acceptodds
Under review as a conference paper at ICLR 2027

Spatial–Illumination Alignment for Controllable Image Illumination Editing

Abstract

Controllable image illumination editing modifies both the luminance intensity and distribution. Pixel-wise alignment measures reconstruction errors but does not explicitly organize illumination variation with regional brightness. Spatial–illumination alignment couples regional luminance levels with spatial arrangement to guide illumination redistribution. For spatial-illumination consistency, we introduce Illumination Intensity Ranking (IIR) loss, constructing luminance rankings over image patches. Combining luminance consistency and positional consistency, IIR supervises how bright each region should be and where regions of corresponding brightness should occur. We further integrate spatial-illumination alignment with latent bridge matching and enhance informative illumination injection by condition-confirmed initialization. Extensive experiments show that the proposed Illumination-Conditioned Latent Bridge Matching (IC-LBM) achieves state-of-the-art performance on indoor scene relighting, improving the controllability of image illumination editing. Ablations confirm that luminance and positional consistency contribute complementary improvements to illumination reconstruction. IIR provides a structured objective for predicting target illumination without intrinsic decomposition or 3D scene reconstruction. Spatial–illumination alignment introduces explicit illumination supervision for image illumination editing, providing a pioneering perspective on modeling illumination transformations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Controllable Relighting, Diffusion Models, Conditional Image Generation, Latent Bridge Matching, Spatial-Illumination Alignment

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_2MatOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work tackles the challenging task of image relighting where a 2D image is relit to a given input lighting condition. The authors propose modelling illumination variation in 2D image by jointly considering regional luminance and spatial information. The contributions are twofold. First, they propose an Illumination Intensity Ranking (IIR) loss which instead of pixel-wise global information operates on image patches. It combines luminance and positional consistency and encourages the model to reproduce both the intensity and spatial distribution of illumination. Secondly, they build upon Latent Bridge Matching and propose Illumination Conditioned Latent Bridge Matching (IC-LBM) that injects illumination tokens through cross-attention layers instead of directly concatenating conditioning information with noisy latents. It is evaluated on Multi-Illumination in the Wild (MIIW) dataset. It consists of indoor scenes captured under multiple illumination condition with specular and diffuse light probes capturing the lighting conditions. They demonstrate improvements in reconstruction quality, illumination controllability and relighting through quantitative visualisation and evaluation.

Critical strengths
  1. Well motivated problem formulation: Pixel wise reconstruction objectives do not model spatially varying illumination in images which the paper motivates well. The proposed patch-wise illumination strategy provides a reasonable approach to addressing this issue.
  2. Clear methodology: The figures are well designed and the text makes the proposed framework easy to follow. The formulation of the IIR loss is well constructed.
  3. Detailed ablation studies of the loss function: The contribution of luminance and positional consistency, the influence  of patch count and the weighting strategy are well evaluated. These experiments provide useful evidence for understanding how different components of IIR contribute to reported improvements.
  4. Effective visual presentation: The visualisations of luminance heat maps and gradient directions help illustrate the limitations of existing approaches and the motivation behind the paper.
Critical weaknesses
  1. Related work section: It provides a broad overview of existing image relighting approaches but does not discuss the limitations or how the papers fail to address the research gap which the proposed work does. A more critical discussion would help position the paper. 

  2. Lack of existing illumination aware works: Several existing works explore the use of patch based illumination aware supervision such as :

    1. Intrinsic Image Harmonization (Guo et al., CVPR 2021)
    2. Deep Image-Based Illumination Harmonization (Bao et al., CVPR 2022)
    3. Zero-Reference Deep Curve Estimation (Guo et al., CVPR 2020)
    4. Intrinsic Image Decomposition via Ordinal Shading (Careaga & Aksoy, 2023)

    While these methods operate in different settings, they explore related objectives  involving spatial illumination consistency. The paper should address how the proposed IIR loss differs from these works and what advantages the ranking based formulation provides over simpler patch wise luminance reconstruction loss. 

    Additionally they should also include A Diffusion Approach to Radiance Field Relighting using Multi-Illumination Synthesis (Poirier-Ginter et al., 2024) which leverages a diffusion model to perform image relighting.

