acceptodds
Withdrawn from ICLR 2027

ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing deep learning-based low-light enhancement methods are typically trained on limited datasets with single enhancement targets, which restricts their generalization ability and controllability in real-world applications. To overcome these limitations, we propose ControlLight, a controllable, consistent, and generalizable framework for low-light enhancement. We first construct a large-scale dataset of real-world degraded images with continuous illumination-strength supervision. To further ensure consistent outputs under different control strengths, we introduce a misalignment-aware weighted flow matching loss that preserves image structure across continuous enhancement strengths. ControlLight allows users to edit real-world degraded low-light images toward satisfactory enhancement results by flexibly controlling the strength while preserving visual consistency and realism. Extensive experiments show that ControlLight achieves state-of-the-art performance against existing low-light enhancement approaches while demonstrating strong continuous controllability and generalization to real-world scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-Light Enhancement

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_N8ujOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ControlLight, a controllable, consistent, and generalizable framework for low-light enhancement. The authors build Light100K with continuous-strength pseudo supervision, design a misalignment-aware weighted flow matching loss to suppress subtle structural misalignment in diffusion-generated targets, and fine-tune FLUX.2-klein-9B with LoRA so that the enhancement strength becomes a continuously adjustable scaling factor.

Critical strengths

The misalignment-aware loss addresses a real problem that prior work left unhandled, since diffusion-generated pseudo-targets carry subtle edge misalignment that becomes visible only after illumination normalization, and the ablations show consistent gains across LI-LPIPS, PSNR, SSIM, and LPIPS with visual confirmation of reduced structural drift.

Critical weaknesses
  1. The main novelty is a combination of existing components, since LoRA-based controllable editing, diffusion-generated pseudo supervision, and Retinex-inspired decomposition have all been explored separately in prior work, and the paper does not demonstrate why this particular combination yields insights beyond the sum of its parts.
  2. The construction of continuous supervision via interpolation between two endpoints has a direct precedent in CLE Diffusion, which also uses intermediate brightness blending targets, and the Retinex-based interpolation is an incremental refinement of that idea rather than a fundamentally new mechanism.
  3. The evaluation relies predominantly on non-reference metrics with known bias toward brighter and sharper outputs, and the model lags compared methods by roughly 10 dB in PSNR on the only two paired benchmarks.
Questions for authors
  1. Can you demonstrate what your combination achieves beyond applying Retinex interpolation directly inside the CLE Diffusion pipeline? 2. Is the gain over Concept Sliders attributable to the per-strength supervision or to the stronger base model and larger data?
Reviewer_wkrzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work addresses continuous low-light image enhancement. Its main contribution is Light100K, a new benchmark. Low-light inputs are fed into FLUX.2-klein-9B with a fixed prompt to recover normal brightness while preserving scene structure, forming paired low/enhanced samples. Pairs with misaligned structures are filtered out, leaving 20K high-quality pairs. Retinex-based interpolation generates four intermediate light-level images to build the benchmark. Enhancement model is trained on Light100K with an improved flow-matching loss-- Lower weights are assigned near edges to tolerate minor structure misalignment between image pairs. LoRA is adopted in training, where the LoRA scale s is the interpolation factor for light levels. Experiments show decent visual quality and smooth transitions across continuous light levels.

Critical strengths
  1. Continuous low-light image enhancement enjoys great practical value.
  2. Comparative experiments are conducted against existing low-light enhancement and image editing methods.
  3. The method achieves favorable performance on various image quality metrics and yields smooth transitions across continuously enhanced images.
Critical weaknesses
  1. The technical novelty is limited. For instance, the pipeline relies on a prompted FLUX model to produce enhanced images and adopts interpolation between low-light/bright image pairs to synthesize samples at intermediate light levels.
  2. The enhanced outputs show noticeable inconsistency with the ground truth. As shown in Figure 6, our result at s=1.0 deviates substantially from the ground truth — the small arc structure at the bottom right is absent in the enhanced image. Similar observations can be found in the appendix.
  3. As reported in Table 2, the proposed method fails to preserve image structures as effectively as many prior approaches. Its improvement over the vanilla FLUX.2 is unsurprising, given that the method is built upon this baseline.
Questions for authors

The advantages of the proposed method require stronger justification. Achieving favorable scores on image quality metrics alone is insufficient. Structure consistency with the input, as well as alignment and fidelity to the ground truth, also need to be carefully considered.

