acceptodds
Under review as a conference paper at ICLR 2027

R2RFlow: Reflectance-to-Restoration Flow with Complementary Evidence for Low Light Image Enhancement

Abstract

Low-light image enhancement aims to recover visually pleasing normal-light images while preserving the scene content concealed by complex illumination degradation. However, existing generative approaches often start from uninformative noise without directly using the scene content preserved in the low-light observation to initialize the transport. Existing flow objectives primarily supervise the underlying flow field, without directly exploiting the known restoration endpoint available in paired restoration. To address these limitations, we propose R2RFlow, a reflectance-to-restoration flow framework integrating complementary evidence for one-step low-light image enhancement. R2RFlow initializes the transport from an observation-derived reflectance state and constructs finite-interval transport toward the paired normal-light image, explicitly leveraging scene content available in the degraded observation. To complement restoration-relevant information not fully retained by the reflectance source, we further incorporate structural, illumination, and frequency cues through role-specific conditioning pathways. In addition, a restoration-aligned multi-path learning strategy supervises the shared transport operator from the deployment source, reference intermediate states, and model-induced rollout states through their final restoration outcomes. Experiments on the LOL-v1 and LOL-v2-Real benchmarks show that R2RFlow achieves competitive performance against existing low-light enhancement methods. When trained exclusively on LOL-v1, R2RFlow also achieves favorable performance on LSRW-Nikon and LSRW-Huawei without target-domain fine-tuning, demonstrating favorable cross-dataset transfer.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-light image enhancement, Image restoration, Flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_aS8sOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies paired low-light image enhancement (LLIE), aiming to recover normally exposed images from low-light observations while preserving scene structure and details. The authors propose R2RFlow. A Retinex decomposition produces a reflectance component R and an illumination component I; R is used as the transport source, and a MeanFlow-style finite-interval operator performs one-step restoration. The model further introduces structural priors, illumination modulation, and wavelet detail fusion. Training supervises deployment, reference-intermediate, and predicted-rollout paths using restoration losses. Experiments are conducted on LOL-v1 and LOL-v2-Real. A model trained on LOL-v1 is also transferred to LSRW-Nikon and LSRW-Huawei. The paper reports main benchmark results, cross-dataset results, and ablations of source initialization, evidence conditioning, and training objectives.

Critical strengths

The problem formulation is task-aware. The paper separates the transport source from the restoration evidence: reflectance is used for initialization, while illumination, structural, and frequency information complement the restoration. This is more interpretable than simply concatenating all inputs, and Figure 2, Figure 3, Table 3, and Table 4 provide initial evidence for the design.

The experiments cover the main design components. Table 3 compares different source states, Table 4 incrementally adds conditioning modules, and Table 5 compares training objectives. The cross-dataset experiments also provide evidence beyond a single benchmark.

Critical weaknesses

The methodological contribution needs stronger justification. Section 3.1 adopts MeanFlow’s finite-interval parameterization, and Eq. (4) is a deterministic residual predictor at deployment. Table 5 provides a useful deployment-only comparison, but both auxiliary paths are added together, improving PSNR from 25.10 to 25.26 dB. This leaves their individual contributions and the benefit of learning interval transport unclear. The related-work discussion also provides little comparison with other approaches to one-step restoration.

Important experimental details are missing. The paper does not identify the Retinex decomposition checkpoint or specify the pretraining data and frozen status of the preprocessing networks. The MeanFlow baseline objective and time-sampling procedure are not fully documented. Although Section 4.1 specifies No GT-mean evaluation, the provenance of each baseline result remains unclear. Figures 2 and 3 omit the data split used for their statistics, and local structural correlation is not defined. The unresolved LOL-v1 citation and missing PSNR header in Table 3 also need correction.

Questions for authors

What distinguishes the proposed interval transport learning from restoration-supervised residual prediction? Please specify the MeanFlow baseline objective and time sampling in Table 5, and clarify the individual effects of the reference-intermediate and rollout losses.

Which checkpoints and pretraining datasets are used for the Retinex decomposition and SMG networks? Are these networks frozen, and do their training data overlap with the evaluation images?

Were the baseline results in Tables 1 and 2 obtained from checkpoints, retraining, or published numbers? Were all outputs evaluated with the same preprocessing and No GT-mean implementation?

Which data split is used in Figures 2 and 3? How is local structural correlation computed and averaged?

7 comments

We are grateful for the reviewer's acknowledgment of our work. We have taken the time to thoroughly address your questions and have included detailed responses to clarify any uncertainties.

Q1. Is LOL-v1 (?) an unresolved citation placeholder? Is the missing PSNR ↑ label in Table 3 a typesetting error? For Figure 2 and Figure 3, which data split is used, and what are the exact definitions and aggregation procedures for source-target MSE and local structural correlation?

We thank the reviewer for pointing out these omissions. The unresolved citation marker “LOL-v1 (?)” and the missing “PSNR ↑” label in Table 3 are citation and typesetting oversights. We will provide the correct citation and fix the table header.

The analyses in Figures 2 and 3 were conducted on the training sets of LOL-v1 and LOL-v2-Real, without using test images. We denote a candidate representation by AA and its paired normal-light target by XX.

  • Pixel-space discrepancy (Figure 2). We use mean squared error:

    MSE⁡(A,X)=13HW∑h,w,c(Ah,w,c−Xh,w,c)2.\operatorname{MSE}(A,X)=\frac{1}{3HW}\sum_{h,w,c}\left(A_{h,w,c}-X_{h,w,c}\right)^2.

    For each image, this metric averages over all pixels and RGB channels. The results are then aggregated through weighted averaging over the LOL-v1 and LOL-v2-Real training sets. A lower value indicates greater pixel-space proximity to the normal-light target.

  • Local structural correlation (Figure 3). We use the local Pearson correlation coefficient:

    ρw(A,X)=cov⁡w(A,X)var⁡w(A)var⁡w(X).\rho_{w}(A,X)=\frac{\operatorname{cov}_{w}(A,X)}{\sqrt{\operatorname{var}_{w}(A)\operatorname{var}_{w}(X)}}.

    Local means, variances, and covariances are computed using a **normalized 11×1111\times11 Gaussian window **. Specifically,

    μA=∑qwqAq,μX=∑qwqXq,\mu_{A}=\sum_{q}w_{q}A_{q},\qquad\mu_{X}=\sum_{q}w_{q}X_{q},

    cov⁡w(A,X)=∑qwq(Aq−μA)(Xq−μX),\operatorname{cov}_{w}(A,X)=\sum_{q}w_{q}(A_{q}-\mu_{A})(X_{q}-\mu_{X}),

    var⁡w(A)=∑qwq(Aq−μA)2,\operatorname{var}_{w}(A)=\sum_{q}w_{q}(A_{q}-\mu_{A})^2,

    with the target variance computed analogously. Correlations are calculated separately for each channel and valid window, averaged within each image, and then averaged over the training images. We retain the sign without taking the absolute value or squaring the coefficient. A higher value indicates greater local structural similarity.

