acceptodds
Under review as a conference paper at ICLR 2027

PCAR-Flow: Physics-Calibrated Allocation of Residuals for Metalens Imaging

Abstract

Metalens measurements exhibit highly nonuniform recoverability across field position, wavelength, spatial frequency, and signal level. Generative priors can recover detail in weakly constrained regions, but may also introduce content that is not uniquely supported by the measurement. We propose Physics-Calibrated Adaptive Residual Flow (PCAR-Flow), which allocates generative freedom according to effective local observability, a calibrated measure combining optical transfer, sensor noise, and local image statistics. PCAR-Flow recovers measurement-supported content through a deterministic path and applies conditional flow matching only to unresolved high-frequency residuals. Under the local Gaussian model, the complement of observability equals normalized posterior band uncertainty and, for channel-separable transfer, approaches the classical null-space limit as measurement noise vanishes. Experiments show that PCAR-Flow improves reconstruction quality while keeping forward inconsistency close to the deterministic solution.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computational Imaging, Metalens Imaging, Residual Flow, Physics Prior, Deep Learning

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_wmJ6OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PCAR-Flow, a physics-calibrated generative method for metalens image reconstruction. Metalens measurements suffer from highly non-uniform recoverability across spatial positions and frequencies. To address this, the authors introduce "effective local observability," a metric combining optical transfer, sensor noise, and local image statistics to quantify measurement support. The pipeline first uses a deterministic base reconstructor to recover well-observed structures. Then, it applies conditional flow matching exclusively to unresolved high-frequency wavelet residuals. Finally, a reliability-gated admission module controls how much generated detail is injected into the final image. Theoretical analysis links observability to posterior uncertainty under a local Gaussian model. Experiments demonstrate that PCAR-Flow outperforms strong baselines (e.g., DMDiff) on the M1 dataset and successfully generalizes to a different metalens system (M2) without retraining.

Critical strengths
  1. Strong physical grounding and elegant framework: The concept of "effective local observability" provides a principled way to bridge physical optics and generative models. By restricting the flow matching to high-frequency residuals and gating them based on physical cues, the method avoids hallucinating details where the measurement already provides strong support. This effectively balances perceptual quality and physical consistency.

  2. Impressive cross-metalens generalization: The experiment transferring the M1-trained model directly to the M2 metalens system (using only the new calibration package) is a critical strength. It proves that the learned prior allocation mechanism is genuinely physics-aware and not simply overfitting to a specific lens's degradation pattern.

Critical weaknesses
  1. Reliance on ground-truth oracle for gate training: The admission gate is supervised using an oracle admission map (Eq. 21), which requires computing the least-squares scale between the proposal and the true target residual. While upstream modules are frozen, relying on a ground-truth-dependent oracle during Stage C training raises concerns about whether the gate will generalize optimally at inference time when no ground truth exists. The gap between oracle supervision and real-world test-time behavior needs more justification.

  2. Fragility of the local Gaussian assumption: The theoretical interpretation of observability relies heavily on a local Gaussian model for both the latent state and noise. However, natural images contain sharp edges and highly non-Gaussian local structures. The paper acknowledges this briefly in the limitations, but given that observability is the core signal driving the entire pipeline, its accuracy in non-Gaussian regions could silently degrade the gating performance.

Questions for authors
  1. Regarding the gate training (Stage C): Have you tried training the admission gate purely via reinforcement learning or differentiable physics losses (e.g., optimizing EphysE_{phys} and LPIPS directly) without the oracle map gs⋆g^\star_s? If so, how does the performance compare?

  2. How sensitive is the overall pipeline to small, realistic errors in the calibration package (e.g., slight misalignment in the PSF grid)? Does the gate naturally suppress bad generations in these cases?

  3. The method excludes the LL band from the flow matching. Could applying a very lightweight flow to the LL band further improve low-frequency color/contrast recovery without significantly harming EphysE_{phys}?

5 comments
AuthorsOpenReview ↗

Clarifications Regarding Gated Supervision, Calibration Robustness, Band Selection, and Observability Assumptions

In response to the reviewer's concerns regarding oracle supervision, calibration errors, and local statistical assumptions, we provide further explanations of the relevant mechanisms and analyze their applicability and limitations through controlled experiments.

