acceptodds
Under review as a conference paper at ICLR 2027

VeilSplat: Separating Changing Smoke from 3D Gaussian Scenes

Abstract

Smoke-free novel-view synthesis seeks to recover a clear, consistent scene from images captured through changing smoke. Existing approaches often rely on pretrained depth models or generative restoration priors. We introduce Veil- Splat, a physics-based Gaussian reconstruction method that jointly recovers scene and medium from posed smoky RGB images without external pretrained net- works. Our key principle is to separate surface coverage from medium trans- mission: smoke attenuates and redistributes radiance while the direct rendering branch retains the scene’s occlusion weights. A compact density field combines shared spatial bases with per-view coefficients to capture spatially nonuniform smoke variation. Path-sampled optical thickness couples this field to attenu- ated direct radiance, broadened forward scattering, and airlight. The recovered scene supports clean novel views through ordinary Gaussian splatting. On five RealX3D smoke scenes, VeilSplat leads all five evaluated baselines with mean PSNR/SSIM/LPIPS of 16.30/0.6288/0.5083, gaining 5.44 dB PSNR over the strongest baseline. Matched component studies establish the benefits of cover- age separation, capture-varying transmission and forward scattering.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian splatting, novel-view synthesis, inverse rendering, participating media, physics-based reconstruction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5FYAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how to recover a clean 3D scene from images taken through smoke that changes during capture. VeilSplat jointly fits a Gaussian scene and a separate smoke model, accounting for attenuation and scattering while preserving the scene’s occlusion weights. The fitted scene can then be rendered without the smoke model. On five RealX3D scenes, it improves clean novel-view quality over the five evaluated baselines. The ablations examine coverage separation and changing smoke density, while a separate experiment evaluates forward scattering by fixing the scene and measuring the fit to smoky observations.

Critical strengths
  1. The improvement on the evaluated real scenes is substantial and consistent. VeilSplat gains 5.44 dB in mean PSNR over the strongest evaluated baseline and has the highest PSNR and SSIM in each of the five scenes (Tables 1 and 3). The visual comparisons also show clearer boundaries and surface detail. These results are measured at held-out clean viewpoints, so they provide direct evidence that the method improves scene recovery.

  2. The varying smoke model has a measurable benefit despite its small size. Adding two shared spatial bases and per-capture coefficients improves mean PSNR by 1.10 dB over the static density field, with better PSNR and LPIPS in all five scenes under matched training settings (Table 4). This supports the decision to model changes in smoke across captures. The extra model is only needed during fitting; clean views still use ordinary Gaussian rendering.

Critical weaknesses

The contribution of forward scattering to clean scene recovery is not directly evaluated. The ablation in Section 4.5 and Appendix G freezes an already fitted Gaussian scene, refits only the medium, and compares fitting errors on the smoky training images. This supports the authors’ claim that the branch improves observation fitting, but does not establish whether it improves clean novel-view quality when the scene and medium are jointly optimized. Since forward scattering is an important part of the method, this gap limits my assessment of its contribution, though it is not sufficient to undermine the experimental results for the method as a whole.

Questions for authors
  1. How exactly is the no-forward-scattering condition constructed? Which model was used to obtain the frozen Gaussian scene? How are ω and the airlight weight handled when the forward branch is removed? These details matter for interpreting the error differences in Figure 7.

  2. DehazeGS also reconstructs clear novel views from multi-view foggy RGB images and is discussed in the paper. Could you clarify why it was not included in the experimental comparison? If a usable implementation is available, a comparison on the same five RealX3D scenes, with its pretrained depth prior retained and clearly noted, would help assess VeilSplat relative to this closely related method. If such a comparison is not feasible, please explain the implementation or evaluation constraints.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 5FYA

Response to Reviewer 5FYA

Thank you for your positive, constructive review and for recognizing our clean novel-view gains and compact time-varying smoke representation.

[W1] Evaluate the contribution of forward scattering directly on clean novel-view recovery under full joint optimization, rather than only refitting the medium on a frozen scene.

Response: Jointly optimizing the 3D scene and smoke model end-to-end (withholding smoky captures 4, 10, 16, and 22 per scene), physical forward and scene-lit weights lift clean test PSNR by +0.35 dB for the full model, and covariance dilation achieves the best SSIM (0.577) and LPIPS (0.531) (Section 4.6, Tables 5, S13).

We jointly train three variants on Hinoki, Koharu, and Shirohana (18, 21, and 22 smoky training captures): (1) Full model (using the physical small-angle weight t1−f−tt^{1-f}-t, which matches fτf\tau at small τ\tau and vanishes as t→0t \to 0; Section 3.2, Figure 2a), (2) No forward lobe (f=ω=0f = \omega = 0, standard airlight (1−t)A(1-t)\boldsymbol{A}), and (3) No dilation (h→0h \to 0), keeping f,ωf, \omega active without covariance broadening:

Jointly trained variant (3-scene mean, Table 5)Clean PSNR ↑SSIM ↑LPIPS ↓
Full model (h>0h > 0)15.410.5770.531
No forward lobe (f=ω=0f=\omega=0)15.060.5600.562
No dilation (h→0h \to 0, weights kept)15.460.5690.553
  • Removing global veil bias (f,ω>0f, \omega > 0 vs. f=ω=0f = \omega = 0): Modeling scattered scene light raises clean PSNR from 15.06 dB to 15.41 dB (+0.35 dB) for the full model.
  • Preventing edge halos from baking into the clean Gaussians (h>0h > 0 vs. h→0h \to 0): At comparable PSNR, enabling covariance dilation improves the edge-sensitive metrics: SSIM 0.569 → 0.577, LPIPS 0.553 → 0.531, because the medium absorbs edge halos instead of the clean Gaussians.

