acceptodds
Withdrawn from ICLR 2027

LAMP: 3D Source-Aware Lighting Estimation at Multiple Points

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing illumination estimation methods primarily recover lighting distributions for rendering, but typically leave the contributions of individual physical light sources at specific scene locations implicit. We address this gap by leveraging the scene understanding capabilities of vision-language models (VLMs) to connect local illumination with its physical sources. We introduce LAMP, a framework for point-conditioned, source-aware illumination reasoning from a single RGB image with marked query points. LAMP constructs a shared inventory of active light sources, ranks contributing sources for each query, and predicts their 3D query-to-source directions. We supervise these relationships using procedurally controlled indoor scenes, reusing independently rendered source responses to derive contribution targets across lighting configurations and obtaining exact directions from source and query positions. On the Infinigen test set, fine-tuning Qwen3-VL-8B increases dominant-source 3D [email protected] from 0% to 70.6%, with 94.5% of directions within of ground truth among correctly localized sources. Combined with a lightweight network predicting dense RGB irradiance, LAMP achieves probe-rendering quality comparable to prior lighting-estimation methods while additionally localizing individual physical sources in 3D and explicitly attributing local illumination to them.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
lighting estimation, Vision-Language Models, Computer Vision, Indoor Lighting, 3d source-aware lighting estimation, Spatially-varying Illumination

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3Q4bOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduce a 3d source-aware lighting estimation technical called lamp. Given a list of 3D queries, LAMP first estimates an inventory of light sources, then estimate the direction from these queries to their most important light directions. LAMP is based a VLM, specifically Qwen3, and trained on synthetic data to make these predictions. The authors also train a separate lightweight ResNet to predict dense RGB irradiance. The method is quantitatively evaluated on synthetic data for light source localization and direction prediction. Qualitative results are shown on downstream probe rendering task on synthetic and real image datasets.

Critical strengths
  1. The technique attempts to separate out the lighting contribution of individual light sources, which can be useful for application such as synthetically turning light on and off.
  2. The technique uses vlm to predict dominant light sources, which could be an interesting component of a lighting editing system that enables editing individual lights semantically.
Critical weaknesses
  1. The most fatal concern is the quantities the authors propose to estimate in the paper is not useful for many if not most downstream applications. Like what the author have shown in section A.1.4. of the supplementary Since neither light direction nor irradiance map can be used to edit lighting, the author have to resort to complicated and fragile heuristics to insert light probe into the scene. The relighting result is only heuristically plausible looking and not physically correct at all. I will solely reject the paper based on this concern alone as this renders the paper almost completely useless. I strong encourage the author evaluate standard lighting estimation metrics such as RMSE/Si-RMSE/SSIM/Ang. Err. for local environment map such as Table 2 of "Lighting in Motion: Spatiotemporal HDR Lighting Estimation".

  2. The technique cannot predict the local environment map at all (which implies weakness point 1 by definition). Therefore, the author have to resort to evaluation of dominant light source accuracy (where energy can arrive through indirect path). However, I do not believe the evaluation is meaningful and I cannot think of a downstream application where knowing the direction of dominant light source (including energy can arrive through indirect path) is very useful. If the author believe otherwise please clarify.

Questions for authors

I think the central concern is clear: given the quantities the authors propose to estimate cannot really be used to light inserted objects, I do not believe the technique is very useful. In particular I do not see how irradiance map and direction of dominant light source(where energy can transport through indirect path) can be used to support downstream applications such as object insertion. Please clarify if there is any misunderstanding.

4 comments

We thank the reviewer for highlighting this concern. We provide additional experiments on the existing test sets to directly assess the usefulness and limitations of our outputs. All changes in the revised paper are marked in blue.

W1 & Q

Our original paper did not explain the connection between the predicted quantities and probe rendering clearly enough. In A.1.4, we construct a compact lighting model with one virtual emitter per ranked source and an isotropic fill term, following the general discrete-light formulation exemplified by Gardner et al.[1]. Predicted source geometry determines emitter placement, while decoded contribution shares determine relative emitter power weights. The calibrated irradiance map supplies the overall brightness and chromaticity. Together, these components specify the lighting used by our physically based renderer. A full environment map is therefore not a prerequisite for using our outputs in object rendering. To directly evaluate the effectiveness of this construction, we provide quantitative rendering comparisons below, testing the resulting rendering accuracy.

