acceptodds
Under review as a conference paper at ICLR 2027

GeoPolar: Few-Step RGB-to-Polarization Diffusion with Geometry Priors

Abstract

Polarization reveals surface orientation and reflection properties that are useful for depth and shape estimation, but capturing it requires dedicated sensors, so the cue is absent from ordinary RGB images. Generative models have therefore been proposed to synthesize polarization from RGB. However, the angle of polarization depends not only on surface orientation but also on whether reflection is diffuse or specular, and the two cases yield angles 90° apart that appearance alone cannot distinguish. We resolve this diffuse–specular ambiguity by leveraging the geometric prior of monocular foundation models. From the same RGB image, such a model recovers the scene geometry, which we convert into the angular coordinates of polarization rather than supplying it as raw normals. The network then only needs to select between the two candidate angles instead of regressing the angle from scratch. Because the foundation-model prior carries the geometric burden of the task, the generator can remain lightweight: a four-step diffusion model operating directly in pixel space. GeoPolar is an order of magnitude lighter than latent-diffusion baselines and remains lighter with its geometry model, yet it surpasses them on benchmark data at native resolution and on most metrics of a set captured with an unseen camera. Further analyses show that the gain follows the geometric content of the prior and persists under moderate errors in the normals.

Primary area:
generative models
Keywords:
polarization imaging, RGB-to-polarization, diffusion models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Vx3jOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses RGB-to-polarization estimation, i.e., predicting AoLP and DoLP from a single RGB image. It argues that appearance-only models suffer from a diffuse–specular branch ambiguity (a 90° AoLP offset). To address this, it converts normals from a frozen monocular geometry model (MoGe-2 used in the paper) into a polarization-coordinate prior. This prior conditions a lightweight four-step pixel-space residual-shifting diffusion model GeoPolar, which outperforms latent-diffusion baselines on a combined benchmark and an unseen-camera set, at much lower cost.

Critical strengths

The analysis of the form of the geometry condition (Sec. 5.3, Table 4, Appendix G) is helpful. By rewriting the condition at test time, the authors show that the azimuth carries most of the useful information. A raw normal map is worth no more than g = 0, so the polarization-coordinate encoding matters, not just the presence of geometry. Replacing g with another image’s geometry hurts further, showing that the model follows the content of the prior.

Critical weaknesses
  • The motivation for the task is unclear. It is not clear why polarization information needs to be estimated from RGB images in the first place. The introduction cites several downstream applications of polarization images to motivate the task. However, the paper provides no experiment showing that the polarization predicted by the proposed method is more faithful and actually benefits any of these downstream tasks.

  • The reliance on estimated normals is not adequately justified. Shape from polarization is typically used precisely to recover surface normals where RGB alone struggles, such as texture-less or reflective surfaces. The proposed method reverses this logic and depends on an input normal map. This dependency directly limits the method. Table 5 shows that once the normal error exceeds 20° (std of 30° and 45°), both AoLP and DoLP errors become worse than the w/o-geometry model. Monocular normal estimators are least reliable on exactly the surfaces where polarization is most useful, such as glass and specular objects (Appendix I). This raises further questions the paper leaves unanswered. What is the purpose of using a geometry prior to estimate polarization? For which downstream polarization applications does geometry-assisted, more accurate polarization actually help, and for which is it redundant? For example, if a normal map is already available, recovering normals from the predicted polarization seems circular. The paper lacks both the discussion and the experiments needed to answer these questions.

  • The comparison with baselines is not fully convincing. All baselines in Tables 1–2 are retrained by the authors on their own split rather than evaluated with released weights, and the retrained models appear substantially weaker than the originals.

    • On the unseen-camera set, which no model has seen, the released PolarAnything (T=20) reaches 38.6° AoLP_m, versus 47.2° for the retrained version (Table 11).
    • On the combined benchmark, the released GenPolar is far stronger than the retrained one: 16.5° vs. 33.4° on RLP and 4.4° vs. 21.9° on RSP (Table 10). The authors attribute this to possible train–test overlap, but this is not verified.

    As a result, the claimed margins over prior methods may partly reflect suboptimal reproduction. The most reliable evidence in the paper is the comparison against its own w/o-geometry control, not against the state of the art.

Questions for authors
  • Clarify the practical motivation for RGB-to-polarization estimation. Evaluate on at least one or two downstream tasks cited in the introduction, using each method’s predicted polarization.
  • Justify relying on estimated normals, given that SfP is mainly needed where monocular normals fail. Discuss and test which downstream applications benefit from geometry-assisted polarization and which become redundant.
  • Strengthen the baseline comparison:
    • Verify the retraining of PolarAnything and GenPolar against their original training settings.
    • Report results with released weights on data confirmed to be disjoint from their training sets.
    • Explain why the retrained models are much weaker even on the unseen-camera set.
9 comments

Overall response

We thank the reviewer for the careful and constructive review, and we appreciate the recognition that our analysis of the form of the condition shows that the polarization-coordinate encoding, rather than the mere presence of geometry, is what matters. In this response, we address each concern as follows.

