acceptodds
Under review as a conference paper at ICLR 2027

FiberR2R: RGB-to-RAW Reconstruction with Learned Forward ISP Sensitivity Guidance

Abstract

RGB-to-RAW reconstruction aims to infer sensor-domain RAW measurements from RGB images, but the image signal processor (ISP) can apply lossy transformations that make the corresponding RAW measurements difficult to recover. To address this challenge, we propose FiberR2R, a three-stage RGB-to-RAW reconstruction framework with learned forward ISP sensitivity guidance. FiberR2R first learns a differentiable forward ISP model from paired RAW–RGB data, then predicts an RGB-conditioned RAW Anchor, and finally trains a residual refiner around the Anchor. We use squared vector–Jacobian product (VJP) responses to provide estimates of the local RGB response energy of the learned forward ISP. These scores jointly parameterize Anchor-centered perturbation sampling and complementary residual supervision. At inference, FiberR2R predicts the Anchor RAW and applies one bounded residual refinement without evaluating the learned forward ISP or its VJPs. Across five camera-specific datasets, FiberR2R has the highest SSIM on all five datasets and the highest PSNR on four, with only 11.51M parameters and 10.38 GFLOPs. These results support learned forward ISP sensitivity as an effective training guidance for RGB-to-RAW reconstruction task.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
RGB-to-RAW reconstruction, learned forward ISP sensitivity, computational photography

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2ADiOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FiberR2R, a framework for reconstructing RAW images from RGB images. It first learns a differentiable forward ISP model, then trains a network to predict an RGB-conditioned RAW anchor, and finally refines the result with a residual network. The main idea is to estimate how sensitive the learned ISP model is to changes in individual RAW coordinates, and use this information to guide the training process. Experiments on several camera-specific datasets show that the method performs well in terms of RAW reconstruction accuracy.

Critical strengths
  1. The paper is well written, the figures are well presented, and the overall structure is good. The authors also provide code and demos in the supplementary material, which helps improve the reproducibility and credibility of the work.

  2. The method achieves strong overall performance and is compared against current state-of-the-art methods, demonstrating its effectiveness. The ablation studies are also carefully designed.

  3. The paper has a clear motivation and addresses an important limitation of existing methods. I believe it could be suitable for a poster presentation after the necessary revisions.

Critical weaknesses
  1. Efficiency problem, the paper reports only parameters and GFLOPs, without any wall-clock runtime. Given the sliding-window inference (64×64 windows with 16-pixel overlap), please report full-resolution inference time and the GFLOPs measurement resolution, as well as the total training cost of the three stages compared with the baselines.

  2. The paper provides limited discussion of related work, and I encourage the authors to include more relevant studies in the revised version to strengthen its technical depth. Since the central motivation is to reuse RGB annotations to train RAW-domain perception models, the discussion should cover recent work on RAW-domain perception, such as Dirty-Pixels (Diamond et al., TOG 2021) [1], ROD (Xu et al., CVPR 2023) [2], RAW-Adapter (Cui & Harada, ECCV 2024) [3], and RAWild (Liu et al., NeurIPS 2026) [4]. For RGB-to-RAW reconstruction, metadata-based approaches (Punnappurath & Brown, WACV 2021) [5] and the AIM 2022 Reversed ISP Challenge (Conde et al., ECCVW 2022) [6] are also relevant.

  3. Appendix A.1 / Fig. 6 claims the generated RAW can “recover high-bit-depth information to eliminate glare.” However, information clipped in 8-bit RGB cannot be reliably recovered; the network likely produces plausible estimates instead. Ground-truth comparisons or quantitative detection results (e.g., on NOD) are needed to support this claim. Otherwise, the authors could consider softening the claims in this section.

