acceptodds
Desk rejected at ICLR 2027

ELM-GS: Environment Lighting Modulation on Gaussian Splatting for Efficient Inverse Rendering

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Inverse rendering aims to recover intrinsic scene properties from images, enabling applications such as scene editing. Recent Gaussian Splatting (GS)-based methods achieve high-quality Physically-based Rendering (PBR) but often incur substantial computational costs due to expensive ray tracing. Ambient occlusion (AO)-based alternatives reduce this cost, yet occlusion estimation overhead and albedo–-shading entanglement still limit their efficiency and editability. Inspired by asset-production pipelines that precompute AO as texture maps for efficient lighting modulation, we propose ELM-GS, an efficient 3DGS framework for inverse rendering with environment lighting modulation. ELM-GS introduces a lighting-aware scene representation that decodes learnable features into primitive-wise AO and environment compensation (EC) terms, which respectively attenuate and enhance local environment illumination. Both modulation terms can be splatted into G-buffer textures for efficient PBR. To improve physical plausibility, we derive multi-view occlusion priors via Screen Space Ambient Occlusion to explicitly supervise AO learning. We further construct Mat-Edit, a synthetic benchmark with pixel-level material annotations to support fine-grained material editing, together with ground-truth AO and material-edited images for comprehensively evaluating the performance of inverse rendering. Experiments show that ELM-GS achieves state-of-the-art efficiency in inverse rendering and novel view synthesis while outperforming existing AO-based 3DGS methods in intrinsic decomposition and scene editing. The code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Inverse Rendering, Ambient Occlusion

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_V1PkOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ELM-GS, an efficient 3D Gaussian Splatting framework for inverse rendering. Primitive-wise learnable lighting features are decoded into directional visibilities, from which AO and bent normals are derived, and view-dependent EC to modulate direct lighting. SSAO-derived multi-view occlusion priors supervise AO learning. The authors also introduce Mat-Edit, a synthetic benchmark with ground-truth AO, material annotations, and material-edited reference images.

Experiments show improved intrinsic decomposition and scene editing over GS-IR and GI-GS, with substantially faster rendering. Compared with RadioGS on synthetic scenes, ELM-GS offers higher rendering efficiency but lower novel view synthesis and relighting quality.

Critical strengths
  1. Clear motivation and efficient design. Fig. 3 intuitively illustrates AO-based lighting ambiguity. ELM-GS models local occlusion through primitive-wise AO and bent normals, while EC compensates for simplified direct lighting. Ablations demonstrate substantial rendering speed gains from learnable AO and improvements in NVS and selected decomposition and editing metrics from EC.
  2. Effective geometry-based supervision. SSAO-derived multi-view occlusion priors constrain AO learning. Ablations show improved AO estimation and relighting quality with this supervision.
  3. Useful benchmark contributions. Mat-Edit complements existing evaluations of NVS, albedo, and relighting with ground-truth AO maps, pixel-level material annotations, and material-edited reference images.
Critical weaknesses
  1. Limited directional expressiveness. Predicted directional visibilities are reduced to scalar AO and a bent normal for a single diffuse irradiance lookup. This approximation cannot fully capture the interaction between visibility distributions and directional environment lighting. Although EC compensates for local illumination errors under the original lighting, it is omitted during relighting, leaving this approximation uncorrected by EC under novel illumination.

  2. Unclear benefit of the visibility parameterization. Deriving AO and bent normals jointly from directional visibilities provides a plausible inductive bias. However, its practical advantage over directly predicting AO and bent normals as a paired output is not evaluated. A comparison between these parameterizations would clarify the benefit of the intermediate directional visibility representation.

  3. Potential confounding in material editing evaluation. FlashSplat assigns each Gaussian a single material label, which may be inaccurate for primitives spanning material boundaries or multiple surfaces. Shared initialization helps control comparison conditions, but editing scores may still combine decomposition, representation, and label-association errors. Their respective contributions are not separately quantified.

