acceptodds
Under review as a conference paper at ICLR 2027

SPIRE: A Differentiable Mesh Renderer for Closing the Export Gap

Abstract

Triangle meshes are the standard representation for graphics and geometry processing, yet high-fidelity novel-view synthesis commonly relies on neural fields or Gaussian primitives. Converting these representations to meshes can reduce rendering fidelity, creating an export gap between native rendering and the reusable asset. Direct mesh optimization is challenging due to hard surface boundaries and occlusion. Conventional mesh renderers make binary coverage and visibility decisions at discrete pixel samples, yielding zero derivatives almost everywhere. A real camera, by contrast, integrates radiance over each pixel. This pixel-area integration can yield continuous, differentiable image values even for opaque meshes with hard boundaries. We introduce SPIRE, a differentiable mesh renderer that uses analytic anti-aliasing to realize pixel-area integration. The renderer integrates clipped triangle footprints, combines adjacent faces without mutual occlusion, and spatially tracks the remaining visible footprint when compositing depth-separated surfaces. A sparse backward pass efficiently differentiates these operations with respect to mesh geometry and spherical harmonic appearance. Given multiview RGB images, SPIRE reconstructs conventional opaque meshes with shared vertices and vertex colors, without external normal, depth, or diffusion priors. Our meshes achieve the highest post-export fidelity across the evaluated Blender and Filament settings on synthetic and real scenes, with small export gaps. Controlled ablations establish the importance of pixel-area coverage and surface-aware compositing for direct mesh optimization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
differentiable rendering, analytic anti-aliasing, mesh optimization, 3D reconstruction, novel view synthesis

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GipeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a differentiable mesh renderer using analytical anti-aliasing to improve exports of opaque meshes from neural fields or 3DGS.

Given posed multi-view images, the proposed approach first extracts a mesh using GOF, then optimizes the mesh using differentiable rendering with an analytical anti-aliasing technique.

There is substantial related work in the field of analytical anti-aliasing (cone tracing, diff rasterization, warped-area sampling, MC edge sampling, ...), and it is not clear to me what the novel contribution is in the proposed AA formulation. This could be clarified.

My main criticism is that geometry and appearance is entangled in the evaluation. The novel contribution here is improved geometric fidelity, so I suggest removing the appearance bit from the evaluation and focus on geometric quality of the extractions. Given that the appearance model does not match the reference anyway (Blender use a lit PBR material, while the exported meshes uses vertex colors AFAIU), there is already an appearance mismatch, which obstructs the evaluation of geometric quality.

Overall, the extracted meshes seem to be of high quality, and this is a very useful tool to have. However, I would have prefered the validation to focus entirely on geometry quality, with rendered views with normal map and/or diffuse shading instead of baked emissive color.

I am on the fence here. In this setting, given that the geometry quality improves over GOF, I am leaning positive, but if the main advantage over the GOF baseline is improved appearance (represented as emissive color/SH DC component, not full PBR), then I am less convinced.

Critical strengths

Improved mesh quality via differentiable rendering with analytical antialising.Using an analytical anti-aliasing approach to improve the extracted mesh makes sense to me.

Critical weaknesses

While I like the approach overall, I would prefer if the validation focus on geometry quality, using Chamfer distance and renderings with normal map shading and/or diffuse shading, to highlight the geometry quality. Baked emissive color is not a very useful appearance representation, I see it as a stop-gap. Similar to GOF and FlexiCubes (https://research.nvidia.com/labs/toronto-ai/flexicubes/), I would recommend to visualize the extracted meshes using normal map shading or diffuse shading, to more clearly see the geometry detail in the extracted meshes.

The only geometry evaluation I found was in Table 13 and Table 23. In Table 13, SPIRE has higher CD than GOF, which is used to initialize SPIRE's optimization. This is slightly concerning. Is SPIRE optimization reducing the geometry quality from the GOP initialization for this dataset? This could be discussed in more detail. Is this only for DTU or also true for NeRF synthetic? Also, other approaches, like MILo performs better than SPIRE here.

This statement at l377: "GOF retains sharper geometry but yields mottled vertex colors" also hints that the difference is not in geometric quality, but exported appearance.

Questions for authors

In Table 14, why is the GOF initialization training time for SPIRE larger than the GOF training time? I thought GOF was used to initialize SPIRE?

Let DΠ be the rendered median-depth image: How is this computed?

This statement at l377: "GOF retains sharper geometry but yields mottled vertex colors" also hints that the difference is not in geometric quality, but exported appearance. Is this generally the case, that GOF retains sharper geometry, and the proposed approach mostly improves appearance (the SH DC component)?

Some citations: Ray Tracing with Cones John Amanatides http://www.cs.yorku.ca/~amana/research/cones.pdf

Analytic Anti-Aliasing of Linear Functions on Polytopes Auzinger et al.

Unbiased Warped-Area Sampling for Differentiable Rendering, Bangaru et al.

Differentiable monte carlo ray tracing through edge sampling. Li et al.

7 comments
AuthorsOpenReview ↗

Official comment for reviewer Gipe (part 1/4: W1)

Thank you for recognizing the quality and usefulness of the extracted meshes and the value of analytic anti-aliasing for mesh optimization. We also appreciate your questions about separating geometry from appearance, interpreting the GOF comparison, and clarifying the relation to prior work. We will incorporate the clarifications, results, and revisions addressing all your points into the camera-ready version. We answer Critical Weaknesses W1-W3 and Questions Q1-Q4 in review order; Q4 addresses your suggested references.