[1]. Dirty Pixels: Towards End-to-End Image Processing and Perception

[2]. Toward RAW Object Detection: A New Benchmark and A New Model

[3]. RAW-Adapter: Adapting Pre-trained Visual Model to Camera RAW Images

[4]. RAWild: Toward Sensor-Agnostic RAW Object Detection via Physics-Guided Curve and Grid Modeling

[5]. Spatially Aware Metadata for Raw Reconstruction

[6]. Reversed Image Signal Processing and RAW Reconstruction. AIM 2022 Challenge Report

Questions for authors
  1. In the submitted supplementary materials, most of the folders in the “example” directory appear to be empty. Could the authors please check this part?

  2. Although the paper already includes a fairly comprehensive set of comparisons, I’m curious whether the proposed method still has an advantage over generative-based models such as RawGen [7]. Could the authors include additional experiments to compare their method with these approaches?

[7]. RawGen: Learning Camera Raw Image Generation (ECCV 2026)

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2ADi: efficiency, validation, and supplementary examples

Thank you for recognizing the reconstruction results and for the concrete suggestions on efficiency, related work, application claims, and the supplementary examples. We have revised each of these areas. The detailed point-by-point response is in Response_to_Reviewers.pdf, pp. 6-9, in the updated supplementary ZIP. All appendix and table references below refer to the revised manuscript uploaded in the PDF field.

W1 - Full-resolution inference and training cost (Appendix A.8, Tables 16-18). We measured five methods on all 947 test images, with three repetitions per image, using one H20 GPU per image, FP32, and TF32 disabled. Timing includes preprocessing, transfers, model execution, and stitching, but excludes file I/O and model loading. FiberR2R's domain-mean latency is 4.140-9.574 seconds, giving 1.87-2.13x speedups over the specified RAW-Flow retest. ReRAW is faster. FiberR2R has the lowest measured native allocated peak memory among these wrappers, at 232.2 MiB; this is not a comparison of separately optimized deployment footprints. We clarify that 11.49M parameters cover the deployed Anchor and Refiner, and 10.63 GFLOPs refers to one 64x64 RGB window, excluding the training-only ISP. We also provide stage-wise epoch/update budgets, GPU counts, and measured update times. Complete historical GPU-hours are unavailable, so these measurements are not presented as a complete training-cost ranking.

W2 - Related work (Section 2). We reorganized the discussion around RAW-domain perception and the information available for RGB-to-RAW reconstruction, incorporating Dirty Pixels, ROD, RAW-Adapter, RAWild, Spatially Aware Metadata, and the AIM reversed-ISP challenge. The revised discussion distinguishes the motivation for labeled RAW data from the separate question of whether generated RAW improves downstream perception.

W3 - Validation near saturation (Appendices A.1-A.2, Tables 5-6). We removed the claims of recovering clipped information and eliminating glare. The new analysis compares predicted gradients with paired measured RAW, screening all 947 images and selecting regions using RGB and ground-truth structural support, not prediction quality. For the 470 images eligible for partial near-saturation analysis, domain-mean gradient NRMSE is 0.269-0.612, with every 95% interval below the analytical zero-gradient reference of 1. Completely clipped cells do not show consistent structural recovery. We therefore claim conditional RAW estimation with measured structure fidelity under the stated conditions, not a demonstrated detection or driving-safety benefit.

Q1 - The empty example directories. The omissions were intentional to satisfy the 100 MB attachment limit, but the original documentation did not make this clear, and we apologize for the confusion. The updated package includes the inference implementation, one Canon EOS 5D checkpoint, its full-resolution RGB input and paired RAW reference, and pre-generated comparisons and per-example metrics for all five domains. The other four checkpoints and input pairs are omitted for size, so those four outputs cannot be regenerated from this compact package alone. Source code/README.md explains this coverage, and the demo/validation entry points process the available Canon example rather than requiring all five. This is an inference demonstration, not a full training or five-domain benchmark reproduction.

Q2 - RAWGen (Appendix A.7, Table 15). We added a camera-specific RAWGen adaptation on all five domains and evaluated both native resizing and tiling on the complete test sets. We disclose its single-RGB-view training and the unverified equivalence of its XYZ codec to the official pipeline, rather than treating the resulting scores as a definitive ranking against a faithful RAWGen reproduction. The existing benchmark comparisons with RAW-Diffusion and SpiralDiff remain unchanged.