These analyses motivate distinguishing the transport source from complementary restoration evidence: a representation closer to the target in pixel space does not necessarily preserve all useful structural information. We will include these data splits, metric definitions, and aggregation procedures in the revised manuscript.

Q2. Eq. (3) defines the finite-interval transport operator, while Eq. (13) only defines the three restoration losses. The “MeanFlow objective” in Table 5 is not fully specified. Please clarify the complete objective for each training variant, including the JVP computation, stop-gradient operation, t/r sampling, whether the original MeanFlow identity is used, and how the proposed model differs from direct endpoint regression.

  • MeanFlow ablation (Table 5). We retain the same reflectance source and conditioning architecture. For zt=(1−t)X+tRz_{t}=(1-t)X+tR and vc=R−Xv_{c}=R-X, the original MeanFlow identity [1] gives the target

    utgt=sg⁡![vc−(t−r)ddtuθ(zt,t,r;C)].u_{\mathrm{tgt}}=\operatorname{sg}\\!\left[v_{c}-(t-r)\frac{\mathrm{d}}{\mathrm{d}t}u_{\theta}(z_{t},t,r;\mathcal{C})\right].

    The JVP tangent is (vc,1,0)(v_{c},1,0), with rr and conditioning fixed; the target is stop-gradient. We use 75% r=tr=t and 25% r<tr<t samples, without an extra boundary loss. The official weighted squared error is minimized against this target. Remaining sampling settings follow the official configuration; its version and complete parameters will be documented.

  • Final objective.

    Ltotal=E![Ldep+0.5Lint+0.5Lroll].\mathcal{L}_{\mathrm{total}}=\mathbb{E}\\!\left[\mathcal{L}_{\mathrm{dep}}+0.5\mathcal{L}_{\mathrm{int}}+0.5\mathcal{L}_{\mathrm{roll}}\right].

    All paths share the same restoration criterion. The MeanFlow JVP loss is used only in the ablation, not the final objective.

  • Difference from endpoint regression. Endpoint regression supervises only restoration from RR with (t,r)=(1,0)(t,r)=(1,0). R2RFlow additionally restores reference and predicted intermediate states using the shared predictor. Gradients through the predicted state jointly optimize both rollout segments through the final restoration loss.

The distinction is training-state coverage, interval conditioning, and differentiable compositional supervision. Deployment retains a one-step residual form with one predictor evaluation.

We sincerely appreciate your helpful suggestion and will clarify the MeanFlow configuration in the revision.

[1]. Geng Z, Deng M, Bai X, et al. Mean flows for one-step generative modeling[C]//Advances in Neural Information Processing Systems. 2025, 38: 75460-75482.

AuthorsOpenReview ↗

Further response serve to adress reiviewer's concerns(Part 1/2)

Q3. What is the result of a direct regression model X=f(R,S,I,Y)X=f(R,S,I,Y) using exactly the same inputs, U-Net, parameter budget, and restoration loss? If direct regression performs similarly, how should the independent contributions of finite-interval transport and multi-path flow training be interpreted?

The “Deployment objective” in Table 5 is the same-architecture direct residual regression baseline. It retains the same inputs, conditioning pathways, U-Net, parameter budget, and restoration criterion, using only deployment endpoint supervision:

Ldep=ρ!(R−uθ(R,1,0;C),X).\mathcal{L}_{\mathrm{dep}}=\rho\\!\left(R-u_{\theta}(R,1,0;\mathcal{C}),X\right).

With fixed time conditions (t,r)=(1,0)(t,r)=(1,0), this is a conditional residual regressor, equivalently written as X^=fθ(R,S,I,Y)\hat{X}=f_{\theta}(R,S,I,Y).

  • In-dataset comparison (LOL-v1).

    Training strategyPSNR ↑SSIM ↑LPIPS ↓
    Residual regression / deployment-only25.100.85120.1071
    Full multi-path supervision25.260.85780.1018

    Compared with direct residual regression, the multi-path-supervised flow model achieves higher PSNR while maintaining favorable SSIM and LPIPS performance.

  • Cross-dataset comparison. Both LOL-v1-trained models are directly evaluated without target-domain fine-tuning:

    Training strategyNikon PSNR ↑SSIM ↑LPIPS ↓Huawei PSNR ↑SSIM ↑LPIPS ↓
    Residual regression / deployment-only15.250.44630.234418.960.59870.3821
    Full multi-path supervision16.150.47280.231220.080.59980.3790

    PSNR improves by 0.90 dB on LSRW-Nikon and 1.12 dB on LSRW-Huawei, with higher SSIM and lower LPIPS on both datasets.

With the source, architecture, restoration criterion, and deployment fixed, these comparisons support the additional contribution of multi-path supervision to restoration and cross-dataset generalization. Inference still requires one predictor evaluation.

We thank the reviewer for this helpful suggestion. We will explicitly identify the residual regression baseline in Table 5 and include the additional cross-dataset comparison in the revision.

Q4. How does R2RFlow differ precisely from existing conditional flow, rectified-flow, one-step restoration, and Retinex-based generative LLIE methods? Why are the closest methods not included as direct baselines or evaluated under a unified protocol?

We compare noise and observation-derived sources for paired LLIE. Among Retinex-based generators, Diff-Retinex [2] restores reflectance and illumination separately, then recombines them. Our analysis finds MSE⁡(R,X)<MSE⁡(I,X)\operatorname{MSE}(R,X)<\operatorname{MSE}(I,X), while II retains local structural evidence. We thus use a single reflectance-to-image transport, with illumination, structure, and high-frequency conditioning.

ComparisonLLFlow [1]R2RFlow (ours)Diff-Retinex [2]
ModelInvertible flow fϕf_{\phi}Interval transport TθT_{\theta}Dual diffusion GR,GI\mathcal{G}_{R},\mathcal{G}_{I}
Source / priorz∼N(μY,I)z\sim\mathcal{N}(\mu_{Y},\mathbf{I})z1=Rz_{1}=RϵR,ϵI∼N(0,I)\epsilon_{R},\epsilon_{I}\sim\mathcal{N}(0,\mathbf{I})
ConditioningYYC=I,S,HF⁡(Y)\mathcal{C}=\\{I,S,\operatorname{HF}(Y)\\}R,IR,I (respective branches)
Restored variablesX^\hat{X}X^\hat{X}R^,I^\hat{R},\hat{I}
OutputX^=fϕ−1(z;Y)\hat{X}=f_{\phi}^{-1}(z;Y)X^=Tθ(R,1,0;C)\hat{X}=T_{\theta}(R,1,0;\mathcal{C})X^=R^⊙I^\hat{X}=\hat{R}\odot\hat{I}
Restoration supervision−log⁡pϕ(X∣Y)-\log p_{\phi}(X\mid Y)Ldep+0.5Lint+0.5Lroll\mathcal{L}_{\mathrm{dep}}+0.5\mathcal{L}_{\mathrm{int}}+0.5\mathcal{L}_{\mathrm{roll}}Ldiff+γctLcontent\mathcal{L}_{\mathrm{diff}}+\gamma_{\mathrm{ct}}\mathcal{L}_{\mathrm{content}} (per branch)

Here, μY\mu_{Y} is LLFlow’s color-map-based prior mean at inference, I\mathbf{I} is identity covariance, R^,I^\hat{R},\hat{I} are restored normal-light components,.

