acceptodds
Under review as a conference paper at ICLR 2027

VolS-GS: Relightable Gaussian Splatting With Volumetric Subsurface Scattering

Abstract

We present VolS-GS, a relightable Gaussian splatting framework that reconstructs objects from one-light-at-a-time (OLAT) captures and renders them under novel lighting and viewpoints. Relightable Gaussian Splatting methods typically model appearance independently at each primitive, which makes non-local effects difficult to represent. This limitation is particularly apparent for subsurface scattering, where light entering the object at one location can emerge at another. Rather than modeling this effect solely with a neural network or a local kernel at each primitive, we use the spatial support of the Gaussian scene as the domain of a differentiable finite-volume transport solver, so that light can propagate through the object's interior. A small network predicts scattering and absorption coefficients for each Gaussian, and the solve redistributes incident light through the resulting field. The coefficients are fit to images rather than measured, so the solve supplies a transport-shaped path for aggregating per- primitive appearance, not a measurement of the material. To keep the learned shadow and specular terms from taking over the other components, our shadow term is predicted from visibility together with the transmittances and the scattering the solve produces, and a regularizer suppresses specular highlights in regions that shadow term predicts to be unlit. Experiments on three OLAT benchmarks show that VolS-GS consistently improves relighting quality on held-out lights and views.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Relighting, 3D Reconstruction, Novel View Synthesis, OLAT, Relightable 3DGS

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_LKnBOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary
  1. VolS-GS decomposes Gaussian appearance into Lambertian diffuse reflection, anisotropic spherical-Gaussian specular reflection, and volumetric subsurface scattering (SSS), with shadow attenuation applied to direct reflection. This separates surface reflection from internal light transport to improve relighting of translucent objects.
  2. A differentiable finite-volume diffusion solver operates on occupancy derived from the Gaussian scene, enabling light propagation across primitives without mesh extraction. Its output is combined with a separate single-scattering term to model the subsurface response.
  3. The shadow model reads cues from the same transport field, while a shadow-aware specular regularizer reduces compensation between appearance branches. Experiments on three benchmarks report improved relighting quality over prior methods.
Critical strengths
  1. The volumetric solver explicitly couples light transport across Gaussians through the occupied medium, capturing nonlocal SSS that depends on object connectivity rather than only per-primitive appearance. A controlled substitution with a per-primitive dipole model reduces PSNR in six of seven scenes, supporting the benefit of this spatial coupling (Appendix G).
  2. The transport domain is derived directly from Gaussian occupancy, requiring neither mesh extraction nor separate trainable grid parameters. Adjoint differentiation avoids storing solver iterations, allowing the transport model to be optimized from images. The method also reports 64.5 FPS on five synthetic SSS objects at 256² resolution on an A6000, after transfer-basis precomputation (Appendix C).
  3. The shadow branch uses the same transport cues as SSS, while shadow attenuation is applied only to direct reflection to avoid attenuating the scattered contribution twice. A shadow-aware regularizer discourages the specular branch from absorbing transported light. Removing transport cues lowers Statue PSNR from 38.35 to 35.09 dB, while qualitative ablations support cleaner component separation (Table 2 and Appendix F).
Critical weaknesses
  1. The SSS transport domain is filled from Gaussians optimized for image reconstruction, whose surface positions and spatial extents may not accurately represent the physical boundary. Such errors could bridge narrow gaps, omit thin connections, or misclassify interior regions, altering SSS propagation. Their impact on the filled geometry and transport results is not directly evaluated (Sec. 2.3).
Questions for authors
  1. Have you considered the geometric inaccuracies of the learned 3DGS representation when constructing the filled transport domain? Please clarify how surface and connectivity errors are handled and evaluate their effects on the interior filling and SSS results.
3 comments
AuthorsOpenReview ↗

First Rebuttal for Reviewer LKnB - W1/Q1

We sincerely thank Reviewer LKnB for the precise and thorough summary, and for the acknowledgement of the spatial coupling through the medium, the mesh-free transport domain and the shared transport cues of the shadow branch. The question about geometric error is an important one, and we address the weakness and the question below.

Reviewer LKnB - W1: Effect of geometric errors on the filled domain and on SSS

"Such errors could bridge narrow gaps, omit thin connections, or misclassify interior regions" & "Their impact on the filled geometry and transport results is not directly evaluated."

We agree that the transport domain inherits any geometric error of Gaussians, so its effect requires a direct evaluation which we now provide.

