acceptodds
Under review as a conference paper at ICLR 2027

Drawing the Light of Image via Material Decoupled Illumination Representation

Abstract

Controllable image relighting is an important problem in image editing, and hand-drawn scribbles provide an intuitive interface for specifying the desired illumination. However, existing methods do not establish a consistent and effective mapping between scribble inputs and relighting results, limiting their ability to control illumination intensity, chromaticity, and complex spatial distributions. We address this limitation by introducing a material-decoupled illumination representation, termed the Lumi Map, which establishes an explicit mapping between user scribbles and the resulting illumination, thereby improving both relighting accuracy and controllability. Specifically, we use a renderer to synthesize source image-Lumi Map-relit image triplets and train the model to predict the target relighting result conditioned on the Lumi Map. To mitigate the domain gap introduced by synthetic data, we further perform reconstruction training on real relighting pairs, improving the model's generalization to real-world images. Finally, we present Dior-Light, an image relighting method controlled by hand-drawn strokes. Extensive experiments demonstrate that our method outperforms existing approaches in relighting accuracy and enables effective control over illumination intensity and chromaticity on in-the-wild images.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image relight, Image Editing

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_EExJOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper is on the topic of controllable image relighting. The authors first introduce a new lighting representation for scenes, Lumi Map, that is material-decoupled and more physics-oriented. This allows for more accurate lighting modeling from synthetic data. Then, a two-stage training strategy is designed to mitigate the domain gap between synthetic data and real data training. Together, the approach, called Dior-Light, allows users to control relighting results on real images using hand-drawn strokes. The method is evaluated on synthetic and real datasets with other relighting methods and alternative lighting representations with reasonable performance.

Critical strengths
  • The proposed Lumi Map representation for lighting has a good theoretical motivation and practical implementation. It models hard-to-model (inter)reflection and shadow effects in complex scenes, better than irradiance and shading that are commonly used in the inverse graphics work (Figure 2). This representation could be useful not only for relighting but also for other relevant research works that require realistic modeling of lighting.
  • I found the two-stage training strategy to bridge the domain gap between synthetic and real data to be useful for inverse problems. This allows researchers to develop advanced models in the synthetic environment and then transfer them to real data. Although the current setting requires real training pairs (source and target), it will be interesting to explore a single real-image setting with a bootstrapped estimated lighting map and some iterative refinement process.
Critical weaknesses
  • The stroke is a key control element for the relighting application. Currently, the way of obtaining the strokes in synthetic data is to use free-form stroke masks to sample from the lighting map (Figure 12 in the supplemental material). This is quite different from how an actual user would use strokes to express their relighting need. Also, there will be no actual (target) lighting map to sample from, so the inference-time stroke control will always be suboptimal for the trained model. This is a significant gap. Btw, I had trouble finding the details in the main text about how the strokes are obtained from the Lumi Map and eventually found it in the supplemental material. Such important information should definitely be in the main text.
  • There are some weaknesses in the quantitative evaluations. 1) While the full relighting results of the proposed Dior Light outperform other methods, Dior w/o stage 2 has a very similar performance to other lighting representations (esp. Irradiance) in Tables 2 and 3. This undermines the significance of the new representation in practice: maybe modeling the interreflections doesn't matter that much? 2) The quantitative or qualitative comparison can be done with more relighting methods (already mentioned in the related work), even if they are not designed for scribbles. 3) With the two-stage training strategy, it is unclear how much real data is needed to properly tune the model to be more generalizable.
Questions for authors
  1. For the stroke mechanism, how can one design the strokes in the synthetic environment to fit more closely to a real user's use case?
  2. Can the authors clarify the performance of Dior w/o stage 2 vs other illumination representations?
  3. How much real data is needed for the two-stage training strategy to be successful?
4 comments
AuthorsOpenReview ↗

Response to Reviewer EExJ

Response to Reviewer EExJ

We thank the reviewer for the thoughtful review and the encouraging comments on the representation and the training strategy.

W1(a) / Q1. Synthetic strokes and hand-drawn controls

The primary purpose of our automatic stroke sampling is to train the model to infer full-image illumination from sparse lighting cues, not to replicate the shapes of human-drawn strokes. Nevertheless, we agree that the difference between sampled and hand-drawn strokes should be examined. We therefore manually annotated hand-drawn strokes for 150 samples, 50 from each of the three benchmarks, and compared the performance obtained with hand-drawn and automatically sampled strokes on the same samples. The results show that this difference does not lead to a noticeable performance drop. Despite covering fewer pixels, hand-drawn strokes achieve comparable performance on ReNe and better performance on MultiIllum and SynthBench. A likely reason is that human annotators draw on prior knowledge of lighting structure and place strokes in more informative regions. This demonstrates the robustness of our method to the distribution of user strokes. Qualitative results are shown in Fig. 9 (right) of the revised main text, and quantitative results are reported below.