We train one shared predictor through deployment, reference, and rollout restoration losses; inference uses one evaluation. LLFlow and Diff-Retinex are already evaluated on the same test images under No GT-mean (Q6); Table 5 includes a fixed-architecture MeanFlow ablation.

We appreciate this helpful feedback, which will improve the clarity of our manuscript.

[1]. Wang Y, Wan R, Yang W, et al. Low-light image enhancement with normalizing flow[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2022, 36(3): 2604-2612.

[2]. Yi X, Xu H, Zhang H, et al. Diff-Retinex: Rethinking low-light image enhancement with a generative diffusion model[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 12302-12311.

AuthorsOpenReview ↗

Further response serve to adress reiviewer's concerns(Part 2/2)

Q5. Which exact models and checkpoints are used for D and SMG? Are they frozen? Do their pretraining data include LOL, LSRW, related test images, or datasets with overlapping scenes?

The two pretrained modules are adopted from the following works:

  • Retinex decomposition. We use the Decom-Net subnetwork of Retinex-Net, introduced by Wei et al. in Deep Retinex Decomposition for Low-Light Enhancement [1], as our decomposition module DD. It extracts reflectance and illumination from the low-light observation:

    (R,I)=D(Y).(R,I)=D(Y).

  • Structural prior extraction. We use the structure modeling branch (edge prediction network) of SMG, introduced by Xu et al. in Low-Light Image Enhancement via Structure Modeling and Guidance [2], to extract the structural prior:

    S=GSMG(Y),S=\mathcal{G}_{\mathrm{SMG}}(Y),

    which provides conditioning information for the main restoration network.

Training data and freezing. We adopt the architectures from these works and pretrain both modules on the corresponding benchmark training split. For LOL-v1, both Retinex and SMG are pretrained exclusively on the LOL-v1 training set; for LOL-v2-Real, both are pretrained exclusively on the LOL-v2-Real training set. After pretraining, both modules are frozen during training of the main restoration network. Their training data are therefore restricted to the respective benchmark training split, without using test images or additional external training data.

Cross-dataset evaluation. For LSRW-Nikon and LSRW-Huawei, we directly evaluate the complete LOL-v1 model: Retinex, SMG, and the main restoration network all retain the weights obtained in the LOL-v1 experiment. No component is retrained or fine-tuned on either target dataset.

We appreciate this helpful suggestion, which will improve the clarity of our manuscript. We will explicitly document the module sources, training-data scope, freezing status, and checkpoint identifiers in the revision.

Q6. Were the baseline results in Tables 1 and 2 obtained by unified retraining, official checkpoints, or numbers taken from prior papers? Were all methods evaluated with the same input resolution, preprocessing, PNG saving procedure, brightness handling, and No GT-mean implementation? Do all supervised methods in Table 2 use LOL-v1 checkpoints?

Except for the case described below, baseline results were obtained by running the publicly released official checkpoints and evaluating them on the same test images using unified evaluation code, rather than directly copying metrics from previous papers.

For RetinexNet on LOL-v2-Real, no official checkpoint trained on this dataset was available. We therefore used the official implementation and followed the training parameters of the official LOL-v1 experiment, including a batch size of 16, to train the model on the LOL-v2-Real training set and evaluate it on the corresponding test set.

All metrics are computed from saved PNG outputs under the same No GT-mean protocol. We do not rescale or correct the brightness of enhanced images using the mean intensity of the normal-light ground truth. LPIPS is consistently computed using AlexNet features. We will document each method’s input resolution, normalization and other preprocessing, output-saving procedure, and brightness handling, distinguishing model-specific input requirements from the evaluation steps shared across all results.

All supervised methods in Table 2 use their LOL-v1 checkpoints and are directly evaluated on the target datasets without target-domain fine-tuning. For R2RFlow, this setting covers Retinex, SMG, and the main restoration network: the entire system uses weights obtained in the LOL-v1 experiment.

[1]. Wei C, Wang W, Yang W, et al. Deep Retinex decomposition for low-light enhancement[C]//British Machine Vision Conference. 2018.

[2]. Xu X, Wang R, Lu J. Low-light image enhancement via structure modeling and guidance[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 9893-9903.

Reviewer_ad59OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents R2RFlow, a reflectance-to-restoration flow framework for one-step low-light image enhancement. It constructs finite-interval transport from observation-derived reflectance and incorporates structural, illumination, and frequency information through dedicated conditioning pathways. A restoration-aligned multi-path learning strategy supervises the shared transport operator through final restoration outcomes from the deployment source, reference intermediate states, and model-induced rollout states.

Critical strengths

A better performance than existing generative-based methods

Critical weaknesses
  1. The motivation is not sufficiently clear. The authors adopt a generative modeling approach but do not adequately explain the unique advantages of this paradigm for low-light image enhancement. Furthermore, it remains unclear what specific limitations of existing low-light enhancement methods the proposed approach further addresses and how it overcomes these limitations.

  2. The experimental results are insufficient to substantiate the claims made in the paper. Specific concerns are detailed in the following Questions For Authors.

Questions for authors

1.In the abstract, the authors describe their proposed method as “R2RFlow, a reflectance-to-restoration flow framework integrating complementary evidence for one-step low-light image enhancement.” However, the finite-interval mechanism that enables one-step inference is largely inherited from MeanFlow. Although Table 5 shows that multi-path training outperforms supervision using only the deployment path, I would still like the authors to distinguish more clearly between the mechanisms adopted from MeanFlow and the designs introduced in this paper. I also suggest reporting the individual contributions of the reference intermediate path and the model-induced rollout path, to establish how the proposed framework improves one-step low-light image enhancement.

2.The authors provide reasonable motivation for incorporating structural, illumination, and frequency information, and Table 4 shows the effects of progressively adding the corresponding modules. However, I do not think these experiments sufficiently establish the individual roles of the three types of information or their complementarity. Figure 3 mainly shows that the illumination component and the original observation retain useful information, while the cumulative ablation in Table 4 does not directly reveal the contribution of each type of information within the complete model. I suggest removing each of the three types of conditioning separately from the full model and providing local visualizations of the structural prior, high-frequency information, and gating weights. In particular, I would like the authors to show how the high-frequency conditioning provides additional details, rather than duplicating structural information or amplifying noise.

