acceptodds
Under review as a conference paper at ICLR 2027

CoLight-GS: Coupled Radiometric Supervision for Low-Light Gaussian Splatting

Abstract

Low-light novel-view synthesis aims to reconstruct a well-exposed 3D scene from dark multi-view images. Exposure correction alters both the reconstruction tar- gets and the scales of image residuals, linking brightness recovery to how scene details are fitted. We present CoLight-GS, a Gaussian reconstruction framework that couples bright-domain structural supervision with dark-domain observation fitting through a fixed invertible tone map. The map lifts dark observations into bright targets, while its analytic inverse maps the same scene rendering back to the observed domain, jointly supervising a single Gaussian scene. A shared gain defines a coherent exposure convention across views with matching capture condi- tions, while source-histogram anchoring accommodates heterogeneous exposures. Derivative-based weighting coordinates the residual scales of the bright and ob- served domains. CoLight-GS requires no pretrained depth or restoration models and renders novel views directly with standard Gaussian splatting. On RealX3D and LOM, CoLight-GS outperforms five compared methods in mean PSNR, SSIM, and LPIPS, reaching 18.10 dB and 23.45 dB PSNR, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian splatting, novel-view synthesis, radiometric supervision, inverse rendering, exposure correction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_uB5JOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents CoLight-GS, a 3DGS-based reconstruction framework that couples bright-domain structural supervision with dark-domain observation fitting through a fixed invertible tone map. Experiments on RealX3D and LOM verify the effectiveness of the proposed method.

Critical strengths
  1. This paper proposes to use a invertiable tone mapping function for low-light novel view synthesis that aims to reconstruct a scene with normal exposure with exposed dark observations, which is interesting. Based on the model design, several regulations and supervisions are proposed as well.

  2. Experiments on two datasets verify the effectiveness of the proposed method. Especially, this model does not require any pretrained models, e.g, monocular depth estimation.

Critical weaknesses
  1. Overall, the paper is not well written and hard to follow. For example, the motivation is not well written, and readers may have a whole picuture about why and how the model is designed. Additionally, related works are not fully discussed, where the authors list several methods for each topic without clarifying the differences and connections between the proposed method and the previous methds. From my point of view, section related work is a good place for authors to clarify the motivation and values of the proposed methods while the authors neglects this part instead.

  2. Low-light images are typically constrained by poisson-gaussian photon shot and read noise, limited dynamic range and non-linear CRF (i.e., sRGB domain). In this case, dark shadows in low-light images are heavily quantized and corrupted by sensor noise. The fixed tone mapping function does not take this feature into account and may amplify the nuse and quantization. In addition, the paper does not provide any noise modeling, denoising modules, or dequantization strategy.

  3. Actually, in real-world cameras, CRF for each channel is a bit diiferent. However, the authors use an identical tone mapping function for all the channels.

  4. In RealX3D, setting a = 0.8 with q* = 0.05186 yields g* = 31.03. Why was a = 0.8 chosen? Ground-truth normal-light test images in RealX3D typically have their 90th percentile luminance around 0.8. Selecting a = 0.8 essentially peeks at the target distribution to manually calibrate the global scene exposure.

Questions for authors
  1. Since Tg(d)=1−exp(−gd)T_g(d) = 1 - exp(-gd) is an elementary monotonic curve applied to 8-bit display sRGB images, how does your method avoid fitting and amplifying Poisson-Gaussian noise and quantization banding in dark shadow areas? Did you observe noisy floaters in regions where low-light sensor noise was amplified?

  2. In Tab. 2, Shared Gain + Inverse-Only achieves 18.10 dB PSNR, which is identical to the full model's 18.10 dB PSNR. Given that the canonical branch contributes 0.00 dB to PSNR, what evidence is there that dual-domain coupling provides a meaningful radiometric constraint rather than simply adding a second SSIM term?

  3. How were the threshold value of 32 and the target mean of 0.49 determined for LOM? Were these parameters tuned directly on the 5 evaluation scenes? How would a user set these parameters on an uncurated, unseen capture?

