acceptodds
Under review as a conference paper at ICLR 2027

SpectralSplat: Appearance-Disentangled Feed-Forward Gaussian Splatting for Driving Scenes

Abstract

Feed-forward 3D Gaussian Splatting methods have achieved impressive reconstruction quality for autonomous driving scenes, yet they entangle scene geometry with transient appearance properties such as lighting, weather, and time of day. This coupling prevents relighting, appearance transfer, and consistent rendering across multi traversal data captured under varying environmental conditions. We present SpectralSplat, a method that disentangles appearance from geometry within a feed-forward Gaussian Splatting framework. Our key insight is to factor color prediction into an appearance-agnostic base stream and an appearance-conditioned adapted stream, both produced by a shared MLP conditioned on a global appearance embedding derived from DINOv2 features. To enforce disentanglement, we train with paired observations generated by a hybrid relighting pipeline that combines physics-based intrinsic decomposition with diffusion-based generative refinement, and supervise with complementary consistency, reconstruction, cross-appearance, and base color losses. We further introduce an appearance-adaptable temporal history that stores appearance-agnostic features, enabling accumulated Gaussians to be re-rendered under arbitrary target appearances. Experiments demonstrate that SpectralSplat preserves the reconstruction quality of the underlying backbone while enabling controllable appearance transfer and temporally consistent relighting across driving sequences.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian splatting, feed-forward network, appearance disentanglement, autonomous driving, novel view synthesis, neural rendering, scene representation, illumination separation, multi-view reconstruction

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_NsnmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SpectralSplat studies how to adjust the weather and lighting appearance of driving scenes after reconstructing them with feed-forward 3D Gaussian Splatting. Built on UniSplat, the method introduces a global appearance code into color prediction: a zero code produces base colors, while the actual code produces colors under the corresponding appearance. To obtain training pairs of the same scene under different conditions, the authors combine physical rendering priors with diffusion relighting. Reconstruction, appearance-swap, base-color consistency, and other losses encourage the separation of scene content and appearance. At inference, a prebuilt appearance library allows different appearances to be applied while keeping the reconstructed geometry fixed. The method also caches features instead of fixed colors, allowing accumulated Gaussians to be recolored using the current code. Experiments evaluate reconstruction on Waymo and nuScenes and demonstrate appearance swapping, library-based control, and code interpolation on Waymo. The method primarily adjusts overall scene appearance rather than providing a complete physical simulation of weather.

Critical strengths
  1. The problem has practical value, and its integration with feed-forward reconstruction is natural. For driving scenes, being able to switch appearances after reconstruction without optimizing each scene again is a useful capability. Beyond demonstrating relighting for an individual scene, this paper focuses on incorporating this capability into a reusable feed-forward reconstruction system. I think this is a valuable direction.

  2. The design of the historical feature cache makes sense. If the cache stores fixed RGB values, changing the appearance code may create color inconsistencies between historical Gaussians and the current frame. Caching features and then predicting colors consistently using the selected code is an intuitive solution that fits the needs of temporal reconstruction.

Critical weaknesses
  1. The contribution beyond existing appearance-conditioned GS methods needs further clarification. Using appearance embeddings to adjust Gaussian colors and using MLPs to predict conditioned colors have appeared in prior work. What stands out here is the integration of these designs with feed-forward driving-scene reconstruction, an appearance library, and temporal caching. However, I am not yet fully convinced that the proposed code extraction and library construction are necessary compared with simpler alternatives. For example, with the same synthetic data and training budget, could directly learning a condition embedding for each weather condition or time of day achieve similar results? The paper also acknowledges that an advantage over this alternative has not been established. I think this would be a useful baseline to help identify which parts of the design account for the gains.

  2. The central capability of reusing the appearance library across scenes still lacks sufficiently direct quantitative validation. The appendix states that the current library is constructed from the augmented Waymo validation split. The existing visualizations therefore do not establish whether appearance can still transfer reliably when the library construction scenes and test scenes are completely disjoint. In addition, Table 3 uses appearance codes encoded from each image pair rather than the averaged codes in the library, so it does not directly evaluate the library as used in practice. Given that the authors also show unreliable transfer between real images, I think the current evidence more strongly supports control within the synthetic appearance distribution. Broader conclusions about content–appearance disentanglement should be made cautiously.

