acceptodds
Under review as a conference paper at ICLR 2027

Generative Lighting Estimation for Multi-View Captures

Abstract

An environment map is the standard representation of the distant illumination arriving at a point in a scene, and many inverse rendering and relighting pipelines on real captures depend on one. Generative methods predict a map from a single image with a learned prior. Everything outside that image's field of view is then inferred from indirect cues rather than observed, so the estimate does not faithfully reproduce what is in the scene. Per-scene inverse rendering instead optimises against a whole multi-view capture, solving for illumination jointly with geometry and materials. It is constrained only where the capture has observations, and the light sources usually lie outside them. We recover a high dynamic range environment map from a sparse set of posed low dynamic range frames, so that the learned prior supplies only what the capture does not. The map is produced as a set of views that together cover all directions, generated jointly by a multi-view diffusion model so they agree with each other and with the input frames, and assembled into a panorama. The single-image setting is the smallest case of this. On it we improve on the state of the art on the Poly Haven and Laval Indoor benchmarks. Additional frames improve the estimate further, with diminishing returns. On captures with a moving camera, the frames need not be taken where the map is centred; the benefit falls off slowly with distance. Each frame enters with its pose, so unlike stitching and outpainting the model is not bound to a single viewpoint. A probe over scenes with near-field landmarks confirms this.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
lighting estimation, HDR environment map, image-based lighting, multi-view diffusion, inverse rendering, novel view synthesis, panorama generation, diffusion models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_VLP8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper estimates an HDR environment map at a query location from sparse images with known camera poses. A diffusion model generates overlapping LDR views, a second model recovers HDR content, and exposure alignment and blending assemble the panorama. The stages train on panoramic videos and HDR stills, respectively. The contribution combines a generative prior with camera geometry to use observations from different positions. Validation includes three lighting benchmarks, experiments varying view count and camera displacement, and controls removing translation information. Results show benchmark improvements and benefits from additional views.

Critical strengths
  1. The method estimates an environment map at a query location from sparse, posed observations, using LDR video with camera translation to learn content generation and HDR panoramas to learn radiometric reconstruction. This decomposition provides a practical way to exploit complementary data sources without requiring every training example to include both camera translation and HDR supervision. The resulting ability to use observations captured away from the query location is potentially valuable for real capture and object-insertion pipelines.

  2. The experiments provide meaningful evidence that additional observations improve predictions. When all input views share the same camera position, increasing their number from one to four improves performance by 2.25 dB on a fixed region outside the coverage of all four views. For inputs captured at different positions, the authors include two controls: a model trained only on views from a single location, and the original model evaluated with camera translations set to zero. Although these comparisons do not fully establish the claimed geometric mechanism, they support the value of additional observations and accurate pose information, providing quantitative evidence for the method’s effectiveness.

Critical weaknesses
  1. The paper repeatedly describes LuxDiT as accepting only a single image, although LuxDiT explicitly supports video-conditioned lighting estimation and compares its video mode with frame-wise inference. This mischaracterization undermines the broad claim that existing generative methods discard additional frames and leaves the most relevant multi-frame comparison missing. Explicit poses, sparse wide-baseline observations, and arbitrary query locations may still distinguish this work, but those differences need to be established through accurate positioning and comparisons in shared settings. Without this, the paper’s advancement over existing methods remains insufficiently demonstrated.

  2. The evaluation establishes improvements in panorama prediction more convincingly than improvements in HDR lighting estimation. The main benchmark tone-maps environment maps before rendering, suppressing the high-intensity illumination that HDR reconstruction is intended to recover. The supplementary evaluation using linear environment maps covers only Laval Indoor and lacks paired significance testing, while the displaced-view experiments evaluate Stage 1 LDR panoramas rather than the complete HDR pipeline. An experiment combining displaced multi-view inputs, HDR ground truth, and linear-light rendering is therefore needed to substantiate the central contribution.