[Q1] In the frozen-scene study, which model provided the frozen scene, and how were ω\omega and the airlight weight handled when removing forward scattering?

Response: Both arms started from the converged scene of the f=0f=0 variant of Eq. (1) and optimized only the smoke parameters for 500 steps; the no-forward arm dropped the dilated pass while keeping (1−ω)(1−t)A(1-\omega)(1-t)\boldsymbol{A} with ω\omega learnable (Section S16; Figure 7 of the original paper is now Figure S8).

Experimental detailFrozen-scene refit setup
Frozen 3D sceneConverged scene of the f=0f=0 variant of Eq. (1)
Optimized variablesSmoke grid, per-capture coefficients, and medium scalars only
Optimization schedule500 Adam steps, lr 0.001, identical camera order and loss
Dataset76 smoky training views across Koharu, Natsume, and Shirohana
No-forward conditionDilated forward render removed; airlight (1−ω)(1−t)A(1-\omega)(1-t)\boldsymbol{A} with ω\omega optimized

[Q2] Why was DehazeGS not included in the experimental comparison?

Response: DehazeGS (Yu et al., 2026) has no public code or checkpoints (checked November 2026), so we retrained two restore-then-reconstruct baselines (DCP → 3DGS and pretrained DehazeFormer → 3DGS) under one protocol, where VeilSplat leads on all metrics (+0.97 dB over DehazeFormer → 3DGS and +2.84 dB over DCP → 3DGS; +5.17 dB over the strongest official RealX3D baseline in Table 2) and stays at or above 13.45 dB on every scene (Section S4, Table S3, Figure 3).

Retrained under one protocol (5-scene mean, Table S3)Clean PSNR ↑SSIM ↑LPIPS ↓
3DGS (retrained)10.020.5260.697
WaterSplatting (retrained)11.950.5170.751
DCP → 3DGS13.110.5480.625
DehazeFormer → 3DGS14.980.6050.531
VeilSplat (ours, one configuration)15.950.6130.508

VeilSplat reaches ≥13.45\ge 13.45 dB clean PSNR on every room, whereas DCP → 3DGS drops to 11.11 dB on Shirohana and DehazeFormer → 3DGS to 11.31 dB on Hinoki (Table S3). If an official DehazeGS implementation becomes public during discussion, we will evaluate it on all five scenes with its pretrained depth prior preserved and noted.

Thank you again for a clear, focused review that strengthened the joint-training forward-lobe evaluation and restoration comparisons of the paper.

Thank you for the reply. I have no major concerns and maintain my recommendation. And I have updated the AI-use documentation field.

Reviewer_u9ThOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents VeilSplat, a physics-based 3D Gaussian Splatting framework for reconstructing clear 3D scenes from images affected by changing smoke. The method separates surface coverage from medium transmission and introduces a capture-varying density field with three-branch rendering. Experimental results on RealX3D demonstrate substantial improvements over existing methods.

Critical strengths

The separation of surface coverage and medium transmission is meaningful, and the proposed dynamic medium representation is relatively compact. The experimental results are promising, and the ablation studies provide useful evidence supporting the effectiveness of the proposed components.

Critical weaknesses
  1. The physical consistency of the three-branch rendering model and the ambiguity between scene appearance and medium transmission require further clarification. Additionally, the authors should better distinguish their contributions from existing scattering-aware reconstruction methods and discuss their relationship to recent image dehazing approaches.
  2. The experimental comparisons are primarily limited to 3D reconstruction methods, with insufficient consideration of recent image dehazing and restoration approaches. Furthermore, the evaluation lacks direct geometric accuracy measurements, and the computational overhead during training is not sufficiently discussed.
Questions for authors

Please refer to the weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer u9Th · Part 1/2

Response to Reviewer u9Th · Part 1/2

Thank you for your constructive, detailed review and for recognizing the value of our opacity/transmission separation, compact dynamic smoke representation, and ablation studies.

[W1a] Clarify the physical consistency of the three-branch rendering model.

Response: Eq. (1) is a single image formation whose three splat passes are its three terms; it reduces to Koschmieder's law for f=ω=0f=\omega=0, matches single scattering (wf=fτ+O(τ2)w_f=f\tau+O(\tau^2)) in thin smoke, and returns no object light once t→0t\to 0 (Section 3.2, Eq. (1), Figure 2a).