  • W1 and Q1. We evaluate four downstream tasks with the polarization predicted by each method (Sec. 5.5, Table 10, Appendix N).
  • W2 and Q2. We clarify the role of the prior and show what it adds beyond the normals (Secs. 5.2 and 5.5, Tables 9 and 18).
  • W3 and Q3. We detail our GenPolar retraining, report the released weights on disjoint data, verify our PolarAnything retraining, and show that our prior also improves GenPolar (Sec. 4.1, Appendices F, G and L, Tables 2, 14–16 and 20).

All changes are highlighted in blue in the revised PDF, and all table, figure, section and appendix numbers below refer to the revision.


W1 and Q1. Motivation and downstream evaluation

This valuable comment is well taken, and we fully agree that the task should be supported by downstream evidence. Polarization cameras are rare compared with RGB cameras, and polarization datasets are small. Estimating polarization from RGB makes polarization-based methods applicable to images and cameras without a polarizer, which is also the goal of PolarAnything and GenPolar. We evaluated four downstream tasks cited in Sec. 1 and replaced the captured polarization by the prediction of each method (Sec. 5.5, Table 10; protocols in Appendix N). The tasks include the protocols of PolarAnything (multi-view reconstruction with PISR on NeRSP) and GenPolar (material classification with SfPUEL and reflection removal with PolarFree).

input polarizationSfP normals, NeRSP (258) ↓multi-view reconstruction, PISR (62 views) ↓material, SfPUEL (overall / balanced %) ↑reflection removal, PolarFree (PSNR ↑ / LPIPS ↓)
captured17.814.092.3 / 88.422.44 / 0.133
GeoPolar25.120.789.8 / 94.722.35 / 0.133
GenPolar26.524.776.5 / 87.722.28 / 0.135
PolarAnything34.425.895.2 / 50.022.24 / 0.134

Normal errors are given in degrees, and bold marks the best result among the predicted polarizations. The balanced accuracy averages the recall of metals and the accuracy of dielectrics, since PolarAnything classifies nearly every pixel as dielectric.

GeoPolar also performs well in these downstream tasks. Among the predicted polarizations, it yields the most accurate normals in single-view and multi-view shape from polarization and the highest balanced accuracy in material classification, and it remains close to captured polarization in reflection removal. Figs. 8, 17 and 18 show examples of material classification and reflection removal.

W2 and Q2. Role of the estimated normals

We are grateful for this insightful comment, which raises a point that we have also considered carefully. When a normal map is available, recovering normals from our polarization is indeed circular, and we do not claim gains in normal estimation (Sec. 5.5, last paragraph). The predicted polarization is, however, not a re-encoding of the normals, because it contains information that the normals alone cannot provide.

Polarization rendered from the normals does not reproduce the measured DoLP. Rendering the same MoGe-2 normals with the diffuse or the specular dielectric model for every pixel gives the following errors on DoFP, RLP and RSP (Table 18, Appendix I).

AoLP$_m$ (°)DoLP$_m$
diffuse model64.2 / 44.3 / 54.00.13 / 0.15 / 0.12
specular model25.8 / 45.7 / 36.00.41 / 0.39 / 0.43
oracle (better model per pixel, chosen with the ground truth)17.2 / 15.8 / 11.60.39 / 0.24 / 0.35
GeoPolar11.7 / 28.2 / 13.40.055 / 0.105 / 0.065

The normals do not determine the material response. On NeRSP, whose objects are all dielectric, the material head of SfPUEL classifies 73–84% of the pixels of the polarization rendered from the MoGe-2 normals as metal, but only 18% of those of GeoPolar. The DoLP depends on the material, which the normals do not carry.

Real estimator errors do not make the prior harmful. On HAMMER, we grouped the pixels by the actual azimuth error of MoGe-2 (Table 9, Appendix M; AoLP$_m$, °).

MoGe-2 azimuth errorpixelsGeoPolarw/o geometryGenPolar
0–5°69.5%20.633.420.9
5–10°15.0%25.833.926.4
10–20°7.5%29.936.433.7
20–30°3.0%34.538.739.2
30–45°2.2%36.339.939.1
45–90°2.8%37.039.338.3

GeoPolar is more accurate than the model without geometry in every bin, including the pixels where the prior deviates by 45–90°. Unlike these real errors, the σ ≥ 30° rows of Table 6 add content-independent noise to every pixel. Fig. 16 shows examples of the real errors of the estimator, and Fig. 15 compares them with the synthetic perturbations (Appendix M).