We have tried to make both the new evidence and the remaining documentation limits easy to inspect. We hope the revised measurements, paired-RAW validation, and clarified supplementary package address the concerns underlying your suggestions.

Thank you for the detailed response and the additional analyses. The efficiency measurements, expanded related work, revised claims on near-saturation recovery, and the RawGen comparison address my main concerns. From my perspective, this is a poster-level paper, and I remain inclined to accept it.

Thank you for your thoughtful follow-up and for taking the time to assess our responses and additional analyses. We sincerely appreciate your recognition of our work and your continued support for acceptance.

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

This paper uses the local observability of the forward model (RAW to RGB) to guide the training distribution and loss design of the inverse problem (RGB to RAW). The authors first train a differentiable forward ISP model from RAW-to-RGB paired data. After that, the authors train a deterministic RGB-to-RAW anchor using the learned forward ISP model. Then the authors use sensitivity-weighted residual to further refine the anchor. The method achieves clear improvements on five datasets.

Critical strengths

This paper uses the local observability of the forward model (RAW to RGB) to guide the training distribution and loss design of the inverse problem (RGB to RAW).

The authors use VJP to estimate coordinate-wise RAW-to-RGB response energy. This VJP sensitivity guides both perturbation sampling and residual loss for refining.

Critical weaknesses

The authors hypothesize "If the perturbation barely changes the RGB output, nearby RAW candidates are hard to distinguish from the observation; if it induces a notable RGB response, the perturbation can be readily identified." in line 080-082, which should be statistically validated.

Questions for authors

ISP is a non-invertible process, and multiple RAWs may produce the same RGB, i.e.,

$ F(x_1) = F(x_2) = y, \quad x_1 \neq x_2. $

But, the model cannot know RAW is x1x_1 or x2x_2. So, in my understanding, a higher PSNR may not imply the model has the power to recover the original RAW info before the ISP. I wonder how the authors view this problem.

3 comments
AuthorsOpenReview ↗

Response to Reviewer pJcv: local response, non-injectivity, and reconstruction

Thank you for highlighting the distinction between local forward response and recovery of a unique RAW measurement. This prompted us to revise the motivating statement and test its relevance to reconstruction directly. The detailed response is in Response_to_Reviewers.pdf, pp. 5-6, in the updated supplementary ZIP. The appendix and table references below refer to the revised manuscript uploaded in the PDF field.

W1 - Statistical validation of the local-response interpretation (Appendix A.3, Tables 7-9). We now state that a small RGB response makes a particular RAW perturbation less distinguishable under the specified forward mapping, rather than implying unique recoverability. We added three checks:

First, we measure the error of the first-order approximation at the native perturbation scale. The relative Taylor remainder is 7.87-12.71% for the learned ISP and 3.83-4.10% for the independent renderer across the five domains. This quantifies the approximation error rather than assuming exact local linearity.

Second, we test the VJP coordinate ranking against finite-difference responses from an independent, fixed renderer. The mean response energy of the bottom-ranked 20% is 0.1725-0.2180 times that of the top-ranked 20%, with all ratio intervals below one.

Third, we test association with actual Anchor error against paired RAW. After conditioning on CFA phase and 16 intensity bins, the VJP-derived ambiguity score exceeds the grouped-shuffle control in correlation by 0.0812-0.1760, with positive paired image-bootstrap intervals in all five domains. Under stricter, exact 256-level intensity conditioning, four domains retain positive intervals and FiveK Nikon's interval includes zero. We report this qualification and the corresponding sample coverage.