Questions for authors
  1. Why is LuxDiT described as supporting only a single input image? Please correct this characterization and clarify how your method differs from its video mode. Could you also provide a comparison in a setting that both methods support?

  2. Do the improvements from multiple views translate into better HDR lighting estimation? Could you provide an evaluation using inputs captured at different camera positions and GT HDR environment maps, with rendering performed using the original linear environment maps? Synthetic scenes could provide an initial test.

  3. How is the unseen mask in Section 5 constructed? Does it account for depth, occlusion, or correspondences across views? If it relies only on rotation, how do you ensure that content in the masked target region is not already visible in images captured from other camera positions?

3 comments

LuxDiT video conditioning

You are correct that LuxDiT supports video inputs, and we will correct its characterization as single-image-only. The distinction is explicit pose conditioning and query-location estimation, not simply accepting multiple frames.

We evaluated LuxDiT video on the same 181 Poly Haven scenes using 25 frames spanning a full yaw rotation at 60-degree vertical FoV. All frames enter the model, and we score its final output panorama. Our method receives four posed views with less angular coverage. Both methods use the standard tone-map-before-render protocol of Table 2.

MethodInputsDiffuse SI-RMSE ↓Matte-silver SI-RMSE ↓Mirror SI-RMSE ↓
LuxDiTSingle image0.0850.1940.403
LuxDiT25-frame video0.07620.20170.4246
OursSingle image0.06620.16650.3633
OursFour posed views0.04360.11250.2934

Our method benefits from additional observations across all three materials, while LuxDiT's SI-RMSE changes are mixed. Compared with LuxDiT video, our four-view model reduces SI-RMSE by approximately 43%, 44%, and 31%. The corresponding absolute reductions have paired 95% scene-bootstrap intervals of [0.0265, 0.0390], [0.0782, 0.1005], and [0.1164, 0.1468]. This supplies an external multi-frame comparison in a shared co-located setting.

HDR fidelity

To examine HDR fidelity separately, we extended the single-view Laval Outdoor evaluation to linear-light rendering, removing tone mapping and clipping of HDR environment maps before rendering. We evaluated the full set and a 138-scene direct-sun subset selected from ground truth and visually checked.

The table reports linear SI-RMSE differences, ours minus LuxDiT, and paired 95% confidence intervals, multiplied by 1,000. Negative differences favor ours.

MaterialFull Laval OutdoorDirect-sun subset
Mirror−0.0880 [−0.1287, −0.0428]+0.0056 [−0.0370, +0.0596]
Matte silver−0.0070 [−0.0498, +0.0370]+0.0625 [+0.0120, +0.1151]
Diffuse−0.0036 [−0.0230, +0.0158]+0.0161 [−0.0086, +0.0405]

Intervals use 50,000 paired scene-bootstrap draws and are pointwise. After Holm correction across the 68 full-set and subset comparisons, only the small full-set mirror improvement is supported among these six comparisons. The ranking depends on the metric: LuxDiT achieves lower percentile-normalized linear RMSE for matte-silver and diffuse spheres on the direct-sun subset. Thus, the linear-light comparison is mixed rather than showing the consistent advantage observed under the standard protocol.

We also report preliminary controlled synthetic experiments in the follow-up below, including the complete input-count and translation-ablation results.

Angular-mask definition

We construct the mask from input-camera orientations and intrinsics, without depth, occlusion, or correspondence reasoning. The scored region lies outside the angular coverage of all eight input views, excludes the nadir below −45°, and remains fixed across input counts.

For displaced cameras, scene content visible in an input can move into this region through parallax. “Unseen” is therefore too strong; we will use “outside the eight-view angular coverage” and make the distinction explicit. The translation-control comparison remains on identical target pixels with identical input images and viewing directions, changing only the supplied translations.