Where geometry-assisted polarization helps and where it is redundant. As stated in Sec. 5.5, geometry-assisted polarization is redundant for recovering normals when the normals are already available. It is beneficial for tasks that require the polarization itself beyond the surface orientation, namely the DoLP and the material response, as in material classification and reflection removal.

W3 and Q3. Baseline retraining and released weights

We sincerely thank the reviewer for examining our reproduction so carefully, and we acknowledge this concern. We intended to follow the setting of GenPolar exactly, but this was not possible, because two of its training datasets (OPID and IPID) were not public at the time of our submission and its exact train/test split is still not specified. We therefore retrained GenPolar with the public code, following the setting of the paper as closely as possible. In our submission, we trained Stage I for 258 epochs, compared with 100 epochs in the GenPolar paper. Since the model still appeared under-trained, we continued Stage I to 800 epochs. This improved GenPolar on all three subsets of our test split (RLP from 33.4° to 31.5°, DoFP from 27.2° to 21.3° and RSP from 21.9° to 18.5°), while GeoPolar remains more accurate on each subset (28.2°, 11.7° and 13.4°). The revision reports this model in Table 2 and all other tables (Appendix F), and Table 14 compares the two training settings.

GenPolar paperours, submissionours, revision
training datapublic sources and OPID, IPID (not public at our submission)our training split (8 public sources)same
train / test split50 random images per group held out, list not releasedour split, identical for all compared modelssame
Stage I / Stage II epochs100 / 30258 / 30800 / 30
prediction targetz0z_0, Stokes ×1 (paper text, code default)ϵ\epsilon, Stokes ×5 (recipe of the released weights)same
effective batch, GPUs4, 4 × A401, 1 × RTX PRO 6000same
GPU-hours–159401

All compared models, namely GeoPolar, the model without geometry and the retrained GenPolar, were trained on the same training split. The released weights were trained with ϵ\epsilon-prediction and a 5× Stokes target, as stated in the inference code of the authors ("Legacy release weights used a 5x Stokes target") and confirmed by reproducing their quick test.

Released weights on fully disjoint data. We evaluated all models on three datasets that none of them was trained on (Table 20, Appendix L; AoLP$_m$, °).

dataset (images)GeoPolarGenPolar retrainedGenPolar releasedPolarAnything retrainedPolarAnything released
HAMMER (146)23.924.950.832.136.5
MCubeS (102)33.938.349.138.736.2
NeRSP (258)18.823.825.529.831.4

On every disjoint dataset, our retrained GenPolar is more accurate than the released weights, which indicates that our retraining is not weaker. GenPolar holds out only 50 random images per group from the same public pool from which our test split is drawn, so a large part of our test split is expected to be in its training set. The strength of the released model on our test split (16.5° on RLP and 4.4° on RSP; Table 15, Appendix G) is therefore likely to result from this overlap.

Our prior also improves GenPolar. We added our prior to GenPolar as an additional ControlNet condition, using the same code, data and recipe (Sec. 4.1, Appendix F). GenPolar with our prior is more accurate in AoLP on all three subsets (Table 2, Sec. 4.2), and its Stage I training loss is lower.

Stage IStage II evaluation, AoLP$_m$ (°) / DoLP$_m$
training loss (AoP / total)RLPDoFPRSP
GenPolar0.110 / 0.09431.5 / 0.12621.3 / 0.09518.5 / 0.087
GenPolar + our prior0.106 / 0.09128.6 / 0.12115.2 / 0.08017.3 / 0.090

The polarization-coordinate prior is therefore a general component rather than a property of our backbone.

Reviewer_JfKsOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces GeoPolar, a four-step pixel-space diffusion model that addresses the 90∘90^\circ diffuse–specular AoLP ambiguity in RGB-to-polarization synthesis. Instead of supplying raw normal maps, GeoPolar converts geometry priors from a monocular foundation model directly into polarization angular coordinates. Consequently, it achieves consistent gains over appearance-only and latent-diffusion baselines across three benchmark subsets.