Questions for authors
  1. Regarding Weakness 1: Could you add a baseline that directly learns a condition embedding for each weather condition or time of day, using the same backbone, synthetic data, and training budget? I would like to see what clear benefits extracting codes from images and averaging them into a library provides over this simpler alternative, to better establish the contribution of this design.

  2. Regarding Weakness 2: Could you clarify whether the library construction scenes overlap with the test scenes, and quantitatively evaluate the averaged library codes when the two sets are completely disjoint? Beyond the appearance swapping in Table 3 using codes from the paired images themselves, I would like to see whether fixed library codes can also reliably control appearance in new scenes.

3 comments

We thank the reviewer for the thoughtful review. We are glad the reviewer sees reusable feed-forward appearance control as a valuable direction and found the feature-cache design sensible.

We ran both experiments you asked for. In short:

  • Q1: a learned per-condition table is on par with our encoder on seen conditions, and 2.4 dB behind on a new condition (Table R4).
  • Q2: a library built from fully disjoint scenes matches the submitted one (20.70 vs. 20.65 dB; Table R1b).
  • Real photos: rain edits are recognized as real rain in 58–72% of unseen scenes (Table R6).

W1 (contribution over appearance-conditioned GS).

"The contribution beyond existing appearance-conditioned GS methods needs further clarification."

We agree that appearance embeddings and conditioned color MLPs are established.

  • Prior work: NeRF-W and WildGaussians optimize each scene. Feed-forward in-the-wild methods (WildSplatter, GenWildSplat, WildSplat) target a few unposed views of static landmarks.
  • Our contribution: we make appearance controllable in feed-forward, multi-camera, streaming driving reconstruction, where each place is observed once, with a reusable library and a history that can be recolored.
  • Evidence:
    • Existing feed-forward methods trail SpectralSplat by 1.6–4.9 dB on driving scenes (Table R5, in our reply to Reviewer 49n4).
    • The library and the history are now evaluated directly (Table R1b in part 2/2; Table R2 in our reply to Reviewer 49n4).

Q1 (learned per-condition table).

"could directly learning a condition embedding for each weather condition or time of day achieve similar results?"

On seen conditions, yes, about the same. On a new condition, no: it is 2.4 dB worse. Thank you for suggesting this baseline.

  • Setup: a matched comparison, fine-tuned from the submitted model. The two variants are identical except for the source of the code:
    • (a) a learned embedding table indexed by condition;
    • (b) our encoder, with a library built from training scenes.
  • Budget: 1,500 steps each, on the same relit pairs, in the same order.
  • New-condition test: we repeated both runs with clear_night removed from fine-tuning. The starting model has seen night-like lighting, so this tests adding a condition without retraining.
  • Evaluation: 202 validation scenes. Both variants are fine-tuned on pre-generated pairs, so their absolute numbers are compared with each other, not with Table R1b.

Table R4:

Code sourceSeen conditions PSNR↑ / SSIM↑ / LPIPS↓Held-out clear_night PSNR↑
Learned table20.24 / 0.660 / 0.47920.09 (no entry; mean of entries)
Learned table, best entry chosen per frame using the target (oracle)—22.49
Encoder + library (ours)20.34 / 0.680 / 0.49122.52 (30 exemplars, unseen in fine-tuning)
Reference: trained with clear_night (table / encoder)—21.96 / 22.25

Optimized table, frozen decoder. We also optimized nine codes directly for reconstruction through the frozen submitted decoder. They match the library centroids in PSNR (−0.06 dB [−0.13, 0.01]) and improve LPIPS by 0.030.

  • On seen conditions, the encoder matches a dedicated per-condition table. The table's PSNR differs by −0.10 dB [−0.23, +0.02]; it has better LPIPS and lower SSIM.
  • Unlike a table, the encoder adds a condition without retraining. This answers the open question noted in the paper.
    • 30 exemplars give 22.52 dB, as good as when the condition was in training.
    • The table has no entry and loses 2.4 dB [1.8, 3.1]. It matches only with an oracle choice of entry.
  • The encoder needs no condition labels during training, and it supports per-image codes from a single reference (+1.9 dB; Table R1b, part 2/2).