I≈tJ+(t1−f−t)(Kτ∗J)+(1−t)[ω(Kτ∗J)+(1−ω)A]I \approx tJ + (t^{1-f}-t)(K_\tau \ast J) + (1-t)[\omega(K_\tau \ast J) + (1-\omega)\boldsymbol{A}]

ComponentPhysical role
Direct (tJtJ)Unscattered object light attenuated by transmittance t=e−τt = e^{-\tau}
Forward lobe ((t1−f−t)(Kτ∗J)(t^{1-f}-t)(K_\tau \ast J))Small-angle multiple scattering (Wells, 1973; McGlamery, 1980; Jaffe, 1990): surviving object light is t1−ft^{1-f}, so the blurred share t1−f−t=fτ+O(τ2)t^{1-f}-t = f\tau + O(\tau^2) vanishes as t→0t \to 0, with blur variance h2τh^2\tau
Airlight ((1−t)[… ](1-t)[\dots])Indoor smoke is illuminated both by nearby room surfaces (ω(Kτ∗J)\omega(K_\tau \ast J)) and by ambient light ((1−ω)A(1-\omega)\boldsymbol{A})
Chromatic extinction (σtk\sigma_t^k)Separate extinction per RGB channel, matching wavelength-dependent attenuation in paired smoky/clean views (Figure 2c)

One dilated splatting pass supplies Kτ∗JK_\tau \ast J for both forward scattering and scene-lit airlight (Eq. (4), Figure S7), and the three rendered passes sum to the composite observation within 1.192×10−71.192 \times 10^{-7} maximum error (Section S14). Setting f=0f=0 gives the purely scene-lit model, and ω=0\omega=0 gives Koschmieder's law with small-angle forward scattering.

[W1b] Clarify the ambiguity between scene appearance and medium transmission, and how it is resolved.

Response: View-shared Gaussian radiance prevents per-capture smoke from leaking into scene appearance, and the single remaining global scale gauge is pinned down by multi-view contrast decay and a room-scale per-channel black-point prior (+0.48 dB mean PSNR, LPIPS 0.517 → 0.508; Section 3.5, Table 3).

Each 3D Gaussian stores one view-shared colour JJ while smoke lives solely in the observation model. For a nearly uniform medium, (t,J)↦(κt,A+(J−A)/κ)(t, J) \mapsto (\kappa t, \boldsymbol{A} + (J - \boldsymbol{A})/\kappa) leaves tJ+(1−t)AtJ + (1-t)\boldsymbol{A} invariant, and camera-to-surface distances vary by a median factor of 1.03–1.06 across views. Two physical cues fix κ\kappa:

  • Multi-view contrast decay seeds smoke colour: Local contrast decays with metric distance dd as e−σtkde^{-\sigma_t^k d} independent of surface texture, allowing log-contrast regression to seed per-channel σtk\sigma_t^k from smoky views alone (Section S6, Table S5).
  • Per-channel black-point prior fixes global scale: Classical 15×1515\times 15 RGB DCP is 0.068 on dark Hinoki and 0.29–0.36 across the other four clean rooms because bright walls and furniture fill whole 15×1515\times 15 patches (Figure 4a, Table S6). Our room-scale per-channel black-point statistic (Eq. (7); W=201W=201 at 4× pooling) is 0.005–0.031 on clean rooms, with hinged loss (m=0.2m=0.2) ≤0.0053\le 0.0053 (inactive on clean rooms), whereas residual smoke raises the per-channel black-point statistic by ≥7×\ge 7\times in every channel (Figure 4b, Section S7). Together with chromatic extinction, this adds +0.48 dB 5-scene mean PSNR (+1.51 dB on Hinoki, +1.15 dB on Shirohana; Table 3, Table S4).

[W1c] Better distinguish contributions from scattering-aware 3D methods and clarify the relationship to image dehazing.

Response: Section 2 distinguishes VeilSplat from scattering-aware NeRF/3DGS reconstructions along three axes (time-varying basis density, strict opacity/transmission decoupling, and small-angle covariance dilation) and contrasts joint 3D inversion against per-frame image dehazing.

Compared to SeaThru-NeRF (Levy et al., 2023), WaterSplatting (Li et al., 2025), SeaSplat (Yang et al., 2025), I2-NeRF (Liu et al., 2025b), ScatterNeRF (Ramazzina et al., 2023), DehazeNeRF (Chen et al., 2024), and DehazeGS (Yu et al., 2026), VeilSplat combines: (1) a capture-varying medium via shared spatial bases, (2) strict opacity/transmission decoupling so smoke never alters Gaussian alpha-compositing weights, and (3) closed-form covariance dilation weighted by small-angle scattering t1−f−tt^{1-f}-t. SmokeSeer (Jain et al., 2026) needs paired thermal images and smoke segmentation; VeilSplat uses posed RGB only. Unlike single-image dehazers, which estimate tt and A\boldsymbol{A} independently per frame and introduce cross-view colour shifts, VeilSplat inverts one 3D medium across all views.

AuthorsOpenReview ↗

Response to Reviewer u9Th · Part 2/2

Response to Reviewer u9Th · Part 2/2

[W2a] Add comparisons with image dehazing and restoration approaches.

Response: Evaluated against two restore-then-reconstruct pipelines (DCP → 3DGS and pretrained DehazeFormer → 3DGS) retrained under one protocol, VeilSplat leads across PSNR (+0.97 dB over DehazeFormer → 3DGS and +2.84 dB over DCP → 3DGS; +5.17 dB over the strongest official RealX3D baseline in Table 2), SSIM (0.613), and LPIPS (0.508), and is the only method at or above 13.45 dB on every scene (Section S4, Table S3, Figure 3).