Critical strengths
  1. The geometry-to-polarization encoding is physically motivated and turns an otherwise difficult angular prediction into a more constrained branch-selection problem.
  2. The ablations isolate the value of geometry, including error analysis and perturbations to estimated normals.
  3. The accuracy and efficiency trade off is compelling: on DoFP, AoLP error falls from 18.4° to 11.7° against the matched no-geometry model, with substantially lower memory use than latent-diffusion baselines.
Critical weaknesses
  1. The method remains sensitive to unreliable monocular normals, particularly on reflective surfaces, painted objects, and regions with invalid geometry. Although the paper investigates normal perturbations and confidence gating, the latter does not consistently improve performance. A more detailed discussion of these failure cases and their implications for practical deployment would help clarify the limitations of the proposed approach.
  2. The unseen-camera evaluation is valuable but covers only six scenes, limiting the evidence for generalization across diverse materials, lighting conditions, and sensor configurations. A more detailed discussion of these limitations would help strength the paper.
  3. The related work discussion is missing a more comprehensive comparison with relevant studies, including RGB-to-Polarization Estimation: A New Task and Benchmark Study (NeurIPS 2025), PolarAnything: Diffusion-based Polarimetric Image Synthesis (ICCV 2025), and PolarScale: A Physics-Grounded Benchmark for Radiometrically Consistent RGB-to-Stokes Estimation (NeurIPS 2026). A clearer discussion of their differences in physical assumptions and evaluation protocols would better establish GeoPolar’s contribution.
Questions for authors
  1. Can the authors better characterize when the geometry prior should be trusted, beyond synthetic perturbations and the current confidence-gating results?
  2. How stable are the gains across more unseen cameras, challenging reflective materials, and different lighting conditions?
  3. How does GeoPolar’s normalized Stokes output satisfy physical consistency constraints, and how should its evaluation be interpreted relative to the benchmarks and Stokes-based methods listed above? Overall, I find the proposed approach technically sound and the experimental results encouraging, although the above concerns warrant further clarification. I am inclined to raise my score if the authors adequately address these concerns in their rebuttal.
4 comments

Overall response

We sincerely appreciate the positive assessment and the constructive review, and we are glad that the reviewer finds the encoding physically motivated, the ablations informative, and the accuracy–efficiency trade-off compelling. In this response, we address each concern as follows.

  • W1 and Q1. We characterize when the prior can be trusted, using the real errors of the estimator and regions without geometry (Sec. 5.4, Tables 8 and 9, Sec. 6, Appendix J).
  • W2 and Q2. We add three held-out datasets captured by other groups with other cameras (Sec. 4.3, Appendix L, Table 20).
  • W3 and Q3. We compare the assumptions and protocols of the related methods and report the physical consistency of our output (Sec. 2, Table 1, Appendix A).

All changes are highlighted in blue in the revised PDF, and all table, figure, section and appendix numbers below refer to the revision.


W1 and Q1. When the prior can be trusted

This is an important question for practical deployment, and we are grateful that the reviewer raised it. We characterize when the prior can be trusted, using the real errors of the estimator rather than synthetic perturbations.

Where the geometry is defined, the prior helps even when it is locally inaccurate. HAMMER provides ground-truth normals from scanned meshes and contains transparent and specular objects. We grouped the pixels of its 146 test images by the actual azimuth error of MoGe-2 (Table 9, Appendix M; AoLP$_m$, °).

MoGe-2 azimuth errorpixelsGeoPolarw/o geometryGenPolar
0–5°69.5%20.633.420.9
5–10°15.0%25.833.926.4
10–20°7.5%29.936.433.7
20–30°3.0%34.538.739.2
30–45°2.2%36.339.939.1
45–90°2.8%37.039.338.3

GeoPolar is more accurate than the model without geometry in every bin, including the pixels where the prior deviates by 45–90°.

Where the geometry is undefined, the prior should be marked as unknown. MoGe-2 marks regions without geometry, such as the sky, as invalid, and we provide them as an unknown token (Sec. 3.2). An estimator that predicts normals everywhere requires the same rule, and masking the sky of Metric3D v2 with an off-the-shelf segmenter recovers the gain (the masked row is in Table 8, Appendix M; AoLP$_m$, °).

priorDoFPRLPRSP
MoGe-2 (invalid regions as unknown)11.728.213.4
Metric3D v217.728.120.1
Metric3D v2 with the sky as unknown13.028.114.7
w/o geometry18.439.718.2

The learned confidence. We acknowledge that the learned confidence does not consistently reduce the error over the polarization mask. Nevertheless, it is effective on strongly polarized pixels, where it reduces the error from 18.7° to 14.6° on RLP and from 15.5° to 11.6° on RSP (Table 19, Appendix J). However, it is not accurate enough to help everywhere (AUC of 0.80–0.88 in separating correct from incorrect geometry), because the reliability of the prior is difficult to learn directly. We expect that direct supervision of this reliability, for example from ground-truth geometry, would allow the confidence to be learned more accurately. These results are now reported in Sec. 6 and Appendix J (Table 19, Fig. 12).


