acceptodds
Under review as a conference paper at ICLR 2027

One Object, Different Corrections: Controllable Low-Light Enhancement with Photometric Actions

Abstract

A vision-language model (VLM) brings semantic understanding to low-light image enhancement, but language guidance alone does not specify how much correction to apply. An instruction to brighten an object leaves open how its shadowed and well-lit parts should be adjusted. We introduce an Executable Photometric Action Space (EPAS) that separates image-dependent correction from semantic execution control. A signed log-gain field specifies correction direction and magnitude, while a separate control field determines where and how strongly it is applied, allowing spatially varying corrections within semantic regions. During training, Metropolis–Hastings sampling explores regional actions under an energy defined on the input image's photometric statistics and VLM-derived preferences, with VLM quality assessment guiding candidate selection. This constructs action targets without paired normal-light references; the selected actions are distilled into a deterministic predictor for inference without action search. Applying the controlled gain produces a preview that guides generative restoration for detail recovery and noise suppression. Experiments show that spatially varying gains reduce the shadow undercorrection and highlight overcorrection caused by object-wise uniform gains, especially when correction requirements vary substantially within an object. Outputs approximately follow supplied actions, with more accurate restoration when using the input’s own gain field. A single trained model achieves competitive reconstruction quality, strong perceptual performance, and flexible global and local control across diverse benchmarks and challenging illumination conditions without dataset-specific retraining or fine-tuning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-Light Image Enhancement, Spatially Controllable Image Restoration, Executable Photometric Action Space

Reviews

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

This paper targets controllable low-light image enhancement. The authors propose an EPAS that separates correction from semantic control through a gain field and a control map. The gain field allows spatially varying corrections, while the control map specifies where and how strongly to apply them. The authors train the gain predictor using targets selected through photometric search and VLM assessment. The controlled preview and spatial conditions then guide an adapted image editor. The authors evaluate the same trained model across several low-light datasets and test its global and local enhancement control.

Critical strengths
  1. The authors propose a straightforward representation that separates image-dependent correction from semantic control through a gain field and a control map. This allows the same inferred correction to be applied to different selected areas at different strengths.

  2. The experimental results demonstrate global and local enhancement control with competitive enhancement performance. The same trained model works across several datasets without dataset-specific retraining or fine-tuning, showing its usefulness under different low-light conditions.

Critical weaknesses
  1. The accuracy of the inferred gain field needs more evidence. The gain predictor is trained using targets selected through photometric criteria and VLM assessment. However, the current experiments do not directly show whether the predicted gain provides an appropriate correction before generative restoration. Can the authors evaluate the predicted gain or the full-gain preview against paired normal-light references? This would help establish whether the gain itself provides a suitable correction or whether the editor also compensates for inaccurate predictions.

  2. The roles of the learned membership maps and the choices in the search procedure need more analysis. The authors use five soft membership maps, but do not show what they learn or explain how the number was chosen. Can the authors visualize the individual maps, analyze what they learn, and explain why five maps are used? The nine-transition MH search proposes updates to the first four parameter vectors twice and the fifth only once. Why does the search use nine transitions, and why does it stop before completing the second pass?

Questions for authors
  1. Please see my question in Weakness 1.

  2. Please see my questions in Weakness 2.

  3. Can the authors compare with an image editor adapted using the preview and fixed prompt, but without the additional spatial conditioning? For example, the editor could be adapted using LoRA. This would clarify the contribution of the explicit conditioning beyond adapting the editor to restore the preview.

  4. Can the authors show examples of more than two entities at different enhancement levels? For example, one object at level 1, another at level 3, and the background at level 2. The formulation supports this setting, but the presented local examples only show two entities.

  5. I would suggest that the authors include an explicit section to discuss the limitations of the method, analyze failure modes, and discuss future directions.

3 comments

We sincerely appreciate your insightful comments, which have greatly contributed to improving our work. We address your concerns as follows.


[W1, Q1] Accuracy of the predicted gain field

To evaluate the correction provided by the predicted gain field GG, we added Table 4 to the main paper (see the manuscript for detailed evaluation settings) and Figure 17 to the appendix, comparing each stage against normal-light references. The full-gain preview substantially improves PSNR (7.773 → 20.098 dB) and reduces low-frequency color error ΔE00LF\Delta E_{00}^{\mathrm{LF}} (36.627 → 8.766). Figure 17 shows reduced noise and color blotches after editing, with broadly similar overall tones, consistent with improved LPIPS (0.301 → 0.161) and reduced high-frequency luminance error (0.028 → 0.014). These results support that the preview provides suitable brightness and color correction, while the editor mainly complements it by suppressing noise and artifacts, with minor photometric refinement.

StagePSNR ↑LPIPS ↓SSIM ↑ΔE00LF\Delta E_{00}^{\mathrm{LF}} ↓MAEHF\mathrm{MAE}^{\mathrm{HF}} ↓EσE_{\sigma} ↓
Input7.7730.5600.19136.6270.0420.036
Preview20.0980.3010.7168.7660.0280.078
Final output20.5160.1610.8248.3130.0140.031

Your question drew our attention to the need for more direct evidence of gain prediction accuracy. The new reference-based evaluation has helped us better establish the effectiveness of photometric correction before generative restoration and clarify how the two stages contribute to the final result.