RMSE/Si-RMSE/SSIM/Ang. Err. Table

We update our Table 5 with a glossy material, comparing our method's rendering quality with prior works.

DatasetMethodLPIPS ↓RMSE ↓si-RMSE ↓SSIM ↑Ang. Err. ↓
InfinigenLiMo0.17448.291.650.797.8
Diff.Light-Turbo0.2172.771.870.7210.5
LuxDiT0.2012.311.920.727.8
GaSLight0.3062.881.770.6514.2
Ours0.1731.921.330.816.5
Laval Indoor SVGaron et al.0.1051.080.720.877.0
LiMo0.1302.621.160.774.5
Diff.Light-Turbo0.1432.291.150.757.7
LuxDiT0.1351.771.140.786.4
GaSLight0.2321.811.110.689.8
Ours0.1291.720.750.838.7
LuxRemixLiMo0.2186.812.330.808.4
Diff.Light-Turbo0.2163.042.350.819.6
LuxDiT0.2043.982.300.828.7
GaSLight0.3673.122.510.6712.0
Ours0.2463.072.240.8011.1
Laval Indoor RFLiMo0.1904.932.550.878.7
Diff.Light-Turbo0.1782.942.470.867.6
LuxDiT0.1733.262.590.874.2
GaSLight0.3173.062.710.736.9
Ours0.1903.062.510.867.4

Our method achieves the best on Infinigen for both diffuse and glossy probes, as well as the best glossy si-RMSE on LuxRemix (2.24 vs. 2.30). It also closely matches LiMo’s LPIPS on Laval Indoor SV (0.114 vs. 0.114 for diffuse and 0.129 vs. 0.130 for glossy) and LuxDiT’s diffuse SSIM on LuxRemix (0.85 vs. 0.86).

References [1] Marc-André Gardner et al. “Deep Parametric Indoor Lighting Estimation.” ICCV, 2019. Paper.

W2 & Q

A local environment map describes incident radiance by direction, whereas our formulation links local lighting contributions to identifiable physical sources. Dominant-source accuracy evaluates which source contributes most at each query, including contributions transported through indirect paths. Contribution shares quantify how much each source contributes. We further evaluate the usefulness of these outputs through probe rendering and controlled lighting changes. We believe such source-referenced outputs could support future source-level editing.

Shading Direction of the composed lighting.

We compare the shading vectors of diffuse probes rendered under predicted and reference lighting. Our method achieves the lowest shading-vector error on Infinigen, Laval-SV, and LuxRemix. On Laval-RF, our error is close to LuxDiT’s (0.124 vs. 0.125), while LiMo performs best at 0.109. These results demonstrate that the composed lighting captures directional shading structure across synthetic scenes and real photographs.

MethodInfinigen vec. err ↓Laval-SV vec. err ↓LuxRemix vec. err ↓Laval-RF vec. err ↓
Garon et al.–0.119––
LiMo0.1810.1780.1930.109
DiffusionLight-Turbo0.2600.1720.2140.136
LuxDiT0.2390.1530.2060.125
GaSLight0.2620.2330.2270.181
Ours0.1670.1110.1690.124

Light Editing and Rendering.

We additionally render probes at selected query locations under paired two-lamp configurations with A:B contribution ratios of 10:90 and 90:10. For each configuration, we compare the probe rendered using our predicted lighting with the corresponding ground-truth render. The paired examples visualize how changes in the relative source contributions affect the inserted object’s brightness and shading. They connect our source-referenced predictions to a concrete rendering task and complement the quantitative evaluation above. The rendering result is shown in Figure 11 in Appendix A.3.10 of the paper.

I thank the authors for substantive additions. My concern, however, is not about the quality of the added evaluation but about what it evaluates.