Q1 - Non-injectivity and the meaning of PSNR. We agree that if F(x1) = F(x2), RGB alone cannot determine which RAW measurement was captured. FiberR2R predicts one conditional estimate using camera-specific paired training data and image context; it does not claim a unique physical inverse. Higher PSNR means closer agreement with the paired RAW reference under the evaluated distribution, not proof that information discarded by the ISP has been recovered. The introduction, problem formulation, and application discussion now follow this distinction.

Appendix A.2 further evaluates predictions against measured RAW near saturation. It finds structure agreement in eligible partially saturated regions, where some RGB cues remain, but not consistent structural recovery in completely clipped cells. This provides a direct check of the scope of the reconstruction claim.

These revisions preserve the practical objective of the work: improving conditional RAW estimation using local forward-response information during training, while requiring only the Anchor and one bounded refinement at inference. We appreciate your question because it helped us separate the useful statistical interpretation from a stronger recoverability claim that the evidence does not support.

Thanks for authors' rebuttal. My concerns are addressed by the authors' rebuttal and revisions. I would lean toward a weak acceptance, slightly above the threshold.

We sincerely thank the reviewer for their positive feedback and recognition of our revisions. We greatly appreciate your constructive comments and thoughtful suggestions, which have helped us improve our manuscript.

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

FiberR2R trains a differentiable forward-ISP surrogate, an RGB-conditioned RAW Anchor, and a bounded residual Transformer. Squared VJPs estimate the surrogate's coordinate-wise RGB response energy at the Anchor. These estimates guide both training perturbations and complementary, separately normalized residual losses. Inference uses the Anchor and one residual correction, without the surrogate or VJPs (pp. 3–5, Eqs. 1–12). Table 1 reports the highest SSIM on all five camera-specific datasets and highest PSNR on four, with 11.51M parameters and 10.38 GFLOPs (p. 6).

Critical strengths
  • Strong reported reconstruction results with compact inference. The method improves over the fixed Anchor by 1.65–5.46 dB on five datasets (p. 7, Table 2). Its parameter count is lower than those of RAW-Flow and SpiralDiff.

  • The guidance has evidence beyond the full-model comparison. Across all datasets, VJP variants employing either "sampling-only" or "loss-only" strategies outperform the "Uniform" sampling baseline; combining these two approaches yields further improvements (page 8, Table 3). Additionally, probe-count diagnostics, surrogate finite differences, and an independently constructed renderer provide effective, complementary means of verification (pages 8–9, Figure 5 and Table 4).

  • The analysis states its scope clearly. Appendix A.6 distinguishes coordinate energies from general correlated directions and derives the expected linearized response under the proposed sampler (pp. 13–14, Eqs. 14–18). The paper acknowledges deterministic estimation and the bounded correction's reachability limit (pp. 2, 13). Three-seed RAW-PSNR standard deviations are 0.0329 and 0.0084 dB on the two evaluated datasets (p. 13, Table 5).

Critical weaknesses
  1. Key guidance settings and implementation details need clarification. Appendix A.6 explains how the various complementary losses are combined; however, I could not find information in the PDF document regarding the settings for "sensitivity smoothing," "perturbation scale," "intensity offset," or "training probe-count." The probe-count sweep experiments do not explicitly state the specific configurations used during training. Disclosing these settings, the implementation scope of the specified normalization, and the observed "sensitivity-weight distribution" would facilitate the reproduction of this method and provide greater clarity in interpreting its performance during training (see pages 5, 8, and 11–14).

  2. The sampling comparison does not fully separate where perturbations are applied from their overall magnitude. Although the variants share the nominal scale and boundary rule, their weights can yield different total RAW perturbation magnitudes after that rule is applied. Comparing VJP-sample-only with Uniform at matched perturbation energy and with the same residual loss would help assess whether the gain extends beyond changes in overall magnitude. Figure 5 already matches RAW norms for local response diagnostics; the proposed control concerns reconstruction training. The loss-only variant improves over Uniform with the same sampler, providing separate evidence for loss guidance (pp. 7–8, Fig. 5 and Table 3).