Effect on the filled domain. We voxelize the ground-truth solid from the original Blender meshes onto the same 64364^3 grid and compare it with our domain cell by cell (Table 13). Please note that real captures have no ground-truth geometry, so this comparison uses the three synthetic scenes whose source meshes are available. Thin parts are kept (78 to 97% of the cells in parts under about three cells thick), and narrow gaps are rarely bridged (0 to 8.6% of the gap cells, excluding the one-cell band around the surface that the footprint covers by design). The largest disagreement is on Candle, whose underside faces away from all cameras and is therefore sparsely covered by Gaussians, so some interior cells near the base fall outside the domain (recall 0.84). Table 14 shows that this has no measurable cost. The domain deliberately extends about one cell past the surface (Table 11 in Appendix H), which accounts for part of the precision gap at this resolution.

Table 13. Learned domain against the ground-truth solid (64364^3).

SceneIoURecallPrecisionThin parts kept ↑\uparrowGaps bridged ↓\downarrow
Statue0.670.910.7189%8.6%
Candle0.730.840.8578%0.5%
Soap0.800.960.8397%0.0%

Effect on SSS. To measure the cost of the remaining geometric error, we replace our domain with the ground-truth solid and retrain from scratch, using it for both training and testing with everything else unchanged. The ground-truth domain changes PSNR by at most 0.17 dB on the three scenes, with no consistent sign (Table 14), even on Statue, whose learned domain agrees least with the ground truth. The ground-truth solid also lacks the one-cell extension past the surface that our domain has by design. The remaining geometric error therefore has no measurable cost on these scenes, because the diffused light varies smoothly and is insensitive to small errors of the domain boundary, while the per-Gaussian SSS amplitude kisss\mathbf{k}^{\mathrm{sss}}_i adjusts the fine detail on top of it.

Table 14. Transport domain learned from the Gaussians vs. the ground-truth solid, test PSNR (dB).

Domain of VolS-GSStatueCandleSoap
Learned from the Gaussians (Ours)38.3543.6441.01
Ground-truth solid38.1843.5041.12

Reviewer LKnB - Q1: How such surface and connectivity errors are handled

"Please clarify how surface and connectivity errors are handled"

We build the domain (Sec. 3.3, Sec. 2.3 in the submission) such that the typically expected errors of Gaussians fitted to images barely change its topology.

  • Holes and noisy surfaces. The interior is filled from the oriented Gaussians themselves (Sec. 3.3), which tolerates holes and noise and does not require a watertight surface. The fill is normalized by a high quantile over the grid, which keeps a few stray Gaussians from shifting it.
  • Thin parts. Thin structures are kept by adding the cells that contain a Gaussian center, dilated by one cell. Without this footprint, PSNR drops by 1.5 to 4.0 dB (Table 11).
  • Floaters and gaps. The grid box is set from the 0.2 to 99.8% quantiles of the Gaussian centers, so distant floaters do not stretch the grid, and a floater inside it stays an isolated cell that exchanges no flux with the object. The source march weights path length by occupancy, so empty space between separate parts adds no extinction (Eq. 3).
  • Drift during training. The domain is rebuilt from the current Gaussians whenever densification changes their count and at least every 300 iterations. Therefore, it follows the geometry as it improves.

Their effects on the interior filling and on SSS are evaluated in our answer to W1 above (Tables 13 and 14).

Tables 13 and 14 are in the new Appendix H, and Appendix B describes the grid box, the footprint union, the dilation and the occupancy cutoff. Thank you again for this insightful question. We hope the rebuttal answers the reviewer’s concern and would be glad to provide further analysis if needed.

The Authors

Thank you for addressing my concerns and clarifying the points I raised. I will maintain my Clear acceptance rating.

AuthorsOpenReview ↗

Official Comment by Authors to Reviewer LKnB

Dear Reviewer LKnB,

Thank you for the constructive suggestions and for maintaining your Clear acceptance rating. We are glad that your concerns have been addressed. Thank you for the time and care that you put into reviewing VolS-GS.

Best regards,

The Authors

Reviewer_D8PeOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper Volumetric GS reconstructs objects with translucent surface elements. We describe the effect of light entering a surface and existing at a different location as subsurface scattering. It is often approximated, as the integral over all incoming sub surface scattered directions requires to path trace rays inside the object, making this computation expensive. While predecessor methods already model this effect for Gaussian Splatting, they only approximate the effect with surface level estimators, this work tries to model the volumetric process through a corse grid.

The paper is excellent written, with lots of thought / care and focus on every detail that matters. I really like the methodological approach and the results look amazing. Given from the state of the submission, I would recommend accepting the paper.