  3. Generalizability of illumination conditioning strategy: Given that one of core contributions is the improved illumination conditioning strategy, the evaluation is restricted to MIIW dataset where illumination is conditioned using specular chrome-probe images. It is unclear whether the conditioning representation generalised to other lighting conditions and datasets. For instance, DiLIghtNet consists of illumination using environment map. Does the conditioning provide similar improvements. 

  4. Generalizability and effectiveness of IIR loss: Given the second core contribution to be the IIR loss, it’s unclear whether the improvements are specific to the proposed IC-LBM framework or extends to other related works. Would incorporating IIR into the original LBM framework yield similar improvements? Does the improvement extend to other relighting datasets? Existing ablations establish benefits of individual components but further clarification of advantages of ranking formulation over simpler regional luminance supervision would strengthen the central contribution.

  5. Implementation details of baseline comparison: The paper lacks the details of how the existing approaches were setup and tested. Given they all have different illumination conditioning strategies, the paper lacks how the inputs to each model were setup.

Questions for authors
  1. Illumination-aware losses: How does the proposed IIR loss differ from existing illumination-aware objectives, such as those in Intrinsic Image Harmonization (CVPR 2021), Deep Image-Based Illumination Harmonization (CVPR 2022), and Zero-DCE (CVPR 2020)? In particular, what advantages does luminance ranking and positional consistency provide over a simpler patch-wise luminance reconstruction loss?
  2. Effectiveness of IIR independently of IC-LBM: Would incorporating the proposed IIR loss into the original LBM framework yield similar improvements? This would help clarify whether the benefits of IIR generalize beyond the proposed conditioning architecture.
  3. Generalizability of illumination conditioning: The proposed conditioning strategy is evaluated using chrome-probe images from MIIW. Would the same approach provide improvements with alternative lighting representations, such as environment maps or spherical harmonics? Is the improvement primarily attributable to the conditioning architecture or the specific lighting representation?
  4. Fairness of baseline comparisons: Could the authors clarify how target illumination conditions were provided to each baseline in Section 4.3? Were all methods trained or fine-tuned on the same MIIW split, and were any architectural modifications required to accommodate the illumination conditions?
  5. Generalizability across datasets: Both central contributions are evaluated exclusively on MIIW. Are there reasons to expect the proposed conditioning strategy and IIR loss to generalize to other relighting datasets with different illumination distributions? A discussion of the expected limitations would be helpful.
  6. Sensitivity to image resolution: Appendix A examines the effect of patch count at a fixed image resolution. How sensitive is IIR to changes in image resolution during training and inference? Does the optimal patch configuration depend on the spatial scale of illumination variations?
12 comments
AuthorsOpenReview ↗

Response to W1, W2

Thank you for your careful review and constructive suggestions. We have revised the related work section, added relevant references, and summarized the limitations of existing methods to clarify the advantages of our approach.

AuthorsOpenReview ↗

Response to Q1, W4

Thank you for your review and question. Intrinsic Image Harmonization (CVPR 2021) isolates illumination changes through intrinsic decomposition. Deep Image-Based Illumination Harmonization (CVPR 2022) uses physical rendering to characterize shadow relationships. Zero-DCE (CVPR 2020) focuses on contrast information for low-light enhancement. Our IIR is designed for controllable image illumination editing and models illumination variation directly in 2D images, without additional geometric conditions or intrinsic decomposition. With the luminance ordering of image patches as labels, IIR cleverly couples the comparison of luminance levels with the comparison of bright-region locations.

A simple patch-wise luminance loss does not explicitly model illumination variation. For controllable illumination editing, the simple patch-wise luminance error tends to favor an average optimum over rather than a solution that matches the condition. We have added a detailed comparison between IIR and simple patch-wise luminance losses to Appendix A. With the same architecture and training configuration, IIR outperforms patch-wise luminance losses in PSNR, SSIM, RMSE, and all three illumination controllability metrics. In particular, Target Recall@1 improves by 20.77%, showing that IIR guides outputs toward the specified condition rather than an average optimum.

