Pith. sign in

REVIEW 3 major objections 6 minor 47 references

Robust Low-light Scene Restoration via Illumination Transition

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read RoSe renders low-light 3D scenes in normal light by learning a world-centered illumination-transition field, reporting 26.28 dB mean PSNR on the LOM benchmark.

desk verdict Promising low-light NeRF method with strong empirical results, but the central product-form rendering assumption is unproven and needs a rigorous test. read the letter →

arxiv 2507.03976 v2 pith:WSWP4QXW submitted 2025-07-05 cs.CV

Ze Li , Feng Zhang , Xiatian Zhu , Meng Zhang , Yanghong Zhou , P. Y. Mok This is my paper · ORCID
classification cs.CV
keywords low-lightscenerestorationnovelviewsynthesisneuralradiancefieldsilluminancetransitionfieldRetinextheorylow-rankdenoisingmultiviewconsistencyunsupervisedenhancement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes RoSe, a framework that synthesizes normal-light novel views from low-light multiview images without paired supervision. It treats the task as estimating an illuminance transition field in 3D space: a scalar field, anchored in the world rather than per view, that maps low-light observations to normal-light ones. The transition field is constrained to be low-rank, which suppresses sensor noise without explicit noise modeling or 2D denoising. On the LOM benchmark, RoSe reports a mean PSNR of 26.28 dB, outperforming the prior end-to-end methods LLNeRF (23.13 dB) and Aleth-NeRF (19.87 dB).

What carries the argument

The central object is the world-centered illuminance transition field $i(\mathbf{x})$, a scalar field defined at each 3D location that depends only on position, not viewing direction, and is volume-rendered along the camera ray with the same density $\sigma$ as the normal-light branch to give $\hat{I}(\mathbf{r})$. The load-bearing identity is the pixel-level Retinex-inherited product $\hat{C}_{\mathrm{low}} = \hat{C}_{\mathrm{nor}} \odot \hat{I}$, extended into the radiance-field setting. The low-rank denoising module (LRD) implements the low-rank assumption through reduction to a $k$-dimensional subspace, similarity-weighted reconstruction against learnable low-rank filters, and reweighting of the original feature. The illumination correction loss with hyperparameter $e$ regulates the global brightness of the rendered normal-light image.

What would settle it

Take a scene with known normal-light radiance field and density, define a smooth illumination transition field, render low-light views by compositing the pointwise product at sample points, then run RoSe and compare the separately rendered normal-light branch to the known ground truth; if the separately rendered $\hat{C}_{\mathrm{nor}}$ deviates from ground truth while the composite matches, the multiplicative identity is a loss artifact rather than a faithful decomposition.

Watch

Extended reading notes

Core claim

RoSe's central claim is that low-light restoration in 3D can be reduced to estimating a single multiview-consistent illuminance transition $I$ such that the observed low-light radiance equals the normal-light radiance multiplied elementwise by $I$. Building on Retinex theory, the authors treat both the normal-light scene and the transition as neural fields that share the same volume density $\sigma$, render each along a ray by standard NeRF integration, and reconstruct the low-light pixel as the product $\hat{C}_{\mathrm{low}} = \hat{C}_{\mathrm{nor}} \odot \hat{I}$. Because illumination is smooth and thus low-rank while noise is high-rank and random, a low-rank denoising module in the transition branch suppresses noise. The network is trained with only low-light images, using an MSE loss with an inverse tone curve and an unsupervised illumination correction loss that steers the normal-light branch to a target brightness level. The paper reports that this yields state-of-the-art rendering quality and multiview consistency on the LOM dataset.

Load-bearing premise

The whole approach hinges on the assumption that a low-light pixel equals the normal-light pixel multiplied, channel by channel, by a single transition value, and that this multiplication survives when both sides are computed by volume rendering along the same ray.