Critical strengths
  • The method is superb, I really like the idea of utilising winding numbers in order to detect the area that is covered by volume. I also like the idea of not optimising the field itself directly, but make it a propagation of first surface gaussians to voxels to spread coefficients and later neighbouring voxels to "flow light" through the object.

  • The results look very good, the reconstruction nicely capture the subsurface scattering effect and also the specularity of materials. Over the board this method performs qualitatively and quantitatively nicely.

Critical weaknesses
  • I'm not a big fan of deferring related work to the appendix. Its a critical part of the paper, it feels a bit like the authors cheated to get more space.
Questions for authors
  • Does this method allow for modelling of inner structure? Imagine having a hand lit by light with veins and arteries. Would this method be able to model the geometric "hard" surfaces of these in substance geometry? While still modelling SSS correctly? In your training strategy you describe that this

  • Did you try higher resolution grids? Whats the impact?

  • Your method requires build and propagation time, whats the time impact compared to other methods?

  • Why does SSS-GS seem suddenly to perform so bad for Pixiu and Cat (it seems like there is a camera issue)

4 comments
AuthorsOpenReview ↗

First Rebuttal for Reviewer D8Pe (1/2) - W1/Q1/Q2

We thank Reviewer D8Pe for the positive assessment and the thoughtful questions. We are glad the reviewer acknowledges VolS-GS's contributions and, in particular, the idea of propagating from the surface Gaussians to voxels instead of optimizing the field directly. Our answers to the weakness and the four questions follow, with Q3 and Q4 in Part 2.

Reviewer D8Pe - W1: Related work in the appendix

"I'm not a big fan of deferring related work to the appendix." (W1)

We agree that reading the related work before the method and the experiments helps the reader place our approach among prior work and follow the later sections. The revised paper moves it back into the main text (Sec. 2) without exceeding the page limit.


Reviewer D8Pe - Q1: Inner structure such as veins

"Does this method allow for modelling of inner structure?" & "While still modelling SSS correctly?" (Q1)

We would like to say yes, in principle. The finite-volume operator takes per-cell coefficients, so a vein could enter as a region with its own σs\sigma_s and σa\sigma_a, or as internal faces with their own boundary term. Our current model does not do this. Its domain has one boundary, the outer surface recovered from the Gaussians, and every interior cell takes the coefficients of the nearest surface cell (Sec. 3.3).

The harder part is identifiability. From outside, an internal structure shows up only through the diffused light it modulates, and that light is blurred over the transport length, so its exact shape would be hard to recover. Since the domain and the medium both come from the Gaussians, transport that follows Gaussians driven by a deformable template could also be explored. With anatomical priors, this points toward relightable animatable avatars, which the new Appendix J discusses.


Reviewer D8Pe - Q2: Higher-resolution grids

"Did you try higher resolution grids? Whats the impact?" (Q2)

We tested this directly by training 32332^3 to 1283128^3 grids on three scenes. Table 10 (in our reply to Reviewer XXch and in Appendix H) reports PSNR, rendering speed and memory. From 64364^3 to 1283128^3, the frame rate falls to 42 to 46% and the memory held while rendering roughly triples, while PSNR shows no consistent gain. Table 11 adds a 1283128^3 grid with the footprint dilation widened to keep its world-space width, and it too stays below 64364^3 on both scenes.

The grid stores the fluence, which diffusion spreads over the transport length. This is the non-local part that per-primitive-only models cannot represent, and because it varies smoothly, 64364^3 already resolves it. Each Gaussian reads the fluence at its own position, so the image keeps the resolution of the Gaussians rather than that of the grid. We therefore keep 64364^3 for every scene.

The rebuttal is continued in the next comment.

The Authors

AuthorsOpenReview ↗

First Rebuttal for Reviewer D8Pe (2/2) - Q3/Q4

Reviewer D8Pe - Q3: Time impact of building and propagating

"Your method requires build and propagation time, whats the time impact compared to other methods?" (Q3)

We measured the two parts separately and compared the full frame with the other methods.

Propagation. The transport solve runs at every frame and is the largest cost our method introduces. The submission already timed each stage of a frame (Table 3, now in Appendix B), and we repeat it below. The solve takes about half of the frame, 8.2 of 14.8 ms on Bunny, 8.9 of 16.3 ms on Statue and 23.8 of 47.1 ms on Drums.

Table 3. Time per stage of one frame on one A6000 (ms). The number of Gaussians is given in parentheses.