Q1. Contribution of Oracle Admission Supervision and Alternative Gate-Training Objectives

Eq. (21) defines the oracle admission target as the clipped local least-squares coefficient between the generated residual and the reference residual. This target is used exclusively during Stage C training, where the base reconstructor and residual flow are frozen and the generated proposals are detached. The gate is therefore trained to estimate how much of each proposed residual should be admitted, with Eq. (22) further incorporating physical-consistency and monotonicity constraints. At inference, the oracle target and reference image are not required; the gate predicts admission coefficients solely from the observable cues Ωs\Omega_s, UbU_b, CbC_b, and rbr_b.

To assess whether the gate can be trained without oracle admission targets, we compare two alternative objectives using the same frozen upstream networks and gate architecture (Rebuttal Table W1). The first directly combines perceptual and physical-consistency losses, while the second uses only physical-consistency and monotonicity losses. The perceptual-physical objective achieves 30.39 dB PSNR and 0.1357 LPIPS, approaching the original oracle-supervised result (30.62 dB and 0.1289, respectively), but with a higher EphysE_{\mathrm{phys}} of 0.0442 versus 0.0398. In contrast, physical-consistency and monotonicity losses alone reduce EphysE_{\mathrm{phys}} to 0.0380, but also lower the mean gate activation to 0.117, with PSNR and LPIPS deteriorating to 29.84 dB and 0.1530. This indicates that optimizing physical consistency without a direct perceptual objective tends to produce overly conservative residual admission, limiting the recovery of useful details.

The comparison shows that differentiable objectives can train the gate without an oracle admission map, but do not fully reproduce the balance achieved by oracle-supervised training. The local least-squares target provides explicit spatial supervision for residual allocation, while the physical-consistency and monotonicity terms constrain its effect on the measurements. Importantly, this supervision is confined to training and introduces no additional reference-image requirement at inference.

Rebuttal Table W1. Gate Training without Oracle Admission Targets.

Gate training objectivePSNR ↑LPIPS ↓EphysE_{\mathrm{phys}} ↓
Original oracle-supervised gate30.620.12890.0398
Perceptual + physics losses30.390.13570.0442
Physics + monotonicity losses only29.840.15300.0380

Q2. Robustness to Calibration Mismatch and PSF-Grid Misregistration

Table 6 and Figure 6 evaluate complementary forms of calibration mismatch. Table 6 considers joint moderate calibration drift, physical remounting, and thermal drift, while Figure 6 reports a controlled PSF-perturbation sweep. Under joint moderate drift, the mean gate activation decreases from 0.382 to 0.255, whereas LPIPS and EphysE_{\mathrm{phys}} increase from 0.1289 to 0.1544 and from 0.0398 to 0.0602, respectively. Physical remounting and thermal drift exhibit similar reductions in gate activation. These results suggest that the gate becomes more conservative under degraded calibration, although the remaining reconstruction errors indicate that admission control cannot fully compensate for an inaccurate forward model.

To specifically examine spatial misregistration of the PSF calibration grid, we introduce controlled coordinate offsets of ±5% and ±10% of the nominal grid spacing (Rebuttal Table W2). The measured inputs and learned network parameters are kept unchanged, while the spatial coordinates associated with the PSF calibration samples are perturbed. For each offset, the calibration-dependent forward operator and reconstruction cues are recomputed, producing the corresponding base reconstruction and residual proposals. We then compare gated and always-on admission using the same base reconstruction and residual proposals within each perturbation condition, with only the admission map changed (gs=1g_s=1 for always-on admission). This paired comparison isolates the effect of residual admission under each calibration error.

For physical-consistency evaluation, we use the nominal, unperturbed calibration operator rather than the perturbed operator supplied to the reconstruction pipeline. Thus, EphysE_{\mathrm{phys}} measures the discrepancy relative to the nominal imaging system, without using the evaluation operator as an inference-time input.

At ±5% displacement, gated admission reduces LPIPS from 0.1511 to 0.1337 and EphysE_{\mathrm{phys}} from 0.0618 to 0.0436 compared with always-on admission. At ±10%, the corresponding values decrease from 0.1616 to 0.1402 and from 0.0724 to 0.0501. Meanwhile, mean gate activation decreases from 0.362 to 0.330 as the displacement increases, indicating a more conservative admission response under larger calibration misalignment.