W2/Q2 (library overlap, centroids).

"Could you clarify whether the library construction scenes overlap with the test scenes, and quantitatively evaluate the averaged library codes when the two sets are completely disjoint?"

Yes, the submitted library overlapped with the test scenes. Following your suggestion:

  • we rebuilt the library from 48 training-split scenes only;
  • we evaluated its fixed centroids on all 202 validation scenes (2,020 frames), with relit targets balanced over the nine conditions.

CIs are 95%, from a bootstrap over scenes.

Table R1b (main rows; the full table is in our reply to Reviewer 49n4):

Code used at render timePSNR↑SSIM↑LPIPS↓
Per-image code from the target (upper bound)22.590.7240.450
Library as submitted (24 validation scenes)20.650.6950.501
Library from 48 training scenes (disjoint)20.700.6980.500
Library from the other half of validation (disjoint, 2-fold)20.700.6970.498
Mean of all codes (no condition information)19.800.6830.504
Centroid of a wrong condition19.070.6690.508
  • The library generalizes to unseen scenes: disjoint libraries score the same as the submitted one.
  • The centroids carry condition information: +0.90 dB [0.82, 0.99] over the mean code and +1.64 dB [1.51, 1.77] over a wrong condition. The gain is largest at night (+1.8 to +2.5 dB).
  • They control the appearance of new scenes:
    • A probe recognizes the intended condition from renders of unseen scenes as often as for per-image codes (9-way: 25.7% vs. 21.3%; chance 11%; Table R1a, in our reply to Reviewer 49n4).
    • Mean luminance tracks the target conditions with r = 0.97.

Control within the synthetic distribution. We agree, and we will phrase the claim this way. To quantify how far the edits reach real appearance, we classified library renders of 325 unseen daytime frames. The classifier was trained only on real Waymo photos and reaches 97% accuracy on held-out real scenes. Table R6:

Library codeReal-photo classifierNo editCLIP zero-shotNo edit
rain (day)57.8%0.6%69.5%4.0%
rain (golden hour)72.0%0.6%60.3%4.0%
clear night0.6%0.0%56.0%13.2%
rainy night0.9%0.0%56.6%13.2%
  • Rain: codes make unseen scenes read as real rain, for a classifier that never saw a synthetic image.
  • Night: CLIP reads 56% of the edited scenes as night, against 13% without an edit. The real-photo classifier rarely does, because real Waymo night scenes are dominated by street lamps and headlights (local light sources).
  • The model reproduces its training distribution faithfully.
    • Our renders are 2–3× closer to the synthetic targets than unedited renders are.
    • Their remaining distance to real photos (KID×10³ 0.61–0.69) matches that of the generator's own images (0.46–0.67), so a stronger generator carries over directly.

We hope these results address your concerns. If they do, we would be grateful if you would consider updating your score. We are happy to answer any further questions.

Thank you for the additional experiments and clarifications. Your rebuttal addresses some of my concerns. After considering your response, I have decided to maintain my original score.

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

The paper aims to disentangle geometry from appearance factors (such as lighting, exposure, weather) for feed-forward gaussian splatting framework tailored for autonomous driving. The author proposes to use appearance codes obtain from DINOv2 to represent base color and weather-effect-controlled color, which are both obtained from the proposed hybrid relighting pipeline. Additionally, the gaussian history also keeps the cache per-pixel features, allowing rendering with relighting supported. The paper is well supported by experiments with a lot of qualitative results, and the ablation are sufficient.

Critical strengths
  • The paper addresses an interesting use case where we want to disentangle geometry and lighting effect for feed-forward gaussian splatting case.
  • The quantitative results are thorough, ablation are well thought-out and complete, and quantitative results are meaningful and sufficient.
Critical weaknesses
  • The hybrid relighting and weather simulation approach is an established field with many concurrent works that are not cited. For example, WeatherWeaver [1] can already do both and its demonstrated performance and controllable weather effect. More comparison results might be needed if the author wants to claim this pipeline as one of the core contributions.
  • The paper’s main motivation is to disentangle lighting, weather, and exposure from input camera image, but the method only seems to partially address the global lighting effect and global exposure, with per-camera exposure, local lighting effects, and weather simulation either unaddressed or deferred into limitations.