Editorial extensions

If this is right

  • Low-light novel view synthesis can be solved without paired normal-light supervision, using only low-light images and a target brightness hyperparameter.
  • Multiview consistency is obtained by making the transition field world-centered, avoiding the per-view enhancement conflicts seen in prior methods.
  • Sensor noise can be suppressed by low-rank regularization of illumination, without explicit noise models or 2D denoising networks.
  • On the LOM benchmark, the reported gains over LLNeRF and Aleth-NeRF are roughly 3.15 dB and 6.41 dB in mean PSNR, respectively.
  • The ablation suggests that the low-rank denoising module sharpens the estimated density distribution along rays, indicating cleaner geometry estimation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the multiplicative identity holds, the same world-centered transition field could be reused for other illumination edits, such as relighting or exposure correction, by swapping the target illumination level $e$.
  • A direct test of the decomposition's physicality would be to train RoSe on scenes with paired normal-light ground truth and compare $\hat{C}_{\mathrm{nor}}$ directly against the ground-truth normal-light radiance field; the paper evaluates only the composite reconstruction.
  • The low-rank assumption suggests a measurable signature: the recovered transition field should have rapidly decaying singular values; measuring that spectrum on held-out scenes would test whether the denoising is truly low-rank or just a learned filter.
  • Because the transition field is world-centered, it may transfer across viewpoints more easily than per-image enhancement; a testable extension is to train on sparse views and see whether transition predictions generalize to unseen viewpoints better than 2D enhancement baselines.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes RoSe, an unsupervised neural radiance field framework for synthesizing normal-light novel views from low-light multiview images. The key idea is to model an illuminance transition field in 3D space and render it jointly with a normal-light radiance field, with the observed low-light pixel expressed as the elementwise product of the two separately rendered quantities (Eq. 3). A low-rank denoising module is introduced in the illuminance branch, and training is driven by an MSE loss with an inverse tone curve plus an unsupervised illumination correction loss. Experiments on the LOM dataset report PSNR 26.28 mean, substantially above prior LLNeRF (23.13) and Aleth-NeRF (19.87), with ablations on the low-rank module and illumination level.

Significance. The problem of low-light novel view synthesis is practically important, and the reported gains are sizable and consistent across most scenes. The dual-branch architecture and the low-rank denoising idea are interesting and are supported by ablations. The release of code and data, and the comparison with a broad set of baselines, are strengths. However, the central modeling assumption in Eq. (3) — that a pixel is the product of separately volume-rendered color and illuminance transition — is not derived or empirically validated, and the evaluation is limited to a single dataset with five scenes. If the product-form relationship is justified, the method is a strong contribution; as it stands, the physical interpretation of the decomposition remains unsubstantiated, and the reported gains could stem from model flexibility rather than a correct Retinex-based model.

major comments (3)
  1. [Section 3.2, Eqs. (3), (6), (8)] The paper defines the observed low-light pixel as the elementwise product of the separately volume-rendered normal-light color and illuminance transition, C_low = C_nor * I. If the pointwise low-light radiance is c_nor(x) * i(x) with shared density sigma(x), the physically correct volume-rendered pixel is the weighted sum of pointwise products, C_low = sum_n w_n (c_nor,n * i_n). The product of the two weighted sums equals this only when i is constant over the support of the rendering weights along each ray. Low-rankness of illumination gives spatial smoothness, not constancy along camera rays, and the paper provides no argument or measurement showing the approximation error is small. The authors should either derive a rigorous condition under which Eq. (3) holds or add an experiment that compares the product of the two rendered integrals with the ray-integrated pointwise product on the trained model. Without this, the interpretation of C_nor as a normal-light radiance field is not established, and the reported improvements may be an artifact of the factorization's flexibility.
  2. [Section 4, Dataset] All quantitative and qualitative results are obtained on a single benchmark, the LOM dataset, which contains only five scenes. Given that the central claim is a physically motivated rendering model for low-light scenes, the evaluation should be broadened to at least one additional dataset or a synthetic environment with controlled ground-truth illumination and geometry. Such an evaluation would test whether the illuminance-transition assumption generalizes and whether the claimed multiview consistency holds outside this specific benchmark. The current single-dataset evidence is insufficient to support the strength of the stated conclusions.
  3. [Section 3.3, Eq. (13)] The loss as written applies the inverse tone curve only to the target term: LMSE = sum_r || C_low(r) - phi(C_low(r)+epsilon) ||^2, where the first C_low(r) is the predicted rendered image and the second is the ground-truth observed image. This asymmetric formulation trains the network to predict a tone-mapped target, which would systematically bias the rendered normal-light output C_nor through Eq. (3). If the intent is to apply the tone curve to both prediction and target for weight rebalancing, the equation should be written accordingly (e.g., || phi(C_low(r)) - phi(C_low_gt(r)) ||). Please clarify the correct form and state whether any inverse transformation is applied to the rendered images before evaluation.
minor comments (6)
  1. [General] There are several typos: 'Exsiting' in Section 1, 'not but' in Section 4.1, and 'RetinexLight' in Eq. (14) should be 'RoSe'. Also, in Table 1 the mean SSIM for Aleth-NeRF is misprinted as '0 754'.
  2. [Section 3.2, Eq. (8)] The illuminance transition is rendered with the same density as the normal-light scene, but the paper does not discuss whether shared density is appropriate when the low-light scene may contain atmospheric effects or view-dependent attenuation. A brief comment on this assumption would be helpful.
  3. [Section 3.2, Low-rank denoising] The dimensions k and M of the low-rank subspace and the embedding are not specified anywhere in the paper, including the ablations. Please provide these values.
  4. [Section 3.3, Eq. (14)] The illumination correction loss weight lambda is fixed at 1e-3 without a sensitivity analysis. While Fig. 4 shows robustness to the illumination level e, the effect of lambda is not investigated.
  5. [Section 4, Tables 1-3] No error bars or multiple random seeds are reported. This is a minor concern given the large mean gains, but for per-scene cases where the gap is small (e.g., 'shrub' in Table 1), knowing the variance would help assess significance.
  6. [Section 5] The conclusion states that RoSe relies on the Lambertian assumption, but the method section includes a view-dependent color branch. The authors should clarify where the Lambertian assumption enters, or remove this statement if it is inaccurate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Eq. (3) is a modeling assumption validated on held-out views, not a fitted value presented as prediction.