Although reconstruction quality still degrades with increasing misalignment, gated admission consistently yields lower perceptual error and forward inconsistency than unrestricted residual injection at both perturbation levels. These results demonstrate that the reliability gate mitigates the adverse effects of PSF-grid misregistration by regulating residual admission, rather than correcting the underlying calibration error.

Rebuttal Table W2. Sensitivity to Spatial PSF-Grid Misalignment.

PSF-grid displacementAdmissionLPIPS ↓EphysE_{\mathrm{phys}} ↓Mean gate
±5%Gated0.13370.04360.362
Always-on0.15110.06181.000
±10%Gated0.14020.05010.330
Always-on0.16160.07241.000

Q3. LL-Band Selection and Reliability-Guided Low-Frequency Generation

We retain the LL band in the deterministic pathway to preserve low-frequency structures supported by the calibrated measurements. As shown in Supplementary Table 13, unrestricted LL generation slightly improves LPIPS (0.1289 to 0.1261), but reduces PSNR from 30.62 to 30.14 dB and increases EphysE_{\mathrm{phys}} from 0.0398 to 0.0619, indicating a trade-off between perceptual improvement and measurement consistency.

To examine whether this trade-off can be mitigated, we further evaluated a lightweight LL residual branch with three admission strategies: always-on, fixed scaling, and reliability-gated admission. The deterministic base and high-frequency flow were kept frozen, with identical LL residual proposals across the three configurations; only the LL admission strategy was varied. As shown in Rebuttal Table W3, reliability-gated admission provides the best balance among the tested LL configurations. Compared with always-on admission, it improves PSNR from 30.38 to 30.59 dB and reduces EphysE_{\mathrm{phys}} from 0.0529 to 0.0416 (21.4%), LL-MAE from 0.0249 to 0.0231 (7.2%), and ΔE00\Delta E_{00} from 3.64 to 3.36 (7.7%), with only a marginal LPIPS increase (0.1268 to 0.1271).

However, compared with the original HF-only PCAR-Flow, the reliability-gated LL variant provides only a modest LPIPS improvement (0.1289 to 0.1271), while slightly reducing PSNR (30.62 to 30.59 dB) and increasing EphysE_{\mathrm{phys}} (0.0398 to 0.0416). Its LL-MAE and ΔE00\Delta E_{00} also remain higher (0.0231 vs. 0.0197 and 3.36 vs. 3.22), with inference time increasing from 0.98 to 1.11 s per image. Thus, although reliability gating substantially mitigates the adverse effects of unrestricted LL generation, the additional branch does not improve all reconstruction criteria. We therefore retain the LL band in the deterministic pathway in the default PCAR-Flow, while demonstrating the potential of reliability-guided admission as a more controlled alternative to unrestricted low-frequency generation.

Rebuttal Table W3. Effect of Lightweight LL Residual Modeling and Admission Strategies.

ConfigurationLL AdmissionPSNR ↑LPIPS ↓EphysE_{\mathrm{phys}} ↓LL-MAE ↓ΔE00\Delta E_{00} ↓Time (s) ↓
Original HF-only PCAR-Flow--30.620.12890.03980.01973.220.98
Lightweight LL flowAlways-on30.380.12680.05290.02493.641.09
Lightweight LL flowFixed scaling30.510.12760.04620.02383.491.09
Lightweight LL flowReliability-gated30.590.12710.04160.02313.361.11

Q4. Reliability of Effective Observability under Non-Gaussian Structures and Spatially Varying PSFs

The posterior interpretation of effective observability in Appendix A.3 is derived under a local plug-in Gaussian model for the latent coefficients and measurement noise, yielding Vs∣μ,c(p)Ps,c(p)=1−Ωs(p,c)\frac{V_{s|\mu,c}(p)}{P_{s,c}(p)}=1-\Omega_s(p,c). This is a local statistical approximation, not a claim that natural images are globally Gaussian or that the identity holds exactly for arbitrary non-Gaussian structures. To examine its empirical reliability beyond relatively smooth regions, we further stratified the evaluation by local image gradients and spatial PSF variation (Rebuttal Table W4).

