acceptodds
Under review as a conference paper at ICLR 2027

Distinguishing Intrinsic Scene Colors from Color Casts for sRGB White Balance Correction

Abstract

White balance (WB) correction for sRGB images requires removing color casts introduced by inaccurate WB settings while preserving intrinsic scene colors. Although existing methods achieve competitive WB correction performance, nonlinear image signal (ISP) processing entangles intrinsic scene colors with color casts, making it difficult to determine which chromatic variations should be corrected and which should be preserved, especially when large regions of intrinsic scene color dominate the image and resemble color casts, which can lead to inaccurate correction. To alleviate this correction ambiguity, we introduce the color cast state (CCS), an image-level representation that explicitly characterizes the tendency and severity of color casts, and propose a CCS-guided framework for sRGB white balance correction. Specifically, we learn a structured CCS representation from paired incorrectly white-balanced and ground-truth images using chromatic tendency and severity supervision together with soft similarity learning, thereby preserving continuous relationships among different color cast states. The learned CCS is then used as an explicit condition for the correction network, enabling correction that better preserves intrinsic scene colors while removing color casts. Extensive experiments on public benchmark datasets demonstrate that our method achieves state-of-the-art performance and consistently performs favorably under different WB settings, delivering more visually pleasing correction results.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
sRGB White Balance Correction, Color Cast State Representation, Color Constancy, Computational Photography

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_vfugOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a novel CCSG-WB framework to address the problem of sRGB WB correction: how to disentangle intrinsic scene colors from color casts. Owing to the strong non-linear operations of the camera ISP pipeline, previous works fail to handle cases when the dominant colors of the scene are similar to the color cast. To resolve this ambiguity, this paper introduce an image-level representation characterizing cast tendency called Color Cast State (zCCSz_{\text{CCS}}). The framework operates in two stage. In Stage I, CCSRL uses paired images and a soft similarity comparison learning setup. It trains a CCS encoder to learn a well-structured CCS space. In Stage II, the setup uses the frozen CCS encoder to extract zCCSz_{\text{CCS}}. It then adjusts a U-shaped Transformer backbone through CCSG blocks. These blocks use FiLM parameters γi\gamma_i and βi\beta_i. The Rendered WB dataset and the Rendered Cube+ dataset show that the method gives the best known accuracy. They also show that the method works well across cameras. They further show that the method has very low delay.

Critical strengths
  1. Well-Motivated Problem Formulation: The paper pinpoints the core ambiguity of sRGB-WB correction: the dominant intrinsic color of the scene, after the non-linear ISP pipeline, looks like a color cast. Modeling the color cast explicitly as a conditional guidance vector zCCSz_{\text{CCS}} is an insightful approach.
  2. Sound Methodological Design: The method combines a tendency loss Lt\mathcal{L}_t, a strength loss Ls\mathcal{L}_s, and a soft comparison loss Lc\mathcal{L}_c. This combination effectively avoids hard label edges. This combination builds a continuous and smooth color tint space.
  3. Strong Empirical Results: The proposed method consistently achieves the best performance compared to state-of-the-art baselines (e.g., ABC-Former, DCLAN) for the MSE, MAE, and ΔE2000\Delta E_{2000} metrics on the Rendered WB and Rendered Cube+ datasets.
  4. Efficiency and Robustness: The method takes about 73 ms73\text{ ms} per image on the Set 1-Test dataset. It gets the lowest ΔE2000\Delta E_{2000} across all color temperatures between 2850K and 7500K. It demonstrates robust adaptation of casting.
Critical weaknesses
  1. Stage Ⅰ of the model strongly relies on pairs of images with incorrect white balance and their corresponding ground truth, which is impractical for most unpaired and in-the-wild images, where rendering paired images is impossible.

  2. The experiments are restricted to global and uniform color casts (2850K-7500K). The method is not evaluated or discussed on scenes with mixed illumination and spatially varying white balance.

  3. The running speed remains competitive. However, the CCS encoder is built on a pre-trained ViT-B model. The method resizes input images to 224×224224 \times 224 during feature extraction. However, the method still keeps a large ViT next to a Transformer repair network. This design may cause problems when people use it on devices that have limited resources.