full rationale

RoSe's derivation chain is not circular. The pixel-space identity Zlow = Znor * I in Eq. (2) follows algebraically from the Retinex assumption of equal reflectance under low and normal light, and Eq. (3) extends that identity to volume-rendered quantities as a deliberate modeling ansatz, not as a consequence of the rendering equations. The paper does not claim Eq. (3) is derived from volume rendering, and its validity is an empirical/physical question about whether the two separately rendered integrals can be multiplied; that is a correctness or robustness concern, not a logical reduction of the method's output to its input. Training uses only low-light observations via the reconstruction loss in Eq. (13), together with a manually chosen global illumination target e = 0.45 in Eq. (12), and evaluation is performed on held-out test views against paired normal-light ground truth in Table 1. No fitted parameter is relabeled as a prediction, and no load-bearing premise is justified by self-citation: the cited Retinex theory [23] is an external, standard result, and the low-rank denoising ablations are benchmarked on held-out views. The closest concern, that Eq. (3) may fail when the illuminance transition varies along a ray, would affect physical interpretability and generalization, but it does not make the empirical PSNR comparison circular, because the reported gains are externally falsifiable rather than guaranteed by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim depends on free parameters e and lambda, the Lambertian and reflectance-invariance assumptions from Retinex, the low-rank property of illumination, and the invented 3D illuminance transition field. No external benchmarks are used to set these; they are chosen or trained on the low-light data.

free parameters (3)
  • e (desired illumination level) = 0.45
    Used in the illumination correction loss (Eq. 12) to set the average brightness of the normal-light rendering. The paper states this value gives balanced performance.
  • lambda (loss weight) = 1e-3
    Weight of the illumination correction loss in Eq. (14).
  • low-rank dimensions k and M = not reported
    Dimensionalities of the low-rank subspace and filter bank in the LRD module (Eqs. 9-11); not specified in the paper, making exact reproduction impossible.
assumptions (4)
  • domain assumption Reflectance is invariant between low-light and normal-light captures (R_nor = R_low)
    Core Retinex assumption (Section 3.1) used to derive the illuminance transition relation. Fails for view-dependent specularities, which the authors acknowledge in the Conclusion as a remaining Lambertian limitation.
  • ad hoc to paper Illuminance transition is world-centered (view-independent), i(x)
    The transition MLP in Eq. (7) takes only 3D location, not view direction, to ensure multiview consistency. This is a modeling choice with no physical derivation.
  • domain assumption Illumination is low-rank while noise is high-rank
    Motivates the low-rank denoising module in Section 3.2. The paper provides no quantitative evidence for the rank separation, though ablations show the module helps.
  • ad hoc to paper Product of rendered C_nor and I equals observed low-light pixel
    Eq. (3) applies the pixel-level Retinex identity to volume-rendered integrals without proof. This is the load-bearing modeling assumption flagged as the weakest assumption.
invented entities (1)
  • Illuminance transition field I(x)
    purpose: A 3D scalar field that maps low-light pixel intensities to normal-light intensities, used to reconstruct C_low = C_nor * I.
    Introduced as a latent neural field; it is not directly measured or falsifiable outside the model. Its only support is the improved rendering quality on the LOM benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Low-light Scene Restoration via Illumination Transition." pith.science (2026). https://pith.science/paper/WSWP4QXW