We apply Dark Channel Prior (He et al., 2011) and DehazeFormer (Song et al., 2023; official DehazeFormer-MCT mixed-dataset weights for real hazy images) to every smoky training frame prior to 3DGS reconstruction under the same cameras, train-only initial points, resolution, and metrics (Section S4, Table S3):

Retrained under one protocol (5-scene mean, Table S3)PSNR ↑SSIM ↑LPIPS ↓
3DGS (retrained on smoky views)10.020.5260.697
WaterSplatting (retrained)11.950.5170.751
DCP → 3DGS13.110.5480.625
DehazeFormer → 3DGS14.980.6050.531
VeilSplat (ours)15.950.6130.508

VeilSplat stays at or above 13.45 dB on all five rooms (worst scene: 13.45 dB vs. 11.11 for DCP → 3DGS and 11.31 for DehazeFormer → 3DGS; Table S3, Figure S2). Because per-frame dehazers predict inconsistent transmission and colour balance across views, 3DGS bakes in colour casts for DCP and residual haze for DehazeFormer (Figure 3).

[W2b] Include direct 3D geometric accuracy measurements.

Response: Evaluated against the clean-capture 3DGS oracle (36.44 dB test PSNR), VeilSplat achieves 7.51 / 7.85 cm mean accuracy / completeness, outperforming smoky-trained 3DGS at 27.78 / 47.33 cm (Section S11, Table S11).

We measure bidirectional median distances, in metric centimeters, between opaque Gaussian centers (α>0.5\alpha > 0.5) and the clean-capture 3DGS oracle:

5-scene meanAccuracy (cm) ↓Completeness (cm) ↓
3DGS (trained on smoky views)27.78 cm47.33 cm
VeilSplat (ours)7.51 cm7.85 cm

VeilSplat improves accuracy from 12.13 to 4.23 cm on dark Hinoki and from 104.08 to 9.65 cm on Natsume, eliminating the severe floater collapse suffered by smoky-trained 3DGS (Table S11).

[W2c] Discuss the computational overhead during training and inference.

Response: Training takes 78–82 minutes per scene on a single RTX 4090 (~20 GB VRAM), while clean novel-view rendering runs at 1.15–1.33 ms per frame using 0.439–0.497 GB peak VRAM with zero medium overhead (Section S12, Figure S3).

Optimization runs for 24k iterations with up to 10610^6 Gaussians (three splat passes and eight density samples per camera–primitive segment). The smoke field (3×1233 \times 12^3 voxel grids plus two scalars per capture) is queried only during fitting and discarded afterward, so clean rendering at about 1600×1054–1065 px is a single standard 3DGS pass.

Thank you again for a constructive review that sharpened the physical derivation, restoration comparisons, and geometric evaluation of the paper.

Reviewer_Bf6DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

VeilSplat recovers a clean static 3D Gaussian scene from posed smoky images by separating surface coverage from medium transmission. A 12³ log-density field with two per-view coefficients supplies optical depth along camera-to-primitive segments, driving three additive splat passes: attenuated direct radiance, a forward-scattered lobe obtained by dilating primitive covariances, and constant airlight. On five RealX3D scenes it reports 16.30 dB mean PSNR, +5.44 dB over the strongest released baseline, plus component studies.

Critical strengths
  1. The coverage/transmission separation is the right decomposition for this problem, and the paper follows it to a genuinely useful consequence: because radiance is shared across views and the medium exists only as an observation model, clean inference collapses to an ordinary 1.15–1.33 ms splat, and the scene cannot absorb unexplained per-view variation into its colours. That is the standard failure mode of transient-appearance models, so getting it right is what makes the recovered scenes qualitatively different from the baselines.

  2. The covariance dilation with determinant-ratio opacity compensation is a cheap, general way to obtain an image-space-constant scattering kernel inside a splatting pipeline, without a second volumetric renderer. Even if the physical parameterization needs revision, the mechanism is reusable beyond this paper.

Critical weaknesses
  1. The headline margin cannot be attributed to the method. SeaThru-NeRF reaches 4.948 dB on Hinoki and I2-NeRF 5.704 dB on Shirohana, below what a constant-image predictor would achieve, and Fig. 3 shows baselines that are either fully washed out or nearly black — consistent with a colour-space or exposure mismatch between the smoky inputs and the separately captured clean GT. The baselines are the benchmark's released predictions rather than retrained models, and the initialization and poses are themselves inherited from clean-capture SfM, which is never disclosed: RealX3D estimates poses on the GT images precisely because COLMAP fails on the degraded ones. Until at least 3DGS and WaterSplatting are reproduced under the authors' own protocol, +5.44 dB is not evidence of much.

  2. The two claims that carry the title are not established. The forward lobe is weighted by ω(1−t) where single scattering gives τe^{−τ}, so it grows with optical depth and approaches the full unattenuated scene radiance as t→0, behaving as de-occlusion rather than scattering; the ablation supporting it removes the dilation, ω and the only spatially varying additive term at once, so it measures capacity rather than broadening, and the decisive control (h→0 with ω(1−t)c retained) is missing. More importantly, the regime the title advertises is never tested. All five sequences are captured after the smoke "diffuses uniformly", and Fig. 5 shows whole-view transmission varying by only ~0.05–0.1 over 25–26 captures. Under smoke that evolves substantially, the entire per-view difference would have to be carried by a rank-2 medium with two scalars per capture on a 12³ grid: advection is representable at best to first order, a plume cannot translate, deform, split or grow, and because the bases are spatially zero-mean the mean log-density is identical for every capture, so the total amount of smoke cannot change. The TV, Gram and coefficient-smoothness penalties push in the same direction, damping exactly the fast changes such a sequence would contain. Since scene radiance is shared across views, the shortfall cannot be hidden in appearance and must surface as photometric residual, so the expected behaviour is underfitting rather than a wrong decomposition — but that is the regime the paper claims to handle, and it is the one regime absent from the evaluation.