StageBunny (82k)Statue (108k)Drums (550k)
Visibility rasterization3.483.648.18
Shading, pre-solve0.961.257.20
Transport solve8.198.9223.81
Composition, post-solve0.380.492.41
Primary rasterization1.801.985.46
Frame14.8116.2947.07

Building. Building the domain is not a per-frame cost. It runs once when a model is loaded, and during training every 300 iterations and whenever densification changes the Gaussian count (Algorithm 1). We newly measured one build at 0.76 s on Bunny, 1.05 s on Statue and 1.38 s on Pixiu.

Compared with other methods. Table 12, newly added to Appendix H and also reported to Reviewer XXch, measures all five methods on the same A6000. Even with the solve, VolS-GS renders at 30 to 64 FPS. GS$^3$ and SSD-GS, which shade each Gaussian locally, are faster, and so is RNG on three of the four scenes. VolS-GS is 4.8 to 74 times faster than SSS-GS, the other method that models subsurface scattering explicitly.

Table 12. Rendering speed (FPS, higher is better) / GPU memory while rendering (GiB, lower is better) on one A6000.

MethodBunnyStatuePixiuTranslucent
GS$^3$169.4 / 0.11141.9 / 0.13117.3 / 0.17135.0 / 0.30
SSD-GS97.6 / 0.1478.7 / 0.2069.5 / 0.3044.3 / 0.77
RNG83.8 / 1.0382.7 / 1.0230.6 / 2.9433.9 / 8.38
SSS-GS13.3 / 0.155.3 / 0.270.67 / 1.021.08 / 0.84
VolS-GS (Ours)64.1 / 0.3758.3 / 0.4349.7 / 0.5430.1 / 1.13

Reviewer D8Pe - Q4: Why SSS-GS performs poorly on Pixiu and Cat

"Why does SSS-GS seem suddenly to perform so bad for Pixiu and Cat" (Q4)

On the same camera files, GS$^3$, RNG, SSD-GS and VolS-GS all train normally (27.76, 26.77, 27.95 and 28.11 dB on Cat), which points to a failure specific to SSS-GS. It is confined to the four scenes from a separate capture set, Cat, Fish, FurScene and Pixiu. These are 5122512^2 with focal lengths of 3100 to 3500 px (an 8 to 9.5° field of view) and principal points 35 to 130 px off center. The three 102421024^2 captures, at 31∘31^\circ and nearly centered, do not show this failure.

On these captures SSS-GS does not optimize stably. On Cat its Gaussian count runs away early (1.68M by iteration 10k), and on Pixiu its training repeatedly hits non-finite gradients.

SSS-GS rasterizes with a centered projection but unprojects its depth buffer with the true principal point. Making the two consistent did not change the training trajectory of Cat over the first 12k iterations, and neither a densification cap nor a wider EWA guard helped. We could not isolate the mechanism further. We therefore report these rows as measured, as stated in Sec. 4.2. SSS-GS trains normally on the other three NRHints captures and on its own benchmark, where VolS-GS is ahead on all eight scenes (Table 1). Our margin over SSS-GS does not rest on these four failures.


Thank you again for the encouraging review and the careful questions. We would be happy to answer anything further.

Best regards,

The Authors

Sorry for the late response, the rebuttal did not change my score. I still think this is a very good written paper, showing good results and an interesting method. I thank the authors for taking the time answering my questions and running new experiments. I vote for acceptance of this paper. Regarding modelling inner structures, I would like to see this as a follow up work.

AuthorsOpenReview ↗

Official Comment by Authors to Reviewer D8Pe

Dear Reviewer D8Pe,

Thank you for the follow-up and for supporting acceptance. Your review led to several improvements in the revised paper. We hope to extend your suggestions in future work. Thank you for the time and care that you put into the review.

Best Regards,

The Authors

Reviewer_XXchOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses subsurface scattering (SSS) in Gaussian-based object relighting. Its central idea is to couple scattering responses across spatial locations through a volumetric diffusion solve. The volume is constructed from Gaussian occupancy, and its medium coefficients are learned from images. The scattering response is combined with surface reflection and learned shadow attenuation. Transport cues and a specular regularizer help separate these components. Experiments compare image quality on three OLAT benchmarks and evaluate the main components through ablations.

Critical strengths
  1. The volumetric formulation accounts for the object's spatial connectivity when modeling subsurface scattering. This provides a geometric constraint that independent per-Gaussian scattering estimates lack.

  2. The method improves performance across three OLAT benchmarks.

Critical weaknesses
  1. The paper does not separately report PBR NVS under the original illumination or quantitative accuracy for intermediate attributes. These results would complement the relighting evaluation and help assess reconstruction and decomposition quality.

