- Rating
- 2(+1)
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents a two-stage inverse rendering method that combines Gaussian Splatting (GS) with mesh-based path tracing. It first reconstructs planar Gaussians using a modified VA-GS with normal and mask supervision, and then extracts a mesh with learnable textures. The inverse rendering stage performs multi-bounce path tracing at each planar Gaussian, using the mesh and its textures for visibility queries and secondary-ray shading, and renders the image by α-blending the shaded Gaussians. The authors argue that this design reduces multi-view inconsistency and the per-pixel shading cost of deferred shading. On Synthetic4Relight and TensoIR, the method achieves the highest novel view synthesis (NVS) and relighting PSNR among the compared methods, with a training time of approximately 40 minutes per scene. The paper also includes ablations on sampling, indirect illumination, and multi-view material consistency.
Critical strengths
- Explicit multi-bounce illumination modeling for GS-based inverse rendering. The method combines mesh-based multi-bounce path tracing with per-Gaussian shading and Gaussian splatting. Explicitly accounting for inter-reflections is a useful contribution toward more accurate material and lighting decomposition in GS-based inverse rendering.
- Strong performance with efficient training. The method achieves the highest NVS and relighting PSNR on both evaluated datasets, with competitive performance on other metrics. The reported training time of approximately 40 minutes is also favorable compared with existing GS-based inverse rendering.
Critical weaknesses
- Insufficient positioning against closely related work. MIRReS[1] already combines a radiance field with mesh-based multi-bounce inverse rendering, but is neither discussed nor evaluated. The contribution of per-Gaussian shading and the hybrid material representation needs clearer justification. A comparison with MIRReS[1] should also account for geometry optimization: MIRReS[1] refines geometry during inverse rendering, whereas the proposed method keeps the geometry fixed after the first stage. In addition, RadioGS [2] is categorized as a deferred-shading method, although it performs forward shading on Gaussian surfels. Please clarify this categorization and whether it affected how this baseline was implemented and evaluated.
- Limited evidence for the advantages over deferred shading. The paper argues that its formulation avoids problems caused by view-dependent attribute blending, but does not directly evaluate robustness to camera-position or field-of-view(FoV) changes mentioned in lines 258-259. Figure 2 is also difficult to assess without ground truth, and some edges appear sharper in the deferred results. Ground-truth crops and quantitative comparisons under controlled camera and FoV changes changes would help support this claim. An analysis of sensitivity to Gaussian size and density would further clarify the trade-off between the two shading approaches.
- Unclear consistency between Gaussian materials and proxy textures. Matching Gaussian material maps with learnable textures in image space does not guarantee that the mesh textures provide consistent material values at secondary intersections. Their convergence and the impact of such discrepancy remains unclear. Intermediate albedo maps and convergence plots over training would help assess convergence and the reliability of this design. Also, note that MIRReS[1] directly queries material of the intesection from the radiance field, thus additional comparison or discussion seems necessary to further justify the design.
- Noisy albedo and environment map. In Figures 3 and 5, the albedo near the contact region between the hotdog and the plate appears noisy compared to existing works, and the noise is also visible in the relighting results. Also, Figure 13 show rather noisy environment map. Could you discuss the possible cause of such noise? In my opinion, the Monte Carlo path tracing may have led the noise. For comparison, MIRReS[1] adopts ReSTIR[3] and RadiosityGS[4] adopts weighted reservoir sampling to reduce variance, while the proposed work seem to use standard importance sampling methods. Additional discussion in sampling strategy might be necessary.
[1] Dai et al. "Inverse Rendering Using Multi-Bounce Path Tracing and Reservoir Sampling." International Conference on Learning Representations (ICLR), 2025.
[2] Han et al. "Radiometrically Consistent Gaussian Surfels for Inverse Rendering." International Conference on Learning Representations (ICLR), 2026.
[3] Bitterli et al. "Spatiotemporal Reservoir Resampling for Real-time Ray Tracing with Dynamic Direct Lighting." ACM Transactions on Graphics (SIGGRAPH), 2020.
[4] Jiang et al. "Differentiable Light Transport with Gaussian Surfels via Adapted Radiosity for Efficient Relighting and Geometry Reconstruction", ACM Transactions on Graphics (SIGGRAPH ASIA), 2025
Questions for authors
For brevity, questions closely related to the weaknesses are included in the section above. Additional questions are listed below.