I believe the technique is fundamentally flawed. There is no principled or physically correct way to use the estimated quantities to render an inserted object or relight the scene. At the end of the day, it is impossible to represent global illumination of a general indoor scene with light position centric representation. The rendering equation ultimate integrate over incident radiance at each query point across solid angles. At best, this representation can be turned into a light probe made of a few Dirac deltas at the estimated point positions plus one constant term; the directional energy arriving through bounces off walls and ceilings is not represented, only averaged.

To be precise about what is estimated: for each query the model produces (a) a ranked list of contributing sources drawn from a shared inventory in which each source is stored as a single 3D position of the emitter origin, or the centre for area lights and is defined regardless of visibility (b) one unit 3D direction per listed source (c) one scalar contribution share per listed source, defined as that source's fraction of the total source luminance at the query and, independently, a dense per-pixel RGB irradiance map. The contribution share is derived from solo-source renders whose transport response explicitly includes visibility, shadows, reflection and inter-reflection. It is energy delivered at the query after occlusion. The direction is the straight line to the emitter's 3D position, taken from an inventory that is defined regardless of camera visibility, including occluded and off-screen sources. The two components therefore describe different light paths, and when the dominant transport is indirect the pair has no physical referent, as it is neither a light position nor an incident direction.

Since the technique is fundamentally flawed, the authors have to resort to fragile heuristics even to render a probe sphere. The explanation they offer for how a probe is rendered is, verbatim, the construction I called a heuristic, which I spell out below:

By "fragile heuristics" I mean the specific sequence by which the tuple is turned into something a renderer can consume (Appendix A.1.4). (i) For each query, one virtual emitter is placed along the predicted direction, but its distance is not predicted: it is taken from the query to the referenced inventory position, floored at 0.5 m, and set to 3 m whenever the source is not localized in 3D. (ii) Emitter power is proportional to the predicted share, with no radiometric value attached. (iii) The light not attributed to any emitter is absorbed by an isotropic fill with a fixed fraction f_fill = 0.65 of the target luminance. (iv) Emitters and fill are then jointly rescaled so that the analytic mean radiance of a white Lambertian sphere at the query matches the irradiance map's luminance and chromaticity i.e. absolute brightness and colour come from the irradiance branch, not from the source estimates. (v) Rendering uses only the query's own 3D position and normal; for photographs these are themselves estimated (monocular metric depth, a 21×21 plane fit, and a gray-world white-balance correction), and scene geometry, occlusion, and any interaction between the inserted object and the scene are absent.

It follows that, of the four quantities that determine the rendered image: absolute luminance, chromaticity, the 65% isotropic component, and the directional component, three are supplied by the irradiance branch and by a constant, and the source estimates control only the direction and relative weighting of the remaining 35 per cent. A rendering produced this way therefore cannot attribute its own fidelity to source awareness. The paper's own Table 18 states the ceiling explicitly: even with ground-truth sources and ground-truth shares, the shading-direction error remains 17.7°, because each source is a point emitter and the rest is a global fill. A sphere rendered under such a model is indistinguishable from an environment map plus a parametric fit.

Finally, the two tables in the response do not follow my requested evaluation or establish downstream utility. In my initial review, I specifically referenced Table 2 of Lighting in Motion (LiMo). While outputting an explicit environment map is not strictly mandatory if a method can render inserted objects directly, evaluating on local environment maps, or, equivalently, on mirror and sharp specular probes, is standard precisely because it is most sensitive to angular and structural errors in the incident light field. Notably, whereas LiMo's Table 2 evaluates four spheres including a perfect mirror (Mirr) and a perfectly glossy sphere, the authors omit the mirror probe in Table 5 and evaluate only a Lambertian sphere and a wide-lobe sphere (roughness = 0.25). Both act as heavy angular low-pass filters that blur away missing scene reflections and conceal the failure of approximating global illumination with point lights plus a 65% uniform isotropic fill.