3.I think the main comparison with generative methods is insufficiently up to date and does not adequately cover the methods most relevant to this work. The most recent generative baseline in Table 1 is GSAD (NeurIPS 2023), whereas the paper focuses on one-step low-light restoration using a flow-based framework. The recent Consist-Retinex also combines Retinex representations with one-step restoration and explicitly aligns its predictions with paired normal-light targets. I would like the authors to discuss the differences from this closely related work more thoroughly, to provide more convincing support for the paper’s novelty and claimed performance advantages.

Minor Weaknesses 1.The paper reports one predictor evaluation (1 NFE), but it does not report the actual end-to-end inference time. Since the pipeline also requires Retinex decomposition and SMG-based structure extraction, I would like the authors to report the full inference latency, including these stages, under a clearly specified hardware and input-resolution setting.

2.The experimental setup and conclusion state that the LOL-v1-trained model is also evaluated on LOL-v2-Real without fine-tuning, but Table 2 only reports results on LSRW-Nikon and LSRW-Huawei. The LOL-v2-Real results in Table 1 come from a separately trained model. I would like the authors to provide the missing cross-dataset results or revise the corresponding statements.

3.For example, Figure 1 uses “Complimentry” and “Eliment-wise,” which should be “Complementary” and “Element-wise,” respectively. The citation for LOL-v1 on page 2 appears as “(?),” and Table 3 is missing the PSNR column header. I suggest proofreading the text, figures, and tables more carefully.

5 comments

We thank you for your constructive comments and provide detailed responses to your questions below.

Q1. I would still like the authors to distinguish more clearly between the mechanisms adopted from MeanFlow and the designs introduced in this paper. I also suggest reporting the individual contributions of the reference intermediate path and the model-induced rollout path, to establish how the proposed framework improves one-step low-light image

Our motivation is to select an observation-derived source closer to the restoration target, supplement useful information weakened in this representation, and improve one-step enhancement through restoration-oriented multi-path supervision.

Our work was initially inspired by MeanFlow [1] and adopts its finite-interval parameterization. Our task-specific designs concern reflectance initialization, complementary evidence conditioning, and restoration-aligned multi-path learning. Table 5 already shows that, under the same initialization and conditioning architecture, our training objective outperforms the MeanFlow loss across PSNR, SSIM, and LPIPS.

The reference path supervises restoration from paired intermediate states, expanding state and interval coverage. The rollout path supervises a differentiable composition through model-predicted states, allowing the final restoration error to optimize intermediate-state generation and encourage states that support subsequent restoration.

Following your suggestion, we conducted additional ablations introducing each auxiliary path separately. All variants retain the same restoration network and one-step inference.

DeploymentReferenceRolloutPSNR ↑SSIM ↑LPIPS ↓
✓——25.100.85120.1071
✓✓—25.100.85040.1065
✓—✓25.120.85700.1020
✓✓✓25.260.85780.1018

✓ indicates that the corresponding path loss is included.

Reference-path supervision alone leaves PSNR unchanged, with mixed effects on SSIM and LPIPS. Rollout supervision improves SSIM and LPIPS. Combining reference-path supervision with rollout supervision increases PSNR from 25.12 to 25.26 dB, with additional improvements in SSIM and LPIPS. The benefit of reference-path supervision is most evident when used jointly with rollout supervision.

Full multi-path supervision achieves the best performance across all three metrics, improving PSNR by 0.16 dB over deployment-only supervision. These results support the joint use of supervision from reference states and compositions through predicted states, while retaining one-step inference. It has been very helpful in enhancing our manuscript.

Q2. I suggest removing each of the three types of conditioning separately from the full model and providing local visualizations of the structural prior, high-frequency information, and gating weights.

The structural prior provides spatial structural cues; illumination guides multi-scale encoding and decoding through LWDM and IGCA; and observation-derived high-frequency information provides local details through GWDF. These sources may contain overlapping information. Since high-frequency responses include both details and noise, GWDF uses learnable gating to control their contribution, aiming to selectively exploit useful high-frequency information.Visualizations of the conditioning information are already included in the method figure.

We conducted additional ablations by removing one conditioning source at a time from the complete model. The corresponding removed inputs or modules are specified below:

Information typePSNR ↑SSIM ↑LPIPS ↓
Without structural conditioning (S)24.800.81040.1517
Without illumination conditioning (LWDM + IGCA)23.720.82870.1368
Without observation-derived high-frequency conditioning (GWDF)24.840.81500.1410
All conditioning information (full model)25.260.85780.1018

Removing structural and high-frequency conditioning causes larger SSIM decreases (0.0474 and 0.0428) than removing illumination (0.0291), consistent with their roles in structure and detail preservation. Illumination removal produces the largest PSNR drop (1.54 dB), versus 0.46 dB and 0.42 dB for structural and high-frequency removal, respectively, highlighting its importance for pixel-wise accuracy. All removals worsen LPIPS. These results complement Table 4’s cumulative ablation, showing that each source contributes with the others retained.

We will clarify the correspondence between the visualizations and conditioning pathways in the revised paper.Thank you for the suggestion, which helped us provide more targeted experimental evidence.

[1] Geng Z, Deng M, Bai X, et al. Mean flows for one-step generative modeling[C]//Advances in Neural Information Processing Systems. 2025, 38: 84032-84054.

AuthorsOpenReview ↗

Further response serve to adress reiviewer's concerns(Part 1/2)

Q3. The recent Consist-Retinex also combines Retinex representations with one-step restoration and explicitly aligns its predictions with paired normal-light targets. I would like the authors to discuss the differences from this closely related work more thoroughly, to provide more convincing support for the paper’s novelty and claimed performance advantages.

Consist-Retinex [1] is closely related to our work. Our motivation is to explore a single transport path from an informative observation-derived source to the normal-light image, while using information retained in other representations as complementary conditioning.

Using unified notation, let (R,I)(R,I) denote the low-light Retinex components and (R⋆,I⋆)(R^\star,I^\star) the normal-light component targets used by Consist-Retinex. The comparison is summarized below:

AspectR2RFlow (ours)Consist-Retinex [1]
Transport sourceRRσmax⁡ϵR, σmax⁡ϵI\sigma_{\max}\epsilon_R,\ \sigma_{\max}\epsilon_I
ConditioningC={S,I,HY}\mathcal C=\{S,I,H_Y\}fR(⋅∣R), fI(⋅∣I)f^R(\cdot\mid R),\ f^I(\cdot\mid I)
Prediction targetXX(R⋆,I⋆)(R^\star,I^\star)
One-step predictionX^=Tθ(R,1,0;C)\hat X=T_\theta(R,1,0;\mathcal C)Z^=fϕZZ(σmax⁡ϵZ,σmax⁡∣Z)\hat Z=f^Z_{\phi_Z}(\sigma_{\max}\epsilon_Z,\sigma_{\max}\mid Z)
Image reconstructionX^=R−uθ(R,1,0;C)\hat X=R-u_\theta(R,1,0;\mathcal C)X^=R^⊙I^\hat X=\hat R\odot\hat I
Training objectiveLdep+βLint+γLroll\mathcal L_{\rm dep}+\beta\mathcal L_{\rm int}+\gamma\mathcal L_{\rm roll}Lconsist+λLfixed\mathcal L_{\rm consist}+\lambda\mathcal L_{\rm fixed}

Here, Z∈{R,I}Z\in\{R,I\} and ϵZ\epsilon_Z is standard Gaussian noise. SS and HYH_Y denote the structural prior and observation-derived high-frequency information, respectively.