Questions for authors
  1. What advantages does decoding directional visibilities and deriving AO and bent normals offer over directly predicting AO and bent normals? Could the authors provide a comparison of reconstruction quality?

  2. In the “SSAO only” and “w/o LAO” variants, is environment irradiance queried using the surface normal or a bent normal derived from SSAO directional visibilities?

10 comments

Thank you for your careful review. We appreciate your recognition of the paper’s "clear motivation and efficient design," "effective geometry-based supervision," and "useful benchmark contributions." In our responses below, we clarify the limitations of the lighting approximation under novel illumination, report an additional ablation on directional visibility and bent normals, and explain the scope of our material editing evaluation and the irradiance lookup used in the SSAO variants. At the end of each response, we describe the corresponding changes and highlight them in the revised manuscript.

W1. Limited directional expressiveness. Predicted directional visibilities are reduced to scalar AO and a bent normal for a single diffuse irradiance lookup. This approximation cannot fully capture the interaction between visibility distributions and directional environment lighting. Although EC compensates for local illumination errors under the original lighting, it is omitted during relighting, leaving this approximation uncorrected by EC under novel illumination.

Thank you for highlighting the directional limitations of the AO and bent normal approximation under novel illumination. The limited directional expressiveness is inherent to the AO and bent normal lighting approximation. We identify this limitation in Figure 3 and introduce EC to account for unmodeled shading effects during inverse rendering, reducing their impact on the estimated albedo. Table 5 shows that adding EC improves albedo estimation on both TensoIR and Mat-Edit. Figure 17 further shows that EC reduces differences in estimated albedo across regions of the same material that experience different local shading.

However, we omit EC during relighting since it is specific to the original illumination. In this setting, although the bent normal accounts for the directionality of local occlusion to some extent, the limitation identified by the reviewer remains. Despite this limitation, the simplified lighting model provides substantial efficiency gains. Compared with ray-tracing-based methods, ELM-GS substantially reduces training time and increases rendering speed. This reflects a trade-off between the directional expressiveness of the lighting model and computational efficiency.

W2. Unclear benefit of the visibility parameterization. Deriving AO and bent normals jointly from directional visibilities provides a plausible inductive bias. However, its practical advantage over directly predicting AO and bent normals as a paired output is not evaluated. A comparison between these parameterizations would clarify the benefit of the intermediate directional visibility representation.

Thank you for drawing attention to the role of directional visibility in our method. We follow common practice in computer graphics by using bent normals to guide environment map lookups toward the more visible regions of the hemisphere. Since our method can efficiently estimate visibility in multiple directions, deriving bent normals from those estimates is a natural choice. We adopt bent normals as an established rendering technique, with our main contribution lying in redesigning how per-primitive AO and EC modulate environment lighting to reduce computational cost and improve the performance of AO-based 3DGS inverse rendering.

We agree that scalar AO can be predicted directly. However, directly predicting the bent normal as an unconstrained direction would undermine its physical interpretation. The purpose of using bent normals is to account for the directionality of local occlusion at a surface point when sampling the environment map. By definition, a bent normal represents the average direction of visible incident light. Direct prediction leaves this defining relationship unconstrained and shifts the objective from characterizing the directions from which environment light can reach the point to selecting a sampling direction that best suits the point.

In an additional ablation, we compare the "Full" model with a variant that directly predicts scalar AO and uses the surface normal for the diffuse irradiance lookup, following common practice in AO-based 3DGS inverse rendering.