[W1] Seeing the geometry without baked colors

We evaluate both the appearance preserved after mesh export and the underlying geometry. (1) We render exported meshes with a common emission material to measure how well they preserve the observed image appearance. (2) We will add the requested normal-map and diffuse-shading comparisons to show geometry without baked colors. These will complement the paper's Chamfer-distance and depth evaluations, with further quantitative evidence from silhouettes and 3D boundaries in Tables G1-G3.

(1) Why we evaluate exported appearance. A central goal of SPIRE is to produce meshes that preserve image fidelity when used in standard graphics software. Rendering the exported vertex colors with an emission material measures this transfer directly. It tests whether the mesh remains a faithful representation of the observed scene after leaving the training renderer. This is why baked-color comparisons are part of the evaluation alongside geometric accuracy.

(2) How we assess geometry separately. The paper already reports DTU Chamfer distance in Tables 13/23 and depth-map visualizations in Table 13. Your suggested normal-map and diffuse-shading views make local shape easier to inspect without baked colors, and we will include these comparisons with matched views and lighting. Tables G1-G3 further measure silhouettes and the 3D locations of visible boundaries. Comparing these with the same initial mesh isolates the change in geometry, directly addressing whether the gain goes beyond appearance. We will add these evaluations to Appendix D.5.

AuthorsOpenReview ↗

Official comment for reviewer Gipe (part 2/4: W2)

[W2] Does refinement improve the actual geometry?

SPIRE improves visible geometry from the same initial mesh. SPIRE combines visible face portions to better represent fine or curved scene structures within the given coarse topology. Chamfer distance (CD) also measures hidden portions, so it does not fully reflect this improvement. We thus evaluate geometry seen by standard mesh renderers. We clarify (1) DTU initialization and show that SPIRE (2) matches object silhouettes more closely (Table G1) while (3) placing their boundaries more accurately in 3D (Table G2), and (4) recovers more depth boundaries between overlapping object parts (Table G3).

(1) DTU comparison. As described in Appendix C, SPIRE starts from a mesh constructed by TSDF fusion of GOF-rendered depth maps. Table 13 evaluates a separately extracted GOF mesh. GOF's lower Chamfer distance therefore does not mean that SPIRE made its initial geometry less accurate. We examine the change directly on NeRF-Synthetic by comparing the same initial mesh before and after optimization in Tables G1-G3.

(2) Visible shape (Table G1). The color-only control fixes the initial geometry, so changes in silhouette require moving the faces. SPIRE improves coverage and outline alignment in held-out views. This directly shows geometric improvement from the initialization, beyond fitting colors.

Table G1. NeRF-Synthetic silhouettes and contours. Held-out views use a common ray caster and reference image alpha masks. Intersection-over-union (IoU) measures mask overlap; contour F-score measures outline alignment at each pixel tolerance. RaDe-GS retains its largest connected component; unfiltered TSDF retains all.

MethodSilhouette IoU ↑Contour F@1px ↑Contour F@2px ↑Contour F@4px ↑
GOF0.9780.9360.9560.971
MeshSplatting0.9820.9340.9600.980
2DGS0.6480.3290.4430.531
nvdiffrast0.9740.9160.9390.958
MILo0.9630.8260.9170.963
RaDe-GS0.7010.4040.5180.655
RaDe-GS unfiltered TSDF0.9290.5370.6980.877
Initial mesh / color-only0.9760.9170.9440.963
SPIRE0.9820.9580.9720.983

(3) Boundary depth (Table G2). We next check whether these outlines lie at the correct depth. SPIRE improves both precision and recall over the initial mesh and baselines in 3D. Thus, the visible boundary is more accurately placed in space, beyond matching its image projection.

Table G2. Outer-boundary 3D accuracy on NeRF-Synthetic. Mesh and original-scene silhouettes are backprojected using first-hit depths from identical rays. Points match within 0.02 scene units in the same view. Precision measures matched predictions; recall measures recovered reference points; F1 balances both. Eight-scene means (%).

MethodOuter precision ↑Outer recall ↑Outer F1 ↑
GOF94.4893.0193.59
MeshSplatting94.7993.1793.79
2DGS47.8545.0343.47
nvdiffrast92.0990.6691.10
MILo92.1592.4891.85
RaDe-GS75.6553.7058.64
RaDe-GS unfiltered TSDF81.1185.8782.72
Initial mesh / color-only94.0991.1992.36
SPIRE94.8994.2894.43

(4) Internal structure (Table G3). Inside the outline, one object part can lie in front of another, creating a depth boundary. SPIRE recovers more of these reference boundaries in 3D at all three tolerances. The geometry updates therefore recover visible internal structure as well as the outer shape.

Table G3. Internal 3D boundary recovery on NeRF-Synthetic. Neighboring foreground depth changes above 2% of the nearer depth are backprojected from mesh and original-scene first-hit maps, excluding the silhouette band. Recall is the fraction of reference points recovered within each 3D distance in the same view. Distances are in scene units; eight-scene means (%).