Questions for authors
  1. How does the global representation zCCSz_{\text{CCS}} perform on images with spatially non-uniform illumination or mixed light sources (e.g., warm light on the left and cool light on the right of the scene)?
  2. The CCS encoder relies on a pre-trained ViT-B/32 fine-tuned with LoRA. Could a lighter CNN or compact Transformer backbone be substituted to reduce parameter overhead? And how much performance degradation in ΔE2000\Delta E_{2000} would result?
  3. In severely under-exposed or over-exposed regions where nonlinear ISP processing leads to hue clipping and detail degradation, how does the global FiLM modulation (γi\gamma_i, βi\beta_i) avoid introducing color distortion and visual artifacts?
1 comment
AuthorsOpenReview ↗

Response to Reviewer vfug

Thank you for the thorough review of our paper and the insightful comments.

W1. Paired Supervision and Real-World Applicability

The paired incorrectly white-balanced and ground-truth images serve only as supervision signals for training stage and are not required during inference. After CCS representation learning is completed, the whole framework takes only a single sRGB image as input at inference time, and the trained CCS encoder directly captures the color cast state from the input.

Furthermore, Rendered WB Set 2 and Rendered Cube+ are exclusively used for testing and excluded from training. Their image formation pipeline closely resembles the process of erroneous on-camera white balance. This benchmark provides a controlled and practically relevant setting for evaluating the WB correction. Our method achieves a mean ΔE2000\Delta E_{2000} of 4.02 on Rendered WB Set 2 and 3.68 on Rendered Cube+. These results demonstrate that our CCSG-WB generalizes to real-world scenes. When in-the-wild retraining is desired, paired training data can be constructed following KNNWB [1], as long as RAW images are accessible.

W2/Q1. Global CCS guidance and spatially varying illumination

Our CCS is designed to characterize chromatic deviations arising from global white balance errors, help distinguish color casts from intrinsic scene colors, and deliver correction guidance for the shared correction backbone. The current formulation mainly targets global color casts, consistent with the evaluation settings of Rendered WB and Rendered Cube+. WB correction under mixed illumination and spatially varying lighting constitutes a related yet distinct task. We will plan to explore spatially varying or region-wise CCS conditioning as a future extension.

W3/Q2. Lightweight CCS Encoder and Accuracy–Efficiency Trade-off

Reducing the computational overhead of the CCS encoder is critical for practical deployment. To this end, we only replace the CCS encoder with a lightweight ResNet-18 encoder while keeping all other settings unchanged.

CCS EncoderMSE ↓MAE ↓ΔE2000\Delta E_{2000} ↓
ResNet-1860.133.06°4.01
ViT-B/32 (Ours)51.702.70°3.68

The experimental results show that after replacing the CCS encoder with this lightweight CNN encoder, the mean ΔE2000\Delta E_{2000} is 4.01, an absolute gap of 0.33 compared with the ViT-B/32 variant (3.68).

Q3. Color Fidelity under Challenging Exposure Conditions

In our method, the global CCS representation modulates intermediate features via FiLM to provide guidance. The final correction result is jointly determined by the global color cast state and local image features. Furthermore, the correction backbone integrates multi-scale feature extraction, attention mechanisms, and skip connections to preserve local image information during reconstruction. Combined with the reconstruction loss, these components enable the network to correct color casts while reducing undesired color distortions and visual artifacts. We will add visual results under challenging exposure conditions to the supplementary material in the revised manuscript.

[1] When color constancy goes wrong: Correcting improperly white-balanced images. CVPR 2019.

Reviewer_xdg7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work addresses sRGB white balance correction, focusing on the ambiguity between intrinsic scene colors and color casts that arises because nonlinear ISP rendering entangles the two. The authors argue that this ambiguity is particularly harmful when dominant intrinsic colors resemble color casts, a situation in which existing end-to-end methods and those relying on global color statistics tend to produce inaccurate corrections. To address this, the proposed method introduces the color cast state (CCS), an image-level representation that explicitly characterizes the tendency and severity of the color cast. Specifically, the proposed framework, CCSG-WB, first learns a structured CCS representation from paired incorrectly white-balanced and ground-truth images, using chromatic tendency and severity supervision. A soft similarity learning objective is further employed to preserve continuous relationships among different color cast states in the representation space. The learned CCS is then injected into a shared correction backbone through CCS-guided blocks, which generate feature-wise modulation parameters that adapt the correction behavior to each input. Experiments on public benchmarks show that the method achieves state-of-the-art performance and remains consistently effective across different WB settings.