Both methods exploit paired normal-light supervision. Consist-Retinex separately generates normal-light reflectance and illumination before reconstructing the image. R2RFlow instead uses reflectance directly as the transported state and predicts the complete normal-light image through one shared operator. The remaining information serves as conditioning, and the multiple training paths optimize this same operator while retaining a single full-interval prediction at deployment.

We will include this comparison and expand the discussion of closely related work. Thank you for the suggestion, which helps us clarify the motivation and contribution boundaries of our method.

Q4. Since the pipeline also requires Retinex decomposition and SMG-based structure extraction, I would like the authors to report the full inference latency, including these stages, under a clearly specified hardware and input-resolution setting.

Thank you for pointing out that predictor evaluation counts alone do not adequately characterize the practical inference efficiency of the complete system.

We measured the full inference latency from the raw low-light image YY to the enhanced image X^\hat X on a single RTX 5090, using an input resolution of 400×600400\times600, batch size 1, and FP32 precision. Timing includes Retinex decomposition, SMG-based structure extraction, conditioning construction, one forward pass of the main restoration network, input/output conversion, and CPU↔GPU data transfers.

Module / PipelineParametersAverage latency
Retinex decomposition207,7484.37 ms
SMG-based structure extraction16,297,20583.01 ms
Conditioning construction00.16 ms
Main restoration network (one forward pass)5,130,64030.19 ms
Input/output conversion and CPU↔GPU transfers01.01 ms
Full pipeline Y→X^Y\to\hat X (measured)21,635,593118.31 ms

The complete system contains 21.636 M parameters, with a measured average end-to-end latency of 118.31 ms/image and a P95 latency of 125.20 ms/image, corresponding to approximately 8.45 images/second based on the average latency. The one-step restoration prediction takes 30.19 ms, while SMG-based structure extraction takes 83.01 ms, accounting for approximately 70.2% of the end-to-end latency and constituting the main computational bottleneck in the current implementation.

End-to-end and per-stage latencies were measured separately. For the per-stage measurements, CUDA synchronization was performed before and after each stage. Due to the additional synchronization overhead and run-to-run variation, the sum of the unrounded per-stage latencies (approximately 118.73 ms) differs slightly from the measured end-to-end latency (118.31 ms), by approximately 0.35%.

We will include this report in the revision and clarify that 1 NFE refers to one evaluation of the main restoration predictor, whereas the complete system also incurs preprocessing and data-conversion costs.

[1] Xu J, Chen W, Li S, et al. Consist-Retinex: One-step noise-emphasized consistency training accelerates high-quality Retinex enhancement[EB/OL]. arXiv:2512.08982, 2025.

AuthorsOpenReview ↗

Further response serve to adress reiviewer's concerns(Part 2/2)

Q5. The experimental setup and conclusion state that the LOL-v1-trained model is also evaluated on LOL-v2-Real without fine-tuning, but Table 2 only reports results on LSRW-Nikon and LSRW-Huawei. The LOL-v2-Real results in Table 1 come from a separately trained model. I would like the authors to provide the missing cross-dataset results or revise the corresponding statements.

We sincerely thank the reviewer for pointing out this issue. We will revise the statements regarding cross-dataset evaluation on LOL-v2-Real so that our cross-dataset conclusions are consistent with the LSRW-Nikon and LSRW-Huawei results reported in Table 2. This suggestion is very helpful for improving our manuscript.

Q6. For example, Figure 1 uses “Complimentry” and “Eliment-wise,” which should be “Complementary” and “Element-wise,” respectively. The citation for LOL-v1 on page 2 appears as “(?),” and Table 3 is missing the PSNR column header. I suggest proofreading the text, figures, and tables more carefully.

We will correct these spelling, citation, and table-labeling errors in the revision. Thank you for your suggestions, which are very helpful for improving our manuscript.

Reviewer_KyqmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

R2RFlow proposes a one-step framework for low-light image enhancement. It initializes restoration from a pretrained Retinex decomposition’s reflectance component rather than noise, and supplements this source with structural, illumination, and wavelet information. An interval-conditioned predictor combines endpoint supervision, reference intermediate states, and model-induced rollouts. Evaluation covers LOL-v1, LOL-v2-Real, and transfer to LSRW.

Critical strengths
  • Under matched conditioning and training objectives, reflectance initialization improves PSNR from 23.58 to 25.26 dB over low-light initialization.

  • Competitive reconstruction performance. The method achieves the highest PSNR among the evaluated methods on both main benchmarks without ground-truth mean-brightness adjustment, alongside favorable cross-dataset reconstruction results.

Critical weaknesses
  • Novelty relative to PVRF needs justification. PVRF already constructs its image-space flow source from an observation-derived restoration estimate with adaptive noise perturbation. Thus, avoiding pure-noise initialization alone does not establish novelty.

  • The flow interpretation remains insufficiently established. At deployment, Eq. (4) reduces to a deterministic residual predictor, X^=R−uθ(R,1,0;C)\hat{X}=R-u_\theta(R,1,0;C). Endpoint supervision alone does not establish meaningful intermediate transport or consistent interval composition. Intermediate-state accuracy and direct-versus-composed transport tests would substantiate this interpretation.

  • Multi-path learning is not sufficiently isolated. The full objective improves LOL-v1 PSNR by 0.16 dB over deployment-only supervision, but the two auxiliary paths are introduced together. Separate ablations, comparisons with matched training compute, and repeated-seed results would clarify their individual value.

  • Performance claims need qualification. Main-benchmark PSNR gains over the strongest listed competitors are modest (0.11 and 0.14 dB), while LPIPS is not consistently best. The results support competitive reconstruction fidelity, with mixed perceptual advantages. Uncertainty estimates would help assess these differences on the small test sets.

Questions for authors

What are the complete inference and training costs?

Why GPP-LLIE is missed in evaluations, which is a recent a diffusion-based enhancement method?

7 comments

We thank the reviewer for the acknowledgement of the main contributions of our work, the useful comments and the relevant feedback provided on the technical side. We have provided detailed responses to your questions.

Q1. Novelty relative to PVRF needs justification. PVRF already constructs its image-space flow source from an observation-derived restoration estimate with adaptive noise perturbation. Thus, avoiding pure-noise initialization alone does not establish novelty.

Thank you for highlighting the connection between PVRF [1] and our work. Both methods use observation-derived source states, with the following differences in design.