We select 150 samples from three benchmarks for manual stroke annotation (50 per benchmark). Cover. is the mean fraction of pixels occupied by strokes.

Table R3. Quantitative analysis of auto-sampled and hand-drawn strokes

BenchmarkStroke StyleCover. (%)RMSE ↓PSNR ↑SSIM ↑LPIPS ↓
ReNeAuto-sampled30.6122.8021.210.74540.1812
ReNeHand-Drawn8.1923.2521.070.74600.1807
MultiIllumAuto-sampled32.9637.1817.330.73010.2690
MultiIllumHand-Drawn14.1632.6118.380.76230.2419
SynthBenchAuto-sampled9.4430.6519.090.84480.1618
SynthBenchHand-Drawn8.5628.8519.590.84610.1557

Although sparse sampling enables the model to infer global illumination from local cues, scribble-driven relighting would further benefit from a human–computer interaction perspective. In particular, studying how users actually draw lighting strokes, including their shapes and colors, could inform the design of training-time stroke sampling. We leave this to future work.

W1(b). Stroke construction details in the main paper

To improve readability, we have moved the key details of the sparse lumi map sampling from Supplementary C.2 to Sec. 4.1 of the main text, while the full implementation details remain in the supplementary material.

W2(1) / Q2. Contributions of the representation and Stage 2

All representation variants in Table 2 are trained with Stage 2. This additional training on real data explains why some of these variants, particularly the irradiance variant, outperform Dior w/o Stage 2 on ReNe and MultiIllum. The representation ablations should therefore be compared with the full model. Under the same training setting, Dior with the lumi map achieves stronger performance, as shown in Tables 2 and 3.

W2(2). Comparison with additional relighting methods

In addition to the comparisons with RGB↔X [1] and Lume-Palette [2] already included in Supplementary D.2, we have added comparisons with DiffusionRenderer [3] and DiLightNet [4]. All methods are evaluated using their official implementations. For DiffusionRenderer and DiLightNet, the target lighting is estimated directly from the target image using the third-party DiffusionLight-Turbo estimator [5].

Table R4. Extended results inventory

MethodReNe RMSE ↓ReNe PSNR ↑ReNe SSIM ↑ReNe LPIPS ↓MultiIllum RMSE ↓MultiIllum PSNR ↑MultiIllum SSIM ↑MultiIllum LPIPS ↓SynthBench RMSE ↓SynthBench PSNR ↑SynthBench SSIM ↑SynthBench LPIPS ↓
DiffusionRenderer41.0015.990.56730.375462.1812.580.47320.577375.6311.200.36860.6651
DiLightNet46.2914.910.29060.492682.1610.070.39400.589259.6912.800.41290.5861
RGB↔X33.2817.750.43950.411565.1412.170.48880.511990.049.440.15470.8667
Lume-Palette24.5820.500.66620.325744.3916.140.73540.357053.1214.850.65040.2735
Dior-Light (ours)20.0322.380.75420.174033.4918.210.76590.227524.8921.120.86370.1166
AuthorsOpenReview ↗

Response to Reviewer EExJ

W2(3) / Q3. Real-data requirements of Stage 2

We thank the reviewer for this question. Our method does not aim to minimize the amount of real data. To ensure sufficient training, the main model uses all available real data (36K images from MultiIllum and ReNe) and trains Stage 2 for 5,000 steps with a batch size of 16.

To investigate the generalization potential of Stage 2 raised by the reviewer, we conducted a small-scale study of how the amount and source of real data affect generalization under a fixed training budget. Starting from the same Stage-1 model, we trained Stage 2 for 2,000 steps under four settings: 100, 500, and 1,000 MultiIllum (MI) samples, and 1,000 ReNe samples. We evaluated each setting with paired metrics on 300 held-out pairs (150 from ReNe and 150 from MultiIllum), with equal weight for the two benchmark means, and with Elo ratings on the in-the-wild (ITW) test set.

