Range-Null Latent Prior-guided Consistency Model for Low Light Image Enhancement

• The null space, say, {x for Ax=0} of a full-rank linear operator A (as the author formulated in Eq.15) only contains x=0. The null-space refinement term $I - \gamma A^{-1}A$ thereby degrades to $1 - \gamma$. Thus the claimed null-space refinement is over-claim.
• The experiments contain unfair comparisons. The light decoder is trained with benchmark images as the author stated in the implementation details. This renders an unfair comparison for all competitors and raises doubts about the proposed method's soundness. Authors shall use COCO, HQ-50k, or other normal-light datasets that are out of the benchmarking dataset for a fair comparison.
• The dataset proposed is not well-motivated. The authors shall highlight the contribution to the community for this evaluation set.
• Please proofread to eliminate typos.
• I suggest the authors to move the related work part to the main manuscript, since this year the page limit of the conference is 10 pages.

Insignificant Main Idea. Given that the range-null space decomposition (Schwab et al., 2019) has been explored with the diffusion model in (Wang et al., 2023b) for image restoration tasks, and consistency models (Song et al., 2023) have offered quality-computation trade-off against diffusion models via direct noise-to-data mapping, using latent consistency model (Luo et al., 2023) to replace the diffusion model in (Wang et al., 2023b) and applies it for a specific task of enhancing low-light images is incremental. The authors may highlight which part of the proposed range-null latent prior-guided framework is novel and justify its significance via analysis.

Unclear Motivation. In addition, the logic of introducing language prompts for refining null-space content is hard to follow. First, I cannot find where in CFG (Ho&Salimans, 2022) mentions about generating language-aligned images. Second, it is not clear how the language-aligned image generation inspires this paper to use contrasted language prompts to refine null-space and how these contrasted language prompts are obtained. The authors may explicitly explain the motivation for using contrasted language prompts for refining the null-space content of low-light images.

Unclear Novelty of the Proposed Lighting Decoder. The differences between the proposed lighting decoder and the bypass decoder of QuadPrior (Wang et al., 2024) are the different training data. The authors may clarify whether it is novel and (if yes) justify how it represents significant novelty.

Unsupported Claims. The paper claims that the global illumination prior “physically” guarantees the reliability of range-space content. The paper also claims that the consistency model (Song et al., 2023) is more effective for generating null-space content. While these two claims are important to understanding the model designs, there is no evidence and analysis that can justify these two claims. The authors may provide analysis and evidence to justify them.

Unconvincing comparisons. The ground truth images in the proposed UAV-LL datasets are still quite dark, for example, the last case in Figure 5, and the result produced by the proposed method is also darker than those of other methods. While the paper criticizes previous methods for not generalizing well to the proposed dataset according to the quantitative comparisons in Table 1, it seems that existing methods produce visually better results but the ground truth images in the proposed dataset favor the results produced by this paper. This makes the comparisons on the proposed dataset unconvincing. The authors may clarify why the evaluation based on such dark ground truth images in the UAV-LL dataset is reasonable.

Missing ablation studies. The results of removing the self-attention swapping mechanism and using different language prompts and guidance scales are missing in Table 3.

Minor Issues. Line 070, Should the reference for LCM be (Luo et al., 2023) instead of (Song et al., 2023) for the consistency models? Line, 086, this -> This. Line, 398-399, we report…is selected as… Line, 462, should RNLP be RLPCM? Line 377, it is not clear how many GPUs are used.

It is not clear whether the authors would release the training data and codes. Without releasing the training data, it may not be possible to reproduce the results as the paper relies on self-collected training data.

see questions below

1. In line 320, the manuscript indicates that the lightening decoder is trained using a benchmark dataset. This introduces an unfair advantage over prior methods that adhere to stricter training protocols.

2. The assertion that the null space of matrix A contains only {x = 0} holds true given that A is consistently a full-rank operator. Under this condition, the proposed “null-space refinement” degrade to a classifier-free guidance mechanism, thereby undermining the novelty and technical robustness of the approach.

3. The motivation for introducing a new UAV-captured low-light dataset is ambiguous and insufficiently justified.

4. The novelty of the proposed method is limited. The application of range-null space decomposition within a diffusion-based framework and the use of text prompts for classifier-free guidance are established concepts in low-level vision.

1. The global illumination factor assumption mentioned in the paper regarding Range-null Space Decomposition is robustness-deficient, especially in real nighttime environments with diverse lighting sources, such as bustling cityscapes with varied lighting. It seems less credible and should consider local lighting properties rather than relying solely on a constant factor.

2. The qualitative visualization results presented by the authors (as shown in Figures 5 to 12) mainly display night scenes with low contrast and lack exploration of more complex night scenes, such as those in the DARKFACE dataset.

3. In Section 2.3, it is unclear how the global illumination factor, which depends on a constant factor, is calculated. It is also unclear what the various subplots in Figure 4 are intended to illustrate, including the definitions of the axes.

4. In the process of Range-null decomposition, noise removal is not considered. Noise caused by high ISO values or insufficient exposure time is an unavoidable issue for low-light enhancement.

5. The setup of training the Lighting Decoder with an additional 7,000 images collected is unfair when compared to other methods. Why is the decoder's training loss based on Equation 13, in a minimax form?