  2. The voxel grid is central to the transport model. A more detailed ablation of its construction and settings, including resolution, would help clarify their effects on image quality and computational cost.

Questions for authors
  1. Could you supplement the evaluation with PBR NVS under the original illumination and reference comparisons for selected intermediate attributes, such as albedo and normals? Results on representative synthetic scenes would be helpful.

  2. Could you add a targeted study of voxel-grid construction and settings, including resolution, on representative scenes? Image quality, rendering speed, and GPU memory results would help illustrate the tradeoff. Reporting speed and GPU memory for the compared methods under matched conditions would also be useful.

  3. Equation (4) uses the light and view directions to predict scattering and absorption coefficients. When the lighting changes, do you predict these coefficients again and update the matrix used to solve light transport? Could you explain how this relates to the statement in Section 2.4 that the operator does not depend on the light?

6 comments
AuthorsOpenReview ↗

First Rebuttal for Reviewer XXch (1/3) - W1/Q1

We sincerely thank the Reviewer XXch for the careful reading and the detailed, constructive suggestions. We are encouraged that the reviewer found that our volumetric formulation accounts for the object's spatial connectivity, which independent per-Gaussian estimates lack, and that it improves performance on all three OLAT benchmarks.

In order to make the submission easier to follow, the revision moves the related work into the main text (Sec. 2), adds a notation table (Appendix A), points each part of Figure 3 to its section and uses unified terminology for each component across the text. We answer the first weakness and question here, and the remaining points in parts (2/3) and (3/3).

Reviewer XXch - W1a/Q1a: Novel-view synthesis under the original illumination

"PBR NVS under the original illumination" (W1a) & "Results on representative synthetic scenes would be helpful." (Q1a)

We thank the reviewer for this suggestion, as results under the illumination seen in training complement the relighting evaluation. Strictly speaking, OLAT capture has no "original" illumination. The closest analogue is the set of test views whose camera is not in the training set but whose light is. Only the SSS-GS synthetic benchmark has such views, 56 to 87 per scene, while in the other benchmarks every test view is either lit by a light unseen in training or taken from a training camera. On these views, VolS-GS is best on all five scenes in PSNR, SSIM and LPIPS (Table 8, Appendix H).

Table 8. Novel-view synthesis under training illumination, PSNR (dB) / SSIM / LPIPS.

MethodBunnyCandleDragonSoapStatue
GS$^3$34.92 / 0.9818 / 0.024238.28 / 0.9878 / 0.017536.07 / 0.9766 / 0.025937.84 / 0.9845 / 0.024935.79 / 0.9833 / 0.0189
SSS-GS36.25 / 0.9806 / 0.032538.22 / 0.9864 / 0.023134.62 / 0.9589 / 0.054635.57 / 0.9738 / 0.030436.04 / 0.9785 / 0.0351
RNG35.36 / 0.9865 / 0.017940.01 / 0.9935 / 0.008134.72 / 0.9718 / 0.030937.89 / 0.9886 / 0.011529.60 / 0.9660 / 0.0353
SSD-GS39.18 / 0.9910 / 0.013041.77 / 0.9941 / 0.008137.59 / 0.9828 / 0.020337.55 / 0.9878 / 0.018636.49 / 0.9867 / 0.0163
VolS-GS (Ours)40.13 / 0.9922 / 0.011543.47 / 0.9959 / 0.006537.77 / 0.9836 / 0.018940.15 / 0.9929 / 0.008338.93 / 0.9915 / 0.0113

Reviewer XXch - W1b/Q1b: Intermediate attributes against a reference

"quantitative accuracy for intermediate attributes" (W1b) & "reference comparisons for selected intermediate attributes, such as albedo and normals" (Q1b)

We agree that comparing intermediate attributes against a reference is the most direct way to assess their accuracy. For the three scenes whose Blender files are available (Statue, Candle and Soap), we render ground-truth normals and compare on ten evenly spaced test views per scene. Normal has an unambiguous definition and is important for both appearance and geometry. It enters the surface shading (Eq. 2) and sets where the SSS term exits, one cell inward along the normal (Eq. 7). We compare against SSD-GS, the closest baseline, which shares our normal parameterization, so the difference does not come from the parameterization.

VolS-GS lowers the mean error from 13.6° to 10.4°, with the largest gains on Statue and Soap (Table 9), although it has no loss term dedicated to the normals (Eq. 13). The gain therefore comes from the appearance model rather than from a normal prior. Figure 12 in Appendix H shows the normals and their error maps, where VolS-GS keeps even the embossed logo of Soap that SSD-GS smooths away.