[1] Lin et. al., Controllable Weather Synthesis and Removal with Video Diffusion Models, ICCV 2025

Questions for authors
  • Some qualitative results exhibit very noticeable artifacts, especially in Figure 4, 5, 13, that could affect downstream usage but are not addressed in the limitations. What could have cause such an issue?
3 comments

We thank the reviewer for the review and for the positive assessment of our experiments and ablations. In short:

  • W1: a controlled generator comparison shows that physics guidance lowers cross-view error by 11% over IC-Light alone (Table R3).
  • Contributions: SpectralSplat is ahead of contemporary feed-forward appearance methods by 1.6–4.9 dB on driving scenes.
  • W2: we make the global scope explicit. Within it, rain edits are recognized as real rain in 58–72% of unseen scenes by a classifier trained only on real photos (Table R6).
  • Q: we explain the artifacts in Figs. 4, 5 and 13.

W1 (hybrid relighting, WeatherWeaver).

"More comparison results might be needed if the author wants to claim this pipeline as one of the core contributions."

Thank you. Table R3 below provides these comparisons.

  • Finding: the physics guidance improves the cross-view consistency of the generated pairs. We will frame the generator's contribution around it.
  • Related work: we will cite WeatherWeaver (Lin et al., ICCV 2025) and related work on weather synthesis and removal.
  • A different goal: our generator targets structurally aligned multi-view training pairs for a feed-forward reconstructor, not general weather synthesis.
  • Modularity: stronger generators, including video models such as WeatherWeaver, can be plugged in.
  • Comparison with WeatherWeaver: its dataset is public, but its code and weights are not yet. We will add it to Table R3 when they are released.

Table R3 (main columns; the full table is in our reply to Reviewer 49n4). Identical prompts, light parameters, seed and solver on 600 validation images; cross-view color error is relative to real photos:

GeneratorCross-view↓Structure↑CLIP prompt↑Time of day↑KID×10³↓
IC-Light only1.800.34420.650.3500.597
Physics only2.470.645†17.410.683†0.475†
Hybrid, uniform (same peak strength)1.730.33720.260.4170.554
Hybrid, uniform (same mean strength)1.580.33320.420.3920.558
Hybrid, frequency-aware (submitted)1.600.35020.440.4000.564

† Near-black night renders inflate these scores; the physics render cannot express weather.

  • What the physics reference adds over diffusion alone (paired tests, all significant):
    • about 11% lower cross-view error;
    • better structure;
    • +5 points in time of day.

Core contributions. We make appearance controllable in a feed-forward, multi-camera, streaming driving reconstructor, where each place is observed only once, through:

  1. factored color prediction, trained on structurally aligned pairs from our physics-guided generator;
  2. a reusable condition library, now validated on scenes disjoint from its construction set;
  3. a feature history that can be recolored (+2.6 dB over a naive color cache).

Contemporary feed-forward in-the-wild methods trail SpectralSplat by 1.6–4.9 dB on driving scenes (Table R5, in our reply to Reviewer 49n4). This shows that the setting needs this design. Items 2 and 3 are evaluated in Tables R1b and R2 of the same reply.

W2 (only global appearance).

"the method only seems to partially address the global lighting effect and global exposure"

We agree that the method targets global, scene-level appearance, and we will make this scope explicit.

  • Scope: the abstract, introduction and conclusion will refer to global, scene-level appearance: tone, time of day, and the global look of weather.
  • Outside this scope: per-camera exposure, local lighting and physical weather effects. We will state this in the main text and discuss per-region codes as future work.

Within this scope, the edits are recognized as real weather.

  • Classifier: a day/night/rain classifier trained only on real Waymo photos (433 frames, CLIP features). It reaches 97% accuracy on held-out real scenes.
  • Test set: library renders of 325 unseen daytime frames.

Table R6 (share of renders classified as the target condition):

Library codeReal-photo classifierNo editCLIP zero-shotNo edit
rain (day)57.8%0.6%69.5%4.0%
rain (golden hour)72.0%0.6%60.3%4.0%
clear night0.6%0.0%56.0%13.2%
rainy night0.9%0.0%56.6%13.2%
  • Rain: codes make unseen scenes read as real rain, for a classifier that never saw a synthetic image.
  • Night: CLIP reads 56% of the edited scenes as night, against 13% without an edit. The real-photo classifier rarely does, because real Waymo night scenes are dominated by street lamps and headlights: local light sources, which are outside the global scope.