TaskMetricTensoIR w/o BNTensoIR FullMat-Edit w/o BNMat-Edit Full
NVSPSNR35.5035.7136.4136.66
NVSSSIM0.9670.9680.9830.983
NVSLPIPS0.0400.0400.0290.029
AlbedoPSNR31.1831.0724.4424.50
AlbedoSSIM0.9310.9290.9150.913
AOPSNR——20.7620.89
AOWMSE——0.0070.007
Material EditingPSNR——23.7923.69
Material EditingSSIM——0.9390.937
Material EditingLPIPS——0.0730.074
RelightingPSNR25.9826.0324.6924.85
RelightingSSIM0.8960.8960.9170.917
RelightingLPIPS0.0970.0970.0880.088

The full formulation improves relighting performance by 0.05 dB on TensoIR and 0.16 dB on Mat-Edit. These modest gains are consistent with our motivation for using directional visibilities to steer the irradiance lookup toward visible directions. This experiment measures the combined effect of deriving AO from directional visibilities and using a bent normal.

To further assess the effect of directly predicting the diffuse irradiance lookup direction, we conduct an ablation study on Mat-Edit in which an MLP predicts both AO and the lookup direction. As shown in the table below, the "PD" variant improves NVS but substantially degrades albedo estimation, material editing, and relighting. The directly predicted direction lacks a clear physical interpretation, exacerbating the entanglement of lighting and material properties during inverse rendering. It is also strongly coupled to the original learned illumination, leading to inaccurate irradiance estimates and a substantial decline in relighting performance when the environment map is replaced.

TaskMetricPDFull
NVSPSNR37.4736.66
NVSSSIM0.9850.983
NVSLPIPS0.0260.029
AlbedoPSNR23.2124.50
AlbedoSSIM0.9080.913
AOPSNR20.8720.89
AOWMSE0.0070.007
Material EditingPSNR22.2323.69
Material EditingSSIM0.9310.937
Material EditingLPIPS0.0770.074
RelightingPSNR21.8524.85
RelightingSSIM0.8810.917
RelightingLPIPS0.1040.088

In the revised manuscript, we add an ablation study in Section Appendix.I to assess the effect of directional visibility and bent normals.

Reviewer_rg8gOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ELM-GS, an ambient-occlusion (AO)-based 3DGS inverse rendering method that targets efficiency and editability. Starting from GS-IR's stage-1 model, each Gaussian carries a learnable feature. One MLP decodes K directional visibilities, which are averaged into AO and also define a bent normal for the diffuse irradiance lookup. A second MLP decodes a view-dependent "environment compensation" (EC) term. AO, EC, and materials are splatted into G-buffers in one pass and shaded with the split-sum model.

To validate the effectiveness, this work establishes Mat-Edit: six synthetic Blender scenes (100 train / 200 test views each) with GT albedo, AO, per-pixel material labels, paired original/edited material parameters, edited references, and relit images. Material editing lifts the labels to Gaussians with FlashSplat and swaps their parameters. Experimental results demonstrate promising efficiency and editability, supporting the claimed efficiency advantage within AO-based approaches.

Critical strengths
  1. The method achieves consistent, fairly measured efficiency gains within the AO-based approaches. GS-IR, GI-GS and ELM-GS start from the same stage-1 model and are timed on the same GPU, with per-scene stage timings reported (Tables 7–9). Against the fastest AO baseline, ELM-GS renders much faster and significantly cuts decomposition-stage training.

  2. Existing inverse-rendering benchmarks offer little support for evaluating occlusion or material editing quantitatively. To this end, the paper builds Mat-Edit to provide GT AO, pixel-level material labels, paired original/edited material parameters, and edited reference images and relighting results. This could become a reusable resource for the community, independent of the method. Its value depends on a public release.

Critical weaknesses
  1. Besides RadioGS, the only 3DGS baselines are GS-IR and GI-GS. Omitted are GUS-IR (an AO-based method the paper cites) and all cited PBR-3DGS methods (e.g., R3DG, GIR, DeferredGS, GeoSplatting, RTR-GS). Yet the abstract claims "state-of-the-art efficiency in inverse rendering." Without a quality–cost comparison against stronger methods, this shows a faster method within one family, not a better overall trade-off.