ComparisonPVRFR2RFlow
Source state constructionμ+δϵ\mu+\delta\epsilon: restoration estimate + adaptive noiseRR: Retinex reflectance
VLM priorsYes: degradation-type and attribute scoresNo: observation-derived structure, illumination, and frequency information
Flow predictor supervisionReference-path velocity matching: ∥vϕ−(X−μ−δϵ)∥22\lVert v_\phi-(X-\mu-\delta\epsilon)\rVert_2^2Endpoint restoration on three paths: ℓ(X^p,X)\ell(\hat X_p,X), p∈{dep,ref,roll}p\in\{\mathrm{dep},\mathrm{ref},\mathrm{roll}\}

Our work explores how decomposing the observation can support image restoration. In our component analysis, we observed that reflectance is closer to the normal-light image in pixel space, while the illumination component and the original observation retain useful structure and detail. This observation motivates us to assign different roles to these sources of information: reflectance provides the restoration starting point, while structural, illumination, and frequency information serves as complementary conditioning, with restoration learned from paired targets. In this setting, the conditioning information comes from the low-light observation and its decomposed representations, and the relevant modules are trained on the corresponding training set, without additional VLM priors. Further details and supporting evidence regarding the pretraining data sources and comparison fairness are provided in Q6.

Q2. Endpoint supervision alone does not establish meaningful intermediate transport or consistent interval composition. Intermediate-state accuracy and direct-versus-composed transport tests would substantiate this interpretation.

Following your valuable suggestions, we conducted three additional tests to examine the predictor’s interval transport behavior: intermediate-state accuracy, endpoint composition consistency, and non-endpoint composition consistency.

(1) Intermediate-state accuracy. Starting independently from the same reflectance source RR, we predict states at different times, z^r=Tθ(R,1,r)\hat z_r=T_\theta(R,1,r), and compare them with the reference states zr=rR+(1−r)Xz_r=rR+(1-r)X. We also construct an endpoint interpolation baseline, br=rR+(1−r)X^b_r=rR+(1-r)\hat X, using the one-step restoration result X^=Tθ(R,1,0)\hat X=T_\theta(R,1,0).

Time rrModel MAE ↓Endpoint interpolation MAE ↓
0.750.0146970.013715
0.500.0285170.027429
0.250.0418140.041144
00.0548580.054858

The intermediate-state errors are close to those of the endpoint interpolation baseline and increase approximately linearly with the interval length 1−r1-r. This trend is consistent with approximately linear temporal scaling of the endpoint correction.

(2) Endpoint composition consistency. We compare one-step restoration with two-step restoration through a model-predicted intermediate state:

\quad\text{and}\quad T_\theta(T_\theta(R,1,r),r,0).$$ | Intermediate time $r$ | Composition error MAE ↓ | |:---:|:---:| | 0.75 | 0.004013 | | 0.50 | 0.006232 | | 0.25 | 0.005133 | Across the three intermediate times, the MAE between one-step and two-step restoration does not exceed 0.006232, indicating that different interval decompositions yield similar final restoration outputs. **(3) Non-endpoint composition consistency.** We further compare direct and composed predictions that both stop at an intermediate time $s$: $$T_\theta(R,1,s) \quad\text{and}\quad T_\theta(T_\theta(R,1,r),r,s).$$ | Intermediate $r$ → stopping time $s$ | Composition error MAE ↓ | |:---:|:---:| | 0.75 → 0.50 | 0.001600 | | 0.75 → 0.25 | 0.002754 | | 0.50 → 0.25 | 0.002786 | The non-endpoint composition errors do not exceed 0.002786, indicating good agreement between direct and composed predictions at the tested intermediate times. These results characterize the temporal variation of intermediate-state errors and the approximate agreement between direct and composed predictions at both final and intermediate times, providing empirical support for our finite-interval restoration formulation. [1]. Dong W, Zhou H, Ji T, et al. PVRF: All-in-one adverse weather removal via prior-modulated and velocity-constrained rectified flow[PP/OL]. arXiv:2605.14045, 2026.
AuthorsOpenReview ↗

Further response serve to adress reiviewer's concerns(Part 1/2)

Q3. Comparison of different training-path combinations.

Following your suggestion, we conducted additional experiments on LOL-v1 to examine the roles of reference-path and rollout supervision through different training-path combinations. All configurations use the same restoration network and a single full-interval prediction at inference. The results are as follows:

Training configurationPSNR ↑SSIM ↑LPIPS ↓
Deployment-only supervision25.100.85120.1071
Deployment + reference-path supervision25.100.85040.1065
Deployment + rollout supervision25.120.85700.1020
Full multi-path supervision25.260.85780.1018

The results show that adding rollout supervision improves both SSIM and LPIPS. Combining it with reference-path supervision further increases PSNR from 25.12 to 25.26 dB, with additional improvements in SSIM and LPIPS. Full multi-path supervision achieves the best performance across all three metrics, improving PSNR by 0.16 dB over deployment-only supervision.

These results support the joint use of the two additional supervision terms: supervising the final restoration outputs from reference states and from compositions through predicted states improves restoration while retaining one-step inference. We will include these comparisons and analyses in the paper. Thank you for your suggestion, which helped us better examine and clarify the roles of the training paths.

Q4. Performance gains and cross-dataset generalization.

We agree that the full model’s PSNR improvement over the direct residual regression baseline is relatively modest in within-dataset supervised evaluation. We conducted additional cross-dataset comparisons to examine how this training design generalizes to unseen data.

Both models are trained on LOL-v1 and evaluated directly on LSRW-Nikon and LSRW-Huawei, without training or fine-tuning on the target datasets. The results are as follows:

Training strategyNikon PSNR ↑SSIM ↑LPIPS ↓Huawei PSNR ↑SSIM ↑LPIPS ↓
Direct residual regression / deployment-only supervision15.250.44630.234418.960.59870.3821
Full multi-path supervision16.150.47280.231220.080.59980.3790

In cross-dataset evaluation, the full model improves PSNR by 0.90 dB on LSRW-Nikon and 1.12 dB on LSRW-Huawei, with improvements in both SSIM and LPIPS. These results show that the benefits of multi-path supervision over direct residual regression are more pronounced in cross-dataset evaluation, improving restoration on unseen data while retaining one-step inference.

AuthorsOpenReview ↗

Further response serve to adress reiviewer's concerns(Part 2/2)

Q5. What are the complete inference and training costs?

(1) Training setup. We train the main restoration network using NVIDIA A100 (40 GB) with a batch size of 64. During this stage, the Retinex and SMG modules remain frozen, and only the restoration network, containing approximately 5.131 M parameters, is optimized. Multi-path supervision is used during training, while inference requires a single full-interval prediction.

(2) Inference efficiency. We measure the complete inference latency from the raw low-light image YY to the enhanced image X^\hat X on a single RTX 5090, using a resolution of 400×600400\times600, batch size 1, and FP32 precision. Timing includes Retinex decomposition, SMG structure extraction, conditioning construction, one forward pass through the main restoration network, input/output conversion, and CPU↔GPU transfers.