Q (artifacts in Figs. 4, 5, 13).

"Some qualitative results exhibit very noticeable artifacts, especially in Figure 4, 5, 13"

  • Figs. 4 and 13 (dark holes and streaks): these do not come from the appearance pathway. They also appear in the UniSplat render of the same frame.
    • Both figures render poses 0.5 s ahead with the history disabled, so regions uncovered by ego motion have no Gaussians.
    • Those regions were never observed, so no history could fill them.
  • Fig. 5 (library renders at the input views): there are no such holes. The renders are softer than the input, and small bright objects can keep a trace of their source color, because one global code recolors the whole scene.

We hope these results address your concerns. If they do, we would be grateful if you would consider updating your score. We are happy to answer any further questions.

Thank the author for the rebuttal. I still have some concerns over the quality. The comparison results compared to other feed-forward methods (Table R5) compares zero-shot results of baselines that has not been trained on Waymo dataset, making it hard to distinguish if the benefit comes from the proposed method.

Furthermore, the style-transfer metrics seem to be wrt the images generated from the generation pipeline, where the proposed method has been trained on. I suggest for the final version, a more fair comparison on a matched training distribution for baselines is necessary. I also think it is worthwhile to evaluate style-transfer with images generated from a different method than the proposed generator pipeline to showcase generalization.

The library renders in Fig. 5 also concerns me, where vertical streaks artifacts are very visible. I also think if local object can leave such visible artifacts, then it demonstrates a significant downside of using a single global code, which needs to be addressed as it significantly limits downstream usecases.

Reviewer_49n4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SpectralSplat extends UniSplat with appearance control for feed-forward reconstruction of driving scenes. The method extracts a global appearance embedding from backbone image features and conditions the Gaussian color heads on this embedding. Each color head is also evaluated with a zero code to produce a canonical base-color output. Paired original and synthetically relit images provide reconstruction, invariance, appearance-swap, and base-color supervision. The relighting pipeline combines intrinsic estimates from MVInverse, Lambertian rendering, and IC-Light refinement with texture-dependent latent guidance.

During inference, appearance codes can be replaced by condition centroids from a reusable library, allowing the reconstructed geometry to remain fixed during appearance edits. The method also proposes caching historical Gaussian features so their colors can be recomputed under a selected condition.

Critical strengths
  • Controllable appearance in reconstructed driving scenes is an interesting topic for simulation and testing. Reconstructing a scene once and subsequently changing its appearance provides a practical capability beyond replaying its observed colors.

  • Reusing backbone features for appearance encoding is also an interesting implementation choice. Diffusion is used for offline training-data generation and is not required for reconstruction or subsequent library-conditioned rendering.

  • The main reconstruction results are competitive with UniSplat on both datasets.

Critical weaknesses
  • The reusable library is central to the motivation, but Table 3 evaluates a different setting. Its appearance codes are extracted from the paired source and augmented images of the same scene. The library itself is constructed from 240 frame codes from the augmented Waymo validation split. A convincing evaluation should construct the library from a separate scene set. Otherwise, the generalization is not comprehensively evaluated.

  • Recoloring accumulated Gaussians is emphasized in the abstract and contribution list. However, the paper states that the cross-appearance evaluation, additional per-camera comparison, qualitative NVS examples, and supplementary sequence videos operate without temporal history.

  • The hybrid generator is presented as a major contribution and as the source of structurally consistent paired supervision. However, Figure 3 is explicitly a qualitative illustration, and the paper provides no quantitative comparison of the relevant alternatives.

  • Regarding the related work, the discussion should additionally cover closely related contemporary feed-forward methods. WildSplatter combines global appearance embeddings with local Gaussian features and demonstrates fixed-geometry cross-scene appearance control. GenWildSplat uses canonical Gaussian colors and an appearance adapter. WildSplat also separates geometry and reference-conditioned appearance within a feed-forward framework.

  • At least, the paper should compare IC-Light alone, physics-based rendering, the hybrid with uniform guidance, and the proposed texture-dependent guidance under controlled generation settings. The uniform-guidance comparison is especially important for isolating the frequency-aware mechanism from the general benefit of adding a reference image.

