CAGGLE: Color-Aware Guidance with Global and Local Prompts for Exposure Correction

The introduction and Figure 1 do not motivate this work well. The Abstract emphasizes that existing methods may fail in scenes with both under- and over-exposure problems and in over-saturated regions. However, the first image shown in Figure 1 is under-exposed and the other image is over-exposed. While they do not contain obvious mixed exposure problems, the results between the existing methods and the proposed one do not show significant differences that can indicate the problems suffered by existing methods. This results in a weak motivation for introducing color-aware guidance for image correction. Including examples in Figure 1 that better demonstrate the mixed exposure and over-saturation issues highlighted in the abstract would strengthen the paper's motivation.

Missing Important Results. While this paper relies on the Color Naming model (Serrano-Lozano et al., 2024) to provide intermediate supervision, the comparisons to this method are necessary (but are missing) as it is an image enhancement method.

The claimed local prompt map conditioned Cross-Attention (LP-CA) is claimed to be novel but there is no justification for its novelty and the corresponding ablation results.

A Minor Issue. It would be better to add references for the methods in Figure 1. Otherwise, it makes it hard to identify those works.

1. The paper contains many subjective statements with unclear references. In the introduction (page 1, lines 38-41), the authors mention limitations of early research, but it is unclear which specific early research is being referred to and what limitations they had. In the method section, the organization is confusing, making the overall pipeline and the proposed modules difficult to read and follow.
2. There are some concerns regarding the proposed CAGGLE. Why does the paper prioritize learning the local prompt before the global prompt? Intuitively, global illumination information should be easier to learn. How would the performance change if the model learned the global prompt first, followed by the local prompt? Additionally, the Color Naming Loss computes loss based on the input color priors. However, based on the qualitative results, after exposure correction, the color of the ground truth may change, and different exposure correction levels result in varying color shifts. The reviewer is curious about how color consistency is ensured after exposure correction.
3. The computational cost of the proposed method, such as parameters, FLOPs, and runtime, should be reported for comparison.

*The introduction of prompt learning to the exposure correction is commendable, but the explanation of local and global prompts is not very convincing, and the promotion effect of the color naming module on local prompt features has not been well explained.

*The explanation of the proposed method is not clear enough. The color naming module relies on existing methods, but there is a lack of clear explanation on how the local and global prompt modules are implemented.

• The proposed model lacks novelty as it is structurally similar to other common models.
• The authors need to visualize the latent space practically.
  • For example, 1) They can verify whether the encoder's output for a given image, when rearranged into pixel-level features, can be well-divided into K large groups by using K-means algorithm. Or 2) For the given dataset, the authors can extract pixel-level features, classify each pixel as over-exposed, under-exposed, or normally exposed, and then visualize these classifications using t-SNE. If this process is too difficult, they should at least perform a similar process at the image level.
  • Particularly, regarding Table 5 shown by the authors, there is a need to analyze how different latent spaces the models have from each other.