Table R5. Analysis of real-data requirements

Training settingPSNR ↑SSIM ↑LPIPS ↓ITW Elo ↑
Stage 1 (synthetic only)17.500.69930.26201057.30
+ MI 10018.750.74880.23281264.61
+ MI 50019.670.75650.22521279.59
+ MI 1,00018.840.74660.23101149.50
+ ReNe 1,00019.120.73030.23171238.83

All settings improve over the Stage-1 model on paired benchmark and in-the-wild benchmark, with ITW Elo gains ranging from 92 to 222 points. We further observe:

  1. Under a fixed training budget, generalization does not improve monotonically with the amount of real data. Increasing the number of MI samples from 100 to 500 improves all four metrics, and the 500-sample setting achieves the best results on every metric. Increasing the number further to 1,000 degrades all four metrics relative to the 500-sample setting; in particular, the ITW Elo drops from 1279.59 to 1149.50.

  2. The source of real data affects ITW generalization. With the same number of samples (1,000), ReNe achieves a higher ITW Elo than MultiIllum (1238.83 vs. 1149.50) and a higher PSNR (19.12 vs. 18.84). MultiIllum achieves a higher SSIM (0.7466 vs. 0.7303), and the two settings have comparable LPIPS (0.2310 vs. 0.2317).

In summary, Stage 2 is data-efficient: with only 100 real samples, PSNR improves by 1.25 dB and the ITW Elo by more than 200 points over the Stage-1 model. Under a fixed training budget, a moderate amount of real data (500 samples in this study) performs best, and the source of the data further affects ITW generalization. A systematic study of data scaling is beyond the scope of this work. We also appreciate the reviewer's suggestion of bootstrapping lighting estimates from single real images, which could further reduce the need for paired real data. We will explore this direction in future work.

References

[1] Zheng Zeng et al. RGB↔X: Image Decomposition and Synthesis Using Material- and Lighting-aware Diffusion Models. ACM SIGGRAPH Conference Papers, 2024.

[2] Chenjian Gao, Linning Xu, and Tianfan Xue. Decoupled Illumination Priors for Spatially Controllable Multi-View Indoor Scene Relighting. arXiv:2607.08879, 2026.

[3] Ruofan Liang et al. Diffusion Renderer: Neural Inverse and Forward Rendering with Video Diffusion Models. CVPR, 2025.

[4] Chong Zeng et al. DiLightNet: Fine-grained Lighting Control for Diffusion-based Image Generation. ACM SIGGRAPH Conference Papers, 2024.

[5] Worameth Chinchuthakun et al. DiffusionLight-Turbo: Accelerated Light Probes for Free via Single-Pass Chrome Ball Inpainting. arXiv, 2025.

Thanks for providing the additional results and analysis. They look good to me. I don't have any further requests or questions for the authors. I recommend acceptance.

We sincerely thank the reviewer for the positive feedback and for recommending acceptance. Your constructive comments have greatly helped us improve the paper.

Reviewer_D8XkOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper aims to achieve accurate and controllable scribble-driven relighting. It introduces a material-decoupled illumination representation called Lumi Map. The authors then train a Lumi Map estimator and a relighting module using rendered data. To achieve photorealistic results, they further fine-tune the model on real data using the trained Lumi Map estimator from the first stage. Experiments show that the proposed method outperforms previous approaches and produces visually high-quality results.

Critical strengths
  • A novel lighting representation, Lumi Map, that decouples materials and enables controllable scribble-driven relighting.
  • Experiments show that the proposed method, equipped with Lumi Map, outperforms existing approaches.
Critical weaknesses
  • The comparison with other methods on in-the-wild images appears unfair, as there is no standard benchmark for scribble-driven relighting. I suggest that the authors select one or two state-of-the-art scribble-driven relighting methods and fine-tune them on the same dataset to enable a fairer comparison.
Questions for authors
  • Typo: Line 297—there is a missing space between “I” and “is.”
  • Could the authors show examples of the Lumi Map after adding lighting scribbles to help readers better understand how scribble editing affects the Lumi Map?
  • It would be helpful to include additional comparisons with scribble-driven relighting methods, such as LightPainter [1].

[1] Mei et al., “LightPainter: Interactive Portrait Relighting with Freehand Scribble,” CVPR 2023.

3 comments
AuthorsOpenReview ↗

Response to Reviewer D8Xk

Response to Reviewer D8Xk

We thank the reviewer for the positive evaluation and the helpful suggestions.