Questions for authors

See above comments in the critical weakness section.

3 comments

We thank the reviewer for the detailed and constructive review. We are glad the reviewer found controllable appearance in reconstructed driving scenes useful for simulation and testing, and the reuse of backbone features for appearance encoding interesting.

Your review led to most of our new experiments. In short:

  • W1: a library built from disjoint training scenes matches the submitted one (20.70 vs. 20.65 dB; Tables R1a/R1b).
  • W2: recoloring the history gains +2.6 dB over a naive color cache on the pixels it affects (Table R2).
  • W3/W5: in the controlled comparison you asked for, physics guidance lowers cross-view error by 11% (Table R3).
  • W4: SpectralSplat is ahead of WildSplatter and GenWildSplat by 1.6 and 4.9 dB on driving scenes (Table R5).

W1 (library evaluated on the same scenes).

"A convincing evaluation should construct the library from a separate scene set."

We agree, and we have now done this.

  • Library: rebuilt from 48 training-split scenes only.
  • Evaluation: its fixed centroids on all 202 validation scenes (2,020 frames), with relit targets balanced over the nine conditions.
  • Statistics: CIs are 95%, from a bootstrap over scenes.

Table R1b (reconstruction against the relit target):

Code used at render timePSNR↑SSIM↑LPIPS↓
Per-image code from the target (upper bound)22.590.7240.450
Library as submitted (24 validation scenes)20.650.6950.501
Library from 48 training scenes (disjoint)20.700.6980.500
Library from the other half of validation (disjoint, 2-fold)20.700.6970.498
Nine codes optimized directly for reconstruction (frozen decoder)20.640.6840.470
Mean of all codes (no condition information)19.800.6830.504
Centroid of a wrong condition19.070.6690.508
No edit13.910.4920.488
  • The library generalizes to unseen scenes. Disjoint libraries score the same as the submitted one. The submitted library is also no better on its own 24 scenes (20.52 dB) than on the other 178 (20.66 dB).
  • The centroids carry condition information. They gain +0.90 dB [0.82, 0.99] over the mean code and +1.64 dB [1.51, 1.77] over a wrong condition. The gain is largest at night (+1.8 to +2.5 dB).
  • The centroids are close to the best fixed codes. Optimizing nine codes directly for reconstruction gives the same PSNR.
  • Per-image codes are the upper bound (+1.9 dB), since they encode each target's exact light direction.

Do fixed codes produce the intended condition on unseen scenes? We rendered 404 validation frames under all nine centroids, with geometry fixed. A linear probe trained on renders of other scenes then predicts the condition. Table R1a:

Codes9-way (chance 11%)Time of day (25%)Weather (25%)
Real relit photos (ceiling; probe trained on photos)36.4%66.1%53.0%
Per-image codes from the target21.3%57.9%35.4%
Fixed library codes, unseen scenes25.7%62.3%38.3%
  • Fixed codes are as recognizable as per-image codes.
  • Mean luminance of library renders tracks real relit photos with r = 0.97.
  • The ceiling of this 9-way task is 36% even for real relit photos, because several conditions look alike.

W2 (temporal history).

"the cross-appearance evaluation, additional per-camera comparison, qualitative NVS examples, and supplementary sequence videos operate without temporal history."

We now evaluate the history directly.

  • Setup: we stream three past frames under their own appearance, switch the current frame to a library code, and score the t+1 render against its relit target (202 scenes, 404 sequences).
  • Color gap: the mean RGB difference where cached and new Gaussians overlap.

Table R2:

HistoryPSNR↑SSIM↑LPIPS↓PSNR↑ on history pixelsColor gap↓
None20.440.6790.51819.95—
Naive cache (colors stored at insertion)20.310.6760.52316.5740.6
Feature cache, recolored with the current code (ours)20.420.6790.52019.1417.2
  • Recoloring keeps the history consistent with the current appearance. On history pixels it gains +2.57 dB [2.28, 2.84] over a naive cache, and it makes the color gap 2.4× smaller.
  • Same picture with per-image codes: +3.16 dB, and the gap falls from 41.5 to 19.3. As a sanity check, the two caches are identical without an edit.
  • Reading the "None" row: the comparison that isolates our contribution is naive cache vs. recolored cache. The submitted model was trained without history, so history pixels do not yet exceed a render without it; training with history enabled is the natural next step.
  • Final version: we will add this table, with history-enabled qualitative results and sequence videos.