ModuleParametersMean latency
Retinex decomposition207,7484.37 ms
SMG structure extraction16,297,20583.01 ms
Conditioning construction00.16 ms
Main restoration network (one forward pass)5,130,64030.19 ms
Input/output conversion and CPU↔GPU transfers01.01 ms
Complete pipeline Y→X^Y\to\hat X (measured)21,635,593118.31 ms

End-to-end and per-stage timings were measured separately, with CUDA synchronization before and after each stage for the latter. The sum of unrounded stage timings (approximately 118.73 ms) differs slightly from the end-to-end latency (118.31 ms), by approximately 0.35%, due to additional synchronization overhead and variation between measurement runs.

The complete system contains 21.636 M parameters. Its measured mean end-to-end latency is 118.31 ms/image, with a P95 latency of 125.20 ms/image, equivalent to approximately 8.45 images/s based on the mean latency. The one-step restoration prediction takes 30.19 ms. SMG structure extraction takes 83.01 ms, accounting for approximately 70.2% of the end-to-end latency and representing the dominant latency contribution in the current implementation.

Q6. Why GPP-LLIE is missed in evaluations, which is a recent a diffusion-based enhancement method?

Thank you for highlighting GPP-LLIE [1]. Our main quantitative comparisons focused on settings without additional externally pretrained VLM priors, so GPP-LLIE was not included in the same comparison group. The key differences are summarized below.

ComparisonGPP-LLIER2RFlow
Source statezT∼N(0,I)z_T\sim\mathcal N(0,I) (latent noise)RR (Retinex reflectance)
VLM priors(Sg,Ml)=PVLM(Y)(S_g,M_l)=\mathcal P_{\mathrm{VLM}}(Y)None
Training resourcesDtrain\mathcal D_{\mathrm{train}} + external VLM priorsDtrain\mathcal D_{\mathrm{train}} only
Restoration inferencezT→CGPP25 stepsz^0→DX^z_T\xrightarrow[\mathcal C_{\mathrm{GPP}}]{25\ \mathrm{steps}}\hat z_0\xrightarrow{D}\hat XX^=Tθ(R,1,0;Cobs)\hat X=T_\theta(R,1,0;\mathcal C_{\mathrm{obs}})

Here, PVLM\mathcal P_{\mathrm{VLM}} denotes prior extraction using Q-Instruct-tuned LLaVA, producing a global score SgS_g and a local quality map MlM_l. CGPP={E(Y),Sg,Ml}\mathcal C_{\mathrm{GPP}}=\{E(Y),S_g,M_l\}, where E/DE/D denote the latent encoder/decoder. Cobs\mathcal C_{\mathrm{obs}} contains our observation-derived structural, illumination, and frequency cues.

Within each R2RFlow checkpoint, the pretrained Retinex and SMG modules and the main restoration network are trained exclusively on the same benchmark training split Dtrain\mathcal D_{\mathrm{train}}, without additional training datasets. This applies separately to the LOL-v1 and LOL-v2-Real checkpoints. Retinex and SMG remain frozen during training of the main restoration network. Cross-dataset evaluation directly uses the complete source-trained checkpoint, including both pretrained modules and the restoration network, without retraining or fine-tuning any component on the target datasets.

Regarding computational efficiency, VLM prior extraction should be included in the complete pipeline cost. Our end-to-end measurements in Q5 include Retinex decomposition, SMG extraction, and the restoration network.

We will add a citation to GPP-LLIE in the related work and clarify the training data and prior settings underlying our main quantitative comparisons. Thank you for your valuable suggestion, which helped us improve the discussion of related work and clarify our experimental settings.

[1]. Zhou H, Dong W, Liu X, et al. Low-light image enhancement via generative perceptual priors[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2025, 39(10): 10752-10760.

Reviewer_E7QnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes R2RFlow, a one-step low-light image enhancement framework that initializes flow-based restoration from reflectance rather than random noise. It integrates complementary structural, illumination, and frequency information with a restoration-aligned multi-path training strategy to improve image reconstruction quality.

Critical strengths
  1. The reflectance-based initialization and complementary evidence conditioning provide a well-motivated framework for efficient one-step image restoration.
  2. The method achieves competitive results on LOL-v1 and LOL-v2-Real, while demonstrating promising cross-dataset generalization and consistent improvements in ablation studies.
Critical weaknesses
  1. The proposed flow formulation closely resembles one-step residual regression, and its advantages over direct regression remain unclear. Existing methods such as InDI, I2SB, and ResShift have explored restoration from degraded observations, while the novelty and independent contribution of the proposed multi-path learning strategy remain insufficiently demonstrated.
  2. The reliance on pretrained modules raises concerns about comparison fairness, while the lack of computational efficiency analysis and controlled ablations limits the experimental evidence.
Questions for authors
  1. Could the authors further clarify the advantages of the proposed flow formulation over a conventional one-step residual regression approach? Could the authors elaborate on how the proposed multi-path learning strategy differs from existing methods such as InDI and I2SB, particularly in terms of its methodological contributions?
  2. Could the authors clarify the training details of the pretrained modules and provide additional evidence regarding comparison fairness and computational efficiency?
7 comments

We sincerely thank you for your insightful comments and for recognizing the unique and significant contributions of our research. Your support is invaluable, and we would be honored to have you as an advocate for our work.

Q1. How does the flow formulation differ from one-step residual regression, and what are its advantages?

We agree that X^=R−uθ(R,1,0;C)\hat{X}=R-u_{\theta}(R,1,0;\mathcal{C}) has a residual form. The distinction lies in training: a shared predictor receives deployment, reference-state, and predicted-state supervision. We omit C\mathcal C below.

  • Reference-state supervision. For zt=(1−t)X+tRz_{t}=(1-t)X+tR,

    ∥zt−t,uθ(zt,t,0)−X∥1=t∥uθ(zt,t,0)−(R−X)∥1.\lVert z_{t}-t\\,u_{\theta}(z_{t},t,0)-X\rVert_{1}=t\lVert u_{\theta}(z_{t},t,0)-(R-X)\rVert_{1}.

    This provides time-weighted supervision of the reference-path velocity R−XR-X.

  • Predicted-state supervision. Let u1=uθ(zt,t,r)u_{1}=u_{\theta}(z_{t},t,r), z^r=zt−(t−r)u1\hat z_{r}=z_{t}-(t-r)u_{1}, and u2=uθ(z^r,r,0)u_{2}=u_{\theta}(\hat z_{r},r,0). Then

    ∥X^roll−X∥1=∥(t−r)u1+ru2−t(R−X)∥1.\lVert\hat X_{\mathrm{roll}}-X\rVert_{1}=\lVert(t-r)u_{1}+ru_{2}-t(R-X)\rVert_{1}.

    This constrains the summed displacement. Gradients through z^r\hat z_{r} optimize the first prediction according to its effect on subsequent restoration.