@misc{pith2026250703976,
  author       = {Pith},
  title        = {Pith review of: Robust Low-light Scene Restoration via Illumination Transition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WSWP4QXW}},
  note         = {Machine review of arXiv:2507.03976}
}
read the original abstract

Synthesizing normal-light novel views from low-light multiview images is an important yet challenging task, given the low visibility and high ISO noise present in the input images. Existing low-light enhancement methods often struggle to effectively preprocess such low-light inputs, as they fail to consider correlations among multiple views. Although other state-of-the-art methods have introduced illumination-related components offering alternative solutions to the problem, they often result in drawbacks such as color distortions and artifacts, and they provide limited denoising effectiveness. In this paper, we propose a novel Robust Low-light Scene Restoration framework (RoSe), which enables effective synthesis of novel views in normal lighting conditions from low-light multiview image inputs, by formulating the task as an illuminance transition estimation problem in 3D space, conceptualizing it as a specialized rendering task. This multiview-consistent illuminance transition field establishes a robust connection between low-light and normal-light conditions. By further exploiting the inherent low-rank property of illumination to constrain the transition representation, we achieve more effective denoising without complex 2D techniques or explicit noise modeling. To implement RoSe, we design a concise dual-branch architecture and introduce a low-rank denoising module. Experiments demonstrate that RoSe significantly outperforms state-of-the-art models in both rendering quality and multiview consistency on standard benchmarks. The codes and data are available at https://pegasus2004.github.io/RoSe.

Figures

Figures reproduced from arXiv: 2507.03976 by the authors.

Figure 1
Figure 1. Image/video enhancement (Zero-DCE [8]+NeRF, RUAS [15] +NeRF, LLVE [43] +NeRF) vs. state-of-the-art models (LL￾NeRF [34], Aleth-Nerf [5]) vs. our RoSe. data acquisition is often not feasible. Recent advancements in techniques such as Neural Radiance Fields (NeRFs) [19] have demonstrated impressive capabilities in 3D scene rep￾resentation; however, they struggle in low-light conditions because of the reliance on direc… view at source ↗
Figure 2
Figure 2. An overview of the proposed RoSe framework. It learns both normal-light scene representation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Normal-light novel view synthesis comparison in low-light conditions. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Ablation study on the effect of illumination level [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Density distribution of sampling points along the camera [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 37 canonical work pages

  1. [1]

    Unprocessing images for learned raw denoising

    Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 11036–11045, 2019. 5

    work page 2019

  2. [2]

    Retinexformer: One-stage retinex- based transformer for low-light image enhancement

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Tim- ofte, and Yulun Zhang. Retinexformer: One-stage retinex- based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504–12513, 2023. 3, 6, 7

    work page 2023

  3. [3]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InEuropean con- ference on computer vision, pages 333–350. Springer, 2022. 8

    work page 2022

  4. [4]

    You only need 90k parameters to adapt light: a light weight trans- former for image enhancement and exposure correction

    Ziteng Cui, Kunchang Li, Lin Gu, Shenghan Su, Peng Gao, Zhengkai Jiang, Yu Qiao, and Tatsuya Harada. You only need 90k parameters to adapt light: a light weight trans- former for image enhancement and exposure correction. arXiv preprint arXiv:2205.14871, 2022. 1, 2, 6, 7

    arXiv 2022

  5. [5]