W3/W5 (generator comparison).

"compare IC-Light alone, physics-based rendering, the hybrid with uniform guidance, and the proposed texture-dependent guidance under controlled generation settings."

Done.

  • Controlled settings: all settings use identical prompts, light parameters, seed and solver, on 600 validation images.
  • Extra setting: uniform guidance at the mask's mean strength. This separates where guidance is applied from how strongly.
  • Metrics:
    • Cross-view: color error between adjacent cameras, relative to real photos.
    • KID: measured against real night and rain frames.

Table R3:

GeneratorCross-view↓Exposure mismatch↓Structure↑CLIP prompt↑Time of day↑KID×10³↓
IC-Light only1.800.2670.34420.650.3500.597
Physics only2.470.4160.645†17.410.683†0.475†
Hybrid, uniform (same peak strength)1.730.2730.33720.260.4170.554
Hybrid, uniform (same mean strength)1.580.2500.33320.420.3920.558
Hybrid, frequency-aware (submitted)1.600.2460.35020.440.4000.564

† The physics render re-shades the source pixel by pixel and gives near-black night images, which inflates these scores. It cannot express weather.

  • Physics guidance improves IC-Light. Paired tests, all significant:
    • about 11% lower cross-view error;
    • better structure;
    • +5 points in time of day.
  • Uniform guidance, as you requested:
    • The frequency-aware mask improves on uniform guidance at the same peak strength (cross-view 1.60 vs. 1.73).
    • It is on par with uniform guidance at the same mean strength; the only significant difference is +0.017 in structure, in favor of the mask.

W4 (contemporary feed-forward methods).

"the discussion should additionally cover closely related contemporary feed-forward methods."

Thank you for the pointers. We will discuss all three in Related Work and Table 1, and we compare against two of them below.

  • WildSplatter (Fujimura et al., arXiv 2026) predicts Gaussians with per-image appearance embeddings.
  • GenWildSplat (Gupta et al., CVPR 2026) predicts canonical Gaussians and modulates their colors with a light encoder and an appearance adapter. It is also trained with synthetic relighting.
  • WildSplat (Zhang et al., ECCV 2026) injects target appearance into content features through a dual-branch design.

We see them as concurrent evidence for feed-forward appearance factorization. They target a few unposed views of static landmarks. SpectralSplat targets posed multi-camera driving streams, where each place is observed once, and its temporal history must be recolored.

Zero-shot comparison (Table R5). WildSplatter and GenWildSplat have public code, so we ran both zero-shot. WildSplat has no public implementation.

  • Data: 40 validation scenes covering all nine conditions. All methods are scored on the same pixels of the front three cameras.
  • Inputs: SpectralSplat uses its five posed cameras. The baselines take unposed views, each in its better setting (WildSplatter one view at a time, GenWildSplat three views jointly).
  • Reference: the relit image of the same view serves as the appearance reference.
  • Zero-shot: the baselines were not trained on driving data or on pairs from our generator, unlike SpectralSplat.
MethodTransfer PSNR↑ / SSIM↑ / LPIPS↓Reconstruction PSNR↑ / SSIM↑ / LPIPS↓
No edit12.65 / 0.455 / 0.470—
WildSplatter21.51 / 0.650 / 0.47421.22 / 0.688 / 0.366
GenWildSplat18.21 / 0.607 / 0.52323.28 / 0.781 / 0.306
SpectralSplat23.13 / 0.744 / 0.44031.46 / 0.894 / 0.124
  • Transfer: SpectralSplat gains +1.6 dB [1.1, 2.1] over WildSplatter and +4.9 dB [4.1, 5.7] over GenWildSplat (paired), and is best in each of the nine conditions.
  • Most of the gap comes from reconstruction: both baselines are designed for a few views of a static scene.
  • What this shows: the driving setting needs a dedicated design. It is not a verdict on their appearance modeling.

We hope these results address your concerns. If they do, we would be grateful if you would consider updating your score. We are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.