Reviewer_Mh7FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a low-light image enhancement pipeline with adjustable brightness control. A synthetic training dataset is built, based on pseudo-enhanced reference by FLUX.2-klein-9B from real low-light images, leveraging retinex-inspired decomposition and interpolation processes, creating multiple intermediate versions with various enhancement intensities. Pixel level misalignment is mitigated using weighted loss function, along with LoRA training that can adjust enhancement intensity during inference. The proposed model show advantages in terms of non-reference metrics on multiple benchmarks.

Critical strengths
  • This paper tackles a meaningful task in real-world application, where fine-grained adjustable enhancement algorithms are preferred. It also offers an intact pipeline for the task, including data synthesis, generating pseudo-GT, Retinex-based interpolation, and pixel alignment. Both training and inference requires only one model, yet is capable of providing multiple or a spectrum of outputs.

  • Comprehensive experiments and ananlysis demonstrated observable advantage of the proposed method, e.g., Table 1 shows that on 3 real-world benchmarks, the proposed method has the best results in terms of NIQE, CLIP-IQA, MANIQA, and MUSIQ. Table 3 also demonstrates consistent advantage of the proposed method in CLIP-Dir. For control variable ss, using new values of 0.25, 0.5, 0.75 that are unseen during training, shows the generalization and interpolation ability of the proposed method.

Critical weaknesses
  • Controllability is not well-defined in the paper.

    • The control factor ss is trained discretely. Even though the paper has reported experiments of LoRA scaling on unseen ss values, several comparison experiments would be preferable in order to offer more concrete evidence: continuous random sampling of ss; diverse interval of ss values; output image brightness vs ss curve, etc.
    • Definition of ss doesn't entail physically grounded illumination trajectory, since ss describes how close to I1I_{1} which is a FLUX generated psedo-GT, not physically defined or consistently maintained exposure.
    • There are discrepencies between how ss is supposed to work and how ss is utilized.
      • Folllowing the definition from L200 to L210, when s→1s \rightarrow 1, there would be: Rs→0.5R0+0.5R1R_{s} \rightarrow 0.5R_{0} + 0.5 R_{1}. Training traget would approch I1~=L1⊙(0.5R0+0.5R1)≠I1=L1⊙R1\tilde{I_{1}} = L_{1} \odot (0.5R_{0} + 0.5 R_{1}) \not= I_{1} = L_{1} \odot R_{1} .
      • On the other end, since ss controls LoRA adaptation strength according to eq. (7), when s→0s \rightarrow 0, the model reverts back to original FLUX, which doesnot guarantee F(I0,s=0)=I0F(I_{0}, s=0) = I_0, thus making ss deviate from its objective of controling enhancement strength.
    • Quantitative results only measure semantic and perceptual variance, not brightness sensitivity and monotonicity.
  • Although the papre focus on perceptual enhancement and achieved substantial results, failure in restoration fidelity are catastrophic, with 7~10dB lower in PSNR than the best.

  • Effects of low-light noises are not fully addressed or evaluated in the paper.

  • Retinex decomposition and illumination interpolation offers limited novelty, e.g. see [1]. [1] H. A. Gasparyan, S. A. Hovhannisyan, S. V. Babayan and S. S. Agaian, "Iterative Retinex-Based Decomposition Framework for Low Light Visibility Restoration"

Questions for authors
  • Is the combination concretely better than simple solutions? E.g., direct Retinex interpolation + FLUX (no finetuning); or endpoint-only LoRA. These are more informative ablation variations.
  • Would the authors explain more explicitly on the two ends of ss, how is it defined approaching end values, why eq. (1)-(2) have discrepancy with psedo-GT. A densely sampled brightness response curve would be preferable to show monotonicity, difference among images, and behavior around endpoint values.
  • How to validate reliability and fidelity of generated pseudo-GT?
  • How robust is the enhancement pipeline against various noise levels of low-light images?

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

Related papers

Discussion (0)

Sign in to comment.