Even without evaluating an explicit environment map directly, equivalent error-sensitive evaluations are straightforward and require no new data. First, render a perfect mirror sphere (roughness = 0) at the same query locations, radius, camera, and exposure settings under LAMP's composed lighting, and compare it against the ground-truth mirror-sphere rendering (rendered under ground-truth scene illumination on synthetic datasets, or compared against the ground-truth local HDR / captured chrome ball at the measured probe location on Laval-SV) using the same metrics as Table 5 (RMSE, si-RMSE, SSIM, LPIPS, and RGB angular error). Second, insert a complex non-convex 3D mesh with sharp normal variations and mirror-like or highly specular materials (such as the Stanford bunny) into the scene with cast shadows and local occlusion enabled, and compare the rendered object against the ground-truth rendering quantitatively and qualitatively.

As for the second table in the response ("Shading Direction of the composed lighting", reporting vec. err), "shading vector" is neither a standard lighting-estimation metric nor mathematically defined anywhere in the response or the revised manuscript (where this table does not appear; Appendix A.3.7 Table 18 instead reports an angular error in degrees). More fundamentally, collapsing an already low-pass-filtered diffuse sphere down to a single 3D vector discards virtually all angular lighting structure and adds no value beyond Table 5. For all these reasons, my assessment remains unchanged.

Reviewer_V3bxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces LAMP, a framework for estimating the contributions of individual light sources at multiple points in an indoor scene from a single RGB image. It uses a fine-tuned VLM to identify active light sources, predict their 3D positions, and estimate their contributions and directions at each query point. A separate network predicts dense irradiance. The authors generate training labels using controlled synthetic scenes and per-source light transport. Experiments on Infinigen and LuxRemix show improved source localization over the zero-shot model, while probe-rendering results are competitive with existing lighting estimation methods.

Critical strengths
  1. A meaningful formulation of source-aware lighting estimation. Rather than only estimating local illumination distributions, this paper explicitly models the relationships between physical light sources and multiple query points, including 3D source positions, contribution rankings, and directions. While previous work has explored 3D light source estimation and individual light decomposition, this paper brings these elements together in a unified multi-query framework, providing a more interpretable lighting representation. I find this problem formulation interesting, as it also offers a new setting for studying spatial and physical reasoning in vision-language models.
  2. Physically grounded supervision and a relatively comprehensive experimental evaluation. The authors render individual light sources separately and use the linearity of light transport to efficiently generate source contribution labels under different lighting configurations. This provides a physically grounded supervision strategy without requiring manual annotations. The paper also evaluates source localization, direction prediction, multi-source attribution, and downstream rendering quality, along with ablation studies and lighting intervention experiments. Overall, the experiments cover several relevant aspects of the proposed method.
Critical weaknesses
  1. The novelty over existing work and the experimental comparisons could be better established. Previous studies have already explored 3D light source estimation from single images, individual light decomposition, and structured 3D light source prediction using vision models. The main novelty of this paper appears to lie in combining these capabilities into a multi-query source attribution framework, while the learning approach itself mainly relies on fine-tuning an existing vision-language model with auxiliary supervision losses. In addition, the baselines in Table 2 predict incident lighting distributions, whereas this paper predicts directions toward physical light sources. These are not exactly the same objectives. Although the authors address this difference using filtered evaluation subsets, the strictest aligned subset contains only 59 query points. A directly supervised baseline trained on the same data and prediction targets is still missing, making it difficult to assess the advantages of the proposed framework over simpler alternatives.
  2. The experiments do not yet fully demonstrate reliable reasoning about individual source contributions, especially under changing lighting conditions. Predicting the contributions of individual light sources at different query points is a central goal of this paper. However, in the lighting intervention experiment in Table 12, the decoded dominant-source accuracy is only 51.4%, and 40.9% of the sequences never switch to the correct source after the ground-truth dominance changes. This raises questions about whether the model has learned the relationship between changes in source intensity and their contributions. The localization performance also drops substantially for secondary sources, off-frame sources, and on the LuxRemix test set. Although direction prediction is relatively accurate when the source is correctly localized, these results leave some uncertainty about the reliability of source attribution in more complex multi-source scenes.
Questions for authors
  1. A critical limitation of LAMP comes from its oversimplified light‑source modeling. Every light source is reduced to a 3D point‑based anchor. Even when physical light fixtures are clearly observable in input images, this point‑source representation fails to capture real‑world emitter geometry and radiation characteristics, creating mismatches between modelled virtual sources and actual environmental illumination.