AuthorsOpenReview ↗

Response to Q2, W4

Thank you for your comment. To isolate the contribution of IIR, we add it directly to the original LBM framework. With retaining the condition injection mechanism, IIR improves PSNR by 2.36 dB and SSIM by 0.0993, reducing RMSE by 0.0399. The additional results have been included in the updated Table 1.

Reviewer_Th9gOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents IC-LBM, a framework for controllable image illumination editing that integrates spatial–illumination alignment into latent bridge matching. The proposed IIR loss jointly models regional luminance and spatial consistency. Experiments on MIIW demonstrate improvements in image reconstruction quality and illumination controllability.

Critical strengths

1.The proposed IIR loss provides explicit supervision for regional illumination distribution and spatial arrangement. 2.Quantitative results and ablation studies demonstrate the effectiveness of the proposed components.

Critical weaknesses

See questions below.

Questions for authors
  1. The methodological novelty is limited. How does IIR fundamentally differ from existing luminance distribution matching and spatial alignment objectives?
  2. The inference process is insufficiently clarified. How is the stochastic term handled in the single-step Euler update at (t=0)?
  3. In Eq. (7), the luminance sequence is described as sorted, whereas (P_{m,k}) represents the correspondence between spatial patches and luminance ranks. Could the authors clarify how the original spatial ordering is retained when computing the soft correspondence?
  4. The conclusion acknowledges local color shifts and limited recovery of fine details. How does the method balance illumination controllability against color fidelity and structural preservation? In addition, what are its inference latency and memory requirements?
8 comments
AuthorsOpenReview ↗

Response to Q1

Thank you for your review and constructive comments. The discussion of these distinctions was insufficient, and we have added a brief clarification to the related work section. We use luminance ranks to identify image patches and explicitly model illumination variation with differences in luminance levels and bright-region locations. IIR turns the target illumination into a directly optimizable objective. For each luminance rank, it jointly supervises whether the corresponding region reaches the required luminance and whether that rank occurs at the required location.

Compared with other objectives for matching luminance distributions, IIR is designed for controllable image illumination editing and models illumination variation directly in 2D images, without requiring additional geometric conditions or intrinsic decomposition.

Compared with simple spatial alignment, IIR enforces constrains on both the luminance distribution and bright-region locations under substantial illumination variation, preventing globally averaged predictions. To verify this, we conduct a detailed comparison between IIR and simple patch-wise luminance losses. IIR achieves better PSNR, SSIM, RMSE, and all three illumination controllability metrics. In particular, Target Recall@1 improves by 20.77%, indicating that IIR guides the model toward outputs that match the condition rather than an average optimum. Detailed experimental results have been added to the updated Appendix A.

AuthorsOpenReview ↗

Response to Q2

Thank you for your careful review. No stochastic term is added during the single-step Euler update at t = 0. We follow the training and single-step inference procedures of LBM. The stochastic term in LBM is the diffusion term of the Brownian bridge SDE, rather than a conditioning signal. During training, the bridge endpoints are fixed to the source and target images, and use four equally spaced discrete timesteps = {0, 0.25,0.5,0.75} (corresponding to four levels of stochastic perturbation) to learn the drift from the source to the target domain. The stochastic term promotes a smoother and more robust drift. At t=0, the training input is z0z_{0}, and the drift target becomes z1−z0z_{1}-z_{0}. Training has explicitly supervised the displacement used by our one-step endpoint predictor.

AuthorsOpenReview ↗

Response to Q3

Thank you for your careful examination. We apologize for omitting the tilde over LmL_{m} in Eq. (7). In this equation, L~k\tilde{L}_{k} denotes the luminance at rank k, and Pm,kP_{m,k} is the soft probability of assigning the patch at rank m to rank k; both indices refer to ranks. The phrases “m-th spatial patch” and “over spatial patches” in Section 3.2.1 were misleading and have been revised to “ranked patches”. We rank the patches in the predicted and target images separately by luminance. Each patch retains its original spatial coordinates. For the patch at rank m, posm\mathrm{pos}_{m} denotes its position in the original image. The sorted luminance values and their spatial coordinates always remain in one-to-one correspondence.

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

The manuscript addresses controllable image illumination editing by proposing spatial–illumination alignment: rather than pixel-wise matching, it divides images into patches, ranks patches by luminance, and compares luminance values and spatial locations at corresponding ranks.