  2. The ablation does not support the roles claimed for each component (Table 5). (a) Variants shading with SSAO maps directly ("SSAO only", "w/o LAO") give the best relighting on both datasets and the best AO. Over "w/o LAO", the full model clearly gains only on NVS (on albedo and editing: little). Quality seems to come mainly from SSAO pseudo-labels. Learnable AO mainly improves speed. (b) EC adds minor albedo PSNR with relighting changed by +0.15 dB (TensoIR) and −0.27 dB (Mat-Edit). Its NVS gain is expected from any view-dependent residual, and it is dropped at relighting time. Learning it jointly with AO can even cause spot artifacts (App. G). Its claimed disentangling role is weakly supported.

Questions for authors
  1. How is LindL_{ind} in Eq. 10 computed: GS-IR's baked volume, GI-GS-style path tracing, or something learned?
  2. What are the choices of K, M, T,β\beta, ζ\zeta, λ\lambda, the SSAO radius, the MLP and feature sizes, the range/activation of c, and the iteration counts?
  3. Why is the baseline FPS on Mat-Edit 27.7 in Table 5 but 36.5 in Table 4? Is "SSAO only" computed on the fly per frame (0.6 FPS)? An optimized SSAO pass, which takes milliseconds in real-time engines, would be a much stronger speed baseline for "learnable AO".
  4. Is EC kept for material editing and dropped only for relighting? Can EC absorb shading that should be in the albedo? How does relighting change when EC is clamped or regularized?
  5. Will the dataset, including the edited parameters and the Blendkit-derived assets, be released?
8 comments

Thank you for reading our paper closely and examining our experiments in detail. We appreciate your recognition of the "consistent, fairly measured efficiency gains" and your observation that Mat-Edit "could become a reusable resource for the community." In our responses below, we present additional comparisons with R3DG and GIR on Mat-Edit and report rendering performance using an optimized CUDA implementation of SSAO. We also clarify the roles of learnable AO and EC, report how the EC range affects relighting, provide the requested implementation details, correct the Mat-Edit baseline FPS, and specify what the public release of Mat-Edit will include if the paper is accepted. At the end of each response, we describe the corresponding changes and highlight them in the revised manuscript.

W1. Besides RadioGS, the only 3DGS baselines are GS-IR and GI-GS. Omitted are GUS-IR (an AO-based method the paper cites) and all cited PBR-3DGS methods (e.g., R3DG, GIR, DeferredGS, GeoSplatting, RTR-GS). Yet the abstract claims "state-of-the-art efficiency in inverse rendering." Without a quality–cost comparison against stronger methods, this shows a faster method within one family, not a better overall trade-off.

Thank you for raising this concern about baseline coverage and the quality–cost trade-off. Our original evaluation includes recent methods from both families, with RadioGS as an RT-based baseline and GS-IR and GI-GS as AO-based baselines. To broaden the comparison on Mat-Edit, we select two closely related inverse rendering methods with publicly available code. Specifically, we evaluate R3DG and GIR, both of which use ray tracing in their pipelines.

TaskMetricRadioGSR3DGGIRGS-IRGI-GSELM-GS
—Method familyRT-basedRT-basedRT-basedAO-basedAO-basedAO-based
NVSPSNR40.4238.9837.3635.7034.6836.66
NVSSSIM0.9930.9920.9910.9820.9780.983
NVSLPIPS0.0090.0140.0110.0300.0370.029
AlbedoPSNR25.4324.7825.6923.6522.6224.50
AlbedoSSIM0.9400.9180.9260.9120.9000.913
AOPSNR—21.4120.6615.7016.0720.89
AOWMSE—0.0050.0070.0290.0250.007
Material EditingPSNR—24.3423.6623.0620.9623.69
Material EditingSSIM—0.9450.9390.9350.9210.937
Material EditingLPIPS—0.0700.0650.0750.0840.074
RelightingPSNR28.3425.7925.1821.7022.4824.85
RelightingSSIM0.9590.9140.9260.8840.8900.917
RelightingLPIPS0.0460.1060.0660.1040.0990.088
EfficiencyFPS4.012.016.736.514.798.1
EfficiencyTraining314024965060677740495