  • Experimental evidence. With the source, conditioning architecture, and deployment fixed, multi-path training improves LOL-v1 PSNR by 0.16 dB, maintaining SSIM and LPIPS (Table 5). Direct evaluation of LOL-v1-trained models on LSRW gives:

    Training strategyNikon PSNR ↑SSIM ↑LPIPS ↓Huawei PSNR ↑SSIM ↑LPIPS ↓
    Residual regression / deployment-only15.250.44630.234418.960.59870.3821
    Full multi-path supervision16.150.47280.231220.080.59980.3790

    PSNR gains are 0.90 dB on Nikon and 1.12 dB on Huawei, with improved SSIM and LPIPS.

These results support better restoration and generalization through multi-path training, while deployment requires one predictor evaluation.

Q2. How does multi-path learning differ from InDI, I²SB, and ResShift?

We acknowledge prior work on observation-based restoration. Our contribution concerns the source representation and training strategy for one-step low-light enhancement.

  • Motivation. Which observation-derived representation should serve as the source, and how should the remaining information contribute? Comparing low-light images, reflectance, illumination, and noise motivates selecting reflectance; Table 3 experimentally validates this choice. Learnable conditioning branches incorporate complementary structural, illumination, and high-frequency evidence.

  • Formulation and training. Below, X,YX,Y denote clean and degraded images or, for latent-space ResShift, their corresponding representations; n∼N(0,I)n\sim\mathcal N(0,I).

    MethodFormulation and training statesSupervision
    I²SB [1]Analytical Schrödinger bridge: q(zt∣X,Y)=N(μt,Σt)q(z_{t}\mid X,Y)=\mathcal N(\mu_{t},\Sigma_{t})Clean-endpoint-derived score targets
    ResShift [2]Residual-shifting diffusion: zt=X+ηt(Y−X)+κηtnz_{t}=X+\eta_{t}(Y-X)+\kappa\sqrt{\eta_{t}}nHigh-resolution target prediction
    InDI [3]Interpolation with optional noise: zt=(1−t)X+tY+tϵtnz_{t}=(1-t)X+tY+t\epsilon_{t}n; residual-flow ODE interpretationClean-target prediction; iterative inference
    R2RFlowReference states zt=(1−t)X+tRz_{t}=(1-t)X+tR and predicted states z^r=Tθ(zt,t,r;C)\hat z_{r}=T_{\theta}(z_{t},t,r;\mathcal C)Joint deployment, reference-state, and differentiable rollout supervision

    Our finite-interval operator is

    Tθ(zt,t,r;C)=zt−(t−r)uθ(zt,t,r;C).T_{\theta}(z_{t},t,r;\mathcal C)=z_{t}-(t-r)u_{\theta}(z_{t},t,r;\mathcal C).

    InDI is closest to our reference-state supervision. We additionally train through a model-predicted intermediate state, backpropagating the terminal error through both segments. Both auxiliary paths share the deployment predictor.

R2RFlow combines reflectance initialization, complementary evidence, and multi-path training for restoration with one full-interval prediction.

[1] Liu G H, Vahdat A, Huang D A, et al. I²SB: Image-to-image Schrödinger bridge[C]//Proceedings of the 40th International Conference on Machine Learning. PMLR, 2023: 22042-22062.

[2] Yue Z, Wang J, Loy C C. ResShift: Efficient diffusion model for image super-resolution by residual shifting[C]//Advances in Neural Information Processing Systems. 2023, 36: 13294-13307.

[3] Delbracio M, Milanfar P. Inversion by direct iteration: An alternative to denoising diffusion for image restoration[J]. Transactions on Machine Learning Research, 2023.

AuthorsOpenReview ↗

Further response serve to adress reiviewer's concerns

Q3. Could the authors clarify the training details of the pretrained modules and provide additional evidence regarding comparison fairness and computational efficiency?

R2RFlow uses Retinex decomposition to extract reflectance and illumination, and SMG to extract structural information. Below, we clarify their training data and report the complete inference cost, including both modules.

  • Training data and comparison fairness. For the LOL-v1 and LOL-v2-Real experiments, Retinex and SMG are pretrained exclusively on the respective training split and then frozen during training of the main restoration network. For cross-dataset evaluation on LSRW-Nikon and LSRW-Huawei, both pretrained modules and the main restoration network directly use the LOL-v1 weights, without retraining or fine-tuning on either target dataset.

  • Computational efficiency. We measure the complete inference latency from low-light input YY to enhanced output X^\hat X on a single RTX 5090 at 400×600400\times600 resolution, batch size 1, and FP32 precision, including all modules, input/output conversion, and CPU↔GPU transfers.

    ModuleParametersMean latency
    Retinex decomposition207,7484.37 ms
    SMG structure extraction16,297,20583.01 ms
    Condition construction00.16 ms
    Main restoration network (one forward evaluation)5,130,64030.19 ms
    Input/output conversion and CPU↔GPU transfers01.01 ms
    Complete pipeline Y→X^Y\to\hat X (measured separately)21,635,593118.31 ms

    End-to-end and per-stage timings were measured separately, with CUDA synchronization before and after each stage for the latter. The sum of unrounded stage timings (approximately 118.73 ms) differs slightly from the end-to-end latency (118.31 ms), by approximately 0.35%, due to additional synchronization overhead and variation between measurement runs.

    The complete system has 21.636 M parameters, a mean latency of 118.31 ms/image, and a P95 latency of 125.20 ms/image, corresponding to approximately 8.45 images/s. One-step restoration takes 30.19 ms; SMG accounts for approximately 70.2% of the end-to-end latency and is the main computational bottleneck.

Dear Reviewer,

Thank you again for your thoughtful review and constructive feedback. We would appreciate your further assessment of our responses and the additional evidence addressing your main concerns:

  • The controlled comparison in Q1 shows that adding supervision on reference and model-predicted intermediate states improves PSNR over endpoint-only residual regression by 0.16 dB on LOL-v1, 0.90 dB on LSRW-Nikon, and 1.12 dB on LSRW-Huawei, while inference requires only one evaluation of the restoration predictor.

  • Our comparison with InDI, I²SB, and ResShift in Q2 clarifies how R2RFlow uses reflectance as the transport source and draws on complementary structural, illumination, and high-frequency cues to guide restoration. The source–target analysis motivates this choice of starting point, and the initialization ablation in Table 3 validates it.

  • In Q3, we clarify that we pretrained Retinex and SMG only on the corresponding benchmark’s training split and froze both during restoration training. Including both modules, the full pipeline averages 118.31 ms per image on an RTX 5090 (400 × 600, batch size 1, FP32).

We would like to emphasize that R2RFlow uses complementary cues from the low-light image and endpoint supervision across multiple paths to learn finite-interval transport from reflectance to the restoration target. R2RFlow is the first method to explore this formulation for one-step low-light image enhancement with paired supervision.

If our responses have sufficiently addressed your concerns, we would also be grateful if you would consider revising your rating accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.