The method couples an Illumination Intensity Ranking (IIR) loss with Illumination-Conditioned Latent Bridge Matching (IC-LBM), where a chrome-probe condition is encoded into tokens and injected through cross-attention after condition-confirmed initialization.

On the MIIW benchmark, the manuscript reports PSNR 21.17, RMSE 0.0932, and SSIM 0.7701, outperforming LBM, LumiNet, IC-Light, and LDM under the reported metrics.

Critical strengths
  1. A simple, cheap, plug-in objective that targets a real failure mode of relighting models. Pixel/perceptual losses in bridge/diffusion relighting often produce outputs that stay close to the source lighting. The paper's motivation is sound: supervising where the brightest regions should be and how bright they should be at a coarse (8×8) scale directly addresses this. The loss is differentiable, adds negligible compute, and could be attached to other relighting backbones.

  2. Clear evidence that the conditioning mechanism matters. Table 1 shows that naive cross-attention without CCI is only marginally better than no condition (16.95 vs. 16.47 dB), while cross-attention + CCI gives 19.61 dB. This is a useful, if somewhat engineering-level, finding about condition collapse in single-step bridge models.

Critical weaknesses
  1. A simple, cheap, plug-in objective that targets a real failure mode of relighting models. Pixel/perceptual losses in bridge/diffusion relighting often produce outputs that stay close to the source lighting. The paper's motivation is sound: supervising where the brightest regions should be and how bright they should be at a coarse (8×8) scale directly addresses this. The loss is differentiable, adds negligible compute, and could be attached to other relighting backbones.

  2. Clear evidence that the conditioning mechanism matters. Table 1 shows that naive cross-attention without CCI is only marginally better than no condition (16.95 vs. 16.47 dB), while cross-attention + CCI gives 19.61 dB. This is a useful, if somewhat engineering-level, finding about condition collapse in single-step bridge models.