Controlled synthetic evaluation

We also conducted preliminary experiments in three PBRT scenes using object-centred camera orbits, known poses, and independently rendered target panoramas. Nested input sets contain one view, four adjacent orbit views, and all sixteen orbit views. We evaluate three fixed query locations: above the selected object, at input camera 0, and at an “away” location outside the input orbit, corresponding to the original scene camera. We compare correct translations with zeroed translations while holding images, rotations, intrinsics, weights, and inference settings fixed.

Each cell reports standard-protocol sphere SI-RMSE averaged across three materials: correct translations / zero translations. Lower is better; bold marks the lowest error in each row across both translation settings and all input counts.

SceneQuery location1 input4 inputs16 inputs
BistroAbove object0.4258 / 0.43690.3888 / 0.42580.3424 / 0.3435
Input camera 00.3775 / 0.37750.3658 / 0.37320.3021 / 0.2936
Away0.3994 / 0.38140.3865 / 0.37890.2924 / 0.2930
San MiguelAbove object0.5847 / 0.62820.6138 / 0.63950.5451 / 0.5895
Input camera 00.6065 / 0.60650.5589 / 0.59530.6195 / 0.7086
Away0.7539 / 0.81120.6953 / 0.77850.8504 / 0.9378
WatercolorAbove object0.5986 / 0.62920.4665 / 0.54520.4602 / 0.4561
Input camera 00.3284 / 0.32840.2668 / 0.31610.2699 / 0.4073
Away0.2362 / 0.24880.2253 / 0.30350.3386 / 0.4084

Correct translations help: at four inputs, they reduce SI-RMSE at eight of the nine scene–query combinations. At sixteen inputs, the benefit remains consistent across San Miguel's three query locations and substantial at two Watercolor locations, while Bistro shows little benefit.

Additional observations help: with correct translations, increasing from one to four inputs reduces SI-RMSE at eight of nine scene–query combinations. Increasing to sixteen further improves all three Bistro queries, but the trend is not uniformly monotonic: four inputs remain better at the input-camera and away queries in San Miguel and Watercolor.

The results support using posed multi-view evidence while showing that its benefit depends on the scene, query location, and input configuration. Metric agreement is not uniform: San Miguel's SI-RMSE gains are not consistently accompanied by lower RGB angular error. These preliminary results cover three selected scenes, three query locations per scene, and one generation seed.

Thanks to the authors for the clarifications and additional experiments. The LuxDiT video comparison addresses my concern about the missing external baseline, and I appreciate the correction to the interpretation of the angular mask.

My main concern remains whether the improvements translate into better HDR lighting estimation. The new linear rendering results are mixed, particularly on scenes with direct sunlight. The synthetic experiments provide useful evidence that correct translations help, but they still use the standard protocol that compresses dynamic range before rendering, so they do not close the gap between the multiview results and HDR fidelity.

The response strengthens the evidence for using posed observations to improve panorama prediction, but I remain unconvinced that the central HDR lighting contribution is sufficiently validated. I therefore maintain my rating of 2.

Reviewer_pmiDOpenReview ↗
Rating
3
Confidence
2
Informativeness
2(−1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper discusses a method that takes any number of LDR images with known extrinsics as input and reconstructs the HDR envmap at a given location p. Previous work usually only considers a single image and then guesses the lighting of the remaining unseen part, or some inverse rendering methods use all views but cannot complete the unseen part. This work aims to fill this gap. The paper fine-tunes Stable Virtual Camera. In the first step, using YouTube panoramic videos as training data, the paper fine-tunes a model that, at the input location p mentioned above, generates 20 perspective crops, and then uses a traditional panorama assembly algorithm to stitch them into one complete panorama. The authors then fine-tune another Stable Virtual Camera model to convert these LDR output panoramas into HDR, as the final output.