W1. Fine-tuning a competing method and comparison fairness

We selected ScribbleLight for this comparison because it is a recent scribble-driven relighting method (CVPR 2025) with publicly available official training code, which allows it to be retrained on our data. We fine-tuned ScribbleLight for 20,000 steps with a batch size of 32, using its official training code and the same training data as our method. Its training inputs, including the lighting conditions, were constructed with ScribbleLight's default data processing pipeline. The quantitative results after fine-tuning are reported below. Although fine-tuning improves the performance of ScribbleLight, it still lags considerably behind our method.

Table R2. ScribbleLight official vs Fine-Tune under the paper evaluation protocol

MethodReNe RMSE ↓ReNe PSNR ↑ReNe SSIM ↑ReNe LPIPS ↓MultiIllum RMSE ↓MultiIllum PSNR ↑MultiIllum SSIM ↑MultiIllum LPIPS ↓SynthBench RMSE ↓SynthBench PSNR ↑SynthBench SSIM ↑SynthBench LPIPS ↓ITW Elo ↑
ScribbleLight47.4614.640.28190.563256.2913.390.53550.583476.1010.900.39450.5356409.56
ScribbleLight (Fine-Tune)37.67 (−9.79)16.68 (+2.04)0.2583 (−0.0236)0.5258 (−0.0374)51.08 (−5.21)14.22 (+0.83)0.5774 (+0.0419)0.4966 (−0.0868)64.32 (−11.78)12.41 (+1.51)0.4325 (+0.0380)0.4537 (−0.0819)583.66
Dior-Light (ours)20.0322.380.75420.174033.4918.210.76590.227524.8921.120.86370.11661410.65

Q1. Missing space at Line 297

Thank you for pointing out the spacing issue. We have corrected the spacing around “I is” in the training description.

Q2. Visualizing the Lumi map

We have added Fig. 9 to show hand-drawn controls, the resulting relit images, and the estimated illumination distributions before and after relighting. These additions clarify the relationship between the strokes, the lighting representation, and the output.

Q3. Comparison with LightPainter

We discuss the relationship between our method and LightPainter in the main text. Since LightPainter has not released its model weights or training code, we cannot compare with it directly. Our Phong ablation, however, can serve as a strong proxy for LightPainter. It adopts the same Phong-based scribble representation as LightPainter while building on a stronger pretrained backbone. We have clarified this point in the revised main text.

I thank the authors for providing the additional results, which address my questions. I maintain my recommendation for acceptance.

Thank you very much for your consistent support and positive assessment of our work throughout the review process.

Reviewer_6YXyOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Dior-Light, a scribble-guided image relighting method that introduces a material-decoupled illumination representation called Lumi Map to enable controllable illumination intensity, chromaticity, and spatial distribution. The authors adopt a two-stage training strategy combining synthetic rendering supervision and real-image reconstruction, demonstrating improved relighting accuracy and controllability on both benchmark datasets and in-the-wild images.

Critical strengths
  1. Each component of the proposed method is well designed and has a clear and reasonable motivation.

  2. The visual results are impressive, particularly those in Figure 8, which demonstrate effective control over both the chromaticity and intensity of illumination.

  3. The paper is well written, clearly structured, and easy to follow.

Critical weaknesses
  1. Alternative representations for scribble-guided relighting. If the user-drawn scribbles could be mapped to an environment map, wouldn't existing relighting methods conditioned on environment maps also be able to achieve scribble-level lighting control? I wonder why the authors chose to introduce the Lumi Map instead of learning a mapping from scribbles to an existing illumination representation, such as an environment map. Is this because learning such a mapping is more challenging?

  2. Reliability of the estimated Lumi Map in Stage 2. In Stage 2, the Lumi Map is estimated directly from the target image and then used to reconstruct the same target. How do the authors ensure that this estimated representation remains material-decoupled and does not encode target-specific appearance information? Could the authors provide visualizations or additional analyses of the estimated Lumi Maps on real images?

  3. Details of the baseline comparisons. When comparing against multimodal image editing methods, did the authors explicitly explain in the prompts that the scribbles represent desired lighting conditions, or were the scribbles provided separately as conditioning inputs? The paper states that "each baseline is evaluated using the conditioning format best suited to its design." Could the authors provide more details about the exact conditioning inputs and prompts used for each baseline?

  4. Quantitative evaluation of illumination controllability. The results in Figure 8 are very impressive and demonstrate promising control over illumination intensity and chromaticity. However, a quantitative evaluation on synthetic scenes with known target illumination would further strengthen the claim of precise illumination control.