  3. The “beyond intensity” interpretation needs finer controls. Table 4 includes useful inverse noise-factor and grouped-shuffle comparisons; the shuffle preserves region, CFA channel and intensity group. However, the within-intensity correlation statistic pools CFA channels within four broad, equal-width intensity bins. Finer intensity matching within each CFA channel would strengthen the interpretation. The independent-renderer test supports transfer of sensitivity rankings to that fixed renderer. Table 3 does not report reconstruction training guided by inverse noise-factor scores or by sensitivity estimates from the frozen ISP with its learned convolutional residual omitted. These comparisons would help assess alternative guidance signals and the contribution of learned spatial processing (pp. 8–9, 12).

  4. Some evaluation details and application claims need clearer reporting. Although the paper reports retraining official implementations, the PDF does not specify each baseline's camera-specific training and full-resolution inference configurations, making comparability harder to assess (pp. 6, 12). The claims in Appendix A.1 regarding the "recovery of high-bit-depth information" and "enhancement of driving safety" go beyond what is supported by the qualitative examples in Figure 6; these assertions require quantitative evidence or should be phrased with greater caution (p. 11). There is also a specific reporting mismatch: Appendix A.3 attributes bootstrap intervals to Table 3, which contains only point estimates, while Table 4 reports intervals (pp. 8–9, 12). Additionally, the reference to Figure 6(a)/(b) in the introduction should instead point to Figure 1 (pp. 1–2).

Questions for authors
  1. Please specify τ, σ_s, ε_s, training K, and the stabilizer ε in Eq. 10. Are the means in Eqs. 7 and 10 computed per patch/image or across a minibatch, and over which axes? What are the per-dataset distributions of the mean sensitivity weight, and how often is the numerical safeguard in the normalization used?
  2. With F and the Anchor frozen, uniform residual losses, and p0 = 0.75, does VJP-sample-only retain its advantage over Uniform when expected RAW perturbation energy is matched after Eq. 8's boundary constraint? Separately, how does reconstruction perform when guidance uses SD-q⁻¹ or derivative energies from the same frozen ISP's branch excluding its convolutional residual?
  3. For Table 4, can you report VJP-versus-shuffle and VJP-versus-SD-q⁻¹ differences within CFA channels and finer intensity strata, including sample counts and image-bootstrap intervals for those differences? Please specify the existing shuffle groups and how stratum-level correlations are aggregated.
  4. Please provide each baseline's camera-training and windowing/blending configuration, resolve A.3's Table 3 interval reference, and qualify A.1's recovery and safety statements unless quantitative evidence is available.
3 comments
AuthorsOpenReview ↗

Response to Reviewer oJuw: controlled experiments and implementation details

Thank you for your detailed review and for identifying the controls needed to interpret the guidance mechanism. We have added reconstruction-training controls, finer response analyses, and implementation details, and revised the claims to reflect what these experiments establish. The point-by-point response is in Response_to_Reviewers.pdf, pp. 1-5, in the updated supplementary ZIP. All appendix and table references below refer to the revised manuscript uploaded in the PDF field.

W1/Q1 - Settings and normalization (Appendix A.6, Tables 11-13). We specify tau = 0.15, sigma_s = 0.25, epsilon_s = 1e-4, training K = 1, and the loss-denominator floor epsilon = 1e-4. Sensitivity is normalized per patch over CHW; each complementary loss uses its own process-local BCHW mean, not a global-batch denominator. We report coordinate and patch-mean weight distributions from 1,024 training patches per domain. The scale safeguard was inactive in all 5,120 sampled patches, and each denominator safeguard was inactive in all 1,280 sampled microbatches. The probe-count sweep is diagnostic only.