Regarding experiments: 1. The authors make a paired per-scene comparison with the previous LuxDiT and DiffusionLight in the single-view setting, with a large numerical advantage. 2. The authors run an oracle view-count experiment with 2–4 views cut from the ground-truth panorama, and separately score the uncovered part of the sphere, showing the model’s predictions and the gain brought by the number of views. 3. An ablation experiment that trains with the same number of input images but without pose information, showing that the model’s gain comes to some extent from camera translation; and if the translation is removed at inference time for the original model, the lost gain is about the same. This is a fairly rigorous experiment.

Critical strengths
  1. The problem the paper defines is worth solving. Obtaining the environment map at an arbitrary query point p is written as “multi-view generation from N≥1 posed views”, with p entering the NVS backbone as the target camera position, which is a genuine, unfilled interface between generative lighting estimation and posed capture. The appendix is also fairly honest in disclosing the limits of the model’s capability; for example, when the information of a real panorama is available, using it directly is more accurate than the model’s prediction in the vast majority of cases.

  2. The design used for evaluation is also fairly rigorous and convincing. The single-view results improve over prior work, and the comparison is fair. Table 1 shows that it leads or ties LuxDiT on the three benchmarks. The appendix has a direct pasting reference that does not use the model; the results in Appendix I show that, with the same information, the model’s approach does indeed have a considerable advantage over traditional direct stitching.

Critical weaknesses
  1. The biggest problem is that the motivation does not match the method, the multi-view claim lacks external baselines, and the final experimental evidence is small-sample LDR. The paper’s story in the introduction is motivated by “tens to hundreds of views”, but the actual training method only uses 4, and 4->8 views shows no gain or a negative gain, so it can only work on sparse inputs; moreover, a distant real panorama directly resampled beats generation on the uncovered region at every distance level.

  2. Near the poles of the panorama, and for some highlights such as the sun, the reconstruction does not seem to be done properly, for example rosendal plains 2 and snowy cemetery in Figure 9, and sunset jhbcentral and st peters square night in Figure 10, which visually look worse than LuxDiT’s reconstruction results.

Questions for authors
  1. Why were only four views used in the end? In theory, more views should produce better results, but surprisingly there is no improvement from four to eight views. Could you explain this in more detail?

  2. The single-view results also look better. Could this be due to the base model itself being stronger, rather than the algorithm itself actually becoming better? The selling point of this paper does not seem like it should be that the single-view case is also much better than prior work.

  3. Are there more multi-view baseline comparisons?

2 comments

Thank you for emphasizing the multi-view evaluation.

External multi-frame comparison

We added a comparison against LuxDiT's video mode on the same 181 Poly Haven scenes. LuxDiT receives 25 frames spanning a full yaw rotation at 60-degree vertical FoV; our method receives four posed views with less angular coverage. All 25 frames enter LuxDiT, and we evaluate its final output panorama under the same standard benchmark protocol.

MethodInputsDiffuse SI-RMSE ↓Matte-silver SI-RMSE ↓Mirror SI-RMSE ↓
LuxDiTSingle image0.0850.1940.403
LuxDiT25-frame video0.07620.20170.4246
OursSingle image0.06620.16650.3633
OursFour posed views0.04360.11250.2934

Additional observations improve our SI-RMSE across all three materials, whereas LuxDiT's video gains are mixed: diffuse improves, while matte-silver and mirror errors increase. Our four-view model reduces SI-RMSE over LuxDiT video by approximately 43%, 44%, and 31%, respectively. The paired 95% scene-bootstrap intervals for these absolute reductions are [0.0265, 0.0390], [0.0782, 0.1005], and [0.1164, 0.1468]. Our method has lower SI-RMSE on 146, 166, and 175 of the 181 scenes, respectively.

Backbone and task-specific adaptation

We chose SEVA because its camera-conditioned multi-view pretraining matches our formulation: generating coherent perspective views from posed observations. SEVA's backbone is smaller than LuxDiT's, with 1.3B parameters versus CogVideoX-5B-I2V's 5B.

