acceptodds
Under review as a conference paper at ICLR 2027

Zero-Shot Low-Light Image Enhancement via Diffusion-Learned Photometric Priors

Abstract

Pre-trained diffusion models encode natural-image statistics that can be exploited as priors for zero-shot low-light image enhancement (LLIE). Many existing diffusion-based LLIE methods, however, directly use the generative process for image reconstruction, coupling photometric correction with content reconstruction. This coupling may lead to structural drift and synthesized details inconsistent with the low-light observation. We instead exploit the complementary information provided by the observation and the diffusion prior, using the former to anchor scene structure and the latter to provide learned statistics of plausible normal-light appearance. Based on this principle, we propose a structure–photometry decoupled framework that restricts diffusion to photometric guidance while deriving image structure from the observation. We first recover a noise-aware structure estimate from the low-light input, while a short reverse trajectory of a frozen Stable Diffusion model produces a stochastic normal-light reference. Rather than using this reference for direct reconstruction, we project its photometric relationship with the structure estimate into global and spatially varying RGB gain–bias transformations. A globally anchored spatial trust region further stabilizes local corrections against unreliable local statistics and sampling-induced variability in the diffusion reference. The resulting transformations are applied to the observation-derived structure, combining input-anchored content with diffusion-learned normal-light statistics. Experiments on paired and unpaired LLIE benchmarks demonstrate leading zero-shot performance, with the best results on most evaluated metrics and efficient inference. Beyond LLIE, the framework achieves competitive performance on auto white balance without task-specific modification, suggesting the broader applicability of diffusion-learned photometric priors.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-Light Image Enhancement, Zero-Shot Learning, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dBFyOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a zero-shot low-light enhancement framework that uses frozen diffusion outputs as photometric references. It estimates constrained global and spatial RGB gain–bias transformations to enhance an observation-derived structure image while limiting content drift.

Critical strengths
  1. The author addresses a meaningful limitation of diffusion-based restoration.
  2. Clear, easy-to-follow manuscript.
Critical weaknesses
  1. Unclear role of frozen diffusion.Retinex already enhances brightness. What additional benefit does frozen diffusion provide: natural-image priors or exposure correction? If exposure correction, what drives it toward appropriate brightness with an empty prompt?
  2. Dependence on structure quality.If SS loses details through overexposure or clipping, gain–bias correction cannot recover them. How does the global-anchored trust region handle such cases?
  3. Unclear spatial projection. Is photometric projection pixel-wise or grid-based? Please clarify how its spatial adaptivity balances illumination correction against diffusion-artifact transfer, and discuss its limitations.
Questions for authors

see weaknesses.

Reviewer_6PmzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a zero-shot low-light image enhancement framework that decouples observation-derived structural information from diffusion-learned photometric priors. Experiments on multiple paired and unpaired low-light enhancement benchmarks demonstrate competitive zero-shot performance and improved structural fidelity compared with direct diffusion reconstruction. The authors also extend the framework to automatic white balance to illustrate its potential applicability to other image enhancement tasks.

Critical strengths

(1). The proposed idea of extracting photometric information from diffusion outputs while preserving observation-derived structures provides a meaningful alternative to directly using generative models for restoration.

(2). The proposed photometric bottleneck converts high-dimensional diffusion outputs into explicit RGB gain–bias transformations, offering an interpretable mechanism for controlling the influence of generative priors.

(3). The paper evaluates the proposed method on multiple paired and unpaired LLIE benchmarks, covering both full-reference and no-reference image quality metrics.

Critical weaknesses

(1). The central hypothesis of this paper is that diffusion-generated images provide reliable illumination and color statistics even when their structural content is inaccurate. However, the systematic evidence provided in this paper is limited and does not fully support this hypothesis across different illumination conditions, degradation levels, and scene categories. Although Table 5 shows that the method outperforms other reference sources, the classical baseline and the random statistical baseline are relatively simple.

(2). The confidence weights used for local gain-bias estimation depend solely on the local variance of the observation-derived structure, without explicitly considering the reliability or correspondence accuracy of the diffusion reference. Since diffusion-generated images may contain hallucinated textures or spatially inconsistent content, high local variance does not necessarily imply a reliable photometric correspondence.

(3). Preserving the original scene structure is one of the core motivations of this paper. However, the structural fidelity evaluation in Table 6 is limited to comparisons between direct diffusion reconstruction and the proposed projection strategy. Since the final output is explicitly generated based on the structure derived from the observed data, the improvement in structural consistency is, to some extent, expected. More rigorous evaluation of edge preservation, texture consistency, and structural distortion would provide stronger support for the claimed structural advantages.

Questions for authors

Could the authors further clarify the weakness mentioned above?

1 comment
Reviewer_6PmzOpenReview ↗

Response to Author's Reubttal

Thanks for your rebuttal. The response has properly addressed my concerns. I would like to maintain my acceptance rate.