W2 and Q2. Other cameras, materials and lighting

We value this helpful suggestion, which has allowed us to broaden the evaluation substantially. We added three held-out datasets that were captured by other groups with other polarization cameras and were not used to train any of the compared models. They cover laboratory and outdoor lighting as well as transparent, specular and glazed objects (Table 20, Appendix L, Sec. 4.3; AoLP$_m$, °).

dataset (images)lighting, materialsGeoPolarw/o geometryGenPolarPolarAnything
HAMMER (146)lab, transparent and specular objects23.934.224.932.1
MCubeS (102)outdoor street33.939.538.338.7
NeRSP (258)lab, glazed objects18.829.723.829.8

On all three datasets, the prior improves on the same model without geometry by 5.6–10.9°, and GeoPolar is the most accurate of all compared methods. Fig. 14 shows two objects of NeRSP.

W3 and Q3. Related methods and physical consistency

We thank the reviewer for pointing us to these closely related works. We have added this comparison to the related work (Sec. 2) in condensed form as Table 1.

inputconditionoutputgeneratorevaluation
Lin et al. (NeurIPS'25)RGB as S0S_0appearanceS1,S2,S3S_1, S_2, S_3 (incl. circular), normalized–PSNR / SSIM / LPIPS per Stokes component, all pixels
PolarAnything (ICCV'25)RGBappearanceDoLP, (cos⁡2ϕ,sin⁡2ϕ)(\cos 2\phi, \sin 2\phi)latent diffusion, 20 stepsPSNR / SSIM of polarizer images, DoLP and AoLP errors, no mask
PolarScale (NeurIPS'26)linear intensity, normalized per sceneappearancenormalized s1,s2,s3s_1, s_2, s_3, AoLP / DoLP / DoCP, scene scale–normalized and full Stokes, scale error, AoLP on GT DoLP > 0.1, self-consistency, bound violations
GenPolar (ECCV'26)RGB as S0S_0appearancelinear S1,S2S_1, S_2 per color channellatent diffusion, 1 stepper-channel DoLP PSNR / SSIM, AoP error on GT DoLP > 0.05
GeoPolar (ours)RGBappearance + geometry in polarization coordinatesDoLP and direction of (S1,S2)/S0(S_1, S_2)/S_0pixel-space diffusion, 4 steps (129.9M, 0.85 GB)AoLP on brightness ≥ 0.1 and GT DoLP ≥ 0.1, DoLP error, PSNR / SSIM of polarizer images

Like PolarAnything and GenPolar, GeoPolar predicts linear polarization only, because most sources in the common data pool provide linear polarization only. It also predicts scale-invariant quantities and does not estimate the absolute radiometric scale that PolarScale makes explicit.

Physical consistency. The DoLP output is bounded to [0, 1], and the angle is given by the direction of the predicted (S1,S2)(S_1, S_2). Consequently, the four polarizer images Iθ=S02(1+DoLPcos⁡(2θ−2ϕ))I_\theta = \tfrac{S_0}{2}\bigl(1 + \mathrm{DoLP}\cos(2\theta - 2\phi)\bigr) are non-negative, satisfy I0+I90=I45+I135=S0I_0 + I_{90} = I_{45} + I_{135} = S_0, and never violate S12+S22≤S0\sqrt{S_1^2 + S_2^2} \le S_0. Since the network predicts the DoLP and the normalized Stokes components in separate channels, we also measured their agreement on our test split, which corresponds to the self-consistency of PolarScale (Table 11, Appendix A).

DoFPRLPRSP
|DoLP − DoLP recomputed from (S1,S2)(S_1, S_2)|, pixels with intensity ≥ 0.10.0210.0210.019
correlation between the two0.940.920.87
pixels violating S12+S22≤S0\sqrt{S_1^2 + S_2^2} \le S_0000

The two representations agree closely, and no pixel violates the physical bound. Our angle metric (Appendix C) also uses the same wrap-aware definition and the same DoLP level (0.1) as PolarScale, and additionally requires a brightness of at least 0.1.

The rebuttal has addressed all of my concerns. I will raise my score to 4 (accept).

Thank you very much for your positive evaluation and for raising your score. We are glad that our rebuttal addressed your concerns. Your constructive feedback has been invaluable in improving the paper, and we will carefully incorporate the discussed revisions into the final version.

Best regards, The Authors

Reviewer_DgZEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoPolar, an RGB-to-polarization generation framework guided by monocular geometric priors obtained from a pretrained surface normal estimator. The method transforms predicted monocular surface normals into a polarization-aware coordinate space to alleviate the diffuse–specular ambiguity in polarization synthesis. The proposed geometric encoding strategy also demonstrates a certain degree of novelty. Experimental results show that the method achieves improvements over existing RGB-to-polarization approaches. However, several issues remain regarding the reliability of geometric priors, applicability under complex optical conditions, practical effectiveness in downstream tasks, and clarification of methodological novelty, which require further discussion and validation.