[W2, Q2] Learned memberships and search configuration

(i) Learned memberships. Figure 15 visualizes the five learned membership maps (K=5K=5) in different colors. They capture photometric variation beyond semantic boundaries, distinguishing differently illuminated areas within an object and grouping similarly illuminated areas across objects.

(ii) Search configuration clarification. Regarding the search configuration, our original wording was imprecise: “five transitions” meant five burn-in updates per region (B=5B=5), totaling 25 transitions, while “four consecutive states” referred to the last four retained states (N=4N=4), which join the initial prediction to form five candidates. We also failed to explicitly report the 240 post-burn-in transitions (T=240T=240). We apologize for the resulting confusion and have corrected Appendix B.2 to clearly specify the full MH sampling procedure.

(iii) Trade-off and sensitivity analyses. We added a performance–cost trade-off analysis for K,TK,T and a sensitivity analysis for B,NB,N in main-paper Figure 7 and Table 5, respectively. Figure 7 shows diminishing returns near K=5K=5 and T=240T=240, supporting this balance between correction flexibility and search cost. Table 5 shows that increasing BB beyond five or doubling NN provides no consistent improvement.

[Q3] Contribution of explicit spatial conditioning

We added Figure 18 and Table 11 (reproduced below) in Appendix F.1, comparing full EPAS with the same base editor adapted using LoRA with the preview and fixed restoration prompt, without additional spatial conditioning. Full EPAS improves all reported metrics, reducing LPIPS from 0.216 to 0.161 on LOLv1 and from 0.391 to 0.303 on LSRW. Figure 18 shows hallucinated details and color distortion in the preview-only LoRA outputs. These results support the value of explicit spatial conditioning in improving restoration quality, suppressing hallucinated details, and reducing color distortion.

No.MethodLOLv1LSRWUnpaired
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓NIQE ↓MUSIQ ↑CLIP-IQA ↑
1Preview-only LoRA18.9510.6610.21616.0860.5320.3915.55656.1620.540
2Full EPAS (ours)20.5160.8240.16118.7060.5400.3033.07167.4270.554

[Q4] Multi-region control at different enhancement levels

Thank you for the suggestion. We added Figure 14 to the appendix to demonstrate simultaneous brightness control over three semantic regions at different enhancement levels in indoor and outdoor scenes.

[Q5] Limitations and future directions

We appreciate this recommendation. We added Appendix H to discuss grounding and segmentation errors under severe darkness, local detail hallucinations, and future directions for improving region identification and restoration fidelity.

Thank you again for your thoughtful and constructive suggestions. These additions have improved both the clarity of our methodology and the completeness of our experiments. Your feedback has been genuinely helpful in bringing the paper to a more mature form, and we hope the revised version merits your continued support.

Reviewer_WhZuOpenReview ↗

Follow-up

First, thank you for the clarifications and additional experiments addressing my questions. These additions strengthen my recommendation for weak acceptance.

I have three follow-up suggestions regarding Q2 and Q4:

  1. Q2: For the membership maps in Fig. 15, can the authors visualize the individual soft masks instead of the unified overlaid map?

  2. Q2: Following the clarification of the search strategy, I am wondering whether more exploration of the search space would provide further benefits, since the current strategy outperforms greedy search. Could the authors consider: (1) choosing four candidates at post-burn-in transitions 60, 120, 180, and 240 instead of 237, 238, 239, and 240; and (2) always sampling around the initial prediction instead of the previously accepted state?

  3. Q4: For multi-region control, can the authors report the distribution of region counts in training and testing?

AuthorsOpenReview ↗

Follow-up on Membership Visualization, Search Strategy, and Region-Count Statistics

Thank you for your positive assessment and for acknowledging the clarifications and additional experiments. We are pleased that these additions have addressed many of your concerns and strengthened your recommendation. We also appreciate your further suggestions, which help us better examine the search strategy and clarify the evidence for regional control. We address your follow-up comments below.

[Q2-1] Visualization of individual soft membership masks

Thank you for suggesting this clearer visualization. We have updated Figure 15 in the revised manuscript to include the five individual soft membership masks for each example, shown in the second and fourth rows. All masks use a shared color scale from 0 to 1, with dark purple indicating low membership weights and yellow indicating high weights. This presentation makes the spatial contributions and overlap of the individual memberships easier to inspect. We have also revised the accompanying discussion and figure caption accordingly.

[Q2-2] Further exploration of the search space

Thank you for these suggestions. Figure 7 shows diminishing returns from increasing KK or TT around the adopted setting under the original search strategy. Your suggestions examine complementary aspects of exploration: candidate retention and proposal centering.

To investigate these aspects within the rebuttal time budget, we trained the first-stage predictor on the same 50% subset of the DIV2K-derived training data under matched training settings, while keeping the second-stage editor fixed. The table below reports final-output performance on the LOLv1 and LSRW test sets.

(i) Spaced retention. Retaining states at transitions 60, 120, 180, and 240 yields performance close to the original strategy, with no consistent advantage across metrics. In the spaced-retention pool, the state at transition 240 was selected for 74% of the evaluated images. This suggests that selection often favors the final state despite access to earlier candidates, helping explain why broader temporal coverage provides limited additional benefit.