-
Ablation interpretation. On TensoIR, removing indirect illumination during training improves albedo PSNR (+ 0.42dB). Removing multi-view consistency also improves relighting PSNR (+0.08dB) and albedo PSNR (+0.13dB). Is there additional analysis explaining these results?
-
Rendering cost and relighting efficiency. Could you provide a rendering-cost analysis, separating path tracing and Gaussian rasterization time and reporting how the cost changes with image resolution, samples per Gaussian, and number of bounces? A comparison with RadioGS's finetuning-based relighting would also be useful.
5 comments
Response to Reviewer Hfq5 (2)
W4: Noisy albedo and environment maps
We thank the reviewer for raising concerns about noise. We agree that the noticeable noise in the estimated environment maps originates from Monte Carlo integration. During optimization, we employ BRDF importance sampling and environment-map importance sampling, but do not incorporate weighted reservoir sampling for variance reduction.Spatial and temporal resampling are not directly applicable to our method.
Specifically, our per-Gaussian shading formulation does not provide the screen-space neighborhoods typically required for spatial resampling, while temporal resampling is incompatible with our optimization process. MIRReS performs resampling across different samples, producing a noise-reduction effect similar to increasing the sample count.
The relatively high learning rate used to optimize material parameters makes material updates more sensitive to Monte Carlo sampling noise. Consequently, the estimated albedo near the contact region between the hotdog and the board appears noisier than that of existing methods, and this noise is also visible in the relighting results. The noise in the estimated environment maps also originates from Monte Carlo sampling, and the relatively high learning rate used to optimize these maps makes the noise more pronounced.
Our experiments show that lowering the environment map learning rate reduces noise but leads to less complete disentanglement of material and illumination within the same training budget, thereby degrading relighting quality. Similarly, lowering the material learning rate also results in less complete disentanglement within the same training budget. We therefore adopt relatively high learning rates for both material parameters and environment maps as a trade-off, accepting some visible noise in the estimated materials and environment maps to preserve the quality of material estimation and relighting.
Q1: Ablation interpretation
We thank the reviewer for their questions regarding our ablation study. The previously reported improvement in albedo PSNR on TensoIR when removing indirect illumination during training was due to an error in compiling the experimental results. We apologize for this error and have corrected the results in the revised manuscript. Removing multi-view consistency yields slight improvements in the average relighting PSNR and albedo PSNR . These gains are primarily driven by the Lego scene, whereas retaining multi-view consistency improves performance on the other scenes.
Response to Reviewer Hfq5 (3)
Q2: Rendering cost and relighting efficiency
We thank the reviewer for raising concerns about the efficiency of our method. We evaluated rendering efficiency on the Synthetic4Relight and TensoIR datasets at a resolution of 800 × 800, varying the number of samples per pixel and the maximum number of ray bounces, and reported the time spent on path tracing and rasterization separately. Since our method performs rendering on a per-Gaussian basis, the rendering cost is dominated by path tracing, whose runtime depends on the number of Gaussians.
In addition, we compared the rendering times of our method, IRGS, RadioGS, and the fine-tuned version of RadioGS using 256 samples per pixel. Although the fine-tuned version of RadioGS renders faster, it requires precomputation and does not support fully dynamic lighting. In contrast, our method supports fully dynamic lighting without such precomputation. We also provide a comparison of the training times of IRGS, RadioGS, and our method on both datasets. Our method achieves the shortest total training time on both datasets.