Critical strengths

This work explores RGB-to-polarization synthesis through a geometry-conditioned generative framework. Transforming predicted monocular surface normals into polarization-aware coordinates provides a promising approach to mitigating diffuse–specular ambiguity. The authors conduct relatively comprehensive experimental evaluations, including comparisons with existing methods, ablation studies on geometric priors, robustness analyses, efficiency evaluations, and generalization tests. These experiments demonstrate, to some extent, the potential of incorporating geometric foundation models into polarization generation.

Critical weaknesses

This study explores the promising direction of geometry-guided RGB-to-polarization synthesis. However, several key concerns remain. First, the framework relies heavily on the accuracy of monocular geometric priors and assumes that polarization formation follows idealized diffuse and specular reflection models. Second, robustness analyses based on synthetic perturbations of surface normals may not adequately reflect the failure modes of monocular geometry estimators in real-world complex scenes. Third, the practical effectiveness of the generated polarization information for downstream geometric tasks has not been validated. Furthermore, the fundamental distinctions between the proposed method and existing RGB-to-polarization generation approaches are insufficiently clarified. Finally, the effects of illumination variations, complex reflection behaviors, and sensor noise during real-world polarization acquisition require further investigation.

Questions for authors
  1. The proposed method relies on the accuracy of a monocular surface normal estimator. The evaluation should include different geometry estimators to verify whether the reported improvements hold true across different models. Furthermore, the normal perturbation experiments in Table 5 may underestimate real-world uncertainty, as angular errors of approximately 10° are already common in existing monocular normal estimators. Adding synthetic perturbations to potentially inaccurate estimated normals may not faithfully reflect realistic failure modes.

  2. The proposed method primarily assumes idealized diffuse and specular dielectric reflection models, whereas real-world polarization measurements often involve mixtures of reflection, transmission, and complex optical interactions.

  3. Polarization synthesis is typically motivated by downstream applications, such as descattering / dehazing, shape reconstruction and surface normal estimation. The impact of the generated polarization information on these tasks might be further evaluated.

  4. The methodological novelty of the proposed approach compared to existing RGB-to-polarization generation methods has not been sufficiently clarified.

  5. Real-world polarization acquisition is affected by variations in illumination, reflection characteristics, and sensor noise. The effects of these factors on the reconstructed polarization ground truth and model performance require further analysis.

8 comments

Overall response

We are grateful for the constructive review and for the recognition that converting monocular normals into polarization-aware coordinates is a promising way to resolve the diffuse–specular ambiguity. In this response, we address each concern as follows.

  • W1a, Q1 and W2. We evaluate four further normal estimators and the robustness to the real errors of the estimator (Sec. 5.4, Tables 8 and 9, Appendix M).
  • W1b and Q2. We clarify that the reflection models only motivate the coordinates of the condition (Sec. 3.1, Sec. 5.2, Table 18).
  • Q4. We summarize the novelty and compare the methods, including their cost (Secs. 1 and 2, Tables 1, 2 and 4).
  • W3 and Q3. We evaluate four downstream tasks with the polarization predicted by each method (Sec. 5.5, Table 10, Appendix N).
  • W4 and Q5. We analyze the effect of sensor noise and brightness and evaluate on three held-out datasets (Sec. 4.3, Appendices L and O, Tables 20 and 22–24).

All changes are highlighted in blue in the revised PDF, and all table, figure, section and appendix numbers below refer to the revision.


W1a, Q1 and W2. Dependence on the normal estimator

We appreciate this insightful observation. As the reviewer points out, monocular normals already contain errors of about 10°, and further perturbation does not represent the uncertainty of a real estimator. We therefore measured how strongly GeoPolar depends on the choice of estimator and how robust it is to the real errors of the estimator.

Dependence on the estimator. We replaced MoGe-2 at test time with four other normal estimators without retraining. The resulting AoLP$_m$ errors (°) are reported below and in Table 8.

prior at test timeDoFPRLPRSP
MoGe-2 (paper)11.728.213.4
Metric3D v2 ViT-L13.028.114.7
Metric3D v2 giant213.827.914.0
Marigold-normals v1.114.528.415.3
StableNormal16.129.615.3
w/o geometry18.439.718.2

Even without retraining, all four estimators retain the improvement over the model without geometry on all three subsets. As MoGe-2 does natively, regions without geometry such as the sky are given as unknown, and for the other estimators the sky is masked with an off-the-shelf segmenter (Appendix M).