Critical strengths

S1. The paper proposes the color cast state (CCS), an image-level representation that explicitly characterizes the tendency and severity of color casts, together with a learning scheme that preserves continuous relationships among different color cast states. This directly targets the ambiguity between intrinsic scene colors and color casts, which is not explicitly modeled by prior methods.

S2. The proposed CCSG-WB framework effectively exploits the learned CCS by injecting it into a shared correction backbone through feature-wise modulation, allowing a single network to adapt its correction behavior to different color cast conditions while preserving intrinsic scene colors.

S3. Extensive experiments on multiple public benchmark datasets show that the proposed method consistently outperforms existing approaches across diverse WB settings.

Critical weaknesses

W1. The framework mainly combines established components, namely a LoRA-adapted ViT encoder, SCE-based soft contrastive learning, FiLM conditioning, and a Restormer-style backbone. CCS itself is an embedding that regresses the image-wide mean chromatic offset. Additional novelty would strengthen the contribution, for example by demonstrating CCS as a plug-in condition for existing sRGB-WB methods, or by showing that its continuous structure improves robustness to diverse or unseen color casts.

W2. All test sets are rendered from RAW data with the same preset WB settings used for training. It therefore remains unclear whether the claimed continuity of the CCS space generalizes to unseen cast types or to real-world images with genuine AWB failures.

W3. The learned CCS space is visualized only for the final model (Fig. S2), without loss-wise comparisons or quantitative measures of its structure. Moreover, L_c alone accounts for most of the gain in Table 3, which suggests a marginal contribution from the tendency and severity supervision.

W4. The core motivation, resolving cases where dominant intrinsic colors resemble color casts, is supported only by selected qualitative examples. Fig. 4(B) further shows that the output depends strongly on the supplied CCS, yet CCS prediction accuracy is not reported, and controllable editing of CCS is not explored.

W5. Because CCS is an image-level vector and the FiLM parameters are broadcast spatially, the correction is spatially uniform by design. The method is evaluated only on single-illuminant data, and its applicability to spatially varying casts under mixed illumination (e.g., “Attentive Illumination Decomposition Model for Multi-Illuminant White Balancing,” CVPR 2024) is not discussed.

Questions for authors

Q1. Clarification is required on whether CCS improves existing sRGB-WB methods when used as a plug-in condition, along with the parameters and FLOPs.

Q2. I am curious about the performance on unseen cast types and on real-world AWB failures.

Q3. Loss-wise CCS visualizations for Table 3, along with a quantitative measure of their structure, would be helpful.

Q4. It would be informative to report the CCS prediction accuracy, especially when intrinsic colors resemble color casts.

Q5. I wonder whether CCS can be interpolated or edited for controllable WB adjustment.

Q6. A discussion on how the method handles spatially varying casts under mixed illumination is needed.

4 comments
AuthorsOpenReview ↗

Response to Reviewer xdg7 (Part 1/2)

We sincerely thank you for the careful evaluation of our manuscript and the constructive feedback. We hope our clarifications and additional experiments resolve your concerns, we would appreciate your favorable evaluation.

W1/Q1. Contribution and Transferability of CCS

Our core contribution lies in defining and learning the Color Cast State (CCS) representation, which helps distinguish color casts from the intrinsic scene colors. We leverage this CCS to guide white balance (WB) correction. CCS is trained on paired incorrectly white-balanced and ground-truth images, and its training objective builds representations centered on the cast discrepancy to be removed.

To further demonstrate that the learned CCS is not restricted to our proposed correction backbone, we spatially broadcast the CCS representation and concatenate it with the input image as an auxiliary conditioning signal, which is then fed into the DeepWB [1] correction network. The comparison results are presented below.

MethodMSE ↓MAE ↓ΔE2000\Delta E_{2000} ↓
DeepWB92.153.21°4.34
DeepWB + CCS83.482.96°4.14

As shown in the table, incorporating CCS into DeepWB brings consistent improvements on all three metrics. This performance gain is obtained merely by concatenating CCS with the input image as the input to the correction network, further confirming that CCS is a transferable, correction-oriented representation instead of being restricted to a particular architecture. But using CCS as a plug-in conditioning signal for existing sRGB-WB methods introduces an overhead of 88.48 M parameters and 9.30 GFLOPs.