As illustrated in the table above, ELM-GS achieves the highest rendering speed and the shortest decomposition stage training time among the evaluated methods. At 98.1 FPS, it renders approximately 5.9× faster than GIR, the fastest ray-tracing baseline in this comparison. ELM-GS also improves performance over both AO-based baselines across all five tasks, narrowing the quality gap with the ray-tracing methods. These results characterize the quality–cost trade-off of ELM-GS across both method families.

In the revised manuscript, we add additional quantitative comparisons with R3DG and GIR in Section Appendix.F.

W2. The ablation does not support the roles claimed for each component (Table 5). (a) Variants shading with SSAO maps directly ("SSAO only", "w/o LAO") give the best relighting on both datasets and the best AO. Over "w/o LAO", the full model clearly gains only on NVS (on albedo and editing: little). Quality seems to come mainly from SSAO pseudo-labels. Learnable AO mainly improves speed. (b) EC adds minor albedo PSNR with relighting changed by +0.15 dB (TensoIR) and −0.27 dB (Mat-Edit). Its NVS gain is expected from any view-dependent residual, and it is dropped at relighting time. Learning it jointly with AO can even cause spot artifacts (App. G). Its claimed disentangling role is weakly supported.

Thank you for examining the ablation results in Table 5. We would like to clarify the roles of learnable AO and EC.

Learnable AO. We identify three benefits of learnable AO over using SSAO maps in our paper. (1) Learnable AO substantially improves rendering efficiency, which is our primary motivation. In Table 5 in the original manuscript, replacing per-frame PyTorch SSAO with learnable AO increases rendering speed from 0.6 to 96.6 FPS on TensoIR and from 0.6 to 98.1 FPS on Mat-Edit. Even with an optimized CUDA implementation of SSAO, the full rendering pipeline with per-frame SSAO runs at approximately 50 FPS on both datasets. Learnable AO still reduces end-to-end rendering time per frame by roughly half. (2) Learnable AO improves view consistency. SSAO is computed separately for each view, so the same surface point can receive different AO estimates as the camera moves. Our method assigns AO to each Gaussian primitive and uses the same value across views. (3) Learnable AO can use image reconstruction supervision to recover occlusion effects missed by the SSAO priors. We agree that this benefit is limited when the SSAO maps are already accurate. It becomes more apparent when the rendered geometry buffers are inaccurate. As illustrated in Figure 9 in the revised manuscript, the learned AO maps are visibly cleaner. We agree that faster rendering is the primary benefit of learnable AO and our main motivation for introducing it. Our analysis and experiments also support its benefits for cross-view consistency and for AO quality when the SSAO priors are less reliable.

Environment compensation. We acknowledge that jointly learning AO and EC can produce spot artifacts. Since both terms modulate diffuse irradiance, they are prone to coupling during optimization. However, in inverse rendering, the contributions of individual components remain underdetermined, making coupling between them difficult to avoid entirely. The presence of such coupling does not invalidate the beneficial role of EC in mitigating it. We introduce EC to compensate efficiently for local shading variations that the simplified lighting model does not fully capture, reducing the entanglement of shading effects and albedo during inverse rendering.

Table 5 shows that the "Full" model achieves better albedo estimation than "w/o LEC" on both datasets. To isolate the effect of EC from learnable AO, we also compare "SSAO only" with "w/o LAO". Both settings use the same SSAO maps directly for shading, and "w/o LAO" additionally uses EC. Adding EC improves albedo PSNR by 0.82 dB on TensoIR and 0.19 dB on Mat-Edit. Figure 17 also shows that EC yields more consistent albedo estimates across regions of the same material under different local occlusion. Together, these results support the role of EC in disentangling lighting and material properties, reducing baked-in shading in the estimated albedo.