  5. Minor suggestion. The related work section could be more comprehensive by discussing additional recent diffusion-based relighting methods from 2026, such as UniRelight, PI-Light, and HorizonRelight.

Questions for authors

See weakness.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 6YXy

Response to Reviewer 6YXy

We thank the reviewer for the positive assessment and the constructive comments. We address each point below.

W1. Lumi Maps versus environment maps

For the reviewer's final question, our choice was guided by the requirements of stroke-based relighting. We adopt the lumi map for three reasons.

(1) End-to-end learning. We aim to learn, in an end-to-end manner, the mapping from the source image and the lighting condition to the relit image, which avoids error accumulation from intermediate results. We consider predicting a complete environment map from sparse hand-drawn strokes to be a considerably harder task, and errors in this prediction would accumulate in the subsequent relighting stage. Our model avoids this step by mapping sparse strokes directly to the relit image. Dense lumi maps are used during training to help the model learn this mapping implicitly. At inference, no dense lumi map needs to be explicitly reconstructed, so this source of error accumulation is also avoided.

(2) Pixel-aligned conditioning. Prior work, including ControlNet [1] and OminiControl [2], shows that pretrained diffusion backbones can effectively use spatially aligned conditions. The lumi map is aligned with image pixels. In contrast, an environment map is defined over directions, so the model would additionally need to learn how directional illumination maps to image space.

(3) Intuitive interaction. Drawing directly on a 2D image is an intuitive way for users to specify where and how light should fall. The lumi map is defined in the same image space, so sparse strokes can be sampled from it during training. In addition, a single global environment map represents distant illumination and does not fully specify spatially varying near-field lighting.

W2. Reliability of the estimated Lumi Map in Stage 2

We ensure that the estimated representation remains material-decoupled in two ways. First, the lumi map estimated from a real target image is detached using stop-gradient (Eq. 5), which blocks the reconstruction gradient through the estimation path. Second, throughout Stage 2, 10% of the synthetic samples continue to train the estimation mode with rendered lumi maps. These targets use a canonical white material and exclude the original albedo and texture. This supervised signal anchors the estimation mode to the material-decoupled definition in Eq. (3).

We have added real-image estimation examples to Sec. 5.2 and Fig. 9 of the revised manuscript, showing the input photographs, user strokes, relit outputs, and estimated lumi maps before and after relighting. Textures and specular highlights are removed, while spatial structure and shading variations that reflect the illumination distribution are preserved. In the in-the-wild example in the third row, the estimated illumination chromaticity is consistent with the visible light color, and the effects of specular materials are eliminated. These results indicate that the two mechanisms keep the lumi maps estimated from real images material-decoupled, without encoding target-specific appearance.

W3. Conditioning inputs and prompts for the baselines

The exact conditioning inputs and prompts for each baseline are listed in Supplementary Section D.1. We evaluate each baseline using the input interface it natively supports.

On the paired benchmarks, the baselines receive relatively dense lighting condition, whereas our method receives only sparse strokes. This setting, if anything, favors the baselines. For the multimodal editing models (Gemini-Banana, GPT-Image-2, Qwen-Image-Edit, and FLUX.2 Klein), we supply the lighting condition as an additional reference image and describe its intended use in detail in the prompt, so that these models can make full use of their relighting capabilities.

In the in-the-wild evaluation, all image-conditioned methods receive the same hand-drawn strokes as the lighting condition, while IC-Light, which does not accept a control image, receives a textual description of the target lighting.

AuthorsOpenReview ↗

Response to Reviewer 6YXy

W4. Quantitative evaluation of intensity and chromaticity control

We evaluate controllability on 50 synthetic samples with known target illumination. For each sample, we keep all other inputs fixed and vary only the input lighting condition, using 5 intensity levels and 5 light colors. We then examine whether the relit results follow these changes linearly. We report four metrics:

  • Intensity linearity: the R² of a linear fit between input and output intensity.
  • Intensity error: the deviation from the target intensity, measured in photographic stops.
  • Chromaticity linearity: the R² between input and output chromaticity.
  • Chromaticity error: the angular deviation from the target light color, measured in degrees.

To isolate the effect of the lighting representation, we compare the representation variants from Table 2. All variants share the same backbone, training data, and two-stage training procedure.