Table 9. Mean angular error of normals (degrees, lower is better).

MethodStatueCandleSoapMean
SSD-GS11.6015.0914.2013.63
VolS-GS (Ours)9.2615.196.7610.40

We considered adding an albedo comparison, but a single number would be misleading for every method that separates a subsurface term, including SSD-GS and SSS-GS. In the Blender files, the wax, soap and statue materials have Subsurface Weight 1, so all non-specular reflection goes through the subsurface lobe and there is no Lambertian lobe. Base Color sets the color of that lobe, whereas ai\mathbf{a}_i is the coefficient of our local Lambertian term, with the light transported through the volume carried separately. Images constrain only the sum of the two, so an error against Base Color would score how each method divides this sum, which the data do not determine. We therefore evaluate the sum itself, under the training illumination (Table 8) and under novel illumination (Sec. 4), and Appendix G shows the diffuse and SSS components of VolS-GS, SSD-GS and SSS-GS side by side.

We continue our rebuttal in the next comments.

The Authors

AuthorsOpenReview ↗

First Rebuttal for Reviewer XXch (2/3) - W2/Q2

Reviewer XXch - W2/Q2: Grid construction and settings, including resolution

"A more detailed ablation of its construction and settings, including resolution" (W2) & "Image quality, rendering speed, and GPU memory results would help illustrate the tradeoff." (Q2)

We agree that voxel grid is central to our transport model and deserves a dedicated study. The grid is built from the Gaussians and introduces no learnable parameters. Thus, its resolution and construction change how transport is discretized without changing the model capacity. The resolution study uses Bunny, Statue and Pixiu, the construction study uses Bunny and Statue, and every variant is retrained from scratch with all other settings fixed. Both are reported in Appendix H.

Ablation on Resolution (Table 10). Below 64364^3, Bunny, the scene with the strongest subsurface scattering, loses up to 0.56 dB. Above 64364^3, PSNR does not improve consistently, while at 1283128^3 the frame rate drops to 42 to 46% and the memory used while rendering roughly triples on all three scenes.

This is expected because the grid stores the fluence, which diffusion spreads over the transport length inside the medium. This non-local transport is the part that a per-primitive-only model cannot represent. Since it varies smoothly, 64364^3 already resolves it. Each Gaussian reads this field at its own position, so the image keeps the resolution of the Gaussians rather than that of the grid. We therefore keep 64364^3 for every scene.

Table 10. Grid resolution. Memory is the GPU memory allocated while rendering (GiB).

SceneMetric32332^348348^364364^3 (default)96396^31283128^3
BunnyPSNR (dB)39.5939.9340.1540.2540.27
BunnyFPS75.169.564.143.529.5
BunnyMemory0.260.300.370.651.14
StatuePSNR (dB)38.1837.8538.3537.3537.28
StatueFPS67.562.658.339.226.2
StatueMemory0.330.380.430.761.31
PixiuPSNR (dB)31.6231.6131.3831.6731.69
PixiuFPS55.352.749.733.620.9
PixiuMemory0.390.440.540.941.71

Ablation on Construction (Table 11). Our footprint keeps every cell that contains a Gaussian center, dilated by one cell, so thin parts such as the ears of Bunny stay in the domain. Dropping it costs 1.5 to 4.0 dB. Without the dilation, Gaussians on thin parts fall outside the domain (−0.8-0.8 to −0.9-0.9 dB), and two cells of dilation are also worse (−0.17-0.17 to −0.45-0.45 dB). The dilation is counted in cells and was set at 64364^3, where one cell is twice as wide as at 1283128^3. The last row therefore uses two cells at 1283128^3, and it still falls below the default on both scenes at less than half the frame rate. Appendix B now states the full construction, including the grid box, the occupancy of each cell and the cutoff that defines Ω\Omega.

Table 11. Domain construction. Each cell: PSNR (dB) / FPS / memory while rendering (GiB).