Method[email protected] ↑[email protected] ↑[email protected] ↑
GOF69.1780.9291.46
MeshSplatting61.5373.1081.72
2DGS36.9551.4065.58
nvdiffrast64.8876.6185.64
MILo62.6575.0984.35
RaDe-GS35.7144.0651.85
RaDe-GS unfiltered TSDF54.4970.4183.22
Initial mesh / color-only66.4978.8390.79
SPIRE73.2784.6193.44

These visible-shape controls complement the DTU whole-surface comparison, including MILo's 0.68 mm CD. We will add Tables G1-G3 and clarify the initialization in Appendix D.5.

AuthorsOpenReview ↗

Official comment for reviewer Gipe (part 3/4: W3)

[W3] The wording about sharper GOF geometry

By "sharper geometry," we meant that GOF produces a more continuous-looking surface than MeshSplatting in this example. We explain (1) why this visual difference does not necessarily mean more accurate surface positions, and (2) how SPIRE improves visible geometry beyond GOF (Tables G1-G3).

(1) What the visual comparison means. MeshSplatting transitions from semitransparent to opaque faces, which can make previously less visible triangles stand out. GOF instead extracts a level set of its opacity field, producing a more continuous-looking surface in this example, but this level set can still be displaced from the true surface. Since GOF evaluates colors at the extracted vertex positions, this displacement can also lead to incorrect vertex colors. This provides a possible link between surface placement and mottled appearance, motivating direct measurements of the geometry as well as visual inspection.

(2) How SPIRE improves the geometry. Tables G1-G3 show better silhouette alignment, outer-boundary 3D accuracy, and internal-boundary recall than GOF. These tests assess the shape and position of the surface seen by the cameras. Together, they show that SPIRE improves visible geometry alongside exported appearance, addressing the concern that the difference is only in vertex colors. We will clarify the original comparison and connect it to these measurements in Section 4.2.

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

This paper introduces a differentiable renderer designed for directly optimizing fully opaque triangle meshes from multi-view RGB supervision. Instead of softening triangle visibility or relying on surrogate boundary gradients, SPIRE derives geometry gradients from pixel-area integration. It combines analytic triangle–pixel footprint integration, surface-aware accumulation of adjacent faces, approximate propagation of residual visibility across depth-separated surfaces, and a sparse backward pass. The resulting shared-vertex mesh remains opaque throughout optimization and can be exported directly to standard renderers such as Blender and Filament. The experiments show post-export rendering quality on NeRF-Synthetic and Mip-NeRF 360, together with controlled renderer comparisons and downstream mesh editing/simulation examples.

Critical strengths
  1. The paper identifies a meaningful and under-emphasized evaluation target. The distinction between native novel-view rendering quality and fidelity after exporting to a conventional mesh is practically important. The experiments make this distinction explicit rather than treating mesh extraction as an incidental post-processing stage.

  2. The main technical idea is conceptually clean and reasonably novel. Obtaining useful geometry gradients from pixel-area coverage while keeping faces opaque throughout optimization is an attractive alternative to soft/semitransparent rasterization. The combination of exact face-footprint integration, same-surface accumulation, visibility propagation, and sparse differentiation forms a coherent renderer rather than a collection of unrelated heuristics.

  3. The controlled comparisons are particularly useful. Under identical initialization and fixed topology, SPIRE substantially improves joint geometry/appearance optimization over the compared differentiable renderers. On NeRF-Synthetic, for example, the fixed-topology comparison reports Blender PSNR of 29.03 dB for SPIRE versus 27.83 dB for nvdiffrast, while several triangle-based alternatives deteriorate considerably during joint optimization.

  4. The post-export results are strong and consistent. SPIRE achieves 29.03 dB Blender64 PSNR on NeRF-Synthetic and 23.28 dB on Mip-NeRF 360, with smaller export gaps than most baselines.

Critical weaknesses
  1. The paper should more carefully qualify what is exact in the differentiable rendering formulation. The per-face pixel-area integration is analytic, but inter-surface visibility is not tracked exactly. The method replaces the potentially disconnected residual visible footprint with a moment-matched axis-aligned rectangle. Moreover, the backward pass applies stop-gradient to footprint centers and widths. Therefore, the overall renderer is not simply an exact differentiation of the complete forward visibility computation. This matters because one of the paper's motivations is precisely the distinction between differentiating the rendering operation and using surrogate gradients. I find the central idea convincing, but the current wording occasionally makes the exactness claim broader than what is implemented. A small finite-difference gradient validation around occlusion boundaries would substantially strengthen this claim.

  2. Reconstruction quality remains dependent on the initialization topology. SPIRE initializes its mesh using GOF and then preserves connectivity and vertex count during optimization. Thus, the method can refine geometry that already exists, but cannot readily create a missing component, open a missing hole, or recover topology that was absent from the extracted initialization. This limitation is especially relevant because the paper is framed partly as direct mesh reconstruction rather than only a differentiable rendering contribution. An initialization-sensitivity study, e.g. using increasingly decimated or lower-resolution initial meshes, would clarify where SPIRE's reconstruction advantage comes from and where topology becomes the limiting factor.

  3. The strongest result is post-export appearance fidelity rather than geometric accuracy. On DTU, SPIRE's mean Chamfer distance is 0.80 mm, compared with 0.68 mm for RaDe-GS and MILo and 0.74 mm for GOF. Thus, the geometry is competitive but not state of the art. I therefore think the conclusions should emphasize that SPIRE improves post-export photometric fidelity while maintaining competitive geometry, rather than suggest uniformly superior reconstruction quality.