Table R1. Quantitative evaluation of intensity and chromaticity control

RepresentationIntensity R² ↑Intensity Err. (stops) ↓Chroma. R² ↑Chroma. Err. (°) ↓
Irradiance0.88720.20440.98067.2185‾\underline{7.2185}
Shading0.91150.1765‾\underline{0.1765}0.9862‾\underline{0.9862}8.2446
Phong0.70640.39340.958513.4563
Lumi Map (Ours)0.9082‾\underline{0.9082}0.16910.98695.6149

The lumi map achieves the lowest error in both intensity and chromaticity, and its linearity is comparable to that of shading (intensity R²: 0.9082 vs. 0.9115; chromaticity R²: 0.9869 vs. 0.9862).

Compared with irradiance and shading, the largest difference lies in the chromaticity error, which the lumi map reduces by 22% and 32%, respectively. This is consistent with the design of the representation. The irradiance condition contains material-dependent interreflections, and the shading condition is derived under an ideal Lambertian assumption, which can introduce color shifts. As a result, the color of these conditions does not correspond purely to the light color. In contrast, the lumi map excludes material dependence, so the stroke color directly specifies the light color.

W5. Recent related work

We have added UniRelight [3], PI-Light [4], and HorizonRelight [5] to Sec. 2 and discussed the differences between environment-map-based relighting methods and our approach.

References

[1] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. ICCV, 2023.

[2] Zhenxiong Tan et al. OminiControl: Minimal and Universal Control for Diffusion Transformer. ICCV, 2025.

[3] Kai He et al. UniRelight: Learning Joint Decomposition and Synthesis for Video Relighting. NeurIPS, 2025.

[4] Zhexin Liang et al. PI-Light: Physics-Inspired Diffusion for Full-Image Relighting. ICLR, 2026.

[5] Jing Yang et al. HorizonRelight: Relighting Long-horizon Videos Consistently via Diffusion Transformers. ECCV, 2026.

Thank you again for your time and thoughtful review. We would be grateful for your feedback on whether our rebuttal adequately addresses the concerns you raised.

Comments on the paper

2 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 the Rebuttal and Discussion

Dear AC and Reviewers,

We sincerely thank all three reviewers for their careful reviews, constructive suggestions, and timely engagement during the discussion, and we thank the AC for coordinating the process. Below we briefly summarize the discussion for the AC's convenience.

Final reviewer assessments

All three reviewers are positive about the paper. Reviewer D8Xk maintains Clear Acceptance. Reviewer EExJ explicitly recommends acceptance after the rebuttal and has no further questions. Reviewer 6YXy maintains Weak Acceptance, noting that our responses largely address their concerns.

Strengths recognized by the reviewers

(a) The reviewers recognize the Lumi Map as a novel lighting representation with a good theoretical motivation and practical implementation (D8Xk, EExJ). It handles interreflections and shadows better than irradiance and shading and could benefit research beyond relighting (EExJ).

(b) They also find each component of the method well designed and clearly motivated (6YXy) and the two-stage training strategy useful for inverse problems (EExJ).

(c) In terms of results, the method outperforms existing approaches (D8Xk) and achieves impressive control over illumination intensity and chromaticity (6YXy).

(d) All three reviewers rated the paper as clear and easy to follow.

Additional experiments and clarifications

To further validate the Lumi Map:

(a) We added a quantitative controllability evaluation. It shows that the Lumi Map yields the lowest intensity and chromaticity errors among all representations, which demonstrates its advantage in precise and consistent illumination control (6YXy).

(b) Visualizations of Lumi Maps estimated from real images show that the estimation remains reliable on real photographs and that stroke edits translate intuitively into changes in illumination (D8Xk, 6YXy).

(c) A hand-drawn stroke study further shows that user strokes achieve comparable or better performance than sampled strokes while covering fewer pixels, which confirms that our method is robust to real user input (EExJ).

To strengthen the evaluation:

(a) We fine-tuned ScribbleLight on the same training data, and our method still outperforms it by a large margin. This indicates that our advantage does not come solely from the training data but also from the proposed representation and training framework. Additional comparisons with DiffusionRenderer and DiLightNet further confirm that our method outperforms these relighting approaches (D8Xk, EExJ).

(b) Moreover, under the same training setting, the Lumi Map outperforms all alternative representations, which verifies the benefit of the representation itself. In a controlled study with a reduced training budget, Stage 2 already brings clear gains with as few as 100 real pairs, suggesting that it does not rely on large amounts of real data (EExJ).

We are grateful to the reviewers for helping us substantially strengthen the paper, and we thank the AC for considering our work.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.