Questions for authors
  1. Are the camera poses and sparse tracks taken from RealX3D's COLMAP run on the clean GT images? What is the PSNR when poses are estimated from the smoky frames, or when pose noise and dropped views are injected?
  2. Why is the forward weight (1−t) rather than τe^{−τ}? What happens to §4.5's 25× MSE reduction if the physical weight is used, and what does the h→0 control (keeping ω(1−t)c without dilation) give?
  3. Why do SeaThru-NeRF and I2-NeRF fall below 6 dB on some scenes? Were the released predictions produced at the same resolution, colour space and exposure as the clean GT, and can the authors reproduce 3DGS and WaterSplatting themselves?
  4. What are the fitted σ_t (in m⁻¹), ω and h for each scene, and do they lie in a physically plausible range for smoke?
  5. Can the method be shown on a sequence in which the smoke visibly and substantially evolves (or a synthetic sequence with known motion), and does the recovered scene radiance stay stable while the medium absorbs the change? Without such a case, what supports the "changing smoke" claim?
  6. What are the two singular values of the coefficient matrix (or the angle between b1 and b2) — is the fitted model actually rank 2 rather than rank 1 in the five scenes, and what is the per-pixel temporal variance of transmission there?
  7. Does the forward lobe still help on held-out smoky views after full joint refitting, rather than on the 76 training captures after 500 medium-only steps?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Bf6D Part 1/3

Response to Reviewer Bf6D Part 1/3

Thank you for a perceptive, technically sharp review, and for recognizing both why separating surface coverage from smoke transmission prevents per-view leakage and how determinant-compensated covariance dilation provides a reusable image-space scattering kernel inside splatting.

[W1 & Q3] Why do released baselines such as SeaThru-NeRF and I2-NeRF fall below 6 dB? Are the test views matched in resolution, color space, and exposure to the clean GT, and can the authors retrain 3DGS and WaterSplatting under their own protocol?

Response: Retrained under one protocol (same cameras, smoky training views, train-only initial points, quarter resolution, display-referred sRGB; Section S4, Table S3), 3DGS reaches 10.02 dB and WaterSplatting 11.95 dB against 15.95 dB for VeilSplat (main Table 2 reports the official RealX3D results, where VeilSplat leads WaterSplatting by +5.17 dB), and a clean-capture oracle reaches 36.44 dB on the same test views, confirming that cameras, colour pipeline, and clean test images are mutually consistent (Section S3).

  • Matched resolution, colour space, and exposure: All official and retrained methods are scored at RealX3D's quarter resolution in display-referred sRGB ([0, 1]) on the same clean test views, where clean-trained 3DGS reaches 36.44 dB / 0.952 SSIM and retrained 3DGS (10.02 dB) matches RealX3D's published 3DGS (9.76 dB; 9.88 dB re-scored).
  • Why official baselines score below 6 dB on dark rooms: Those rows in main Table 2 are the official RealX3D results (Liu et al., 2025a); scoring RealX3D's distributed predictions reproduces the published table within 0.71 dB across all methods and scenes (Section S3, Table S2; e.g., SeaThru-NeRF on Hinoki is 4.90 dB in RealX3D and 4.95 dB here; WaterSplatting mean is 10.78 vs. 10.86 dB). These methods model backscatter as a near-uniform ambient colour, whereas indoor smoke is lit by the room itself; their medium-free renders either retain a bright veil or assign dark surfaces to the medium branch and output near-black frames (Figure S1). On dark Hinoki (mean clean linear RGB 0.19/0.10/0.05), a veiled render is wrong by a large fraction of the dynamic range (Section S3).
  • Retrained under one protocol and one VeilSplat configuration across all five scenes (+5.94 dB over 3DGS, +4.00 dB over WaterSplatting; Section S4, Table S3, Figure 3):
Retrained under one protocol (5-scene mean, Table S3)PSNR ↑SSIM ↑LPIPS ↓
3DGS on clean captures (oracle)36.440.9520.061
3DGS (retrained)10.020.5260.697
WaterSplatting (retrained, medium-free output)11.950.5170.751
DCP → 3DGS13.110.5480.625
DehazeFormer → 3DGS14.980.6050.531
VeilSplat (ours)15.950.6130.508

[Q1] Are the camera poses and sparse tracks taken from RealX3D's COLMAP run on the clean GT images? What is the PSNR when poses are estimated from the smoky frames, or when pose noise and dropped views are injected?

Response: Yes: all methods in Tables 2 and S3, including RealX3D's released predictions, share the benchmark calibration from COLMAP on the clean captures (Section 4.1). Re-estimating cameras and initial points from the smoky training images alone and retraining VeilSplat unchanged gives 15.85 dB PSNR vs. 16.02 dB for the benchmark calibration (both with frozen-model test-camera refinement; main protocol: 15.95 dB) while surpassing clean-capture poses in SSIM (0.663 vs. 0.625) and LPIPS (0.460 vs. 0.506) in every room (Section S10, Table S9).