W2 / Q2. Generalization Across Diverse Color Cast Conditions

Although the images in the dataset are rendered with the same set of WB settings, the actual color cast state of each image is not determined solely by the WB settings. Different RAW images correspond to distinct scene illumination and intrinsic colors. As a result, rendering different RAW images under an identical WB setting can yield chromatic deviations that differ substantially in both tendency and severity relative to their ground-truth images. The training objective of our CCS representation builds representations centered on the cast discrepancy to be removed. This allows our evaluation to examine the model’s correction performance across diverse color cast states, even when the dataset shares the same set of WB settings.

Moreover, Rendered WB Set 2 and Rendered Cube+ are only used for testing and are excluded from training. They follow an image formation pipeline highly similar to that of erroneous on-camera white balance. In both cases, an inappropriate WB setting is first applied, followed by subsequent nonlinear ISP rendering, which introduces color casts into the final sRGB images. Thus, the benchmark provides a controlled and practically relevant setting for evaluating the WB correction.

W3/Q3. Loss-wise CCS Visualizations

We further conduct loss-wise CCS visualizations, which will be included in the supplementary material. The visualizations reveal different patterns of the learned representation spaces. Under tendency and severity supervision alone, samples with different chromatic features are still partially mixed. When only the soft similarity loss is used, samples with alike color cast attributes form more coherent neighborhoods. Jointly trained with the full CCS learning scheme, the representation exhibits structured distribution over varied scene contents, spanning warm, cool and transitional color cast states.

W4/Q4. Color Cast Target Prediction Accuracy

We evaluate the prediction accuracy of the color cast target to verify that the CCS representation captures correction-relevant chromatic information. We adopt mean absolute error (MAE) and root mean squared error (RMSE) to assess the auxiliary regression head used in Stage I. This regression head predicts the color cast target from the CCS representation. Note that the regression head is not trained under LcL_c alone, so this evaluation cannot be performed for this setting.

Loss FunctionMAE ↓RMSE ↓
Lt+LsL_t+L_s0.048600.07402
LcL_c--
Lt+Ls+LcL_t+L_s+L_c0.022520.03640

As illustrated in the table, the representation learned under the full CCS learning scheme encodes the chromatic tendency arising from color casts with higher precision. Since this prediction target is defined by the discrepancy between paired WB renderings, our experiment directly validates the representation’s ability to capture the color casts requiring correction — the key to distinguishing color casts from the intrinsic scene colors.

AuthorsOpenReview ↗

Response to Reviewer xdg7 (Part 2/2)

Q5. CCS Interpolation and Continuous Correction Guidance

As shown in Figure 4(B), the correction results change smoothly when we vary the input CCS representation while keeping the input image fixed. To further verify whether intermediate color cast states can be obtained through interpolation, we conduct an additional quantitative experiment on the Rendered Cube+ dataset. Specifically, we extract the CCS representations corresponding to the 2850K and 7500K WB settings and interpolate between them to approximate the intermediate states for 3800K and 5500K. Following the interpolation strategy used in DeepWB [1], we compare two sets of correction results for each WB setting: corrections guided by the interpolated CCS, and corrections guided by the CCS directly extracted from the corresponding input image.

WB SettingCCS GuidanceMSE ↓MAE ↓ΔE2000\Delta E_{2000} ↓
2850 KDirect67.873.35°4.46
2850 KInterpolated67.873.35°4.46
3800 KDirect55.052.96°4.14
3800 KInterpolated67.333.47°5.60
5500 KDirect49.732.47°3.32
5500 KInterpolated63.583.01°4.07
7500 KDirect61.582.80°3.73
7500 KInterpolated61.582.80°3.73

At intermediate color temperatures, the interpolated CCS can also guide the correction network without additional training. Naturally, the CCS directly extracted from the image delivers correction guidance with higher accuracy. Combined with the observations in Figure 4(B), these results demonstrate that CCS representations can be manually manipulated to alter WB correction outcomes.

W5/Q6. Applicability to Spatially Varying Illumination

Our CCS is designed to characterize chromatic deviations arising from global white balance errors, help distinguish color casts from intrinsic scene colors, and deliver correction guidance for the shared correction backbone. The current formulation mainly targets global color casts, consistent with the evaluation settings of Rendered WB and Rendered Cube+. WB correction under mixed illumination and spatially varying lighting constitutes a related yet distinct task. We will plan to explore spatially varying or region-wise CCS conditioning as a future extension.