  4. Why was λtail\lambda_{tail} set to 0 on LOM? Does enabling LtailL_{tail} on LOM reduce NVS metrics? Please provide the quantitative results of the proposed method on LOM with λtail\lambda_{tail} with different values.

  5. Since T_g applies the exact same curve across all RGB channels independently, how does CoLight-GS handle scenes where low illumination causes severe chromatic shifts or white balance errors?

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

The paper reconstructs a bright 3DGS scene from posed low-light images. A fixed exponential curve, T_g(d) = 1 − exp(−g·d), brightens the training images, and its closed-form inverse maps the rendering back to the dark domain. The loss compares the rendering with both the brightened image and the original dark image, and the weight between the two terms comes from the derivative of the inverse curve. On RealX3D, one gain g is shared by all scenes and an opacity–footprint penalty is added; on LOM, a histogram rule picks the gain for each scene. No pretrained enhancement or depth network is used. The method reports 18.10 dB PSNR on RealX3D and 23.45 dB on LOM, the best among the six methods compared in Table 1.

Critical strengths
  1. The method is simple and needs no pretrained enhancement or depth network. On RealX3D it improves PSNR from 15.57 dB (LITA-GS, the best listed baseline) to 18.10 dB and LPIPS from 0.4880 to 0.3709. On LOM it also has the best numbers among the compared methods, although these baseline numbers are taken from published papers rather than rerun.

  2. The paper includes several controlled comparisons. Table 2 compares shared and per-view gains with the loss fixed, and Table 3 fixes the exposure map on LOM to test the coupled loss. Appendix B is also careful to claim only a local scale argument for the derivative weight, not full gradient balancing.

Critical weaknesses
  1. The paper's own ablation does not support its main claims. In Table 2, inverse-only fitting with the shared gain already reaches 18.10 dB. This is essentially standard 3DGS with the inverse curve applied to the render, and it beats every baseline in Table 1 by at least 2.5 dB. Adding the three claimed contributions (canonical branch, derivative weighting, tail penalty) leaves PSNR unchanged at 18.10 dB and changes SSIM/LPIPS by +0.037/−0.018. On LOM (Table 3) the gain is 0.18 dB. SSIM improves in every scene, so the direction is consistent, but with 14 scenes, 73 test views and a single seed, a 0.18 dB PSNR gain is within normal seed-to-seed variation and the size of the SSIM gain is uncertain. So most of the PSNR gain over the baselines comes from the tone curve itself, not from the coupling. The ingredients themselves, a fixed exponential curve, its inverse and a loss weight, are standard, so the contribution rests on the coupling. The obvious control is missing: canonical-only fitting, i.e. tone-map the images and train 3DGS on them. Within the range of the map the two branches have the same exact optimum, so this control is needed to show that the inverse branch helps at all. An enhance-then-reconstruct baseline with an off-the-shelf low-light enhancer is also missing; prior work on this task usually reports one. Finally, the weighting ablation is mixed: with per-view gains, derivative weighting at τ = 1 improves SSIM over equal weighting but is worse on PSNR and LPIPS (17.47/0.5788/0.3959 vs. 17.56/0.5407/0.3834). Only τ = 0.75 improves all three. The paper claims only structural gains for the coupling (Section 4.3), and SSIM does improve in every scene. But the headline 2.5 dB PSNR lead over all baselines comes entirely from the inverse-only variant, and without the canonical-only control the structural gain cannot be attributed to the coupling rather than to training on brightened images. Even if that control favored the coupling, the contribution would be a 0.037 SSIM gain from combining standard components, which I do not consider sufficient for ICLR.

  2. The calibration constants are not justified. RealX3D uses anchor 0.8, τ = 0.75 and the tail penalty. LOM uses a mean target of 0.49, a switching threshold of 32, τ = 1 and no tail penalty. These constants set the overall brightness of the output, and PSNR against normal-light ground truth depends heavily on brightness. The paper does not say how they were chosen. In Table 3, the proposed switching rule also loses to a plain fixed-mean anchor on PSNR (23.45 vs. 23.94 dB), although it is better on SSIM and LPIPS.