We acknowledge that adding EC reduces relighting PSNR on Mat-Edit. However, relighting is a composite downstream task. Its performance alone cannot establish how well the method disentangles lighting and materials. Without EC, the estimated albedo absorbs more shading effects. When local shading remains similar under the original and novel illumination, shading baked into the albedo can improve the relighting result even though the decomposition is less accurate. The material editing results are also consistent with this observation. Replacing the estimated albedo removes the shading effects baked into it, and the model with EC performs better in this task.

Overall, our main evidence for the disentangling role of EC is the reduction in baked-in shading effects and the resulting improvement in albedo estimation. We attribute the higher relighting PSNR without EC on Mat-Edit to regions where the baked-in shading remains similar after the illumination changes. The improvement in material editing with EC further supports this interpretation.

In the revised manuscript, we update Figure 9 to include SSAO results for real-world scenes and analyze how learnable AO improves occlusion estimation in Section Appendix.I.

Reviewer_8zgYOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents ELM-GS, an efficient 3D Gaussian Splatting (3DGS) framework for inverse rendering. At its core is a lighting-aware scene representation that decodes learnable features into per-primitive ambient occlusion (AO) and environment compensation (EC) terms. Both terms are rendered into G-buffer textures in a single splatting pass and integrated into a deferred shading pipeline for efficient PBR rendering. The paper further introduces multi-view occlusion priors derived from SSAO to supervise AO learning and improve physical plausibility. Additionally, the authors construct Mat-Edit, a synthetic benchmark with pixel-level material annotations to enable quantitative evaluation of fine-grained material editing. Experiments show that the method achieves leading efficiency among AO-based 3DGS inverse rendering approaches and outperforms comparable baselines on novel view synthesis, intrinsic decomposition, relighting, and material editing.

Critical strengths

Significant efficiency gains. Encoding AO and EC as per-Gaussian attributes enables single-pass splatting of all G-buffer outputs, yielding substantially faster training and inference than existing AO-based methods. It reaches ~98 FPS on Mat-Edit, scales well to large real-world scenes, and supports real-time PBR editing.

Thorough and rigorous evaluation. Experiments cover synthetic and real-world scenes, compare against NeRF-based, ray-tracing-based, and AO-based baselines, and include comprehensive ablation studies to verify each component.

Critical weaknesses

Limited physical interpretability of EC. The EC term is essentially a data-driven compensation factor without strict physical derivation; part of its performance gain comes from data fitting rather than true physical disentanglement, which deviates somewhat from the core goal of inverse rendering.

Limited scale and diversity of Mat-Edit. The benchmark contains only 6 simple object scenes and lacks complex indoor environments, highly glossy materials, and large-scale scenes, limiting the generalizability of the conclusions. No real-world counterpart with ground-truth material labels is provided.

Incremental technical novelty. The overall framework follows the established AO + deferred shading paradigm for 3DGS inverse rendering; the core modules represent incremental improvements rather than a paradigm shift.

Questions for authors

EC is designed to be view-dependent. Does this cause appearance inconsistencies across view changes during material editing or relighting? Why not use view-independent per-primitive attributes instead?

In the ablation, the “w/o SSAO” variant achieves slightly higher NVS PSNR but much worse AO quality. Does this indicate that without geometric supervision, AO absorbs appearance information to boost rendering fidelity at the cost of physical plausibility?

8 comments

Thank you for your careful review. We appreciate your encouraging comments on the paper’s "significant efficiency gains" and "thorough and rigorous evaluation." In our responses below, we clarify the physical motivation for EC, the scope of Mat-Edit, and our technical contributions within the established AO-based rendering framework. We also present additional results on view-dependent EC and discuss how geometric supervision of AO supports intrinsic decomposition and material editing. At the end of each response, we describe the corresponding changes and highlight them in the revised manuscript.