  2. Comparison with existing methods and methodological novelty Given that previous work has already studied explicit 3D light source prediction and individual light decomposition, could the authors further clarify the main technical advantages of the proposed multi-query attribution framework? Would it be possible to include a simpler, directly supervised baseline trained on the same data and labels, such as a visual encoder with source localization and contribution ranking heads? This would help establish the benefits of the proposed framework.

  3. Performance on the lighting intervention benchmark In Table 12, the decoded dominant-source accuracy is only 51.4%, while the auxiliary attribution head achieves 87.3% accuracy in predicting the direction of contribution-share changes. Although these metrics measure different aspects, the results suggest a gap between the auxiliary predictions and the final decoded outputs. Could the authors analyze the main sources of error and explain why the contribution-share supervision does not lead to more reliable dominant-source switching? This seems particularly important for assessing whether the model captures changes in the actual contributions of different light sources.

  4. Generalization to complex scenes and real-world environments The 3D localization accuracy on LuxRemix and for off-frame sources is substantially lower than that for visible sources in Infinigen. Could the authors further analyze how scene complexity, the number of light sources, and source visibility affect performance? Also, since the real-world experiments mainly evaluate rendering quality, is there additional evidence that the model can accurately recover physical source positions and their contributions in real scenes?

3 comments

Thank you for recognizing the value of the problem formulation and the evaluation, and for the constructive comments. We address the specific concerns below. All changes in the revised paper are marked in blue.

Q1 & W1. Supervised Baseline

We added a fully fine-tuned DINOv2 ViT-L/14 baseline with three task-specific heads, trained on the same training records and labels as our LAMP model. LAMP outperforms this supervised baseline at every evaluated rank and distance threshold on Infinigen (updated Table 1), including gains of 18.1 and 13.0 percentage points (pp) in rank-1 3D [email protected]/0.5m. On LuxRemix, LAMP achieves higher at [email protected] for ranks 1-3, while the baseline performs better at the looser 0.5/1.0m thresholds.

Model / RankN_q3D PCK 0.25 m ↑3D PCK 0.5 m ↑3D PCK 1.0 m ↑Direction ([email protected] m) 12.5° ↑Direction ([email protected] m) 22.5° ↑Direction ([email protected] m) 30° ↑Direction ([email protected] m) Med. (°) ↓
Infinigen
Supervised baseline, rank 11,03231.157.676.491.897.598.84.0
Supervised baseline, rank 258615.032.646.496.397.998.43.1
Supervised baseline, rank 31698.323.126.092.3100.0100.04.4
Ours, rank 11,03249.270.682.294.598.298.82.7
Ours, rank 258631.449.556.897.699.099.32.1
Ours, rank 316932.040.246.292.697.198.52.6
LuxRemix
Supervised baseline, rank 19,4127.630.651.267.989.593.99.4
Supervised baseline, rank 28,5382.410.721.175.893.798.38.7
Supervised baseline, rank 36,2581.46.713.877.494.598.18.2
Supervised baseline, rank 43,3570.73.68.283.395.096.77.5
Ours, rank 19,41210.128.546.065.787.693.99.1
Ours, rank 28,5383.29.718.577.293.997.07.6
Ours, rank 36,2582.06.211.882.396.799.56.9
Ours, rank 43,3570.62.13.388.4100.0100.06.2

The gains extend beyond source recall. Table 10(also shown below) evaluates queries where both models recall the ground truth source under 3D matching within 0.5 m. On this common subset, LAMP improves correct source selection by 5.5/7.2/22.3 pp for Infinigen ranks 1-3 and by 5.8 pp for LuxRemix rank 1. Among each model’s correctly selected rank-1 sources, localization within 0.25 m is also more accurate: LAMP 79.2% vs supervised baseline 55.4% on Infinigen and LAMP 38.1% vs supervised baseline 29.3% on LuxRemix.