Although the cross-method comparison does not isolate the backbone's contribution, Table 2 shows that the pretrained backbone alone is insufficient. The training-free pipeline performs worse than LuxDiT in SI-RMSE for all three materials. Our adaptation reduces its diffuse, matte-silver, and mirror SI-RMSE from 0.1186/0.2332/0.4184 to 0.0662/0.1665/0.3633. The single-view results evaluate the adapted system; the view-count and translation controls test the multi-view contribution with the backbone fixed.

Training count and additional inputs

We sampled the conditioning count up to four to reserve more of each training sequence for supervised target views. Four is therefore a training allocation, not an architectural input limit. In the submitted spatial evaluation, increasing from four to eight views improves PSNR from 16.23 to 16.84 dB at the nearest displaced distance rung, but the benefit diminishes at larger distances.

During rebuttal, we additionally evaluated nested sets of 1, 4 adjacent, and 16 full-orbit views in three synthetic PBRT scenes, with independently rendered panoramas at fixed query locations. The table reports standard-protocol sphere SI-RMSE averaged across three materials at input camera 0, using correct translations.

Scene1 input4 inputs16 inputs
Bistro0.37750.36580.3021
San Miguel0.60650.55890.6195
Watercolor0.32840.26680.2699

Sixteen inputs improve reconstruction in Bistro, approximately preserve the four-view result in Watercolor, and worsen it in San Miguel. These results suggest that supporting longer input sequences does not by itself ensure effective use of additional observations. SEVA integrates geometry implicitly through pose-conditioned attention; improving this integration, through larger-view-count training or alternative view-conditioned backbones, is a direction for extending the gains beyond sparse inputs.

Complete-panorama reference

The reference in the appendix receives an entire real panorama at another location, whereas our model receives sparse perspective crops. It achieves higher PSNR on the fixed region outside the eight-view angular coverage at every tested displaced distance, by 0.3–1.8 dB. This is an additional-information reference rather than an equal-input baseline: the complete panorama supplies observations beyond the directions covered by the sparse crops. We report it to make this practical alternative explicit when a complete source panorama is available.

Polar regions and concentrated sources

We acknowledge the reconstruction failures identified in Figures 9 and 10. We extended the Laval Outdoor evaluation to linear-light rendering and a direct-sun subset. Unlike the consistent SI-RMSE gains under the standard protocol, the linear-light comparison gives a mixed ranking. A follow-up diagnostic on Rosendal Plains also confirms that LuxDiT retains a more concentrated source near the GT sun, while our prediction spreads illumination over a cloud region. We will distinguish these source-recovery limitations from the aggregate reconstruction improvements in the revised manuscript.

Thank you for the detailed response and the additional experiments. The comparison with LuxDiT’s video mode addresses my question about an external multi-view baseline. With just four posed views, your method achieves lower SI-RMSE than LuxDiT with 25 frames across all three materials under the standard benchmark protocol. The single-view results also improve on prior methods on most benchmark metrics, although it remains difficult to tell how much of that advantage comes from the backbone.

I am still not fully convinced that the current method benefits consistently from adding views beyond four, and I see this as a practical limitation. Since the introduction motivates the work in part with captures containing tens or hundreds of posed views, I would suggest making the connection to the demonstrated sparse-view setting clearer. My intuition is that more views could still improve the results with better training and a setup that can make better use of the additional observations. This may be a worthwhile direction for future work.

Overall, the response has answered my main questions and clarified that the method’s advantage is primarily in the sparse-view setting. Given this narrower scope, my support for the paper is slightly lower than before. I will keep my overall rating at 3 (Weak acceptance) and lower my Informativeness score from 3 to 2.

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

This paper estimates an HDR environment map at a query location from a variable number of posed LDR images. It adapts Stable Virtual Camera to generate overlapping views covering the sphere, uses a separate adapter for HDR recovery, and assembles the predictions through exposure matching and blending. Separating content generation from HDR recovery allows training on both posed LDR panoramic videos and HDR panoramic stills. Evaluation includes single-image lighting benchmarks on Poly Haven and the two Laval datasets, a co-located view-count study, and displaced-camera experiments testing whether the model uses camera translations.