Critical Weaknesses*

  1. The benefit of the IIR luminance term over simpler pooled-luminance supervision is not sufficiently isolated. Under the spatial-patch interpretation of LmL_m in Eqs. 7-8 , both ordered quantities in Eq. 8 use the same target correspondence PgtP^{\mathrm{gt}}. Let rm=Lmpred−Lmgtr_m=L_m^{\mathrm{pred}}-L_m^{\mathrm{gt}}. Then Eq. 9 is exactly
L_{\mathrm{lum}} &=\sum_{k=1}^{N}w_k\left(\sum_{m=1}^{N}P^{\mathrm{gt}}_{m,k}r_m\right)^2\\ &=r^\top A r,\qquad A=P^{\mathrm{gt}}\operatorname{diag}(w)(P^{\mathrm{gt}})^\top. \end{aligned}$$ Thus, $L_{\mathrm{lum}}$ is a target-conditioned quadratic loss on spatially pooled luminance residuals, with coefficients fixed for a given target and temperature. The predicted correspondence $P^{\mathrm{pred}}$ does not enter this term. Assuming distinct target patch luminances, $P^{\mathrm{gt}}$ approaches a permutation matrix as $\tau\to0^+$, and the loss reduces to target-rank-weighted MSE on the $8\times8$ pooled luminance grid: $$L_{\mathrm{lum}}\to\sum_{m=1}^{N}w_{\operatorname{rank}_{\mathrm{gt}}(m)} \left(L_m^{\mathrm{pred}}-L_m^{\mathrm{gt}}\right)^2.$$ With tied target luminance values, the limiting loss instead aggregates residuals within the tied groups. At finite temperatures, the loss can couple residuals across patches according to target-luminance similarity. In the uniform-correspondence limit of Eq. 14, it reduces to the squared mean-brightness error, since $\sum_k w_k=1$. This analysis shows that $L_{\mathrm{lum}}$ does not directly compare independently ranked predicted and target luminance values. It does not establish that the complete IIR objective is equivalent to a patch-wise loss: $L_{\mathrm{pos}}$ also depends on the predicted correspondence and compares rank-conditioned spatial centroids. Nevertheless, the experiments do not show how much of the improvement is attributable to the ranking formulation rather than to ordinary pooled luminance supervision. A direct comparison with plain and target-rank-weighted pooled luminance losses, together with a full-training ablation of the positional term, would help isolate these contributions. 2. The baseline protocols require clarification, and the full-system improvement should be distinguished from the incremental benefit of IIR. (a) In the ablation sequence in Table 1, changing the conditioning mechanism from concatenation to cross-attention with CCI increases PSNR from 17.58 to 19.61 dB (+2.03 dB), and adding IIR increases it from 19.61 to 21.17 dB (+1.56 dB). These are sequential ablation gains rather than independent causal contributions. The 3.6 dB improvement discussed in Sec. 4.3 is therefore a full-system gain over LBM; the discussion should more clearly distinguish this from the incremental gain from IIR. Table 3 does provide meaningful evidence of an additional benefit from IIR, particularly the increase in Target Recall@1 from 26.53\% to 54.39\%. (b) IC-LBM is initialized from an LBM relighting checkpoint and trained on MIIW using chrome-probe conditioning at $512\times512$. The manuscript does not provide sufficient details about the checkpoints, training or fine-tuning, conditioning inputs, preprocessing, and inference settings used for LDM, LumiNet, and IC-Light. Their native conditioning interfaces differ: LumiNet uses a target-lighting reference image and supports cross-scene transfer, while IC-Light has multiple variants with different supported inputs. LumiNet's published training data already include MIIW, so the relevant question is how the reported baseline was run under this paper's specific protocol. LDM (15.92 dB) and LumiNet (16.07 dB) improve only slightly over the Source baseline (15.81 dB), and all compared methods have lower SSIM than Source. These observations make the missing implementation details especially relevant, but do not by themselves demonstrate inadequate baseline adaptation. Without those details, it is difficult to assess the comparability of the results.
Questions for authors
  1. Could you compare IIR with plain L2L_2 and L1L_1 losses on luminance maps average-pooled to an 8×88\times8 grid, using the same IC-LBM backbone, cross-attention with CCI, training data, and training schedule? Please use a comparable budget for loss-weight tuning and specify the loss reductions and selected coefficients. A target-rank-weighted pooled MSE baseline would further help isolate the effects of soft correspondence and the positional term. Please report PSNR, RMSE, SSIM, and the illumination-controllability metrics in Table 3.

  2. For LDM, LumiNet, and IC-Light, please specify the exact checkpoint, any MIIW training or fine-tuning, the conditioning input and how it was constructed, the resolutions used for training, inference, and evaluation, and the inference settings. How was the same target illumination represented through each method's supported conditioning interface? Could you also include Latent Intrinsics under a clearly specified and comparable MIIW evaluation protocol, including whether any output color correction is applied?

  3. What luminance normalization and numerical range are used in Eq. 7? Since correspondence sharpness depends on luminance differences relative to τ\tau, could you report the distribution of patch luminance differences and the average column-wise normalized entropy of both PgtP^{\mathrm{gt}} and PpredP^{\mathrm{pred}} over the training schedule? Do the assignments become sufficiently concentrated as τ\tau decreases from 1.0 to 0.1? If both correspondences approach uniform distributions, all rank-conditioned spatial centroids approach the same grid center and LposL_{\mathrm{pos}} approaches zero. A temperature ablation would help establish whether this is a practical concern.

  4. All reported relighting evaluations use MIIW, which captures each scene under 25 predefined flash orientations. Could you provide results on another relighting benchmark or on indoor scenes with different lighting configurations to assess whether the benefits of IIR and CCI generalize beyond this capture setup?

  5. Please clarify the indexing of LpredL^{\mathrm{pred}} and LgtL^{\mathrm{gt}}. Sec. 3.2.1 states that the patches are sorted in descending order of luminance to obtain LpredL^{\mathrm{pred}} and LgtL^{\mathrm{gt}}, which suggests that these are sorted luminance vectors. However, Eqs. 7 and 8 and the definition of Pm,kP_{m,k} appear to require LmL_m to denote the luminance of spatial patch mm, with L~k\tilde L_k denoting the kk-th ranked value. Is the description of LpredL^{\mathrm{pred}} and LgtL^{\mathrm{gt}} as sorted vectors a notation error? If the vectors in Eq. 8 were already sorted, the correspondence weights and luminance values would use inconsistent spatial and rank indices.