Questions for authors
  1. Could the authors provide a finite-difference gradient check for vertex positions, particularly for depth-separated surfaces and occlusion boundaries? How large is the discrepancy introduced by the moment-matched visibility approximation and the stop-gradient treatment of footprint moments?
  2. How sensitive is the final result to GOF initialization? For example, what happens if the initial mesh is substantially decimated or extracted at a lower resolution?
  3. Could the authors clarify the intended scope of the “exact” terminology throughout the paper—i.e. exact face-footprint integration versus approximate inter-surface visibility propagation?
5 comments
AuthorsOpenReview ↗

Official comment for reviewer DDna (part 1/4: W1)

Thank you for recognizing the importance of post-export fidelity, pixel-area integration, and controlled renderer comparisons. Your comments on exactness, initialization, and geometry help clarify our contribution. We will incorporate the clarifications, results, and revisions addressing all your points into the camera-ready version. We answer Critical Weaknesses W1-W3 and Questions Q1-Q3 in review order.


[W1] Spatial occlusion, accuracy, and efficiency

SPIRE is designed for accurate, stable, and efficient mesh optimization. We show (1) that moment matching closely reproduces colors computed from exact visibility (Table D1), (2) why sparse differentiation preserves reconstruction quality at lower backward cost (Table D2), and (3) the efficiency gained by replacing explicit visibility tracking with moment matching (Table D3).

(1) Accurate spatial visibility (Table D1). The forward pass approximates the remaining pixel region with a moment-matched footprint (the transmittance window's position and size). It already encodes where nearer surfaces occlude the pixel, so gradients of the face weights and remaining transmittance pass spatial-occlusion information to the mesh. SPIRE has 11.24 and 4.74 times lower mean error than nvdiffrast AA 4x4 in the two- and three-surface tests. Updating the footprint also reduces three-surface error by 90.1% over fixed-footprint accumulation. Tracking both the amount and location of remaining area thus improves visibility without dense sampling such as SSAA 64x64 with 4,096 samples per pixel.

Table D1. Forward accuracy at opaque object edges. We place one or two triangulated foreground meshes over a background mesh and measure mean absolute linear RGB error against exact visible-polygon integration. Each column covers 432 configurations: three mesh densities (14/30/62 faces), eight edge orientations, nine positions spanning -0.45 to 0.45 pixels, and constant or linearly varying foreground colors. nvdiffrast AA 4x4 uses four times the resolution along each axis followed by averaging; SSAA labels give samples per pixel.

RendererTwo surfaces RGB MAE ↓Three surfaces RGB MAE ↓
Fixed footprint8.87e-43.67e-2
nvdiffrast AA9.13e-21.23e-1
nvdiffrast AA 4x41.01e-21.72e-2
diffvg, 8x8 samples7.08e-39.74e-3
SSAA 16x161.99e-32.16e-3
SSAA 64x643.59e-43.89e-4
SPIRE8.96e-43.62e-3

(2) Efficient sparse differentiation (Table D2). Stop-gradient on window position and size avoids differentiating successive moment updates, a dependency chain that adds computation and can amplify numerical errors. This stability/efficiency motivation also appears in Gaussian Blending, Appendix B.1. Full derivatives take 6.16 times as long in our backward-cost comparison for only a 0.02 dB mean Blender PSNR gain, supporting nearly the same quality at much lower cost.

Table D2. Image fidelity and backward cost with and without window derivatives. Mean NeRF-Synthetic PSNR in dB uses matched training with the same validated implementation, forward renderer, initial mesh, and objective. PSNR is native; Blender PSNR evaluates export. Backward time in milliseconds is measured separately on one fixed mesh and camera on the same GPU, comparing the default sparse and validated full-derivative implementations.

Window derivativePSNR ↑Blender PSNR ↑Backward time ↓
Stop-gradient32.2829.0358.5
Full derivative32.3129.05360.7

(3) Exact visibility cost (Table D3). For comparison with SPIRE's moment matching, we implemented a prototype that computes the visible region exactly, without any visibility approximation, by subtracting occluded regions and integrating the remaining polygons. This prototype is about 1,714 times slower. Together with Table D1, this shows why moment matching is useful: it captures spatial occlusion with low error while keeping visibility computation practical. We will add Table D1 to Appendix A.2 and Tables D2-D3 to D.6, and clarify the design in Section 3.3 and Appendix B.3.

Table D3. Moment matching versus explicit visible-polygon rendering. Both process the same trained mesh of about 2.35 million faces with the same camera and GPU. Mean forward time in seconds includes projection and rendering.

Visibility representationMean forward time ↓
Moment-matched footprint0.102
Explicit visible polygons174.623
AuthorsOpenReview ↗

Official comment for reviewer DDna (part 2/4: W2)

[W2] Initialization and mesh complexity

SPIRE retains high export quality across different coarse initializations and mesh sizes. We show (1) that GOF is one initialization choice by comparing different starting meshes (Table D4), and (2) that SPIRE's advantage persists after substantial simplification by comparing methods at common face budgets (Table D5).

(1) High quality across initializers (Table D4). GOF is an initialization choice, not a requirement for SPIRE's high export quality. Using meshes from GOF, Gaussian Blending, Gaussian Surfels, or NeuS2 gives similar mean Blender PSNR, ranging from 27.82 to 28.10 dB. This shows that SPIRE retains high export quality when starting from different coarse initial meshes. Joint fitting also outperforms color-only fitting for each initializer, demonstrating the additional benefit of geometry updates. The initializer supplies coarse positions, shapes, and connectivity; SPIRE optimizes vertex positions and appearance while preserving connectivity and vertex count.