(ii) Initial-centered proposals. This variant performs worse on both datasets. Proposals centered at the current chain state allow accepted updates to accumulate, enabling successive moves away from the initialization. Repeatedly centering proposals at the initial prediction restricts this cumulative exploration and may make search quality more sensitive to the initial prediction, particularly when suitable corrections lie beyond its immediate neighborhood. The results are consistent with a benefit from progressively refining the initial prediction through accumulated updates.

Search variantProposal centerRetained post-burn-in transitionsLOLv1 PSNR ↑LOLv1 LPIPS ↓LSRW PSNR ↑LSRW LPIPS ↓
Original strategyCurrent chain state237, 238, 239, 24019.7830.18418.1470.326
Spaced retentionCurrent chain state60, 120, 180, 24019.7460.18218.1640.329
Initial-centered proposalsInitial prediction237, 238, 239, 24019.2180.20717.5920.354

[Q4] Distribution of semantic region counts

Thank you for requesting these statistics. We summarize below the semantic region counts in the Light100K training and held-out testing data used for the second-stage regional-control experiments. Each entry reports the number of images and its percentage within the corresponding split.

Data splitTotal samples1 region2 regions3 regions≥4 regions
Light100K training2756344 (12.48%)226 (8.20%)542 (19.67%)1644 (59.65%)
Light100K held-out testing70090 (12.86%)31 (4.43%)85 (12.14%)494 (70.57%)

Thank you again for your thoughtful suggestions and continued discussion. They have been valuable in identifying where additional evidence and clearer explanations would strengthen the paper.

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

This paper proposes EPAS for controllable low-light enhancement. It factorizes the photometric action into an image-dependent signed log-gain field and a semantic control field that specifies where and how strongly to apply it. Gain targets are built at training time by Metropolis–Hastings search over regional photometric parameters, under an energy combining brightness and TV priors with VLM-derived regional preferences, followed by VLM quality reranking. These targets are distilled into a direct predictor. The resulting preview and conditions guide Qwen-Image-Edit through a trained ControlNet. The method is evaluated on standard LLIE benchmarks, challenging-illumination sets and global/local brightness control, together with mechanistic analyses of within-object gain variation and action execution.

Critical strengths

The decomposition of correction from execution control is clean, interpretable, and supports both global and mask-free local control with a single model.

The mechanistic analyses are well designed: the object-uniform intervention and the gain-shuffling control separate action execution from action suitability.

Critical weaknesses

Reconstruction fidelity is modest, and gains are hard to attribute to EPAS rather than the 40B-parameter foundation pipeline. PSNR/SSIM trail paired methods substantially. Cost is very high. Removing direct gain conditioning changes LOLv2-Real PSNR by only 0.05 dB.

The training-time “MH exploration” is very short: 5 burn-in plus 4 retained transitions across 5 regions, i.e., fewer than two moves per region. Targets depend on a fixed brightness prior and uncalibrated VLM preferences. The gain over greedy search is small.

Control accuracy is under-evaluated. Local control is measured only with no-reference IQA (MUSIQ/NIQE), with no in-region brightness accuracy metric. Full EPAS has worse exterior preservation than the variant without action inputs (0.067). Execution is attenuated. The object-uniform baseline is a weak straw man for the within-object claim.

Questions for authors

How much of the improvement remains if the same ControlNet-adapted editor is conditioned on a strong conventional enhancer’s output instead of the EPAS preview?

Please report in-region brightness accuracy for local control and explain why exterior preservation degrades with action inputs.

How sensitive are the results to MH chain length, b0b_0, and VLM preference noise? Please add variance across seeds or VLM queries.

How often does the generative editor hallucinate content on paired benchmarks? Please add a fidelity-oriented failure analysis.

6 comments

We sincerely thank the reviewer for their valuable feedback and for giving us the opportunity to further explain and improve our work. We address each of your comments in detail below, and we hope that our responses resolve any questions or concerns you may have.


[W1] Contribution of EPAS, reconstruction fidelity, and computational cost

(i) Contribution before generative restoration. We added Table 4 (reproduced below) to the main paper and Figure 17 to the appendix to evaluate the correction achieved before generative restoration. The gain-based preview improves PSNR from 7.773 to 20.098 dB and reduces low-frequency color error from 36.627 to 8.766. The editor then primarily improves detail quality and suppresses noise and artifacts. These stagewise results establish a contribution from EPAS correction before the generative editor is applied.

StagePSNR ↑LPIPS ↓SSIM ↑ΔE00LF\Delta E_{00}^{\mathrm{LF}} ↓MAEHF\mathrm{MAE}^{\mathrm{HF}} ↓EσE_{\sigma} ↓
Input7.7730.5600.19136.6270.0420.036
Preview20.0980.3010.7168.7660.0280.078
Final output20.5160.1610.8248.3130.0140.031

Your concern about the source of performance improvements was particularly helpful. It prompted us to evaluate photometric correction independently of generative restoration, allowing us to better establish what EPAS contributes on its own and clarify the respective roles of the correction and editing stages.

(ii) Training setup and reconstruction fidelity. EPAS uses self-supervised training without the evaluation benchmarks’ training splits. Paired methods generally achieve higher PSNR and SSIM with target-domain supervision, while EPAS delivers strong perceptual quality and cross-dataset generalization using one model without dataset-specific retraining or fine-tuning. As discussed in our response to Q1, comparisons under matched editor training further support the contribution of EPAS corrections to performance on unseen datasets.