VariantBunnyStatue
VolS-GS (Ours)40.15 / 64.1 / 0.3738.35 / 58.3 / 0.43
Without the footprint36.18 / 72.9 / 0.2836.84 / 59.7 / 0.38
Dilation 039.38 / 67.5 / 0.3237.48 / 60.6 / 0.39
Dilation 239.70 / 62.3 / 0.4038.18 / 56.8 / 0.48
1283128^3, dilation 2 (matched width)39.84 / 28.4 / 1.3237.40 / 25.7 / 1.45
AuthorsOpenReview ↗

First Rebuttal for Reviewer XXch (3/3) - Q2/Q3

Reviewer XXch - Q2: Speed and memory of the compared methods

"Reporting speed and GPU memory for the compared methods under matched conditions" (Q2)

We thank the Reviewer XXch for asking for this comparison under matched conditions. Table 12 measures all five methods on the same A6000 at each benchmark's resolution with the protocol of Sec. 4. Each frame is timed with CUDA events, and we report the median after a 20-frame warm-up. GS$^3$ and SSD-GS shade each Gaussian locally and are faster than VolS-GS on every scene, and RNG is faster on three of the four scenes. VolS-GS is the only one of them that solves volumetric transport, which takes about half of each frame (Table 3 in Appendix B), and it still renders at 30 to 64 FPS. The closest comparison is SSS-GS, the other method that models subsurface scattering explicitly. VolS-GS is 4.8 to 74 times faster than SSS-GS, which falls below 2 FPS on Pixiu and Translucent. While rendering, VolS-GS holds 0.37 to 1.13 GiB, less than RNG on every scene, after a one-time setup of 4.1 to 7.3 GiB in the first frames.

Table 12. Rendering speed (FPS, higher is better) / GPU memory while rendering (GiB, lower is better) on one A6000.

MethodBunnyStatuePixiuTranslucent
GS$^3$169.4 / 0.11141.9 / 0.13117.3 / 0.17135.0 / 0.30
SSD-GS97.6 / 0.1478.7 / 0.2069.5 / 0.3044.3 / 0.77
RNG83.8 / 1.0382.7 / 1.0230.6 / 2.9433.9 / 8.38
SSS-GS13.3 / 0.155.3 / 0.270.67 / 1.021.08 / 0.84
VolS-GS (Ours)64.1 / 0.3758.3 / 0.4349.7 / 0.5430.1 / 1.13

Reviewer XXch - Q3: Eq. (4) and the statement that the operator does not depend on the light

"When the lighting changes, do you predict these coefficients again" & "the statement in Section 2.4 that the operator does not depend on the light" (Q3)

The reviewer is right, and this question led to a correction in the PDF since the sentence in Sec. 2.4 of the submission (now Sec. 3.4) was wrong. At every frame we evaluate Eq. (4) for the current light and view directions, splat the coefficients onto the grid and assemble FF from them. Only the domain (where cells are occupied) is rebuilt whenever densification changes the Gaussian count and otherwise every 300 iterations.

Because the coefficients depend on direction, no solve is shared across lights. The point we meant to make is that lights superpose, so a scene lit by several sources is the sum of one-light renders. In shortening the sentence, we gave a wrong reason for this and it is now corrected. Each light has its own solve, which is how Appendix C.3 relights under an environment map. The transfer basis of Appendix B only warm-starts conjugate gradients, and the eight iterations run on the current frame's operator. Against the converged solve, the warm start changes test PSNR by −0.009-0.009 to 0.0000.000 dB.

The revised Sec. 3.4 reads: "The coefficients of Eq. (4) depend on the light and view directions, so FF is assembled per frame. A scene lit by several sources is therefore rendered as a weighted sum of one-light renders, each with its own solve, which is how Appendix C.3 relights under an environment map without retraining." The superposition remark in Appendix A and the text of Appendices B and C are corrected to match. The caption of Algorithm 1 now reads: "Line 1 runs when densification changed the Gaussian count and in any case every 300300 iterations; every other step runs at every evaluation, with the coefficients of Eq. (4) evaluated for the current light and view."

We also noticed that Eq. (4) accidentally omitted the Gaussian's normal ni\mathbf{n}_i, which the implementation uses. The revised Eq. (4) is (σs,σa)(gi)=MLPmat(γ(xi),zi,ni,γ(ωl),γ(ωo))(\boldsymbol\sigma_s, \boldsymbol\sigma_a)(g_i) = \mathrm{MLP}_{\mathrm{mat}}\big(\gamma(\mathbf{x}_i), \mathbf{z}_i, \mathbf{n}_i, \gamma(\boldsymbol\omega_l), \gamma(\boldsymbol\omega_o)\big).

All reported results come from this implementation, so no number changes with these corrections.


The new analyses, the readability improvements and the corrections to Sec. 2.4 and Eq. (4) all came from this review, and we are grateful for it. We hope they address the reviewer's concerns and would appreciate it if the reviewer could reconsider the assessment for VolS-GS in light of them. We are happy to answer any further questions.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response and Summary of Revisions

Dear Reviewers, ACs and PCs,

Thank you so much for the careful reading and the constructive questions. We have taken each review carefully into account, and the expertise you shared has made VolS-GS a more solid paper. We have uploaded a revised PDF and answer each review in its own thread.