VeilSplat, 5-scene mean (Table S9, refined test cameras)PSNR ↑SSIM ↑LPIPS ↓
Benchmark calibration (clean captures; unrefined: 15.95 dB)16.020.6250.506
Smoky training images only15.850.6630.460
  • Real pose shifts, dropped views & why SSIM/LPIPS beat clean-pass poses: COLMAP with fixed intrinsics and CLAHE on the smoky frames registers all 123/123 views and triangulates 33664–62242 points per room (vs. 25,351–30,991 benchmark train-track points), testing real SfM shifts of 0.28°–0.54° median (up to 0.84°) rotation and 0.48%–1.32% (up to 4.13%) position (Table S10). Because RealX3D captures clean and smoky passes separately, bundle-adjusting directly on the smoky frames removes cross-pass pose offsets and sharpens multi-view edges in every room; withholding 4 smoky captures per scene (18, 21, 22 views) likewise retains 15.41 dB clean PSNR (Section 4.6, Tables 5, S13).
  • Test cameras: clean test captures lie outside the smoky SfM, so as in pose-free NVS (BARF, NoPe-NeRF, CF-3DGS), each benchmark test camera is mapped by a training-camera Sim(3) and 6-DoF refined against the frozen model (+0.07 dB on benchmark-calibration models: 15.95 → 16.02 dB).
AuthorsOpenReview ↗

Response to Reviewer Bf6D Part 2/3

Response to Reviewer Bf6D · Part 2/3

[W2a & Q2] Weighting the forward lobe by ω(1−t)\omega(1-t) grows with optical depth and approaches unattenuated radiance as t→0t \to 0, acting like de-occlusion rather than scattering (where single scattering gives τe−τ\tau e^{-\tau}). What happens to §4.5's 25× MSE reduction with a physical weight, and what does the h→0h \to 0 control give?

Response: Agreed on ω(1−t)\omega(1-t): revised Eq. (1) replaces it on the forward lobe with the small-angle multiple-scattering share wf=t1−f−t=fτ+O(τ2)w_f = t^{1-f} - t = f\tau + O(\tau^2), which matches single scattering at small τ\tau and vanishes in both clear air (t=1t=1) and opaque smoke (t→0t \to 0), while ω(1−t)\omega(1-t) acts solely as scene-lit airlight on the low-passed radiance Kτ∗JK_\tau \ast J (Section 3.2, Eqs. (1), (4), Figure 2a); Table 5 evaluates this physical weight and the h→0h \to 0 control under full joint training.

  • Small-angle forward weight (Section 3.2, Eq. (1), Figure 2a): Under classical small-angle scattering (Wells, 1973; McGlamery, 1980; Jaffe, 1990), a fraction ff of extinguished light scatters into a narrow forward cone along the ray. Out of surviving object light e−(1−f)τ=t1−fe^{-(1-f)\tau} = t^{1-f}, a portion tt arrives unscattered and the remainder arrives blurred:

wf(τ)=t1−f−t=fτ+O(τ2),arg⁡max⁡τwf=−ln⁡(1−f)f,wf→t→00.w_f(\tau) = t^{1-f} - t = f\tau + O(\tau^2), \qquad \arg\max_\tau w_f = -\frac{\ln(1-f)}{f}, \qquad w_f \xrightarrow{t \to 0} 0.

This weight matches single scattering to first order, peaks at intermediate optical depth, and decays to zero as t→0t \to 0, returning no more object energy than survives attenuation; accumulated small-angle deflections yield blur variance h2τh^2\tau.

  • Scene-lit airlight: In (1−t)[ω(Kτ∗J)+(1−ω)A](1-t)[\omega(K_\tau \ast J) + (1-\omega)\boldsymbol{A}], the ω(1−t)\omega(1-t) share multiplies only Kτ∗JK_\tau \ast J, never sharp JJ: the kernel width grows as hτh\sqrt{\tau}, so this term carries only low-passed room radiance, never sharp object detail (Section 3.2). Setting f=0f=0 gives the purely scene-lit model, and ω=0\omega=0 gives Koschmieder's law with small-angle forward scattering.
  • Frozen-scene MSE ratios vs. joint training: The frozen-scene MSE ratios (25.98×, 24.55×, and 12.00× on Koharu, Natsume, and Shirohana: 6.000×10−56.000\times 10^{-5}–1.055×10−41.055\times 10^{-4} vs. 1.266×10−31.266\times 10^{-3}–1.874×10−31.874\times 10^{-3}; Section S16, Figure S8) were measured on the f=0f=0 scene-lit formulation and are cited strictly as an observation-fitting diagnostic for that scene-lit term; the clean-view effect of the physical t1−f−tt^{1-f}-t weight and h→0h \to 0 control is measured end-to-end under joint training in Section 4.6 (Tables 5, S13).

[Q7] Does the forward lobe still help on held-out smoky views after full joint refitting, rather than on the 76 training captures after 500 medium-only steps?

Response: With smoky captures 4, 10, 16, and 22 withheld from joint scene–medium training, we evaluate directly on the clean test views—a stricter test than smoky observations, where unmodeled haze can bake into the Gaussians—physical forward and scene-lit weights add +0.35 dB clean test PSNR (15.06 → 15.41 dB), and the full model with covariance dilation achieves the best SSIM (0.577) and LPIPS (0.531) (Section 4.6, Tables 5, S13).