(iii) Direct gain conditioning. The 0.05 dB improvement concerns only adding GG to the editor’s spatial conditioning. That ablation retains the preview already corrected by GG, so it does not measure the overall contribution of gain-based correction. Including GG in spatial conditioning provides information about the applied amplification, which can help distinguish originally bright areas from amplified dark areas with different restoration needs. This additional input improves local MUSIQ from 74.608 to 75.430 and reduces local brightness error from 0.043 to 0.032.

(iv) Computational cost and deployment flexibility. EPAS decouples image-dependent photometric correction from spatial execution control. As shown above, its core correction is established before generative editing and distilled into a compact predictor, avoiding search at inference. The larger models support semantic region selection and further restoration of the corrected preview, enabling enhancement of specified regions at requested brightness levels. A single trained EPAS model serves all evaluated datasets without using their training splits or requiring dataset-specific retraining or fine-tuning.

EPAS also allows component adaptation to suit available resources and quality requirements. Table 14 in Appendix G (reproduced below) shows that the configuration without SAM retains global enhancement while reducing latency from 45.7 to 38.5 s and peak memory from 69.47 to 54.42 GiB. Smaller SAM variants offer limited savings, whereas replacing the editor with FLUX.2-klein-base-4B reduces these costs to 16.8 s and 53.03 GiB, with LPIPS increasing from 0.303 to 0.326 and region-exterior MAE from 0.073 to 0.085. These results demonstrate deployment flexibility with explicit quality–resource trade-offs.

ComponentModelPSNR ↑LPIPS ↓MAEout\mathrm{MAE}_{\mathrm{out}} ↓Time (s) ↓Memory (GiB) ↓
EPASFull18.7060.3030.07345.769.47
EPASwithout SAM18.7060.303N/A38.554.42
Qwen2.5-VL-7B-Instruct18.6050.3080.07640.157.10
VLMQwen2.5-VL-3B-Instruct18.4210.3180.08139.654.79
Qwen2-VL-2B-Instruct18.3280.3250.08739.554.26
SAM 2.1 Hiera-Small18.7060.3030.07445.469.40
SAMSAM 2.1 Hiera-Tiny18.7060.3030.07545.469.39
MobileSAM18.7060.3030.08045.369.33
FLUX.1-Kontext-dev (12B)18.5270.3120.07622.561.37
EditorQwen-Image-2.1 (7B)18.4300.3180.08019.156.12
FLUX.2-klein-base-4B18.2970.3260.08516.853.03

[W2, Q3] MH search configuration and sensitivity

(i) Search configuration clarification. Our original description of the MH procedure was incomplete. “Five transitions” referred to five burn-in updates per region (B=5B=5), totaling 25 transitions across five regions. “Four consecutive states” referred to the last four retained states (N=4N=4), which form five candidates together with the initial prediction. We omitted the 240 post-burn-in transitions (T=240T=240) from that description. We apologize for the confusion and have corrected Appendix B.2 to specify the full procedure.

(ii) Sensitivity and stability. We added Figure 7 and Table 5, reporting mean and standard deviation across three training runs per setting. Returns diminish near T=240T=240, and increasing BB beyond five offers no consistent benefit. The tested brightness priors b0∈{0.45,0.55,0.65}b_0\in\{0.45,0.55,0.65\} yield PSNR between 18.50 and 18.69 dB. Preference noise has little effect at σβ=0.05\sigma_\beta=0.05 but degrades performance at 0.10, indicating tolerance to moderate preference errors.

(iii) Comparison with greedy search. Under matched candidate-evaluation budgets and reranking, MH improves PSNR over greedy search by 0.433 dB and LED MUSIQ by 1.642. The newly added local evaluation in Table 6 further shows a 38.5% reduction in brightness error, supporting improved regional control without test-time search.

No.VariantLOLv2‑RealLEDCOCOLight100K
PSNR ↑LPIPS ↓MUSIQ ↑MUSIQ ↑MUSIQ ↑MAEout\mathrm{MAE}_{\mathrm{out}} ↓B-MAEin\mathrm{B\text{-}MAE}_{\mathrm{in}} ↓MAEout\mathrm{MAE}_{\mathrm{out}} ↓
1Greedy search instead of MH20.9560.18168.95059.24274.1940.0770.0520.074
2Full EPAS21.3890.17569.98260.88475.4300.0730.0320.069

[Q1] Replacing the EPAS preview with strong enhancers

We added this comparison in Appendix F using outputs from two strong enhancers, GSAD and GPP-LLIE, with fixed LOLv1-trained checkpoints. Each preview source receives separate adaptation of the same editor architecture under matched training settings. Table 12 (reproduced below) shows that these alternatives perform better on LOLv1, but their advantage does not transfer to the other three benchmarks, suggesting dependence on the training-data distribution. EPAS achieves better restoration on these unseen datasets and lower local brightness error, supporting stronger cross-dataset generalization without target-benchmark training data or dataset-specific retraining.