Table D4. Refinement from different mesh initializations. We initialize SPIRE with meshes from each listed method, reset their vertex colors to the same neutral gray, and measure mean Blender PSNR in dB on NeRF-Synthetic. Color-only fitting optimizes appearance with fixed vertices; joint fitting also optimizes vertex positions.

Initial mesh sourceColor-only PSNR ↑Joint PSNR ↑
GOF26.3928.10
Gaussian Blending23.7627.89
Gaussian Surfels22.8528.08
NeuS226.5227.82

(2) Common mesh budgets (Table D5). The GOF baseline averages 3.20M faces; our standard initialization already subsamples GOF before extraction to obtain 2.10M (paper Table 14). Further simplification to 1M, 0.5M, and 0.25M retains SPIRE's export advantage over GOF, MeshSplatting, and Triangle Splatting at every common budget. The initializer supplies connected components, holes, and the vertex budget, while SPIRE optimizes from this coarse structure. We will add the initialization and budget studies to Appendix D.2 and clarify the setup in Appendix C.

Table D5. Comparison at common maximum face budgets. Mean Blender PSNR in dB on NeRF-Synthetic, with maximum face budgets from 0.25 million to 1 million faces (M denotes million). These limits apply during extraction for GOF and during mesh optimization for the other methods. All exported meshes use vertex colors as emission colors in Blender.

Method0.25M ↑0.5M ↑1M ↑
GOF21.5923.1324.50
MeshSplatting26.8726.8826.84
Triangle Splatting19.8018.8217.89
SPIRE27.7627.9028.00
AuthorsOpenReview ↗

Official comment for reviewer DDna (part 3/4: W3)

[W3] Geometry beyond the rendered colors

SPIRE improves visible geometry as well as exported image quality. SPIRE combines visible face portions to better represent fine or curved scene structures within the given coarse topology. Chamfer distance (CD) also measures hidden portions, so it does not fully reflect this improvement. We thus evaluate geometry seen by standard mesh renderers. We show that SPIRE (1) matches object silhouettes more closely (Table D6) while (2) placing their boundaries more accurately in 3D (Table D7), (3) recovers more depth boundaries between overlapping object parts (Table D8), and (4) improves image quality through geometry optimization beyond color fitting alone (Table D9).

(1) Visible shape (Table D6). The color-only control fixes the initial geometry, so changes in silhouette require moving the faces. SPIRE improves coverage and outline alignment in held-out views. Its geometry thus better represents the object shape seen by new cameras, beyond fitting colors.

Table D6. NeRF-Synthetic silhouettes and contours. Held-out views use a common ray caster and reference image alpha masks. Intersection-over-union (IoU) measures mask overlap; contour F-score measures outline alignment at each pixel tolerance. RaDe-GS retains its largest connected component; unfiltered TSDF retains all.

MethodSilhouette IoU ↑Contour F@1px ↑Contour F@2px ↑Contour F@4px ↑
GOF0.9780.9360.9560.971
MeshSplatting0.9820.9340.9600.980
2DGS0.6480.3290.4430.531
nvdiffrast0.9740.9160.9390.958
MILo0.9630.8260.9170.963
RaDe-GS0.7010.4040.5180.655
RaDe-GS unfiltered TSDF0.9290.5370.6980.877
Initial mesh / color-only0.9760.9170.9440.963
SPIRE0.9820.9580.9720.983

(2) Boundary depth (Table D7). We next check whether these outlines lie at the correct depth. SPIRE improves both precision and recall over the initial mesh and baselines in 3D. Thus, the visible boundary is more accurately placed in space, beyond matching its image projection.

Table D7. Outer-boundary 3D accuracy on NeRF-Synthetic. Mesh and original-scene silhouettes are backprojected using first-hit depths from identical rays. Points match within 0.02 scene units in the same view. Precision measures matched predictions; recall measures recovered reference points; F1 balances both. Eight-scene means (%).

MethodOuter precision ↑Outer recall ↑Outer F1 ↑
GOF94.4893.0193.59
MeshSplatting94.7993.1793.79
2DGS47.8545.0343.47
nvdiffrast92.0990.6691.10
MILo92.1592.4891.85
RaDe-GS75.6553.7058.64
RaDe-GS unfiltered TSDF81.1185.8782.72
Initial mesh / color-only94.0991.1992.36
SPIRE94.8994.2894.43

(3) Internal structure (Table D8). Inside the outline, one object part can lie in front of another, creating a depth boundary. SPIRE recovers more of these reference boundaries in 3D at all three tolerances. The geometry updates therefore recover visible internal structure as well as the outer shape.

Table D8. Internal 3D boundary recovery on NeRF-Synthetic. Neighboring foreground depth changes above 2% of the nearer depth are backprojected from mesh and original-scene first-hit maps, excluding the silhouette band. Recall is the fraction of reference points recovered within each 3D distance in the same view. Distances are in scene units; eight-scene means (%).