Robustness to real estimator errors. HAMMER provides ground-truth normals from scanned meshes, contains transparent and specular objects, and was not used to train any of the models. We grouped the pixels of its 146 test images by the actual azimuth error of MoGe-2 (Table 9, Appendix M; AoLP$_m$, °).

MoGe-2 azimuth errorpixelsGeoPolarw/o geometryGenPolar
0–5°69.5%20.633.420.9
5–10°15.0%25.833.926.4
10–20°7.5%29.936.433.7
20–30°3.0%34.538.739.2
30–45°2.2%36.339.939.1
45–90°2.8%37.039.338.3

GeoPolar is more accurate than the model without geometry in every bin, including the pixels where the prior deviates by 45–90°. Large real errors are rare, as only 2.8% of the pixels exceed 45°. In contrast, the σ ≥ 30° rows of Table 6 add content-independent noise to every pixel on top of the error of the estimator, which is harsher than real failures. We have added both analyses to Sec. 5.4 (Tables 8 and 9) and state this distinction there. Fig. 15 (top row) compares the synthetic perturbations with the real errors of the estimator, and Fig. 16 shows examples of large real errors (Appendix M).

W1b and Q2. Idealized reflection models

We apologize for the confusion caused by our description and thank the reviewer for the chance to clarify it. The diffuse and specular relations only motivate the coordinates of the condition (Sec. 3.1, last paragraph), in the sense that the angle tends to lie near one of two candidates determined by the surface azimuth. GeoPolar does not impose either model on its output. The network predicts the Stokes components of every pixel and learns them from real captures that contain mixed reflection, transmission and interreflection.

To verify that the output is not a binary choice between the idealized branches, we rendered polarization from the same normals with the dielectric models (refractive index 1.5) and compared it with GeoPolar on DoFP, RLP and RSP (Table 18 in Appendix I, summarized in Sec. 5.2).

AoLP$_m$ (°)DoLP$_m$
diffuse model64.2 / 44.3 / 54.00.13 / 0.15 / 0.12
specular model25.8 / 45.7 / 36.00.41 / 0.39 / 0.43
oracle (better model per pixel, chosen with the ground truth)17.2 / 15.8 / 11.60.39 / 0.24 / 0.35
GeoPolar11.7 / 28.2 / 13.40.055 / 0.105 / 0.065

Even when the appropriate model is chosen for each pixel from the ground truth, the idealized models yield 2.3–7 times the DoLP error of GeoPolar and, on DoFP, also a larger angle error. GeoPolar therefore represents the mixtures and deviations that the idealized models cannot describe. The main difficulty of the problem lies in resolving the ambiguity between the two branches, which appearance alone cannot achieve. The geometric prior, provided as a condition, alleviates this ambiguity, while the network learns the remaining behavior from the data.


Q4. Novelty over existing methods

We welcome the opportunity to clarify the novelty of our approach. Existing RGB-to-polarization methods condition on appearance alone and therefore have to infer an angle that appearance cannot determine, since diffuse and specular reflection produce angles that differ by 90°. GeoPolar differs from these methods in three respects.

First, the geometry is expressed in the coordinates of polarization. We convert the normals of a monocular foundation model into the angular coordinates of polarization and provide them as a condition. The network then learns from data how the polarization of each pixel relates to the candidate angles defined by the geometry, instead of inferring the angle from appearance alone. When the trained model receives the same geometry as a raw normal map at test time, it is no more accurate than when the prior is removed (Table 5, Sec. 5.3), so the model relies on the polarization coordinates rather than on the normal as such.

Second, the prior is a general component. The same prior also improves GenPolar, a two-stage latent diffusion model. With our prior, GenPolar becomes more accurate in AoLP on all three subsets (Table 2, Sec. 4.2), and its Stage I training loss is lower.

Stage IStage II evaluation, AoLP$_m$ (°) / DoLP$_m$
training loss (AoP / total)RLPDoFPRSP
GenPolar0.110 / 0.09431.5 / 0.12621.3 / 0.09518.5 / 0.087
GenPolar + our prior0.106 / 0.09128.6 / 0.12115.2 / 0.08017.3 / 0.090

Third, the generator is lightweight and requires few steps. Because the prior carries the geometric information, a pixel-space diffusion model with 130M parameters and four steps suffices. It requires 11 times fewer parameters and 4.7–14 times less peak memory than the latent baselines, and still 1.7–5.2 times less peak memory when MoGe-2 is included (Table 4, Fig. 5, Sec. 4.4).