No.MethodLOLv1LOLv2-RealLOLv2-SynLSRWLight100K (Local)
PSNR ↑ / LPIPS ↓PSNR ↑ / LPIPS ↓PSNR ↑ / LPIPS ↓PSNR ↑ / LPIPS ↓B-MAEin\mathrm{B\text{-}MAE}_{\mathrm{in}} ↓
1GSAD preview25.683 / 0.10619.327 / 0.23117.864 / 0.28617.263 / 0.3540.046
2GPP-LLIE preview23.417 / 0.09319.786 / 0.21618.293 / 0.27317.138 / 0.3420.044
3EPAS (ours)20.516 / 0.16121.389 / 0.17521.232 / 0.21518.706 / 0.3030.032

[Q4] Content fidelity and hallucination analysis

We added a fidelity-focused failure analysis in Appendix H. Across LOLv1, LOLv2-Real, LOLv2-Syn, and LSRW, hallucinations were observed in fewer than 3% of evaluated images, with dataset-wise counts summarized in the table below. Figure 19(a) illustrates a slight alteration to a book’s printed pattern, visible upon substantial magnification despite favorable PSNR and LPIPS. We also added Appendix H.1 to quantify grounding and segmentation failure rates under severe darkness and discussed future directions for improving region identification and detail preservation.

DatasetTest imagesAddition/deletionStructure/texture alterationsColor deviations/noise
LOLv115010
LOLv2-Real100201
LOLv2-Syn100101
LSRW50100
Total265412

[W3, Q2] Control accuracy and within-object correction

(i) Local brightness accuracy and control behavior. We expanded Appendix E.3 and Table 3 to evaluate local brightness adjustment beyond perceptual quality. EPAS achieves a target-region mean luminance increase (�0�)(�0�) of 0.147 and a level-order success rate (LOSR) of 90.7%, demonstrating effective brightening with consistent ordering across control levels. We also added reference-based local brightness accuracy to Table 6, where EPAS achieves the lowest target-region brightness error of 0.032. Figure 16 in Appendix E.4 further reports tracking errors and control leakage at individual levels.

No.MethodGlobalLocal
B-MAE ↓MUSIQ ↑NIQE ↓MUSIQ ↑NIQE ↓ΔYin\Delta Y_{\mathrm{in}}LOSR (%) ↑
1GSAD+Mask$^{\mathrm{M}}$N/A69.7103.82671.2713.5070.096N/A
2CLE Diffusion$^{\mathrm{M}}$0.13345.3424.57264.2583.5110.11882.0
3ControlLight0.18368.4113.914N/AN/AN/AN/A
4TSECN0.17767.1834.063N/AN/AN/AN/A
5Ours$^{\mathrm{T}}$0.10174.0343.65875.4303.5030.14790.7

(ii) Exterior-preservation trade-off. Regarding exterior preservation, edge-guided control smoothing softens mask transitions and can introduce small changes near region boundaries. We added local brightness accuracy and exterior-preservation metrics to Table 6 in the main paper. The results show that the lower exterior error without action-derived inputs comes with poorer regional accuracy and perceptual quality. On Light100K, full EPAS more than halves target-region brightness error from 0.068 to 0.032, while exterior error increases by 0.005. This shows a substantial improvement in the requested correction with a modest exterior-preservation trade-off.

No.VariantB-MAEin\mathrm{B\text{-}MAE}_{\mathrm{in}} ↓MAEout\mathrm{MAE}_{\mathrm{out}} ↓
1Qwen-Image-Edit-2511 only0.1120.094
2Without action-derived inputs0.0680.064
3Without direct gain conditioning0.0430.078
4Greedy search instead of MH0.0520.074
5Without regional preferences0.0450.071
6Without VLM reranking0.0490.072
7Full EPAS0.0320.069

Your emphasis on local control accuracy has led us to go beyond perceptual quality metrics and examine regional errors, exterior preservation, and stronger within-object comparisons. These evaluations have given us a more complete understanding of both the benefits of spatially varying correction and the trade-offs involved in precise local control.

(iii) Within-object correction beyond uniform gains. The object-uniform baseline isolates the effect of removing within-object gain variation. We additionally compare against luminance-binned gain, which averages EPAS gains within five input-luminance bins per object, and LASQ, an adaptive enhancement method. Both permit different corrections within an object. EPAS reduces object error from 0.052 and 0.046, respectively, to 0.027, with lower shadow and highlight errors as well. These comparisons support the value of its spatial gain allocation beyond the uniform baseline.

No.MethodObject MAE ↓Shadow MAE ↓Highlight MAE ↓
1Object-uniform0.1080.1470.129
2Luminance-binned gain0.0520.0710.043
3LASQ0.0460.0630.037
4Full EPAS0.0270.0340.025

We are especially grateful for the rigor of your review. Your questions encouraged us to look more critically at our own method, from the reliability of action learning to the accuracy of local control and the risks of generative restoration. Addressing these concerns has not only strengthened our experimental evaluation but also helped us present the capabilities and limitations of EPAS more precisely. We truly appreciate the time and attention you have devoted to our work, and we hope these new findings will help resolve your reservations and encourage you to reconsider your rating.

Reviewer_TTBtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces EPAS, a framework for controllable low-light image enhancement. EPAS decouples an image-dependent signed log-gain field from a spatial execution-control field and combines them to construct a photometric preview. The framework trains a deterministic action predictor using targets obtained through training-time Metropolis–Hastings sampling and VLM-based assessment, then uses the preview and spatial conditions to guide an adapted generative editor.

Critical strengths
  • Interpretable formulation. Separating image-dependent correction from where and how strongly it is applied provides a clear control mechanism.