14 comments
AuthorsOpenReview ↗

Response to Q1,W1

Thank you for the detailed analysis and constructive comments. We agree that, for a fixed target and temperature, the luminance term can be written as a target-conditioned quadratic form. The key clarification concerns the residual vector. There is a notation error in Eq. (7), and we detailed explain it in the Response to Q5, W1. Our implementation compares independently sorted luminance values, rather than luminance values at corresponding spatial locations. IIR differs fundamentally from weighted L1/MSE losses computed over fixed spatial patches.

We keep the IC-LBM backbone, cross-attention with CCI, training data, and training schedule unchanged, and replace only the auxiliary luminance supervision. We divided the predicted and target images to 8 × 8 luminance patches, and conduct experiments on patch-wise L1 and MSE losses and weighted versions by ground-truth luminance ranks. The unweighted baselines average the errors over the 64 corresponding spatial locations, while the weighted baselines adopt the same patch-weighting strategy as IIR. To give the losses the same effective weight as IIR, all auxiliary luminance term weights set to 5 with mean reductions. We evaluate PSNR, SSIM, RMSE, and illumination controllability metrics. Plain L1 and L2 on luminance do not gain the same improvements as IIR. Compared with standard supervision on average-pooled luminance patches, IIR achieves better PSNR, SSIM, and RMSE, reduces Mean Luminance Error by 0.005, and improves Bright Region IoU by 6.2% and Target Recall@1 by 20.77%.

For condition-guided illumination editing, supervision based on spatial average pooling tends to the average optimum on the overall dataset rather than the result that best matches the illumination condition. In particular, the substantial improvement in Target Recall@1 indicates that IIR guides predictions toward the specified condition rather than an average optimum. The results demonstrate the need to explicitly model illumination variation for controllable image illumination editing. The updated appendix A provides the detailed experimental setup and complete results.

AuthorsOpenReview ↗

Response to W2(a)

Thank you for your comment. We have added an ablation to Table 1 that introduces IIR without CCI. IIR still improves performance when condition injection is less effective, supporting its effectiveness for image illumination editing. However, since the editing is conditioned on a single reference image, removing effective condition injection results lower performance than the full model with CCI.

AuthorsOpenReview ↗

Response to Q2, W2(b)

Thank you for your suggestions. Space constraints prevented us from detailing the baseline configurations in the main text. We have now added a brief description to Section 4.3.

All methods are trained and evaluated at a resolution of 512 × 512 on a single A800 GPU. We use 30,000 training steps, an exponential moving average (EMA) decay of 0.999, a learning rate of 1 × 10⁻⁵, and the AdamW optimizer with β = (0.9, 0.999). For LDM, we use the U-Net and VAE weights released by DiffHarmony++ (Zhou P et al., ACM MM 2024, a method for foreground relighting). Following the style-conditioning strategy recommended in the original paper, we encode the chrome probe image with a visual encoder and inject the resulting features into the U-Net through cross-attention, without an additional CCI. For LumiNet, we use the officially released cross-scene relighting weights trained on MIIW. We implement the training launch script missing from the official release following the paper. We replace the cross-scene reference image with a chrome probe image and use Latent Intrinsics for illumination control. To maintain a fair comparison, no additional color correction is applied. Our comparison is fair. For LumiNet evaluated quantitatively on MIIW, our reported results (SSIM: 0.6871; RMSE: 0.1682) are comparable to those reported in the original paper (SSIM: 0.647; RMSE: 0.180). For IC-Light, we fine-tune the officially released image-to-image relighting weights iclight_sd15_fbc.safetensors on MIIW. Following the official configuration, we set the MSE and light-transport consistency (LTC) loss weights to 1.0 and 0.1, respectively. During fine-tuning, we leave the text prompt empty and use the chrome probe image as the illumination condition. Following the official implementation, we inject the condition by directly concatenating the condition image with the source image along the channel dimension. For LBM, we fine-tune the officially released pretrained relighting weights on MIIW, using the chrome probe image as the condition. We retain the direct concatenation scheme for condition injection in the original method. All other training and inference settings are identical to those of IC-LBM.