[1] Deep White-Balance Editing. CVPR 2020.

Thank you for detailed response to the review. Most of my concerns are addressed and hope to see a future extension about region-aware CCS conditioning soon. I will raise my score to 3.

We sincerely thank you for your positive feedback and for reconsidering the evaluation of our work. We truly appreciate your time and thoughtful consideration.

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

This paper proposes CCSG-WB for correcting white balance in camera-rendered sRGB images. It introduces an image-level color cast state (CCS), learned in a first stage from incorrectly white-balanced images and their paired references. A LoRA-adapted ViT-B/32 is trained using a four-dimensional Lab/YUV chromatic-offset target, a Lab color-difference severity loss, and a soft contrastive loss that preserves the similarities among cast targets. In the second stage, the CCS encoder is frozen, and its 128-dimensional representation generates feature-wise scale and shift parameters for FiLM layers within a U-shaped Transformer correction network. The correction network is trained with an L1 image loss. Experiments on Rendered WB Set 1-Test, Set 2, and Rendered Cube+ report improved MSE, angular error, and Delta E2000 over the listed baselines, with favorable inference times.

Critical strengths
  1. The motivation is clear and relevant. A dominant blue or warm color in a rendered image may reflect either scene content or an incorrect white-balance setting. This ambiguity is important for sRGB correction, and the paired-rendering setup provides a reasonable way to supervise correction-related information.
  2. The method is simple and coherent. Separating CCS learning from image correction makes the role of the additional representation clear. At test time, the CCS encoder reads only the input image, while the correction network uses its features through lightweight FiLM modulation.
  3. The reported gains are substantial on the public benchmarks. On Rendered WB Set 1-Test, the mean Delta E2000 improves from 2.18 for ABC-Former to 1.67. On Set 2, which contains images from unseen cameras, it improves from 4.55 to 4.02. The results on Rendered Cube+ are also consistently competitive.
Critical weaknesses
  1. The novelty relative to existing illumination-conditioned correction remains moderate. The central ingredients are known: paired-image supervision for color deviations, soft similarity learning, FiLM feature modulation, and a U-shaped image restoration backbone. StyleWB [1] already treats scene lighting as a latent style and uses learned affine conditioning, while HVDualformer [2] and ABC-Former [3] exploit global color information to adapt white-balance correction. CCSG-WB differs in how its cast target is constructed and supervised, but the paper should identify this specific distinction more precisely rather than suggesting that correction-conditioned color representations are entirely new.

  2. The ablations do not fully isolate the contribution of the structured CCS supervision. In Table 3, using the soft contrastive loss alone gives 3.73 Delta E2000 on Rendered Cube+, while the complete three-loss model gives 3.68. Likewise, direct CCS concatenation gives 3.74, compared with 3.68 for FiLM. Supplementary Table S1 shows similarly small gaps on Set 1-Test and Set 2. These experiments support incremental benefits but do not show whether a 128-dimensional representation from a pretrained ViT, trained with a simpler objective or no cast-specific supervision, would work almost as well.

[1] Modeling the Lighting in Scenes as Style for Auto White-Balance Correction. WACV 2023.

[2] HVDualformer: Histogram-Vision Dual Transformer for White Balance. AAAI 2025.

[3] ABC-Former: Auxiliary Bimodal Cross-domain Transformer with Interactive Channel Attention for White Balance. CVPR 2025.

Questions for authors

Please refers to the weaknesses. I think this paper is somewhat interesting and makes sense for this topic.

3 comments
AuthorsOpenReview ↗

Response to Reviewer dpbM

We greatly appreciate your thoughtful assessment of our work and the valuable suggestions.

W1. Comparison with existing methods

Our core contribution lies in defining and learning the Color Cast State (CCS) representation, which helps distinguish color casts from the intrinsic scene colors. We leverage this CCS to guide white balance (WB) correction.

StyleWB [1] models illumination as a style factor for WB correction, while HVDualformer [2] and ABC-Former [3] adopt color distribution statistics or histogram-based cues.

In contrast, CCS is trained on paired incorrectly white-balanced and ground-truth images, and its training objective builds representations centered on the cast discrepancy to be removed. We will further highlight this distinction in the main paper.