  • Well-controlled comparisons. Baselines with matched training data and budgets help isolate the contribution of action-derived inputs beyond editor adaptation.

Critical weaknesses
  • Incomplete positioning against related work. Discussion and comparison with GPP-LLIE (AAAI25, which also uses quantified global and local VLM assessments to guide generative enhancement) would clarify the added value of executable photometric corrections and independent execution control.

  • The response slope of approximately 0.765 suggests attenuated action execution (Table 7). Region-wise target errors, control leakage, and monotonicity across control levels need further evaluation.

  • Direct gain conditioning adds only 0.05 dB PSNR, alongside local MUSIQ improvements. Separate ablations of VLM regional preferences and reranking would strengthen the evidence.

  • High computational cost. Approximately 69.5 GiB peak memory and 45.7 seconds per image limit interactive use (Appendix F). Component-wise profiling and comparisons with simpler variants would clarify the quality-cost trade-off.

Questions for authors
  • How sensitive is action learning to the global brightness prior, VLM prompt, and MH budget?
  • How often do grounding or segmentation failures impair local control in severe darkness?
  • Can a simpler spatial gain predictor achieve comparable control and restoration under matched editor training?
6 comments

We sincerely thank the reviewer for identifying several important concerns and for the opportunity to further clarify our method. We address each point in detail below.


[W1] Positioning against GPP-LLIE

Thank you for highlighting this related work. We added GPP-LLIE to Table 1 and expanded the comparison in Related Work. GPP-LLIE conditions supervised diffusion enhancement on global and local VLM assessments. EPAS instead distills search guided by photometric criteria and VLM preferences into executable photometric corrections. A spatially varying gain field GG specifies correction direction and magnitude, accommodating different correction needs within the same object. Independent execution control allows users to enhance selected regions at specified levels.

Although GPP-LLIE achieves strong performance with paired training, EPAS generalizes across datasets without using their training splits. Table 12 (reproduced below) compares EPAS with fixed LOLv1-trained GSAD and GPP-LLIE previews under matched editor training. Their final outputs perform better on LOLv1, while EPAS achieves better results on all three other benchmarks and lower local brightness error, supporting stronger cross-dataset generalization in this setting without retraining or fine-tuning.

No.MethodLOLv1LOLv2-RealLOLv2-SynLSRWLight100K (Local)
PSNR ↑ / LPIPS ↓PSNR ↑ / LPIPS ↓PSNR ↑ / LPIPS ↓PSNR ↑ / LPIPS ↓B-MAEin\mathrm{B\text{-}MAE}_{\mathrm{in}} ↓
1GSAD preview25.683 / 0.10619.327 / 0.23117.864 / 0.28617.263 / 0.3540.046
2GPP-LLIE preview23.417 / 0.09319.786 / 0.21618.293 / 0.27317.138 / 0.3420.044
3EPAS (ours)20.516 / 0.16121.389 / 0.17521.232 / 0.21518.706 / 0.3030.032

Your suggestion has helped us better position EPAS against existing VLM-guided enhancement methods. In particular, the additional comparison clarifies the value of executable photometric correction and independent execution control, making the distinction between EPAS and related work more explicit.

[W2] Local control accuracy and monotonicity

We added Figure 16 in Appendix E.4 to evaluate regional target errors, control leakage, and brightness monotonicity on 150 COCO images, varying each target across four enhancement levels. Errors relative to the preview and negative bias increase at stronger levels, consistent with the attenuated execution indicated by the 0.765 response slope. Changes outside the target region remain small but also increase with control strength. Adjacent-level brightness ordering holds for 94.7% to 97.3% of targets, and 90.7% preserve the complete ordering across all four levels. Together, the limited control leakage and high brightness-ordering pass rates demonstrate that EPAS effectively adjusts selected regions across enhancement levels while largely preserving their surroundings.

[W3] Direct gain conditioning and VLM ablations

(i) Direct gain conditioning. We would like to clarify that “Without direct gain conditioning” removes GG only from the editor’s spatial conditioning, while retaining the preview already corrected by GG. The 0.05 dB gain therefore reflects the benefit of this additional input, rather than the overall contribution of photometric correction. Table 4 and Figure 17 show that applying GG substantially corrects brightness and color before editing, improving PSNR from 7.773 to 20.098 dB. Adding GG to the editor’s spatial conditioning also indicates how much each location has been brightened, helping distinguish originally bright areas from amplified dark areas that may require more noise and artifact removal. This additional cue improves local MUSIQ from 74.608 to 75.430 and reduces local brightness error from 0.043 to 0.032.

(ii) Regional preferences and VLM reranking. We added separate ablations of regional preferences and VLM reranking in Table 6 (reproduced below), retraining the predictor with each component removed while keeping the adapted editor fixed. Regional preferences improve LED MUSIQ from 56.726 to 60.884. Reranking improves LPIPS from 0.214 to 0.175. Both components also improve local perceptual quality and brightness accuracy, supporting their complementary roles in guiding search and selecting candidates.