Dataset / MethodCorrect source at rank 1 (%)Rank 2Rank 3Rank 4Localized within 0.25 m given the correct source, rank 1 (%)Rank 2Rank 3Rank 4
*Infinigen *
Supervised baseline80.657.732.0–55.449.733.3–
Ours (LAMP)86.1 (+5.5)64.8 (+7.2)54.4 (+22.3)–79.2 (+23.8)71.9 (+22.1)91.1 (+57.7)–
LuxRemix
Supervised baseline65.429.421.114.929.327.328.127.3
Ours (LAMP)71.3 (+5.8)30.0 (+0.6)23.0 (+1.9)9.7 (−5.2)38.1 (+8.8)37.0 (+9.7)35.7 (+7.6)37.2 (+9.9)

Table 11 (also shown below) further contextualizes the LuxRemix results: LAMP achieves comparable in-frame dominant-source [email protected] m (42.0% vs 41.1%), and both models attain 26.4% overall identity-aware success under matching. The baseline’s aggregate PCK advantage is concentrated in off-frame cases, where it recalls more target sources, while LAMP makes fewer wrong-source selections overall (33.7% vs 41.4%). Together, these results support LAMP’s strengths in source attribution and precise localization, while identifying off-frame light recovery as a remaining limitation.

Subset / Method[email protected] (↑)Identity-aware success (↑)Source not recalled (↓)Wrong source selected (↓)Right source, > 0.5 m (↓)
All rank-1 queries
Supervised baseline30.626.413.041.419.2
Ours (LAMP)28.526.418.933.720.8
*Dominant source inside the image *
Supervised baseline41.136.58.036.019.5
Ours (LAMP)42.038.77.134.019.9
*Dominant source outside the image *
Supervised baseline16.312.819.848.618.9
Ours (LAMP)10.29.734.933.322.1

Q2 & W2. Does the model track changes in individual source contributions?

To test whether the model follows the actual contributions rather than scene priors, we built a controlled sweep on 80 Infinigen scenes used neither for training nor for testing. Each query point is lit by two lamps A and B. We rescale the two emissions so that the total contribution at the query is held fixed while the ratio A : B moves through 90:10, 75:25, 60:40, 40:60, 25:75 and 10:90 (50:50 is excluded as a tie), and repeat the sweep at three total brightness levels (×¼, ×1, ×4). Detailed figure including light-sweep visualization is shown in Figures 7 and 8 in Appendix A.3.8 of the paper.

Across all three total-contribution levels (×¼, ×1, ×4), our model consistently outperforms the supervised baseline in endpoint correctness and clean dominant-source switching, supporting its ability to track changing source contributions beyond a single brightness setting. Timely switching remains limited (21.0–22.7%), indicating that accurately resolving the dominance transition remains challenging.

Tier / MethodDominant correct at both ends (%)Clean dominant light switch (%)Timely dominant switch (%)Dominant never switch (↓, %)
gT = 1/4 (total contribution ×1/4)
Supervised baseline70.471.616.411.1
Ours (LAMP)81.8 (+11.4)82.1 (+10.5)22.0 (+5.6)7.6 (−3.5)
gT = 1 (original total)
Supervised baseline65.166.315.913.7
Ours (LAMP)78.5 (+13.5)79.2 (+12.8)22.7 (+6.8)7.6 (−6.1)
gT = 4 (total contribution ×4)
Supervised baseline61.162.316.113.6
Ours (LAMP)73.7 (+12.6)75.7 (+13.4)21.0 (+4.9)6.8 (−6.7)

When the dominant source reverses by a large margin, the model follows it: LAMP is correct at both ends of the sweep for 78.5% of points against 65.1% for the supervised baseline (+13.5 pp). Accuracy falls to 60–62% at 60:40 and 40:60, so the practical limit is discrimination between near-equal sources, which is exactly the regime Table 15 (original Table 12) samples. We state this limit explicitly and add the sweep results in Section 4.2 and Appendix A.3.8 of the paper.