Reviewer_KRmDOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a zero-shot low-light image enhancement method based on a frozen Stable Diffusion. It forms an observation-derived structure SS by max-channel illumination, a Retinex lift, and spectral shrinkage. It further uses two guided DDIM updates to obtain a reference rr. The affine readout still overlaps, at the point of claim, with the latent affine modulation that [Cho et al., 2026] already uses for low-light enhancement and auto white balance. Table 1 also combines numbers from incompatible evaluations, so the reported ranking does not identify a reliable zero-shot advance.

Critical strengths

The split in Fig. 1 is the right question for fidelity-sensitive enhancement, where a diffusion sample can carry normal-light photometry and still invent texture. Eq. (20) makes that constraint concrete by rendering only from SS. Global regression moves LOLv1 PSNR from 17.94317.943 dB for SS to 18.85318.853 dB, the full trust region reaches 19.74719.747 dB, and unconstrained local gains worsen LPIPS to 0.2470.247. Replacing the diffusion reference with the anchor, a VAE round trip, classical gamma and gray-world, or transferred normal-light moments leaves PSNR near 18.018.0 to 18.218.2 dB. Two reverse steps and a reported 0.010.01 min are substantially cheaper than GDP at 8.568.56 min and LLIEDiff at 10.1210.12 min. Fig. 5 also records the failure under extreme underexposure.

Critical weaknesses
  1. Section 2.3 reduces [Cho et al., 2026] to diffusion-based modulation and refinement, while Section 3.3 presents channel-wise gain and bias as the bottleneck. Modulation in [Cho et al., 2026] is already a spatially uniform affine map in frozen Stable Diffusion for illumination and color, and the same model handles auto white balance. The additions are Eq. (15), Eq. (17), Eq. (18), and output y=a⊙S+by=a\odot S+b. This delta is absent from the claim. Eq. (1) follows the max-channel map of LIME [Guo et al., 2016], cited only as a test set. Eq. (12) has the DPS form [Chung et al., 2023]. Wiener shrinkage in Eq. (24) and Shades-of-Gray at p=5p=5 in Appendix C have no bibliography entries.

  2. Table 1 mixes protocols. Retinexformer at 13.02613.026 dB and 0.4260.426 SSIM is the MIT5K checkpoint in [Cho et al., 2026]. The same row lists LOLv2-real at 22.79422.794 dB and 0.8400.840 SSIM, the in-domain number in [Cai et al., 2023]. In-domain LOLv1, 25.1625.16 dB and 0.8450.845 SSIM, is absent. SNR-aware LOLv2-real matches [Cai et al., 2023] at 21.4821.48 dB and 0.8490.849 SSIM. LOLv1 here is 21.88821.888 dB and 0.8480.848 SSIM, against 24.6124.61 dB and 0.8420.842 SSIM in that table. MR. Illuminate appears at 19.73919.739 dB and 0.7150.715 SSIM, against published 21.73921.739 dB and 0.8150.815 SSIM [Cho et al., 2026]. NL is undefined. Table 6 largely repeats SS (SSIM 0.7790.779 versus 0.7930.793, Table 3).

  3. G(⋅)G(\cdot) and TT in Eq. (3) are undefined in the main text. TT, gtg_t, and ϵ\epsilon are overloaded with Eq. (23), Eq. (11), and Eq. (6). Eq. (15) shrinks gain toward 00, and Eq. (18) reverses when ag<0a_g<0. Linear-RGB SS is fit to VAE output rr in an unstated domain. γa\gamma_a, λg\lambda_g, λl\lambda_l, κ\kappa, δa\delta_a, and δb\delta_b lack numerical values, and Eq. (7) to Eq. (12) lack an ablation.

Questions for authors
  1. Please re-evaluate Table 1 and label the training set, code version, and GPU of every baseline. Please reconcile the Retinexformer row, which pairs a MIT5K result of 13.02613.026 dB [Cho et al., 2026] with an in-domain LOLv2 result of 22.79422.794 dB [Cai et al., 2023], and please place the published MR. Illuminate numbers (21.73921.739 dB, 0.8150.815 SSIM, LPIPS 0.1770.177) beside the footnoted rerun. Which ranking survives that correction?

  2. Please give the values of γa\gamma_a, γt\gamma_t, λg\lambda_g, λl\lambda_l, κ\kappa, δa\delta_a, δb\delta_b, and the patch size, and state whether SS and rr are both linear RGB or both sRGB at Eq. (15). A numerical walk-through on one LOLv1 image would settle the domain question. These values are required for the reproducibility claim.

  3. If the correction in Eq. (12) is removed, how much of Table 3 and Table 5 remains for an unguided two-step sample? Please also report the four Table 6 metrics on SS, on rr, and on yy.

  4. Table 2 lists angular error 7.24∘7.24^{\circ} for this method and 5.75∘5.75^{\circ} for MR. Illuminate, against 4.16∘4.16^{\circ} for mixed-illuminant correction [Afifi et al., 2022]. Should the white-balance claim be read from ΔE\Delta E and MSE alone?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.