Critical strengths

The contribution is incremental, but the quantitative improvements are meaningful and reasonably broad. In the single-image setting, Table 1 reports lower scale-invariant rendering error than the best evaluated baseline on both diffuse and mirror spheres across all three datasets. On Poly Haven, the diffuse and mirror errors decrease from 0.085 and 0.403 with LuxDiT to 0.066 and 0.363. Improvements also transfer to both Laval datasets. These results support the usefulness of the proposed adaptation despite its reliance on existing components.

The multi-view experiments provide evidence that the model uses additional observations and camera geometry. Four co-located views improve PSNR on a fixed region unseen by every input from 17.29 to 19.54 dB. The displaced-camera study further compares against both zeroed translations at inference and a separately trained co-located control. Their agreement supports the value of translation conditioning within the tested baseline range, beyond simply supplying more images.

Critical weaknesses

My main concern is whether the improved aggregate scores translate into better recovery of concentrated light sources. In the first row of Figure 9, LuxDiT produces a compact sun-like source, while the proposed method spreads the bright region over a broader cloud-like area. The latter can still reconstruct the overall scene better, but source extent and intensity matter for specular highlights and shadow sharpness. The displayed images are tone mapped, so this observation does not establish which prediction has the more accurate HDR intensity.

The evaluation does not fully resolve this concern. Appendix F explains that the main protocol tone maps and clips the environment maps before rendering, suppressing the dynamic range at issue. I appreciate the corrected linear-light rendering evaluation on Laval Indoor, where most reported metrics favor the proposed method, but it does not test outdoor sun recovery. Appendix H also acknowledges that the method does not reconstruct the sun. This limits how strongly the benchmark gains support improved HDR illumination recovery, although the broader reconstruction gains still support acceptance.

Questions for authors
  1. Can you clarify the apparent loss of a compact sun in Figure 9 and whether similar behavior occurs across sunlit scenes? A focused comparison with LuxDiT on sunlit test scenes, rendering from the linear HDR maps before applying a common display transform, would help establish the consequences for highlights and shadows. Please account for the arbitrary global radiance scale consistently and report paired uncertainty. This would clarify the scope of the HDR improvement; I am not requesting a new large benchmark.

  2. You already promise implementation and data-processing scripts after publication. Will you also release the trained checkpoints? Since the video training corpus will not be released, checkpoints would be particularly valuable for independent evaluation and use.

1 comment

Thank you for raising the distinction between overall reconstruction accuracy and concentrated-source recovery. We extended the Laval Outdoor comparison to linear-light rendering, removing the standard benchmark's tone mapping and clipping of HDR environment maps before they illuminate the scene. We also evaluated a 138-scene direct-sun subset as a stress test, selected from ground truth and visually checked. The Figure 9 illustrations already use linear-HDR illumination; tone mapping is applied afterward to visualize the rendered spheres.

Outdoor HDR evaluation

Our method improves rendering SI-RMSE over LuxDiT across all three materials under the standard protocol. Under linear-light rendering, the comparison is mixed. The table reports linear SI-RMSE, which fits one global scalar per prediction in each rendered-sphere comparison, and paired 95% scene-bootstrap intervals for the difference. Negative differences favor our method; all numerical entries are multiplied by 1,000 for readability.

Evaluation setMaterialLuxDiTOursOurs − LuxDiT [95% CI]
Full Laval OutdoorMirror22.356922.2689−0.0880 [−0.1287, −0.0428]
Matte silver2.03252.0255−0.0070 [−0.0498, 0.0370]
Diffuse0.47210.4685−0.0036 [−0.0230, 0.0158]
Direct-sun subsetMirror32.272032.2777+0.0056 [−0.0370, 0.0596]
Matte silver2.58872.6512+0.0625 [0.0120, 0.1151]
Diffuse0.55170.5678+0.0161 [−0.0086, 0.0405]

Intervals use 50,000 paired scene-bootstrap draws and are pointwise; significance statements use Holm correction across the 68 full-set and subset comparisons.