We are encouraged that reviewers found that our volumetric formulation "provides a geometric constraint that independent per-Gaussian scattering estimates lack" (XXch), "capturing nonlocal SSS that depends on object connectivity" (LKnB), and that it "improves performance across three OLAT benchmarks" (XXch). The reviewers also valued "the idea of not optimising the field itself directly, but make it a propagation of first surface gaussians to voxels" (D8Pe), the transport domain "requiring neither mesh extraction nor separate trainable grid parameters" (LKnB), and the shadow branch that "uses the same transport cues as SSS" (LKnB). D8Pe also found the paper "excellent written, with lots of thought / care and focus on every detail that matters" and the results to "look amazing".

Revisions are highlighted in green and new or moved section titles in orange. Minor wording and spelling edits are not marked. Section, table and figure numbers in our replies refer to the revised PDF, and each reply will point directly to the updated numbers.

Main Changes

  • The Related Work section is now in the main text. (Sec. 2)
  • Sec. 3.4 and Eq. (4) are now corrected regarding the medium coefficients and the operator. (Sec. 3.4 and Eq. (4))
  • The new Appendix H adds
    • novel-view synthesis under training illumination (Tab. 8),
    • normal accuracy against ground truth (Tab. 9 and Fig. 12),
    • studies of grid resolution and domain construction (Tab. 10 and 11),
    • rendering speed and memory of all methods (Tab. 12),
    • and a comparison of the learned domain with the ground-truth solid (Tab. 13 and 14).
  • Appendix A adds a notation table for improved readability, and Figure 3 points each part to its section. (Appendix A and Fig. 3)
  • The new Appendix J discusses the SSS modeling of internal structure. (Appendix J)

The revised PDF already reflects all of these changes, and our reply to each review is posted in its own thread.

Best Regards,

The Authors

AuthorsOpenReview ↗

Final Remarks by Authors

Dear Area Chair and Reviewers,

Thank you for overseeing our submission. We would like to briefly summarize the discussion.

All three reviewers replied during the discussion period. Reviewer XXch noted that "All my questions have been addressed" and supports acceptance, and Reviewers D8Pe and LKnB kept the highest rating. The main results added during the discussion are as follows:

  1. Novel-view synthesis and normals (XXch). Under the training illumination, VolS-GS is best on all five available scenes in PSNR, SSIM and LPIPS (Table 8). Against ground-truth normals, it lowers mean error from 13.6° for SSD-GS to 10.4° without introducing dedicated normal loss (Table 9).
  2. Grid resolution and construction (XXch, D8Pe). Grids that are finer than 64364^3 do not give a consistent gain since our volumetric SSS carries the smoothly varying light based on diffusion theory while each Gaussian keeps the fine detail, and FPS falls to 42 to 46% and the memory nearly triples. The footprint that keeps thin parts in the domain is worth 1.5 to 4.0 dB (Tables 10 and 11).
  3. Cost (XXch, D8Pe). VolS-GS renders at 30 to 64 FPS on one A6000 and is 4.8 to 74 times faster than SSS-GS. The transport solve takes about half of each frame, and building the transport domain is not a per-frame cost (Tables 3 and 12).
  4. Robustness of the learned domain (LKnB). Against the available ground-truth solid mesh, the learned domain keeps 78 to 97% of thin parts and bridges at most 8.6% of gaps. Replacing it with the ground-truth solid changes PSNR by at most 0.17 dB (Tables 13 and 14).
  5. Correction of Sec. 2.4 and Eq. (4) (XXch). Reviewer XXch rightly pointed out that the coefficients of Eq. (4) depend on the light. The operator is assembled per frame and not shared across lights. We corrected the corresponding description and added the omitted normal input to Eq. (4). Both describe the implementation that produced every reported result.

The discussion supports our core contribution. By letting light propagate through the interior of the Gaussian scene, VolS-GS models subsurface scattering as differentiable transport across the object, and this transport outperforms a per-primitive dipole on six of seven scenes, a result Reviewer LKnB cited as a strength (Appendix G). VolS-GS improves relighting on all three OLAT benchmarks with a transport domain built from the Gaussians themselves, without mesh extraction or learnable grid parameters.

Every analysis above came from the reviewers' questions, and we are grateful for how much they have strengthened the paper. We also thank you for your time in handling our submission.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.