Q3 & W2. Scene complexity, number of sources, visibility, and real scenes

On Infinigen, LAMP maintains accuracy across object-count terciles, blocked direct paths, and indirect-dominated illumination, outperforming the supervised baseline in each stratum (Fig. 9 a-d). Both models show lower accuracy with stronger source competition and off-frame dominants (Fig. 10). LuxRemix contains substantially more off-frame dominants (42% vs 5%) and queries with at least three contributing sources (67% vs 16%). (Figure 10 e-h)

On real photographs (Laval-SV/RF), we evaluate probe rendering, where ours and prior works are indistinguishable.

Figures 9 and 10 are in Appendix A.3.9 of the paper.

I appreciate the authors’ rebuttal, which addresses the rendering quality concerns and responds to most of the raised technical questions. Nevertheless, consistent with the comments from other reviewers, I still hold the view that the light source modeling is overly simplified and far from realistic lighting conditions (e.g., natural light, large-area area lights). Besides, the lack of evaluations in real-world environments greatly undermines the practical value of this work. Therefore, I maintain my score of 2.

Reviewer_GfT4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a framework LAMP for estimating which physical light sources contribute to illumination at multiple queried locations in a single image. Its main contributions are three-fold: a shared 3D source inventory with query-specific source rankings and directions; procedural supervision derived from emission-normalized, independently rendered light contributions; and a hybrid architecture combining a fine-tuned VLM for source attribution with a dense irradiance network for brightness and color. It addresses a meaningful gap between estimating illumination distributions and explicitly identifying their physical sources. Its strongest contribution is the task formulation and associated supervision pipeline, supported by various experiments.

Critical strengths
  1. The proposed method links multiple query locations to a shared inventory of physical emitters and provides interpretable information that conventional environment maps do not explicitly expose.
  2. The method shows substantial improvement over the unadapted VLM.
  3. The evaluation is thorough.
Critical weaknesses

The source-aware formulation and supervision pipeline are worthwhile, and the supplementary evaluation is informative. I think it is a nice paper. More challenging real-life cases with scene-object occlusions and interilluminations are expected.

I would like to see the other reviewers' comments.

Questions for authors

None.

2 comments

We thank the reviewer for the positive assessment and comments on occlusion and inter-illumination. These are important aspects of source-aware lighting estimation, and we appreciate the opportunity to examine them more closely. The reviewer’s feedback motivated additional analyses that clarify LAMP’s performance under indirect illumination and occlusion, as well as its remaining limitations. We present these results below and have incorporated them into the revised manuscript. All changes are marked in blue.

Inter-illumination

We stratify the Infinigen test set by how light reaches the query (A.3.9, Fig. 9). Dominant-source [email protected] m is 70.3% on indirect-dominated queries (inter-reflections carry more than 50% of the irradiance, n = 633) versus 71.2% on direct-dominated queries (n = 399), and 71.0% when the dominant source's direct path to the query is blocked (n = 396) versus 70.4% when it is open. Attribution therefore does not degrade when the query is lit indirectly.

Occlusion

To test cases where occlusion changes which source dominates, the test queries for which an unoccluded power/distance² prior selects a nearer or brighter lamp whose direct light is blocked, while another lamp is the true dominant source. Our model identifies the true dominant source in 52.0% of these queries, compared with 42.0% for the supervised baseline. This 10-percentage-point advantage supports LAMP’s ability to attribute illumination when occlusion makes the nearer or brighter source an incorrect choice, although these cases remain challenging.

Dominant-source identification under occlusion. Accuracy (%) on queries where occlusion makes the source favored by the unoccluded power/distance² prior no longer dominant. Higher is better.

MethodDominant-source accuracy (%) ↑
Supervised baseline42.0
LAMP (Ours)52.0

Rendering under occlusion and real-life scenes

Our real-scene evaluation uses probe rendering on Laval Indoor SV/RF (Table 5), where LAMP achieves competitive rendering quality. Rendering under strong occlusion remains challenging. Improving performance in these conditions and evaluating source attribution on real scenes with per-source annotations are priorities for future work.

Thank you for the response. Any visual results of such cases?

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

Related papers

Discussion (0)

Sign in to comment.