No.VariantLOLv2‑RealLEDCOCOLight100K
PSNR ↑LPIPS ↓MUSIQ ↑MUSIQ ↑MUSIQ ↑MAEout\mathrm{MAE}_{\mathrm{out}} ↓B-MAEin\mathrm{B\text{-}MAE}_{\mathrm{in}} ↓MAEout\mathrm{MAE}_{\mathrm{out}} ↓
1Without regional preferences21.1170.19168.86456.72672.8170.0730.0450.071
2Without VLM reranking21.4010.21466.58257.94371.6380.0740.0490.072
3Full EPAS21.3890.17569.98260.88475.4300.0730.0320.069

[W4 Computational cost and deployment flexibility]

We added Table 14 (reproduced below) in Appendix G to quantify the quality and efficiency trade-offs of component substitutions. Enhancing selected regions at specified levels requires semantic grounding and segmentation, adding inference cost beyond global enhancement. Removing SAM retains global enhancement while reducing latency from 45.7 to 38.5 s and peak memory from 69.47 to 54.42 GiB. Smaller SAM variants offer limited savings, whereas replacing the editor with FLUX.2-klein-base-4B reduces latency to 16.8 s and memory to 53.03 GiB, with higher perceptual and region-exterior errors.

EPAS provides a restoration framework that separates photometric correction from execution control, without requiring a fixed combination of models. Search-derived corrections are distilled into a compact predictor, avoiding MH search and candidate quality assessment at inference. Components can be substituted with appropriate adaptation to suit available computational resources and quality requirements. Moreover, a single trained EPAS model achieves strong perceptual quality across datasets without using their training splits or requiring dataset-specific retraining or fine-tuning.

ComponentModelPSNR ↑LPIPS ↓MAEout\mathrm{MAE}_{\mathrm{out}} ↓Time (s) ↓Memory (GiB) ↓
EPASFull18.7060.3030.07345.769.47
EPASwithout SAM18.7060.303N/A38.554.42
Qwen2.5-VL-7B-Instruct18.6050.3080.07640.157.10
VLMQwen2.5-VL-3B-Instruct18.4210.3180.08139.654.79
Qwen2-VL-2B-Instruct18.3280.3250.08739.554.26
SAM 2.1 Hiera-Small18.7060.3030.07445.469.40
SAMSAM 2.1 Hiera-Tiny18.7060.3030.07545.469.39
MobileSAM18.7060.3030.08045.369.33
FLUX.1-Kontext-dev (12B)18.5270.3120.07622.561.37
EditorQwen-Image-2.1 (7B)18.4300.3180.08019.156.12
FLUX.2-klein-base-4B18.2970.3260.08516.853.03

Your suggestion to examine the computational cost of individual components has helped us better understand the practical trade-offs of our framework. The additional experiments provide a clearer picture of how different model choices affect efficiency and restoration quality, making our discussion of deployment flexibility more concrete and balanced.

[Q1] Sensitivity of action learning

We added Figure 7 and Table 5 to the main paper to evaluate these factors on LSRW, reporting mean and standard deviation over three training runs. For the global brightness prior b0b_0, the tested values of 0.45, 0.55, and 0.65 yield PSNR between 18.50 and 18.69 dB, with the default 0.55 balancing reconstruction and perceptual quality. Paraphrasing either VLM prompt changes PSNR by at most 0.11 dB from the default, indicating limited sensitivity to wording within the tested variants. For the MH budget, Figure 7 shows diminishing returns near 240 post-burn-in transitions (T=240T=240). Increasing burn-in sweeps (BB) beyond five offers no consistent benefit, while doubling retained states (NN) brings little improvement despite additional candidate assessments. These results support the chosen settings as a balance between quality and training-time search cost.

[Q2] Grounding and segmentation under severe darkness

To reduce grounding and segmentation failures, EPAS performs these steps on the preliminary photometric rendering (Iph(μ))(I_{\mathrm{ph}}(\mu)), where dark content is more visible. We added experimentss (see the manuscript for detailed settings). in Appendix H using 604 LIS test images and 106 ACDC-night validation images. Table 15 (reproduced below) shows that photometric correction reduces mask failure rates from 34.11% to 12.75% on LIS and from 44.34% to 18.87% on ACDC-night, while improving grounding accuracy and mask overlap. Performance on LIS approaches that obtained with normal-light references, supporting more reliable region identification before local control. These results indicate that preliminary photometric correction substantially reduces the risk of grounding and segmentation errors impairing local control under severe darkness.

No.DatasetInput imageGrounding [email protected] ↑Mean target IoU ↑Mask [email protected] ↓
1LISInput72.68%0.61134.11%
2Photometric rendering92.72%0.77612.75%
3Normal-light reference94.04%0.79810.76%
4ACDC-nightInput66.04%0.54044.34%
5Photometric rendering89.62%0.75618.87%

[Q3] Comparison with simpler gain predictors

We added comparisons in Appendix F using three simpler CNN predictors for global gamma, regional gamma, and dense gain, with matched editor architecture, initialization, training targets, and optimization budget. Table 12 (excerpt below) shows that all three fall short of EPAS across the four restoration benchmarks and local brightness accuracy. Even the strongest variant, dense gain prediction, achieves 20.648 dB PSNR on LOLv2-Real versus 21.389 dB for EPAS, with higher local brightness error (0.038 versus 0.032). These results support the value of our structured photometric parameterization for both restoration and regional control.