W2. The contribution of CCS representation learning

To directly verify the contribution of CCS representation learning, we adopt pretrained ViT features without CCS learning scheme supervision.

Conditioning RepresentationMSE ↓MAE ↓ΔE2000\Delta E_{2000} ↓
Pretrained ViT Features120.863.69°5.08
CCS Representation (Ours)51.702.70°3.68

The above presents our comparison results on the Rendered Cube+ dataset. Our learned CCS representation achieves substantially lower values across all three mean metrics compared with pretrained ViT features. This demonstrates that the learned representation can capture chromatic variations induced by color casts, making it better suited for the white balance correction task, as opposed to generic visual features.

[1] Modeling the Lighting in Scenes as Style for Auto White-Balance Correction. WACV 2023.

[2] HVDualformer: Histogram-Vision Dual Transformer for White Balance. AAAI 2025.

[3] ABC-Former: Auxiliary Bimodal Cross-domain Transformer with Interactive Channel Attention for White Balance. CVPR 2025.

Thanks for the author's response. I have no more questions. Although the response for the novelty is not so satisfying, it is also not bad. So, I keep my score.

We sincerely thank you for taking the time to carefully consider our response and for your thoughtful feedback. We truly appreciate your careful assessment and constructive suggestions throughout the review process. Thank you again for your time and effort in reviewing our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank the Area Chair for coordinating the review process and all three reviewers for their careful assessments and constructive feedback. We appreciate their recognition of the motivation behind CCSG-WB, the effectiveness of our CCS-guided correction framework, and the strong experimental results. The reviewers' suggestions have helped us further clarify the contribution of CCS representation learning and strengthen the experimental validation of our method.

The main concerns raised by the reviewers and our corresponding responses and revisions are summarized below.

Reviewer dpbM – Novelty of CCS and effectiveness of representation learning: We further clarified how CCS differs from existing illumination-conditioned and histogram-based WB correction approaches. Our core contribution is defining and learning the Color Cast State (CCS) representation to distinguish color casts from the intrinsic scene colors for guiding WB correction. To isolate the benefit of CCS-specific learning, we add a baseline using pretrained ViT features without CCS supervision (Table 3). This experiment shows that CCS-oriented representation learning provides extra gains beyond generic pretrained visual features.

Reviewer xdg7 – Transferability, generalization, structured CCS learning, and controllability: We have supplemented a set of experiments to further validate the proposed CCS representation. We integrate CCS into the existing DeepWB network via feature concatenation, and the performance gains verify that CCS is not limited to our correction backbone. We also clarify the distinction between predefined WB settings and image-dependent actual color cast states, and demonstrate that our adopted test benchmarks enable practically relevant evaluation under diverse color cast conditions. Moreover, we add loss-wise CCS visualizations to show how tendency, severity and soft similarity supervision shape the structure of the learned representation space (Figure 4). We further introduce quantitative evaluation on color cast target prediction, which examines the representation’s capacity to capture color casts requiring correction (Table 5). In addition, we conduct CCS interpolation experiments to explore continuous correction guidance and its potential for controllable white balance adjustment. We also clarified the applicability of the current image-level CCS formulation to global color casts and discussed spatially varying illumination as a future extension.

Reviewer vfug – Practical applicability, lightweight encoders, and color fidelity: We clarify that paired WB images are only required for representation learning, and only a single sRGB image is needed during inference. We also investigate the accuracy degradation introduced by the lightweight CCS encoder. We further explain how global CCS guidance interacts with spatially varying image features through FiLM modulation, multi-scale feature extraction and reconstruction supervision. Additional qualitative comparisons under challenging illumination and brightness conditions are added to the supplementary material of the revised manuscript to demonstrate local color fidelity (Figure S2).

Overall, these revisions provide a more comprehensive evaluation of CCSG-WB by demonstrating the benefits of CCS representation learning, validating its effectiveness as a transferable conditioning representation, examining the structure and prediction accuracy of the learned CCS space, and exploring lightweight encoding and continuous correction guidance. Together, these additional experiments and analyses further substantiate the central contribution of our work: explicitly learning correction-relevant color cast states to guide sRGB white balance correction while preserving intrinsic scene colors.

We sincerely appreciate the reviewers' insightful comments and suggestions, which have helped us improve the clarity, experimental completeness, and scientific rigor of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.