We jointly train three variants on Hinoki, Koharu, and Shirohana (18, 21, and 22 smoky training captures; Section 4.6, Table 5): (1) Full model (h>0h > 0), (2) No forward lobe (f=ω=0f=\omega=0) with standard airlight (1−t)A(1-t)\boldsymbol{A}, and (3) h→0h \to 0 control (weights f,ωf, \omega kept, spatial dilation disabled):

Variant (3-scene mean, Table 5)Clean PSNR ↑SSIM ↑LPIPS ↓
Full model (h>0h > 0)15.410.5770.531
No forward lobe (f=ω=0f=\omega=0)15.060.5600.562
h→0h \to 0 control (weights kept, no dilation)15.460.5690.553
  • Scattered scene light (f,ω>0f, \omega > 0 vs. f=ω=0f = \omega = 0): Modeling scattered scene light removes global veil bias, lifting clean PSNR from 15.06 dB to 15.41 dB (+0.35 dB).
  • Covariance dilation (h>0h > 0 vs. h→0h \to 0): At comparable PSNR, enabling covariance dilation improves the edge-sensitive metrics: SSIM 0.569 → 0.577, LPIPS 0.553 → 0.531, because the medium absorbs edge halos instead of the clean Gaussians.
AuthorsOpenReview ↗

Response to Reviewer Bf6D · Part 3/3

Response to Reviewer Bf6D · Part 3/3

[Q4] What are the fitted values of σt\sigma_t (m−1\text{m}^{-1}), ω\omega, and hh for each scene, and are they physically plausible for smoke?

Response: Fitted extinctions of 0.22–0.38 m−1\text{m}^{-1} (meteorological visibility 10.2–18.1 m), forward fractions f=f = 0.09–0.14, blur widths h=h = 31–33 px, and scene-lit shares ω=\omega = 0.25–0.36 all lie in a physically plausible range for indoor smoke/fog (Section S8, Table S7).

Physical quantityFitted range across the 5 scenes
Per-channel extinction σtk\sigma_t^k0.22–0.38 m−1\text{m}^{-1}
Meteorological visibility (3.912/σtk3.912/\sigma_t^k)10.2–18.1 m
View-mean transmittance tˉ\bar t (min–max across captures)0.257–0.607 (Table S8)
Small-angle forward fraction ff0.09–0.14
Scattering blur width hh at τ=1\tau=131–33 px
Scene-lit airlight fraction ω\omega0.25–0.36

[W2b & Q5] Zero-mean rank-2 bases keep mean log-density constant across views, TV/Gram/smoothness penalties may damp rapid change, and RealX3D has mild temporal variation. How does the method perform when smoke visibly and substantially evolves?

Response: Under a 7×7\times transmittance swing with moving, growing and splitting plumes, the rank-2 medium beats a static grid by +2.26 dB and 3DGS by +7.94 dB; the per-capture global level (Section 3.3), which lets the total amount of smoke change, gives the best SSIM (0.764) and LPIPS (0.343 vs. 0.757 for 3DGS) (Section 4.5, Table 4).

  • Controlled evolving-smoke benchmark: On Koharu's 25 clean training views, we composite procedural 3D volumetric smoke from 4 puffs that cross up to 80% of the room and grow by 2.5×, one of which splits in two, on a background whose level swings by 60%. Mean view transmittance swings by 7×7\times (0.094–0.654), and evaluation is directly against the real clean test captures. Because the bases are zero-mean, Section 3.3 adds a per-capture global level so the total amount of smoke can change:
Strong evolving smoke (clean test views)PSNR ↑SSIM ↑LPIPS ↓
3DGS (no medium model)7.270.5620.757
Static 12312^3 medium grid12.950.7110.433
VeilSplat (rank-2 bases)15.210.7620.351
VeilSplat (rank-2 + global level)15.140.7640.343
  • Regularizer check: Every run in Table 4 uses the same TV, basis-Gram, and temporal coefficient-smoothness weights as all RealX3D runs (Section 4.5, Section S2), and the rank-2 medium still gains +2.26 dB over the static grid, so these penalties do not suppress plume translation, growth or splitting.

[Q6] Is the learned RealX3D medium rank-2 rather than rank-1 (singular values, basis angle), and how much does transmission vary over time?

Response: Both spatial bases are active in every RealX3D scene, with singular-value ratios γ2/γ1=\gamma_2/\gamma_1 = 0.37–0.76 and near-orthogonal bases (∣cos⁡(bˉ1,bˉ2)∣≤0.028|\cos(\bar b_1, \bar b_2)| \le 0.028), producing per-primitive temporal transmittance standard deviations up to 0.103 at the 90th percentile (Section S9, Table S8).

Holding the middle camera and 3D scene fixed while stepping through the learned per-capture smoke states, the temporal standard deviation of per-primitive transmittance has a median of 0.013–0.068 and a 90th percentile of 0.026–0.103 across scenes (highest on Koharu at 0.068 and Hinoki at 0.052; Table S8, Figures S5–S6).

Thank you again for a thorough, incisive review whose physical and experimental questions substantially strengthened the paper.