On the full set, our method has lower mean linear SI-RMSE for all three materials, but only the small mirror improvement is statistically supported after correction. On the direct-sun subset, LuxDiT has lower means, but none of these three differences passes the same correction. The ranking also depends on the metric: LuxDiT achieves lower percentile-normalized linear RMSE for matte-silver and diffuse spheres on the direct-sun subset. Thus, the consistent advantage under the standard protocol does not extend uniformly to linear-light evaluation.

Figure 9 source example

We also examined rosendal_plains_2, the first row of Figure 9, in a follow-up run using the submitted checkpoints. LuxDiT predicts a compact sun-like source, whereas our prediction replaces it with a broad bright cloud region. At a common 128×256 resolution, the peak-to-median luminance ratio is approximately 78,450 for GT, 860 for LuxDiT, and 9.2 for our rerun. LuxDiT therefore captures the source better in this example, although its peak-to-median ratio remains about 91 times below GT. The low ratio in our prediction reflects the missing concentrated source, not merely a different global exposure.

Both methods generate bounded image representations of HDR through latent diffusion, then reconstruct radiance from those representations. We suspect that preserving compact, high-intensity sources through this representation is a shared difficulty: spatial smoothing can broaden a source, while errors in its encoded intensity can become large radiance errors after decoding. We will distinguish the observed reconstruction gains from concentrated-source recovery in the revised manuscript.

Checkpoint release

Yes. We will release the trained checkpoints for both stages and the inference code, enabling independent evaluation without reconstructing the video training corpus.

Comments on the paper

3 comments

We thank the reviewers for their feedback on multi-view comparisons, HDR fidelity, and the interpretation of our spatial evaluation. We provide additional evidence and clarify these points in the individual responses.

Multi-frame comparison. LuxDiT supports video conditioning, and we will correct its characterization as single-image-only. We evaluated its video mode on 181 Poly Haven scenes using 25 frames spanning a full yaw rotation. Our method uses four posed views with less angular coverage and achieves approximately 43%, 44%, and 31% lower standard-protocol rendering SI-RMSE for diffuse, matte-silver, and mirror spheres, with paired uncertainty supporting these improvements. Additional views improve our SI-RMSE across all three materials, whereas LuxDiT's video gains are mixed.

Linear-light evaluation. We extended the Laval Outdoor comparison to rendering without pre-render tone mapping or clipping of HDR environment maps, including a direct-sun subset. The results give a mixed ranking and expose concentrated-source limitations in both methods.

Spatial conditioning. Preliminary synthetic experiments with known poses and up to sixteen inputs provide further evidence that correct translations and additional observations can improve predictions, with benefits depending on the scene and query location.

Together, these results support adapting a posed multi-view generative prior for environment-map estimation. The method improves standard lighting-benchmark results and benefits from additional views, while the linear-light evaluation identifies radiometric limitations that those benchmarks obscure. We will release both trained-stage checkpoints and inference code.

Dear reviewers,

Thanks for your continued efforts. The authors have provided a rebuttal to the initial reviews. Please review it carefully, along with all other reviews and the corresponding author responses, and consider whether you have any further questions for the authors or whether the rebuttal has addressed your concerns. There are only a few days remaining before the discussion period closes, so please use this time to seek any further clarifications or raise any additional questions about this paper.

Best,

AC

Dear Reviewers,

With the discussion deadline approaching, we would appreciate any remaining feedback on our responses and whether the additional experiments address your concerns.

We have uploaded a revised manuscript with changes addressing the reviews highlighted in blue. We also made unhighlighted technical clarifications and editorial improvements to the draft.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.