W2/Q2 - Separating allocation, energy, and weighting (Appendix A.13, Tables 23-24). We performed the requested reconstruction-training comparison with frozen ISP and Anchor, uniform residual supervision, p0 = 0.75, and no auxiliary losses. MatchedUniform matches VJP's expected RAW perturbation energy per patch after the boundary constraint, without retaining its coordinate-wise allocation. Across three training seeds from a shared initialization, VJP minus MatchedUniform is +0.1664 dB in Canon and +0.0069 dB in NOD Nikon; both 95% intervals include zero. We therefore do not claim an established sampling-only advantage beyond energy. The separate factorial study finds positive weighting effects under MatchedUniform: +0.0554 and +0.0072 dB, respectively, with intervals above zero. We report all contrasts, including those favoring controls, and distinguish these exploratory effects from a universal component-level gain.

W3/Q2/Q3 - Finer conditioning and alternative scores (Appendix A.12, Tables 20-22). The new analysis uses 16 intensity bins within each CFA phase. VJP minus grouped-shuffle correlation is positive at domain level in all five domains (+0.2775 to +0.3961), with paired image-bootstrap intervals above zero. The appendix specifies shuffle groups, eligibility, aggregation, and counts for all 320 nominal strata; per-stratum estimates and intervals are supplied in stratified_response.csv in the supplementary ZIP. The inverse noise-factor score has higher diagnostic correlation under this finer conditioning, which we report explicitly. Direct reconstruction training nevertheless favors VJP over SD-q^{-1}: +0.1685 dB in Canon (95% CI [0.0164, 0.3262]) and +0.1159 dB in NOD Nikon ([0.0956, 0.1369]) across three seeds. VJP-noConv intervals include zero. Thus, these experiments support reconstruction utility relative to the tested noise-factor alternative, but do not establish an additional benefit from the learned convolutional residual. noConv retains fixed demosaicing and global color/tone processing.

W4/Q4 - Reporting and application scope (Appendices A.1-A.2, A.7, A.9, and A.11). We document available baseline configurations, distinguish historical quality evaluation from the new timing procedures, and identify unrecovered settings rather than infer them. We corrected the Table 3/Table 4 bootstrap attribution and the introduction's Figure 1 reference. Paired measured-RAW comparisons now quantify structure agreement in eligible partially saturated regions, while completely clipped cells are reported separately. We no longer claim recovery of uniquely lost sensor information or measured safety gains.

The original reconstruction benchmark results are unchanged. We believe the revision better supports the paper's central contribution: learned forward-response guidance for a compact, one-step conditional RAW estimator, with clearer evidence about where individual mechanisms help. We would welcome any remaining questions about these controls or their interpretation.

Dear Reviewer oJuw,

We sincerely thank you for your thoughtful review and constructive feedback on our submission.

We have carefully considered your concerns and provided detailed responses on November 13, including additional controlled experiments, statistical analyses, comparisons with alternative guidance signals, and full-resolution efficiency measurements. We have also clarified the implementation details and refined our claims accordingly.

We would be very grateful if you could kindly take another look at our responses and additional experimental evidence. We hope these additions have adequately addressed your concerns. If any questions or reservations remain, we would sincerely appreciate the opportunity to provide further clarification.

Thank you again for your valuable time and consideration.

Thank you the authors for the detailed response. I have raised my rating to 3.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions and Additional Evidence

We thank all three reviewers for their careful reading and constructive feedback. We have revised the manuscript and supplementary materials, added experiments directed at the concerns raised, and provided a point-by-point response to every comment. We summarize the main changes here to help the Area Chair and reviewers assess the revision. The updated supplementary ZIP contains Response_to_Reviewers.pdf, the inference example package, and the per-stratum statistics. Appendix references below refer to the revised manuscript.

Guidance mechanisms and experimental controls (Reviewer oJuw; Appendices A.6, A.12, and A.13). We added the requested energy-matched reconstruction-training control, a sampler-by-loss factorial study, finer CFA-phase/intensity-conditioned analyses, and reconstruction comparisons with inverse noise-factor and noConv guidance. The repeated training comparisons use three seeds in Canon and NOD Nikon. VJP guidance improves PSNR over the tested inverse noise-factor alternative by 0.1685 dB and 0.1159 dB, respectively, with positive paired 95% intervals. We also report the inconclusive matched-energy sampling-only and noConv contrasts and have narrowed the corresponding component-level claims. Exact guidance parameters, normalization axes, weight distributions, and safeguard frequencies are now specified.

