in [2] is similar as QS.

[1] PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes

[2] Gaussian Billboards

  1. Though claiming that QS can provide clearer normal maps, some comparisons on geometry reconstruction would be more meaningful.
Reviewer_pZKoOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper tackles the task of novel view synthesis and real-time rendering in computer vision. It builds upon "Triangle Splatting", where triangles are learned to compose scenes. Instead of triangles, the paper proposes to use quadrilaterals, because they are claimed to allow more flexibility in their representation capability. Since four points in space do not always form a planar quadrilateral (contrary to triangles), a correction mechanism is introduced. Besides, instead of using unicolor shapes, the paper introduces a continuous bilinear appearance field that textures the quadrilaterals. The results show comparable performances to triangles (sometimes slightly better), and the ablations show the importance of the appearance field.

Critical strengths

The main idea of the paper (using quadrilaterals) is clear, the paper is well-written and easy to read, and the structure follows a standard route for a splatting-based computer vision paper. Using quadrilaterals is a legitimate question since triangles were shown to work well, and, for the NVS community, the paper brings answers that deserve to be known.

Critical weaknesses

[W1] Unfortunately, the paper brings little novelty or significant contribution. Triangle splatting does the heavy lifting, and adapting it to quadrilaterals seems like an almost direct variation. The correction mechanism (planar-convex retraction) is more of a hack to make it work than a contribution and is not needed with triangles. The results are said to be better or comparable to triangle splatting but Table 1 shows barely any difference overall. Table 2 is slightly more convincing, and Table 4 shows that efficiency is lower (for, as said, negligible gains). If the focus is on the shape of the primitives themselves, I believe the proposed work might have delivered a stronger message by studying how the performances evolves with the number of sides per shape considered, maybe revealing that the lower, the better (or maybe drawing a performance-efficiency Pareto frontier curve).

[W2] The continuous bilinear appearance field could have been the core contribution of the paper. The ablation conducted in Appendix shows that, in fact, this field matters a lot for performance. Its principle is quite generic, so it could be applied to triangles (or likely to any convex polygon) as well and reveal a similar usefulness there. However, in its current form, the paper does not disentangle well the contributions of quads vs. appearance field, and this might be one of the most important points to clarify. Still, the paper correctly states that it is the combined use of quadrilaterals and appearance field that makes the method effective.

[W3] The paper repeatedly motivates its existence by the fact that "triangles just have three vertices, thus have limited flexibility for modeling fine-scale local geometry and spatial appearance variation" (abstract, introduction, related work, at least). However, this is not supported by concrete evidence, nor even by intuition: every convex polygon (including quadrilaterals) can be decomposed into triangles, not into quadrilaterals; thus, triangles appear to be better suited for flexibility and fine-grained modeling. Furthermore, three points have the advantage of always being planar, while four are not, which is why the correction mechanism is needed. Other than that, the differences or similarities with triangle splatting are not clearly expressed, making the reader potentially wonder what comes from which paper, in particular in sections 3.2 and 3.3 (most comes from TS I believe, so the original contributions should be made clear).

All in all, I think the paper missed the chance to study rigorously how "polygon" splatting scales with the number of sides of the polygon and to evaluate clearly how much the continuous bilinear appearance field helps that kind of splatting (and it could be the main contribution itself). The proposed pipeline seems appropriate for that kind of larger-scale study, but it focused only on quadrilaterals and did not clearly disentangle the benefits of its contributions. In that vein, extensions to meshes, as done, for instance, with Mesh Splatting, would have been valuable as well. In its present form, I believe the paper is a decent contribution for the novel view synthesis community and would fit in specific workshops or venues, but not so much for the general ICLR main track.

Misc:

  • In Eq. (8) the terms of the loss are undefined, and then different notations are used in appendix.

  • Figures showing normals (e.g. Figure 6) should also be compared to ground truth to better appreciate the differences in the results.

  • 8 out of 9 results for TS (retrained) of Table 1 are slightly worse than those reported in the original TS paper. Can this be explained by some changes in the code or is it just "bad luck"?

  • Line 417: it is wrong to claim that results are superior in Tanks and Temples than the baselines, as BBSplat is largely better in PSNR and SSIM there. By the way, can't BBSplat be used in Table 2 as well?

  • Table 4 reports number of elements, which I understand as number of primitives. This does not indicate the number of trainable parameters, which might be another useful number to report in this table to better show the differences between the methods.

  • Line 465: the line about Table 3 repeats line 419.

Questions for authors

I believe the changes I envision would be too substantial to be conducted in a short amount of time, and might change the paper too much.

Still, in case some results are already available, I would be curious to know:

  1. Is the method indeed scalable to any convex polygon (not just quadrilaterals). If so, how well does it work as the number of vertices grows? If not, why, and what should be done to complete the study?

  2. Is the continuous bilinear field indeed usable for triangles? And if so, what impact does it have on TS?

  3. Is it possible that most differences in the results between TS and QS fall within noise range, or that, for a same matched number of trainable parameters, consistent differences appear?

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

Related papers

Discussion (0)

Sign in to comment.