    Aleth-nerf: Illumination adaptive nerf with concealing field assumption

    Ziteng Cui, Lin Gu, Xiao Sun, Xianzheng Ma, Yu Qiao, and Tatsuya Harada. Aleth-nerf: Illumination adaptive nerf with concealing field assumption. In Proceedings of the AAAI Conference on Artificial Intelligence , pages 1435– 1444, 2024. 1, 2, 5, 6, 7

    work page 2024

  6. [6]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2

    2022

  7. [7]

    Fastnerf: High-fidelity neu- ral rendering at 200fps

    Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neu- ral rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision , pages 14346– 14355, 2021. 2

    2021

  8. [8]

    Zero-reference deep curve estimation for low-light image enhancement

    Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1780–1789, 2020. 1, 2, 6, 7

    work page 2020

Show all 47 references
  1. [9]

    Lime: Low-light im- age enhancement via illumination map estimation

    Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light im- age enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2):982–993, 2016. 6

    2016

  2. [10]

    Putting nerf on a diet: Semantically consistent few-shot view synthesis

    Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885–5894, 2021. 2

    2021

  3. [11]

    Enlightengan: Deep light enhancement without paired supervision

    Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing, 30:2340–2349, 2021. 2

    2021

  4. [12]

    Up-nerf: Unconstrained pose prior-free neural radiance field

    Injae Kim, Minhyuk Choi, and Hyunwoo J Kim. Up-nerf: Unconstrained pose prior-free neural radiance field. Ad- vances in Neural Information Processing Systems, 36, 2024. 2

    2024

  5. [13]

    A variational framework for retinex

    Ron Kimmel, Michael Elad, Doron Shaked, Renato Keshet, and Irwin Sobel. A variational framework for retinex. In- ternational Journal of computer vision , 52:7–23, 2003. 7, 8

    2003

  6. [14]

    Learning to enhance low-light image via zero-reference deep curve esti- mation

    Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve esti- mation. IEEE transactions on pattern analysis and machine intelligence, 44(8):4225–4238, 2021. 2

    2021

  7. [15]

    Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement

    Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongx- uan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10561–10570, 202...

    2021

  8. [16]

    Mbllen: Low-light image/video enhancement using cnns

    Feifan Lv, Feng Lu, Jianhua Wu, and Chongsoon Lim. Mbllen: Low-light image/video enhancement using cnns. In Bmvc, page 4. Northumbria University, 2018. 3, 6

    2018

  9. [17]

    Toward fast, flexible, and robust low-light image enhancement

    Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongx- uan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 5637– 5646, 2022. 1, 2, 3, 6, 7

    2022

  10. [18]

    Nerf in the wild: Neural radiance fields for uncon- strained photo collections

    Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duck- worth. Nerf in the wild: Neural radiance fields for uncon- strained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogn...

    2021

  11. [19]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM, 65(1):99–106, 2021. 1, 2, 5, 6

    2021

  12. [20]

    Nerf in the dark: High dynamic range view synthesis from noisy raw images

    Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 16190–1619...

    2022

  13. [21]

    Instant neural graphics primitives with a mul- tiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1–15, 2022. 8

    2022

  14. [22]

    Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs

    Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition,...

    2022

  15. [23]

    Retinex processing for automatic image enhancement

    Zia-ur Rahman, Daniel J Jobson, and Glenn A Woodell. Retinex processing for automatic image enhancement. Jour- nal of Electronic imaging, 13(1):100–110, 2004. 2, 3

    2004

  16. [24]

    Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps

    Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision , pages 14335– 14345, 2021. 2

    2021

  17. [25]

    Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild

    Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8931–8940, 2024. 2

    2024

  18. [26]

    Nerf for outdoor scene relighting

    Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Com- puter Vision, pages 615–631. Springer, 2022. 2

    2022

  19. [27]

    Robustnerf: Ig- noring distractors with robust losses

    Sara Sabour, Suhani V ora, Daniel Duckworth, Ivan Krasin, David J Fleet, and Andrea Tagliasacchi. Robustnerf: Ig- noring distractors with robust losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20626–20636, 2023. 2

    2023

  20. [28]

    Nighttime visibility en- hancement by increasing the dynamic range and suppression of light effects

    Aashish Sharma and Robby T Tan. Nighttime visibility en- hancement by increasing the dynamic range and suppression of light effects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11977– 11986, 2021. 2