Statistical motivation and reconstruction scope (Reviewer pJcv; Appendix A.3). We added tests of the local linear approximation, sensitivity-ranking transfer to a fixed independent renderer, and association with actual Anchor reconstruction error under finer intensity and CFA controls. We also clarified throughout the paper that FiberR2R performs conditional RAW estimation, not recovery of a unique physical inverse. Higher PSNR measures closer agreement with paired RAW references under the evaluated distribution.

Practical evaluation and supporting materials (Reviewer 2ADi; Section 2 and Appendices A.1-A.2 and A.7-A.9). We added full-resolution latency and memory measurements for five methods on all 947 test images, with three repetitions per image, and clarified the resolution used for GFLOPs and the available stage-wise training budgets. We reorganized Related Work and replaced unsupported application claims with paired measured-RAW evaluation, distinguishing partial saturation from complete clipping. We also evaluated a camera-specific RAWGen adaptation while explicitly documenting its differences from the original pipeline. Finally, we clarified the size-limited supplementary package and adjusted its entry points to use the included Canon checkpoint and example.

The original benchmark results remain unchanged: among the compared methods, FiberR2R achieves the highest SSIM on all five camera-specific datasets and the highest PSNR on four. Its deployed model uses 11.49M parameters and one bounded refinement step, without evaluating the forward ISP or its derivatives at inference. The new analyses complement these results by examining the training guidance more directly and making the scope of the conclusions clearer.

We have incorporated the substantive clarifications and additional evidence into the manuscript itself, not only into the response. We respectfully ask the Area Chair and reviewers to consider the revised submission as a whole, taking this additional evidence and the clarified claims into account alongside the reconstruction results and compact inference design. We hope the revision provides a stronger basis for a favorable assessment of the work, and we remain available to address any further questions.

AuthorsOpenReview ↗

Request for an evidence-based assessment of the revised submission

Dear Area Chair, Senior Area Chairs, and Program Chairs, We would like to ask for your careful consideration of our revised submission, FiberR2R (Submission 20591), and the additional evidence provided during the rebuttal. We took all three reviews seriously and devoted substantial time and GPU resources to addressing the questions raised. Our responses on November 13 include new controlled experiments, statistical analyses, comparisons with alternative guidance signals, and full-resolution efficiency measurements. We also clarified implementation details and revised claims where the evidence called for a more cautious interpretation. Our main concern is that we have not yet received a public follow-up from Reviewers oJuw or pJcv explaining whether these additions resolve their concerns. Both reviewers reported confidence 2, indicating limited direct research experience on this topic. Their reviews acknowledge strengths in the reconstruction results and the guidance mechanism, but their visible recommendations remain weak rejection. We understand that a lower confidence score does not invalidate a review. Still, without further discussion, we are unsure which concerns they consider unresolved or what further clarification would help. We would appreciate it if the Area Chair could assess our responses and the revised evidence directly, and ask the two reviewers to clarify any remaining technical objections. Reviewer 2ADi has already confirmed in their November 18 follow-up that our revisions address their main concerns and that they remain inclined to accept the paper. We believe FiberR2R makes a useful contribution by using learned forward-ISP sensitivity to guide reconstruction training while keeping inference compact. In the reported camera-specific comparisons, it achieves the highest SSIM on all five datasets and the highest PSNR on four (Table 1). We have also applied the method to large-scale data processing in real-world autonomous-driving workflows, which gives us confidence in its practical value. All reported results come from actual experiments, and we stand behind their accuracy and reproducibility. We have documented the experimental settings and the scope of the supplementary materials, and we are willing to provide further supporting details if needed. We hope the final decision will reflect the paper's contributions and the full evidence now available, including our responses to the original concerns. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.