Method[email protected] ↑[email protected] ↑[email protected] ↑
GOF69.1780.9291.46
MeshSplatting61.5373.1081.72
2DGS36.9551.4065.58
nvdiffrast64.8876.6185.64
MILo62.6575.0984.35
RaDe-GS35.7144.0651.85
RaDe-GS unfiltered TSDF54.4970.4183.22
Initial mesh / color-only66.4978.8390.79
SPIRE73.2784.6193.44

(4) Image fidelity (Table D9). In Table D9, color-only and joint fitting both optimize appearance fully. Only joint fitting moves vertices. Its higher native and Blender PSNR shows that geometry updates also improve exported image quality. Together, these tests establish gains in visible shape and its rendered appearance. We will add the geometry tests to Appendix D.5.

Table D9. Geometry contribution from main-paper Table 4(a). Mean NeRF-Synthetic PSNR in dB. Both versions start with the same mesh and its initial colors; Table D4 resets colors to gray. Color-only optimizes all spherical-harmonic coefficients with fixed geometry; joint also updates vertices. PSNR is native; Blender PSNR evaluates export.

Fitting modePSNR ↑Blender PSNR ↑
Color-only29.0926.78
Joint geometry and appearance32.2829.03
Reviewer_YLzSOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes SPIRE (Surface-aware Pixel-area Integration REnderer), a differentiable mesh renderer designed to reduce the export gap between native novel-view synthesis quality and the rendering fidelity of meshes exported to standard graphics software. The authors identify the difficulty of obtaining meaningful geometry gradients from hard surface boundaries and occlusion as a key obstacle to directly optimizing opaque triangle meshes. Motivated by the observation that real cameras integrate radiance over finite pixel areas, SPIRE introduces analytic triangle-pixel integration, which computes exact coverage and mean color over clipped triangle footprints. This formulation enables geometry optimization without softening mesh boundaries, introducing semitransparency, or relying on surrogate edge gradients. To handle visibility, the authors further propose surface-aware compositing, which accumulates adjacent faces belonging to the same surface without mutual attenuation while preserving occlusion between depth-separated surfaces. Rather than explicitly tracking complex visible regions, the method approximates the remaining visible pixel footprint through moment matching using its spatial mean and variance. A two-sweep sparse backward pass efficiently computes gradients with respect to mesh geometry and spherical harmonic appearance. For multiview reconstruction, the authors initialize a shared-vertex mesh using Gaussian Opacity Fields (GOF) and jointly optimize vertex positions and spherical harmonic coefficients under RGB supervision, while keeping mesh connectivity fixed and all faces fully opaque. The optimized mesh is subsequently exported with view-independent vertex colors for rendering in standard graphics pipelines. Quantitative and qualitative experiments are conducted on NeRF-Synthetic, Mip-NeRF 360, and multiscale benchmarks, evaluating both native rendering quality and post-export fidelity in Blender and Filament. The results demonstrate that SPIRE achieves the highest post-export rendering fidelity among the evaluated methods while maintaining a relatively small export gap. The authors also provide ablation studies examining the contributions of analytic pixel-area integration, surface grouping, residual-footprint propagation, and regularization, along with controlled comparisons against existing differentiable renderers under identical mesh initialization. Downstream mesh-editing, deformation, and physics-simulation experiments demonstrate the practical utility of the reconstructed meshes for standard graphics workflows.

Critical strengths

1. Well-motivated and practically important problem.

The paper addresses the export gap between native novel-view synthesis quality and the fidelity of meshes rendered in standard graphics software. Unlike methods primarily targeting native rendering performance, SPIRE focuses on producing reusable triangle meshes, making the research relevant to downstream graphics applications. The proposed analytic pixel-area integration provides meaningful geometry gradients without introducing semitransparency, softened boundaries, or surrogate edge gradients. The combination of same-surface accumulation and spatial transmittance propagation addresses both mutual attenuation between adjacent triangles and occlusion between depth-separated surfaces.

2. Strong quantitative results across multiple benchmarks.

SPIRE achieves the highest post-export rendering fidelity among the evaluated methods on NeRF-Synthetic and Mip-NeRF 360, with consistent improvements across Blender and Filament. The evaluation also covers multiple rendering configurations and image resolutions, supporting the robustness of the reported improvements. Beyond reconstruction quality, the paper evaluates mesh deformation, scene editing, and physics simulation, which shows its applicability to downstream graphics workflows.

Critical weaknesses

1. Strong dependence on mesh initialization and fixed topology.

Although SPIRE directly optimizes mesh geometry, it relies on GOF to generate the initial mesh and preserves vertex connectivity and vertex count throughout optimization. The method also cannot explicitly recover missing topology or adaptively allocate vertices to geometrically complex regions.

2. Large computational and memory requirements.

Although the exported meshes support efficient rendering, the optimization process is relatively expensive. On NeRF-Synthetic, SPIRE requires approximately 78.3 minutes including GOF initialization, compared with 27.0 minutes for MeshSplatting. Its native renderer also operates at 37 FPS, substantially slower than MeshSplatting's 171.1 FPS, and its training representation requires considerably more memory.

3. Limited improvement in geometric reconstruction accuracy.

Despite the substantial gains in post-export rendering fidelity, the DTU geometry evaluation (Table 13) reports a mean Chamfer distance of 0.80, comparable to 2DGS and MeshSplatting but worse than GOF (0.74), RaDe-GS (0.68), and MILo (0.68). This suggests that the improved visual fidelity does not necessarily translate into more accurate 3D geometry. The distinction is particularly important given the paper's emphasis on downstream geometry processing and simulation.