    2021

  21. [29]

    The lambertian assumption and landsat data

    JA Smith, Tzeu Lie Lin, KJ Ranson, et al. The lambertian assumption and landsat data. Photogrammetric Engineering and Remote Sensing, 46(9):1183–1189, 1980. 8

    1980

  22. [30]

    Vip-nerf: Visibility prior for sparse input neural radiance fields

    Nagabhushan Somraj and Rajiv Soundararajan. Vip-nerf: Visibility prior for sparse input neural radiance fields. In ACM SIGGRAPH 2023 Conference Proceedings , pages 1– 11, 2023. 2

    2023

  23. [31]

    Neural underwater scene representation

    Yunkai Tang, Chengxuan Zhu, Renjie Wan, Chao Xu, and Boxin Shi. Neural underwater scene representation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11780–11789, 2024. 2

    2024

  24. [32]

    Relight my nerf: A dataset for novel view synthesis and relighting of real world objects

    Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

    2023

  25. [33]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5

    2017

  26. [34]

    Lighting up nerf via unsupervised decomposition and en- hancement

    Haoyuan Wang, Xiaogang Xu, Ke Xu, and Rynson WH Lau. Lighting up nerf via unsupervised decomposition and en- hancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12632–12641, 2023. 1, 2, 6, 7

    2023

  27. [35]

    Deep retinex decomposition for low-light enhancement

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. In British Machine Vision Conference 2018, BMVC 2018, New- castle, UK, September 3-6, 2018 , page 155. BMV A Press,

    2018

  28. [36]

    Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement

    Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wen- han Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 5901–5910, 2022. 2, 3

    2022

  29. [37]

    Snr-aware low-light image enhancement

    Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17714–17724, 2022. 2

    2022

  30. [38]

    Renerf: Re- lightable neural radiance fields with nearfield lighting

    Yingyan Xu, Gaspard Zoss, Prashanth Chandran, Markus Gross, Derek Bradley, and Paulo Gotardo. Renerf: Re- lightable neural radiance fields with nearfield lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22581–22591, 2023. 2

    2023

  31. [39]

    Implicit neural representation for coopera- tive low-light image enhancement

    Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for coopera- tive low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision , pages 12918–12927, 2023. 2

    2023

  32. [40]

    Sparse gradient regularized deep retinex network for robust low-light image enhancement

    Wenhan Yang, Wenjing Wang, Haofeng Huang, Shiqi Wang, and Jiaying Liu. Sparse gradient regularized deep retinex network for robust low-light image enhancement. IEEE Transactions on Image Processing, 30:2072–2086, 2021. 2

    2021

  33. [41]

    Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model

    Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 12302–12311, 2023. 3

    2023

  34. [42]

    pixelnerf: Neural radiance fields from one or few images

    Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4578–4587, 2021. 2

    2021

  35. [43]

    Learning tem- poral consistency for low light video enhancement from sin- gle images

    Fan Zhang, Yu Li, Shaodi You, and Ying Fu. Learning tem- poral consistency for low light video enhancement from sin- gle images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4967–4976,

  36. [44]

    Dual illu- mination estimation for robust exposure correction

    Qing Zhang, Yongwei Nie, and Wei-Shi Zheng. Dual illu- mination estimation for robust exposure correction. In Com- puter graphics forum, pages 243–252. Wiley Online Library,

  37. [45]

    Beyond nerf underwater: Learning neural reflectance fields for true color correction of marine imagery

    Tianyi Zhang and Matthew Johnson-Roberson. Beyond nerf underwater: Learning neural reflectance fields for true color correction of marine imagery. IEEE Robotics and Automa- tion Letters, 2023. 2

    2023

  38. [46]

    Ner- factor: Neural factorization of shape and reflectance under an unknown illumination

    Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul De- bevec, William T Freeman, and Jonathan T Barron. Ner- factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1–18, 2021. 2

    2021

  39. [47]

    Zero-shot restoration of underexposed im- ages via robust retinex decomposition

    Anqi Zhu, Lin Zhang, Ying Shen, Yong Ma, Shengjie Zhao, and Yicong Zhou. Zero-shot restoration of underexposed im- ages via robust retinex decomposition. In 2020 IEEE Inter- national Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2020. 2

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.