  3. Noise is not discussed. With g ≈ 31, brightening also amplifies sensor noise by up to about 31 times. Fitting in the dark domain down-weights this noise, which is probably the best reason to keep the inverse branch. The paper's argument in Eq. 5 and Appendix B is only about residual scale, and there is no comparison with denoising.

  4. The paper is hard to read. Several terms are used without a clear definition, such as "carriers" for Gaussian primitives and the "lift" into a "canonical coordinate". Figure 2 has too many small labels and equations. Figures 4–6 look like unedited script output and add little beyond the tables: the per-scene dot plots in Figures 4 and 5 are cramped, show no values and make the mean markers hard to find, and Figure 6(f) repeats the tail and top-1% bars already shown in Figure 5(d). A compact per-scene table would be clearer.

Questions for authors
  1. How does canonical-only fitting (tone-mapped images, plain 3DGS) compare with inverse-only and coupled fitting, using the same initialization, optimizer, primitive budget and regularization?

  2. Please add an enhance-then-3DGS baseline with at least one off-the-shelf low-light enhancer, under the same training budget.

  3. In the final shared-gain RealX3D setting, can you change one component at a time (equal vs. derivative weighting, with and without the tail penalty)? Results over a few random seeds would help, since some differences are small.

  4. How were 0.8, 0.49, 32, the τ values and the dataset-specific use of the tail penalty chosen? Were the normal-light test images used in any of these choices? I am not suggesting they were; I would just like the procedure stated.

  5. LL-Gaussian (Sun et al., 2025) is cited in Section 2 but not compared. Its code, preprocessed LOM data and checkpoints are public. Its paper uses a different LOM protocol (it reports 23.85 dB for Aleth-NeRF, vs. 19.87 dB in your Table 1), so could you run it under your protocol? Also, as concurrent work that does not affect my score: arXiv 2606.11841 (June 2026, code public) replaces only the pseudo-target in LITA-GS with a tone curve of the same 1 − exp(−g·t) form and reports 17.82 dB on the same nine RealX3D scenes at 15K iterations. This is close to a canonical-only version of your method, so a short discussion of the difference would be useful.

  6. The evaluation covers 14 scenes and 73 test views with a single seed. Could you add the LLNeRF dataset (at least qualitatively if it lacks paired references) or a synthetic varying-exposure set such as the one used by Luminance-GS, where the darkness level can be controlled?

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

This paper proposes CoLight-GS, a framework for low-light novel view synthesis with 3D Gaussian Splatting. CoLight-GS first use a fixed invertible tone map to lift dark observations into bright targets, and its analytic inverse maps the same rendering back to the dark domain. The 3DGS optimization is supervised by both bright and dark domain targets. The method further introduces a shared exposure gain, a derivative-based weight between the two losses, and an opacity-footprint penalty for broad, opaque primitives.

Critical strengths

The method is simple and lightweight. It needs no pretrained depth or restoration models and adds no inference cost.

Critical weaknesses
  1. The novelty of the proposed method is limited. The core design of Colight-GS is the supervision from a curve-mapped pseudo bright-domain target and its dark-domain observation. This design is already used in Luminance-GS and LITA-GS, yet the paper does not acknowledge this shared structure.
  2. The performance gain appears to come mainly from the choice of exposure anchor rather than from the coupled supervision. In Table 2, under the shared gain, inverse-only fitting (dark-domain supervision only) and the full model (bright- and dark-domain supervision) reach the same PSNR (18.10 dB). In contrast, in Table 3, changing the anchor alone can result in a PSNR drop of more than 2 dB. PSNR is highly sensitive to global intensity, and Figure 3 shows that Luminance-GS, LITA-GS and I2-NeRF also recover the structure reasonably well, with the main difference being the overall brightness. I therefore suspect that the comparison in Table 1 largely reflects exposure calibration rather than reconstruction quality.
  3. The hyperparameters require dataset-specific tuning and components are used inconsistently. According to section 4.1, for RealX3D and LOM, hyperparameter selection has different protocol. The opacity-footprint penalty, listed as a contribution, is not used on LOM at all and gives only marginal, mixed changes on RealX3D.
Questions for authors

See Weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.