inputconditionoutputgeneratorparamstime (s)peak memory (GB)
Lin et al. (NeurIPS'25)RGB as S0S_0appearanceS1,S2,S3S_1, S_2, S_3 (incl. circular), normalized––––
PolarAnything (ICCV'25)RGBappearanceDoLP, (cos⁡2ϕ,sin⁡2ϕ)(\cos 2\phi, \sin 2\phi)latent diffusion, 20 steps1427.5M87.4711.82
PolarScale (NeurIPS'26)linear intensity, normalized per sceneappearancenormalized s1,s2,s3s_1, s_2, s_3, AoLP / DoLP / DoCP, scene scale––––
GenPolar (ECCV'26)RGB as S0S_0appearancelinear S1,S2S_1, S_2 per color channellatent diffusion, 1 step (distilled)1427.5M2.273.99
GeoPolar (ours)RGBappearance + geometry in polarization coordinatesDoLP and direction of (S1,S2)/S0(S_1, S_2)/S_0pixel-space diffusion, 4 steps129.9M (460.8M with MoGe-2)5.200.85 (2.28 with MoGe-2)

We have revised the related work (Sec. 2) and the contributions (Sec. 1) accordingly, and added a condensed version of this comparison as Table 1.

W3 and Q3. Downstream evaluation

We found this suggestion particularly valuable, as it has helped us strengthen the evaluation considerably. We evaluated four downstream tasks cited in Sec. 1 and replaced the captured polarization by the prediction of each method (Sec. 5.5, Table 10; protocols in Appendix N). The tasks include the protocols used by PolarAnything (multi-view reconstruction with PISR on NeRSP) and GenPolar (material classification with SfPUEL and reflection removal with PolarFree).

input polarizationSfP normals, NeRSP (258) ↓multi-view reconstruction, PISR (62 views) ↓material, SfPUEL (overall / balanced %) ↑reflection removal, PolarFree (PSNR ↑ / LPIPS ↓)
captured17.814.092.3 / 88.422.44 / 0.133
GeoPolar25.120.789.8 / 94.722.35 / 0.133
GenPolar26.524.776.5 / 87.722.28 / 0.135
PolarAnything34.425.895.2 / 50.022.24 / 0.134

Normal errors are given in degrees, and bold marks the best result among the predicted polarizations. The balanced accuracy averages the recall of metals and the accuracy of dielectrics, since PolarAnything classifies nearly every pixel as dielectric.

GeoPolar also performs well in these downstream tasks. Among the predicted polarizations, it yields the most accurate normals in single-view and multi-view shape from polarization and the highest balanced accuracy in material classification, and it remains close to captured polarization in reflection removal. Figs. 8, 17 and 18 show examples of material classification and reflection removal.


W4 and Q5. Sensor noise, brightness and held-out data

We are thankful for this careful question, which addresses an important aspect of real-world acquisition. We analyzed the effect on the ground truth and on the model separately.

Ground truth. The angle is undefined where the light is unpolarized, and sensor noise dominates it where the DoLP or the intensity is low. To quantify this effect, we used PDD, which provides captures of the same 40 scenes at three noise levels together with noise-free references. The angle errors of the noisy captures (°) are reported below and in Table 22 (Appendix O).

noise levelGT DoLP < 0.020.02–0.050.05–0.1evaluation mask (≥ 0.1)
low27.513.65.82.5
mid36.027.016.77.6
high40.233.321.810.9

Noise makes the measured angle nearly random at low DoLP, mostly in dark pixels. We therefore evaluate the angle only on pixels whose intensity and ground-truth DoLP are at least 0.1 (Sec. 4.1, Appendix C), and we report the error per DoLP bin (Fig. 7, Table 21).

Model performance. On every subset, the error of GeoPolar is largest in the darkest bin, at intensities of 0.1–0.2 (32.9° on RLP, 21.6° on DoFP and 18.8° on RSP; Table 24). This is partly because the ground truth itself is noisier in dark pixels, as the angle measured from the PDD captures with high noise also errs most there (18.3°, against 5.1° above 0.5). When the noisy captures of nine PDD scenes that were not used for training are given as input, the error of GeoPolar changes by at most 1.2° (from 13.4° to 13.7–14.6°), whereas the angle measured from the same captures degrades from 2.8° to 12.8° (Table 23, summarized in Sec. 5.4). Since GeoPolar predicts from intensity and geometry, the sensor noise does not propagate into its angle. Finally, on three held-out datasets captured by other groups with other cameras, under different illumination and with different materials, the prior improves on the model without geometry for every dataset (Table 20, Appendix L, Sec. 4.3; AoLP$_m$, °).

dataset (images)lighting, materialsGeoPolarw/o geometryGenPolarPolarAnything
HAMMER (146)lab, transparent and specular objects23.934.224.932.1
MCubeS (102)outdoor street33.939.538.338.7
NeRSP (258)lab, glazed objects18.829.723.829.8

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.