Rendering times under varying sample counts and maximum bounce counts
| Sample Count | Max Bounces | Path Tracing (ms) | Rasterization (ms) |
|---|---|---|---|
| 16 | 1 | 1.81 | 1.28 |
| 16 | 3 | 2.49 | 1.29 |
| 16 | 5 | 2.67 | 1.30 |
| 64 | 1 | 7.42 | 1.30 |
| 64 | 3 | 10.19 | 1.32 |
| 64 | 5 | 10.96 | 1.33 |
| 256 | 1 | 27.77 | 1.30 |
| 256 | 3 | 35.87 | 1.33 |
| 256 | 5 | 44.10 | 1.27 |
Rendering time (ms) at 256 spp
| Method | Synthetic4Relight: Air Balloons | Synthetic4Relight: Chair | Synthetic4Relight: Hotdog | Synthetic4Relight: Jugs | TensoIR: Armadillo | TensoIR: Ficus | TensoIR: Hotdog | TensoIR: Lego |
|---|---|---|---|---|---|---|---|---|
| IRGS | 1934.63 | 1003.31 | 2547.82 | 1337.98 | 695.67 | 530.38 | 2043.06 | 1442.4 |
| RadioGS | 118.07 | 210.85 | 86.59 | 136.18 | 148.67 | 167.39 | 70.92 | 257.5 |
| RadioGS (finetuned) | – | – | – | – | 5.08 | 5.13 | 5.01 | 5.92 |
| Ours | 47.8 | 38.35 | 28.78 | 32.59 | 28.08 | 107.76 | 29.78 | 56.61 |
Training time (h)
| Method | Synthetic4Relight: Stage I | Synthetic4Relight: Stage II | Synthetic4Relight: Total | TensoIR: Stage I | TensoIR: Stage II | TensoIR: Total |
|---|---|---|---|---|---|---|
| IRGS | 0.72 | 0.50 | 1.22 | 0.75 | 0.50 | 1.25 |
| RadioGS | 0.75 | 1.10 | 1.85 | 0.80 | 1.10 | 1.90 |
| Ours (deferred shading) | 0.50 | 0.25 | 0.75 | 0.50 | 0.25 | 0.75 |
| Ours (forward shading) | 0.50 | 0.15 | 0.65 | 0.50 | 0.15 | 0.65 |
Response to Reviewer Hfq5 (1)
W1: Missing related work
We sincerely thank the reviewer for the valuable suggestions, which have helped us improve the quality of our manuscript. We have expanded the related work section to discuss MIRReS and other relevant methods, and have also compared our method with MIRReS in terms of relighting quality. In the original manuscript, RadioGS was incorrectly categorized as a deferred-shading method. RadioGS performs surfel-level forward shading and uses a forward rasterizer to alpha-blend the resulting radiances. We have corrected this categorization in the revised manuscript and moved RadioGS to the discussion of forward-shading-based methods. This miscategorization in the text did not affect our baseline implementation or experimental evaluation. We used the official RadioGS implementation and preserved its original forward-shading and rasterization pipeline.
W2 & W3: Consistency between Gaussian and Mesh Representations and Limited evidence for the advantages over deferred shading.
We thank the reviewer for pointing out this important issue. We agree that matching Gaussian-rendered and mesh-rendered material maps in image space alone does not guarantee exact material consistency at all secondary intersections.
In our method, we use a modified VA-GS, a variant of PGSR, for geometry reconstruction. During this stage, multi-view consistency constraints encourage consistent surface geometry across different input views. This provides a basis for establishing a relatively stable spatial correspondence between the Gaussian representation and the extracted mesh proxy, supporting subsequent material alignment. In the second stage, we keep the mesh geometry and its UV parameterization fixed and optimize the learnable mesh textures by matching mesh-rendered and Gaussian-rendered material maps across multiple training views. Thus, the same surface region receives material supervision from multiple viewpoints whenever it is visible in those views. Moreover, secondary rays query the mesh material using the UV coordinates of their actual intersection points, providing a view-independent material lookup mechanism. Additionally, following GHPT, we discard intersections at very short distances from the ray origin and reject back-facing intersections. We have also added visualizations of Gaussian-based and mesh-based albedo estimates for different scenes during training, together with loss convergence curves for the second training stage. Across the displayed iterations, the Gaussian-based and mesh-based albedo estimates remain closely aligned in their surface color distributions, demonstrating good material consistency between the two representations. The loss convergence curves further show that our second-stage optimization converges stably across different scenes.
We thank the reviewer for raising questions about the effect of the camera’s field of view on deferred shading and the convergence of our method. In the Appendix, we provide visual comparisons of deferred- and forward-shading results under different fields of view (FoVs). These comparisons show that the results produced by deferred shading vary as the FoV changes. We have added ground truth images to Figure 2 in the revised manuscript to facilitate a direct visual comparison.