W1. Limited physical interpretability of EC. The EC term is essentially a data-driven compensation factor without strict physical derivation; part of its performance gain comes from data fitting rather than true physical disentanglement, which deviates somewhat from the core goal of inverse rendering.

Thank you for raising this important question about the physical interpretability of EC. We analyze the rendering equations for environment lighting to explain the physical motivation for EC and the quantities it is designed to approximate.

Motivation: Figure 3 illustrates a limitation of the AO-based lighting approximation. Two surface points with the same cubemap lookup direction and AO value can receive different amounts of direct environment lighting. For example, one point may occlude a bright region of the environment while the other occludes a dark region. Scalar AO retains the overall amount of occlusion but discards which directions are occluded. During inverse rendering, the optimizer can absorb shading variations that the lighting model fails to capture into the estimated albedo, limiting intrinsic decomposition and affecting downstream editing.

W1. Limited physical interpretability of EC (continued).

Approximation errors in direct lighting: To reduce rendering cost, AO-based 3DGS inverse rendering methods typically approximate each component of direct lighting, diffuse and specular, with a single directional lookup in the environment map. By analyzing the rendering equations, we show how a spatially local scaling factor can compensate for the approximation error in the diffuse and specular components.

Diffuse direct component. In Equation 7, the diffuse irradiance integral LDL_D is approximated by oxExdiro_{\boldsymbol{x}}E_{\boldsymbol{x}}^{dir}, where ExdirE_{\boldsymbol{x}}^{dir} is the unoccluded environment irradiance obtained from a directional lookup. For a given point x\boldsymbol{x}, we define the reference radiance E=ExdirπE=\frac{E_{\boldsymbol{x}}^{dir}}{\pi} and the directional relative radiance ci(ωi)=Lidir(ωi)Ec_i(\boldsymbol{\omega}_i)=\frac{L_i^{dir}(\boldsymbol{\omega}_i)}{E}. Substituting Lidir(ωi)=Eci(ωi)L_i^{dir}(\boldsymbol{\omega}_i)=E c_i(\boldsymbol{\omega}_i) into the integral gives:

L_D &= E \int_{\Omega} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} c_i(\boldsymbol{\omega}_i) (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i \cr &= E \frac{1}{\pi}\int_{\Omega} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i \frac{ \int_{\Omega} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} c_i(\boldsymbol{\omega}_i) (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i }{ \frac{1}{\pi}\int_{\Omega} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i } \cr &= E\pi o_{\boldsymbol{x}} \frac{ \int_{\Omega} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} c_i(\boldsymbol{\omega}_i) (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i }{ \int_{\Omega} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i } \cr &= E_{\boldsymbol{x}}^{dir} o_{\boldsymbol{x}} \frac{ \int_{\Omega} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} c_i(\boldsymbol{\omega}_i) (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i }{ \int_{\Omega} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i }. \end{aligned}$$ Since $V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i}$ is binary, we partition the hemisphere $\Omega$ into the visible region $\Omega_{vis}$, where $V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i}=1$, and the occluded region $\Omega_{occ}$, where $V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i}=0$. This gives: $$\begin{aligned} L_D &= E_{\boldsymbol{x}}^{dir} o_{\boldsymbol{x}} \frac{\int_{\Omega_{vis}} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} c_i(\boldsymbol{\omega}_i)(\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i+ \int_{\Omega_{occ}} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} c_i(\boldsymbol{\omega}_i) (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i}{\int_{\Omega_{vis}} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i}(\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i+\int_{\Omega_{occ}} V_{\boldsymbol{x}}^{\boldsymbol{\omega}_i} (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i} \cr &= E_{\boldsymbol{x}}^{dir} o_{\boldsymbol{x}} \frac{ \int_{\Omega_{vis}} c_i(\boldsymbol{\omega}_i) (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i }{ \int_{\Omega_{vis}} (\boldsymbol{\omega}_i\cdot\boldsymbol{n}_{\boldsymbol{x}}) d\boldsymbol{\omega}_i } \cr &= o_{\boldsymbol{x}} E_{\boldsymbol{x}}^{dir} c_{\boldsymbol{x}}^D. \end{aligned}$$ Here, $o_{\boldsymbol{x}}$ scales the irradiance according to the cosine-weighted fraction of the hemisphere that is visible. The factor $c_{\boldsymbol{x}}^D$ is the **cosine-weighted mean relative radiance over the visible directions**. It adjusts the irradiance according to how bright the visible region is relative to the reference radiance $E$.