Questions for authors

The weaknesses I wrote are probably elements that are difficult to address in a rebuttal. However, I am more inclined to accept it because I believe this paper is well-developed and addresses the core issue of GS to mesh export. Therefore, there is no need to show additional experiments in the rebuttal, but please double-check the citations (bibtex) to confirm which papers were accepted by specific conferences. Some papers published in conferences are listed as arXiv:

  • Meshsplatting arXiv:2512.06818 → CVPR 2026

  • Neus arXiv:2106.10689 → NeurIPS 2021

  • Mesh splatting for end-to-end arXiv:2601.21400 → ICLR 2026

3 comments
AuthorsOpenReview ↗

Official comment for reviewer YLzS (part 1/2: W1, W2)

Thank you for recognizing the importance of reducing the export gap, the consistent benchmark results, and the downstream uses of the meshes. Your comments on initialization, cost, geometry, and citations help improve the paper. We will incorporate the clarifications, results, and citation corrections addressing all your points into the camera-ready version. We answer Weaknesses W1-W3 and Question Q1 in review order.


[W1] Initialization and fixed topology

SPIRE retains high export quality across different coarse initializations and mesh sizes. We show (1) that GOF is one initialization choice by comparing different starting meshes (Table Y1), and (2) that SPIRE's advantage persists after substantial simplification by comparing methods at common face budgets (Table Y2).

(1) High quality across initializers (Table Y1). GOF is an initialization choice, not a requirement for SPIRE's high export quality. Using meshes from GOF, Gaussian Blending, Gaussian Surfels, or NeuS2 gives similar mean Blender PSNR, ranging from 27.82 to 28.10 dB. This shows that SPIRE retains high export quality when starting from different coarse initial meshes. Joint fitting also outperforms color-only fitting for each initializer, demonstrating the additional benefit of geometry updates. The initializer supplies coarse positions, shapes, and connectivity; SPIRE optimizes vertex positions and appearance while preserving connectivity and vertex count.

Table Y1. Refinement from different initializers. We initialize SPIRE with meshes from each listed method, reset their vertex colors to the same neutral gray, and measure mean Blender PSNR in dB on NeRF-Synthetic. Color-only fixes vertices; joint fitting updates vertices and appearance.

Initial mesh sourceColor-only PSNR ↑Joint PSNR ↑
GOF26.3928.10
Gaussian Blending23.7627.89
Gaussian Surfels22.8528.08
NeuS226.5227.82

(2) Coarser meshes (Table Y2). The GOF baseline averages 3.20M faces; our standard initialization already subsamples GOF before extraction to obtain 2.10M faces (paper Table 14). At further-reduced budgets of 1M, 0.5M, and 0.25M, SPIRE retains higher export quality than GOF, MeshSplatting, and Triangle Splatting. The initializer supplies components, holes, and the vertex budget; SPIRE optimizes from that coarse structure. We will add these studies to Appendix D.2 and the setup to Appendix C.

Table Y2. Comparison at common maximum face budgets. Mean Blender PSNR in dB on NeRF-Synthetic, with maximum face budgets from 0.25 million to 1 million faces (M denotes million). These limits apply during extraction for GOF and during mesh optimization for the other methods. All exported meshes use vertex colors as emission colors in Blender.

Method0.25M ↑0.5M ↑1M ↑
GOF21.5923.1324.50
MeshSplatting26.8726.8826.84
Triangle Splatting19.8018.8217.89
SPIRE27.7627.9028.00

[W2] Reconstruction cost and the exported asset

SPIRE's reconstruction effort produces a high-fidelity mesh for efficient reuse in standard graphics systems. Three points matter: (1) the exported mesh uses standard graphics renderers such as Blender and Filament, (2) export reduces model storage by about 83% (paper Table 14), and (3) smaller meshes retain strong quality (Table Y2), with initializer flexibility (Table Y1).

(1) Deployment value. The goal of reconstruction is a mesh that can be reused in standard graphics software. Its image quality and rendering speed in standard renderers, such as Blender and Filament, determine the final asset's value. Native FPS measures SPIRE during optimization; the exported asset runs in the target renderer without SPIRE. The paper's Blender and Filament evaluations show that the learned image fidelity transfers to this deployment setting.

(2) Compact exported storage. About 74% of SPIRE's training-model storage holds spherical-harmonic (SH) appearance coefficients. Export replaces them with vertex colors and removes auxiliary data such as per-face importance and projected-size statistics. Storage falls from 273.1 MB to 47.3 MB. We will clarify this distinction in Appendix D.6.

(3) Resource choices. Reducing the face budget from 1M to 0.25M changes Blender PSNR only from 28.00 to 27.76 dB, retaining higher quality than the compared methods at common budgets. This supports compact assets without giving up SPIRE's export advantage. Table Y1 also shows that GOF is one initialization choice, allowing future efficiency work to use an initializer suited to available resources.

AuthorsOpenReview ↗

Official comment for reviewer YLzS (part 2/2: W3, Q1)

[W3] Geometric accuracy and downstream use

SPIRE improves visible geometry for downstream use as well as image fidelity. SPIRE combines visible face portions to better represent fine or curved scene structures within the given coarse topology. Chamfer distance (CD) also measures hidden portions, so it does not fully reflect this improvement. We show that SPIRE (1) matches object silhouettes more closely (Table Y3) while (2) placing their boundaries more accurately in 3D (Table Y4), (3) recovers more depth boundaries between overlapping object parts (Table Y5), and (4) improves image quality through geometry optimization beyond color fitting alone (Table Y6).