Regarding the low SSIM values raised in W2(b), we recomputed the source-image SSIM with the data range of [0, 1] and confirmed the reported value. To examine why the source images have relatively high SSIM, we consulted the September 2025 corrected version of LumiNet (https://arxiv.org/pdf/2412.00177v3). Table 1 of the version reports an SSIM of 0.438 for the source images on MIIW and 0.440 for LumiNet without color correction, with metrics computed on the data range of [-1, 1]. We measure the source-image SSIM of 0.4508 with the data range of [-1, 1]. Source images indeed exhibit relatively high average SSIM on MIIW. When pixel values are substantially modified to match the target illumination, relighting methods need to improve the preservation of structure and fine texture details. Nevertheless, all methods substantially reduce RMSE and improve PSNR compared to the source images, indicating matching for illumination toward the target.

Comments on the paper

3 comments

Dear reviewers,

Thanks for your continued efforts. The authors have provided a rebuttal to the initial reviews. Please review it carefully, along with all other reviews and the corresponding author responses, and consider whether you have any further questions for the authors or whether the rebuttal has addressed your concerns. There are only a few days remaining before the discussion period closes, so please use this time to seek any further clarifications or raise any additional questions about this paper.

Best,

AC

AuthorsOpenReview ↗

Summary of revised PDF

We made the following main changes in the revised PDF:

  1. We corrected the notation in Eq. (7) and revised the wording to clarify the image patches ranked by luminance.

  2. We added an ablation to Table 1 to isolate the contribution of IIR.

  3. We added relevant studies to the Related Work section and explained the advantages of our approach over prior methods to highlight our contribution.

  4. In Section 4.3, we briefly describe the configurations used to reimplement the competing methods and clarify how we ensured a fair comparison.

  5. We added an experimental comparison between IIR and plain L1 and L2 luminance losses to the Appendix A to show how IIR differs from the plain losses in performance.

  6. We added an analysis to the Appendix A showing how the soft sorting temperature affects the distribution of soft luminance correspondences during training. We also replaced the temperature schedule with a fixed temperature in each ablation to evaluate its effect on performance.

  7. We added evaluation results on the BigTime benchmark to the Appendix B to confirm the method’s generalization across datasets.

  8. We added training and evaluation results on the VIDIT benchmark to the Appendix B to confirm the effectiveness of the condition injection mechanism and the model’s ability to perform cross-scene relighting.

Dear Area Chairs and Reviewers,

We propose an approach explicitly models illumination changes by spatial–illumination alignment. Our method cleverly conducts luminance ranks as labels for image patches and describes both regional luminance levels and shifts in the locations of bright regions. The formulation effectively constrains illumination redistribution without requiring extra geometric conditions or intrinsic decomposition.

During the rebuttal, we clarify some issues and conduct extensive additional experiments to provide further support for our claims with the help of the reviewers' comments. Existing methods for image illumination editing rely on various forms of physical decomposition or latent representation learning, and lack an objective that explicitly models illumination variations directly in 2D images. Our proposed objective substantially improves relighting controllability at minimal computational cost. Beyond image quality, a central question in controllable image editing is constraining a model to produce outputs that match the specified condition. Our work initiates an inquiry into the controllability and proposes a novel solution.

All the three reviewers conducted a thorough review of our work and ultimately expressed their recommendation and approval. We have put considerable care into this work. We believe the novelty of our approach and its contribution to the community are clear. We respectfully look forward to a careful consideration. ​

​Finally, we would like to thank the responsible AC for their timely reminders and assistance. We are grateful to all the reviewers for their detailed analysis of our work and patient guidance. We sincerely thank Reviewer wMYa for the careful review and detailed guidance. We also appreciate Reviewer Th9g for the continued attention, discussions, and encouragement. We are thankful to Reviewer 2Mat for taking time out of the busy schedule to respond to our efforts.

The review process has been very rewarding for us. We look forward to exchanging ideas with more researchers. We sincerely apologize for the inconvenience caused by emails over the past days.

Wishing you all the best!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.