Comments on the paper

1 comment

We thank all three reviewers for their thoughtful, rigorous reviews and for recognizing our core design choices: separating surface opacity from smoke transmission (Bf6D, u9Th, 5FYA), modeling forward scattering via covariance dilation with determinant-ratio compensation (Bf6D), and capturing time-varying smoke with a compact basis field (u9Th, 5FYA). We have uploaded the updated main.pdf and supp.pdf; prefix S refers to supp.pdf, and Table 1 lists all notation.

1. Image Formation and Smoke–Radiance Gauge (Bf6D, u9Th; Sections 3.2, 3.5, Eq. (1))

I≈tJ+(t1−f−t)(Kτ∗J)+(1−t)[ω(Kτ∗J)+(1−ω)A]I \approx tJ + (t^{1-f}-t)(K_\tau \ast J) + (1-t)[\omega(K_\tau \ast J) + (1-\omega)\boldsymbol{A}]

  • Small-angle forward scattering and scene-lit airlight: When a fraction ff of extinguished light scatters into a narrow forward cone, total surviving object light is t1−ft^{1-f}, splitting into unscattered direct light tt and blurred forward-scattered light (t1−f−t)(Kτ∗J)(t^{1-f}-t)(K_\tau \ast J). The weight t1−f−tt^{1-f}-t matches single scattering (fτf\tau) at small τ\tau and vanishes as t→0t \to 0, eliminating de-occlusion in thick smoke (Figure 2a). In Eq. (1) the ω(1−t)\omega(1-t) share is scene-lit airlight multiplying only the low-passed radiance Kτ∗JK_\tau \ast J, never sharp JJ, with per-channel extinction σtk\sigma_t^k seeded from multi-view contrast decay on smoky training views (Figure 2c, Section S6, Table S5).
  • Per-channel black-point prior: Because clean indoor rooms contain dark shadow structure in each channel at room scale, Eq. (7) fixes the global transmission–radiance scale. Its hinged loss is ≤0.0053\le 0.0053 on clean ground-truth captures (inactive on clean rooms), whereas residual smoke without the prior raises the per-channel black-point statistic by ≥7×\ge 7\times across every channel and scene (Figure 4, Section S7).

2. Official RealX3D Benchmark and Retrained Protocol (Bf6D, u9Th, 5FYA; Section 4.1, Supp. S4)

Replacing per-scene tuning in the submission with one shared configuration across all five scenes (Supp. S2) under Eq. (1), VeilSplat reaches 15.95 dB mean PSNR (+5.17 dB over the strongest published RealX3D baseline in Table 2, +5.94 dB over retrained 3DGS, +4.00 dB over retrained WaterSplatting in Supp. S4) and ≥13.45\ge 13.45 dB on every scene:

Official RealX3D benchmark (Table 2, 5-scene mean)PSNR ↑SSIM ↑LPIPS ↓
3DGS on clean captures (oracle)36.440.9520.061
3DGS (Liu et al., 2025a)9.760.4990.659
SeaThru-NeRF (Liu et al., 2025a)7.550.4640.683
WaterSplatting (Liu et al., 2025a)10.780.4450.723
SeaSplat (Liu et al., 2025a)10.420.4460.774
I2-NeRF (Liu et al., 2025a)8.400.2830.699
VeilSplat (ours, one configuration)15.950.6130.508

Scoring RealX3D's distributed predictions reproduces the published table within 0.71 dB on every method and scene (Supp. S3, Table S2). Supp. S4 retrains 3DGS, WaterSplatting, and two dehaze-then-reconstruct pipelines under one protocol (same cameras, smoky training views, train-only sparse points coloured from smoky views, quarter resolution, and sRGB metrics):

Retrained under one protocol (Supp. S4, Table S3)PSNR ↑SSIM ↑LPIPS ↓
3DGS (retrained)10.020.5260.697
WaterSplatting (retrained)11.950.5170.751
DCP → 3DGS (He et al., 2011)13.110.5480.625
DehazeFormer → 3DGS (Song et al., 2023)14.980.6050.531
VeilSplat (ours, one configuration)15.950.6130.508

3. Controlled Evolving-Smoke and Joint-Training Diagnostics (Bf6D, 5FYA)

  • Strongly evolving smoke: Under a 7×7\times transmittance swing with moving, growing, and splitting plumes, the rank-2 medium beats a static grid by +2.26 dB and 3DGS by +7.94 dB, while the per-capture global level gives the best SSIM (0.764) and LPIPS (0.343) (Section 4.5, Table 4).
  • Joint-training forward-lobe ablation (h→0h \to 0 control): With scene and medium fitted jointly and smoky captures 4, 10, 16, and 22 withheld, physical forward and scene-lit weights raise clean test PSNR by +0.35 dB, and covariance dilation yields the best SSIM (0.577) and LPIPS (0.531) (Section 4.6, Tables 5, S13).
  • Additional diagnostics: Retrained end-to-end from COLMAP on the smoky images alone (all 123/123 views registered), VeilSplat reaches 15.85 dB vs. 16.02 dB for the benchmark calibration (both with frozen-model test-camera refinement; main protocol: 15.95 dB) while beating it on SSIM (0.663 vs. 0.625) and LPIPS (0.460 vs. 0.506) in every room—avoiding cross-pass pose offsets and yielding denser CLAHE seeds (Section S10)—alongside metric parameters (Section S8), rank-2 variation (Section S9), and 3D geometry (Section S11).

Thank you again to all three reviewers for the detailed feedback that sharpened both the physics and the empirical evaluation of VeilSplat.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.