(1) Visible shape (Table Y3). The color-only control fixes the initial geometry, so changes in silhouette require moving the faces. SPIRE improves coverage and outline alignment in held-out views. Its geometry thus better represents the object shape seen by new cameras, beyond fitting colors.

Table Y3. NeRF-Synthetic silhouettes and contours. Held-out views use a common ray caster and reference image alpha masks. Intersection-over-union (IoU) measures mask overlap; contour F-score measures outline alignment at each pixel tolerance. RaDe-GS retains its largest connected component; unfiltered TSDF retains all.

MethodSilhouette IoU ↑Contour F@1px ↑Contour F@2px ↑Contour F@4px ↑
GOF0.9780.9360.9560.971
MeshSplatting0.9820.9340.9600.980
2DGS0.6480.3290.4430.531
nvdiffrast0.9740.9160.9390.958
MILo0.9630.8260.9170.963
RaDe-GS0.7010.4040.5180.655
RaDe-GS unfiltered TSDF0.9290.5370.6980.877
Initial mesh / color-only0.9760.9170.9440.963
SPIRE0.9820.9580.9720.983

(2) Boundary depth (Table Y4). We next check whether these outlines lie at the correct depth. SPIRE improves both precision and recall over the initial mesh and baselines in 3D. Thus, the visible boundary is more accurately placed in space, beyond matching its image projection.

Table Y4. Outer-boundary 3D accuracy on NeRF-Synthetic. Mesh and original-scene silhouettes are backprojected using first-hit depths from identical rays. Points match within 0.02 scene units in the same view. Precision measures matched predictions; recall measures recovered reference points; F1 balances both. Eight-scene means (%).

MethodOuter precision ↑Outer recall ↑Outer F1 ↑
GOF94.4893.0193.59
MeshSplatting94.7993.1793.79
2DGS47.8545.0343.47
nvdiffrast92.0990.6691.10
MILo92.1592.4891.85
RaDe-GS75.6553.7058.64
RaDe-GS unfiltered TSDF81.1185.8782.72
Initial mesh / color-only94.0991.1992.36
SPIRE94.8994.2894.43

(3) Internal structure (Table Y5). Inside the outline, one object part can lie in front of another, creating a depth boundary. SPIRE recovers more of these reference boundaries in 3D at all three tolerances. The geometry updates therefore recover visible internal structure as well as the outer shape.

Table Y5. Internal 3D boundary recovery on NeRF-Synthetic. Neighboring foreground depth changes above 2% of the nearer depth are backprojected from mesh and original-scene first-hit maps, excluding the silhouette band. Recall is the fraction of reference points recovered within each 3D distance in the same view. Distances are in scene units; eight-scene means (%).

Method[email protected] ↑[email protected] ↑[email protected] ↑
GOF69.1780.9291.46
MeshSplatting61.5373.1081.72
2DGS36.9551.4065.58
nvdiffrast64.8876.6185.64
MILo62.6575.0984.35
RaDe-GS35.7144.0651.85
RaDe-GS unfiltered TSDF54.4970.4183.22
Initial mesh / color-only66.4978.8390.79
SPIRE73.2784.6193.44

(4) Image fidelity (Table Y6). In Table Y6, color-only and joint fitting both optimize appearance fully. Only joint fitting moves vertices. Its higher native and Blender PSNR shows that geometry updates also improve exported image quality. Together, these tests establish gains in visible shape and its rendered appearance. We will add the geometry tests to Appendix D.5. Editing, deformation, and simulation illustrate downstream use.

Table Y6. Geometry contribution from main-paper Table 4(a). Mean NeRF-Synthetic PSNR in dB. Both versions start with the same mesh and its initial colors; Table Y1 resets colors to gray. Color-only optimizes all spherical-harmonic coefficients with fixed geometry; joint also updates vertices. PSNR is native; Blender PSNR evaluates export.

Fitting modePSNR ↑Blender PSNR ↑
Color-only29.0926.78
Joint geometry and appearance32.2829.03

[Q1] Publication venues and BibTeX

We verified the publication venues and will update the outdated arXiv BibTeX entries to cite the published conference versions.

Thanks for the authors’ response. I have read it as well as the other reviews.

The rebuttal has addressed my questions.

Since I was already in favor of this submission, I will maintain my previous rating 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Response to the Reviewers

Thank you all for your thoughtful reviews and for taking the time to discuss the additional results with us. We are glad that our responses helped address the main concerns you raised. If you feel that your concerns have been resolved, we would be grateful if you would consider raising your rating.

We have incorporated your feedback and the additional experiments into the revised manuscript. For the camera-ready version, we plan to use the additional main-text space to present the key visual comparisons alongside our main findings and make the paper easier to follow. The current revision includes finite-difference gradient checks in Appendix B.3 and comparisons across initial meshes and face budgets in Appendix D.2. Appendix D.5 brings together the requested geometry evaluations, including the normal and diffuse-shading comparisons in Figures 6 and 7 and the Chamfer distance, silhouette, and 3D boundary results in Tables 19–22.

Thank you again for helping us improve the paper. We enjoyed discussing SPIRE with you and hope you found the work interesting as well.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.