Comments on the paper

1 comment

We thank the AC and the reviewers for their careful reading and detailed feedback. We are delighted that they appreciate the following aspects of our work: "Significant efficiency gains" and "thorough and rigorous evaluation" (Reviewer 8zgY), "consistent, fairly measured efficiency gains" and the observation that Mat-Edit "could become a reusable resource for the community" (Reviewer rg8g), and "clear motivation and efficient design," "effective geometry-based supervision," and "useful benchmark contributions" (Reviewer V1Pk).

The main concerns raised by the reviewers include technical novelty and the role of EC (Reviewers 8zgY, rg8g, and V1Pk), directional visibility and AO supervision (Reviewers 8zgY, rg8g, and V1Pk), comparison with additional baselines (Reviewer rg8g), scope and release of Mat-Edit (Reviewers 8zgY, rg8g, and V1Pk), and implementation details (Reviewers rg8g and V1Pk). Our responses to these questions and suggestions are summarized below:

  • Technical novelty and the role of EC. In our response to Reviewer 8zgY, we clarify that the technical contribution of ELM-GS lies in redesigning how learnable per-primitive AO and EC modulate environment lighting within a single splatting pass. We derive the physical interpretation of EC as the mean relative radiance and evaluate the effect of making EC view-dependent. In our response to Reviewer rg8g, we discuss how EC helps reduce baked-in shading in the estimated albedo and examine how its range affects relighting. In our response to Reviewer V1Pk, we explain how EC compensates for local shading effects that the simplified lighting model omits during inverse rendering and why this correction is specific to the original illumination and cannot be carried over to relighting.

  • Directional visibility and AO supervision. In our response to Reviewer 8zgY, we discuss how removing SSAO supervision weakens the physical interpretability of learned AO and degrades intrinsic decomposition. In our response to Reviewer rg8g, we analyze the benefits of learnable AO for rendering efficiency, cross-view consistency, and occlusion estimation when SSAO priors are less reliable, and compare its efficiency with optimized per-frame CUDA SSAO. In our response to Reviewer V1Pk, we explain why we derive bent normals from directional visibilities and evaluate this design against direct scalar AO prediction with surface normals.

  • Comparison with additional baselines. In our response to Reviewer rg8g, we add comparisons with R3DG and GIR on Mat-Edit to evaluate rendering quality and computational cost across a broader set of ray tracing methods.

  • Scope and release of Mat-Edit. In our response to Reviewer 8zgY, we clarify the benchmark’s scope and its purpose in enabling quantitative evaluation of occlusion and material editing. In our response to Reviewer rg8g, we discuss our plans for the public release of Mat-Edit. In our response to Reviewer V1Pk, we clarify that material editing is designed to assess how well the recovered scene components work together in a downstream task and explain how quantitative evaluations of both albedo and AO support the assessment of intrinsic decomposition capability.

  • Implementation details. In our response to Reviewer rg8g, we clarify how we compute indirect lighting and provide the requested model and training settings. In our response to Reviewer V1Pk, we clarify that the two variants applying SSAO directly query environment irradiance using surface normals.

We have highlighted the main changes in the revised manuscript and described the specific changes in our response to the reviewers.

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

Related papers

Discussion (0)

Sign in to comment.