No.MethodLOLv1LOLv2-RealLOLv2-SynLSRWLight100K (Local)
PSNR ↑ / LPIPS ↓PSNR ↑ / LPIPS ↓PSNR ↑ / LPIPS ↓PSNR ↑ / LPIPS ↓B-MAEin\mathrm{B\text{-}MAE}_{\mathrm{in}} ↓
1Global gamma CNN18.763 / 0.22619.274 / 0.24918.416 / 0.29417.086 / 0.3690.049
2Regional gamma CNN19.236 / 0.20320.063 / 0.21819.287 / 0.26717.718 / 0.3460.042
3Dense gain CNN19.781 / 0.18820.648 / 0.19720.154 / 0.24118.137 / 0.3290.038
4EPAS (ours)20.516 / 0.16121.389 / 0.17521.232 / 0.21518.706 / 0.3030.032

We sincerely appreciate the careful thought behind your comments. Your questions have helped us clarify the novelty of EPAS, better explain the contributions of its components, and provide a more balanced discussion of controllability and computational cost. We believe these additions have made the paper clearer and more convincing, and we are grateful for your role in guiding these improvements. We hope our new results address your concerns and would sincerely appreciate it if you could reconsider your assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Reviewer Discussions and Key Rebuttal Evidence

Summary of Reviewer Discussions and Rebuttal Evidence

We thank all reviewers and the Area Chair for their feedback. All three initial reviews positively recognized EPAS's core correction-control separation: TTBt highlighted its "clear control mechanism"; rVcd described it as "clean, interpretable"; and WhZu recognized that it allows the same inferred correction to be applied to different selected areas at different strengths. Our rebuttal provides further evidence for these contributions, summarized below.

Reviewer WhZu

Main concerns: Gain accuracy before restoration, membership maps and search configuration, spatial conditioning, multi-region control, and limitations.

Our response:

  • Added stagewise evaluation: the photometric preview improves LOLv1 PSNR from 7.773 to 20.098 dB before generative restoration (Table 4; Figure 17).
  • Visualized learned memberships, clarified the full MH procedure, and evaluated region counts and search budgets (Figures 7 and 15; Table 5; Appendix B.2).
  • Compared full EPAS with preview-only LoRA adaptation of the same editor. Full EPAS improves all reported metrics, including LOLv1 LPIPS from 0.216 to 0.161 (Table 11).
  • Added three-region control examples and failure analysis (Figure 14; Appendix H).

Reviewer's follow-up:

"These additions strengthen my recommendation for weak acceptance."

We also provided individual soft masks, alternative search comparisons, and training/testing semantic-region counts. Under matched training on a 50% data subset, spaced candidate retention showed no consistent advantage, while proposals centered at the initialization performed worse on both evaluated datasets.

Reviewer TTBt

Main concerns: GPP-LLIE positioning, local control accuracy, gain conditioning and VLM contributions, computational cost, search sensitivity, grounding reliability, and simpler predictors.

Our response:

  • Added GPP-LLIE comparisons and matched-editor experiments with GSAD/GPP-LLIE previews. EPAS performs better on the three benchmarks outside these baselines' LOLv1 training domain and achieves lower local brightness error (Table 12).
  • Evaluated tracking errors, leakage, and monotonicity on 150 COCO images; 90.7% of targets preserve the full four-level brightness ordering (Figure 16).
  • Clarified that removing direct gain conditioning retains the gain-corrected preview; the 0.05 dB difference therefore does not measure the full correction contribution. Added separate regional-preference and VLM-reranking ablations (Tables 4 and 6).
  • Added sensitivity analyses across three training runs, component quality-cost trade-offs, and grounding/segmentation evaluation under severe darkness (Table 5; Appendices G and H).
  • Compared global-gamma, regional-gamma, and dense-gain CNN predictors under matched editor training. EPAS outperforms them across all four restoration benchmarks and local brightness accuracy (Table 12).

Reviewer rVcd

Original assessment: Beyond recognizing the correction-control separation, the reviewer stated:

"The mechanistic analyses are well designed: the object-uniform intervention and the gain-shuffling control separate action execution from action suitability."

Main concerns and our responses:

  • Contribution and cost [W1, Q1]: Stagewise evaluation, matched-editor comparisons with alternative enhancement inputs, and component-level resource measurements.
  • Search reliability [W2, Q3]: Full search-procedure clarification, sensitivity analysis, and matched-budget greedy comparisons; local brightness error decreases from 0.052 with greedy search to 0.032 with MH.
  • Local and within-object accuracy [W3, Q2]: Regional brightness, ordering, and exterior-preservation measurements. Stronger baselines yield object MAE of 0.052 for luminance-binned gain and 0.046 for LASQ, versus 0.027 for EPAS.
  • Fidelity [Q4]: Categorized failure statistics and analysis of content deviations.

Follow-up discussion: The reviewer maintained the rating and raised related-work concerns. We clarified that ControlLight is already cited and compared (Section 2.1; Table 3; Figure 4), and provided mechanism-level distinctions from ControlLight, ReCoRo, InstructRestore, and VLM-IMI. These concern EPAS's prompt-specified regional and level control without user-provided masks, explicit photometric actions, and separate assessment of execution and correction suitability. We committed to expanding this discussion.

The additional evidence supports the core formulation recognized in the initial reviews: image-dependent correction and semantic execution control are distinct, experimentally testable components. Thank you for considering the revised manuscript and discussion evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.