The calibration accuracy degrades in more challenging regions. Specifically, the calibration MAE increases from 0.0408 in low-gradient regions to 0.0718 in high-gradient regions, accompanied by a decrease in Spearman correlation from 0.937 to 0.846. In regions with rapid PSF variation, the MAE further reaches 0.0826 and the correlation decreases to 0.785. The corresponding degradation rates are 10.9%, 19.4%, and 22.7%, respectively. These results expose the limitations of the local approximation under complex image and optical conditions, rather than implying uniformly reliable observability estimates across all regions.

Importantly, the admission gate does not rely on Ωs\Omega_s alone. It jointly incorporates baseline uncertainty UbU_b, confidence CbC_b, and forward disagreement rbr_b (Eq. 16), providing complementary reliability information for residual allocation. Table 5 supports their respective contributions: removing UbU_b, CbC_b, or rbr_b increases EphysE_{\mathrm{phys}} from 0.0398 to 0.0460, 0.0452, and 0.0493, respectively. Thus, observability serves as one component of a multi-cue admission mechanism, rather than a standalone estimate that determines residual injection. We will clarify the scope of the local Gaussian interpretation and include the region-wise analysis to make these limitations explicit.

Rebuttal Table W4. Region-Wise Observability Calibration and Reconstruction Fidelity.

RegionCalibration MAE ↓Spearman rsr_s ↑Degradation Rate ↓
Low-gradient (bottom 80%)0.04080.93710.9%
High-gradient (top 20%)0.07180.84619.4%
Rapid PSF variation (top 10%)0.08260.78522.7%
Reviewer_LwehOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PCAR-Flow combines calibrated image reconstruction, high-frequency residual generation, and an observability-guided gating mechanism for metalens imaging. The goal is to improve perceptual quality while limiting changes unsupported by the measurement. Experiments on M1 demonstrate a favorable trade-off between reconstruction fidelity and perceptual quality, and the model transfers to independently calibrated M2 optics without network retraining.

Critical strengths

The problem is relevant, and separating reliable reconstruction from generated detail is a sensible design. Table 4 shows a useful balance between perceptual quality and physical consistency. Transfer with frozen network weights is also valuable, while the component ablations and supplementary efficiency analysis provide useful supporting evidence.

Critical weaknesses
  1. Insufficient discussion of related work. Recent approaches to learned null-space priors (e.g., NPN and GSNR), physics-conditioned generative reconstruction (e.g., PG-ControlNet), and wavelet-domain flow matching (e.g., FlowLet) are relevant. The paper would benefit from a clearer discussion of the methodological differences.
  2. Figure 3 presents many reconstruction results at a relatively small scale, making it difficult to examine fine-grained details and distinguish the differences between methods, particularly DMDiff and PCAR-Flow. Moreover, the absolute error maps lack a clearly specified color scale, making their interpretation less straightforward. Enlarged crops and consistent, labeled error-map scales would improve the clarity and informativeness of the qualitative comparisons.
  3. Insufficient isolation of observability conditioning. Although Tables 3–5 provide useful ablations, the independent contribution of explicit observability conditioning within the learned admission gate remains unclear. A controlled comparison against an otherwise identical learned gate without observability input would strengthen the central claim.
  4. Methodological complexity and clarity. The framework involves multiple stages and conditioning signals, making it difficult to identify the essential contributions. Additional intuitive explanations of the interactions between observability estimation, residual flow, and reliability gating would improve clarity.
  5. Limited failure-case and perceptual analysis. Although robustness under calibration mismatch is evaluated, systematic failure cases involving inaccurate calibration and hallucinated details are insufficiently explored. A user study could further strengthen the perceptual evaluation.
Questions for authors

The paper addresses a relevant problem and reports promising quantitative results. However, the contribution relative to existing approaches, the necessity of the proposed observability-guided mechanism, and the qualitative evidence for improved detail recovery remain insufficiently convincing.

I therefore lean toward rejection at this stage. I am open to increasing my rating if the authors can provide stronger comparisons, clearer methodological justification, and more convincing visual evidence during the rebuttal.

  1. How does PCAR-Flow differ conceptually from recent learned null-space priors (e.g., NPN and GSNR) and physics-conditioned generative reconstruction methods (e.g., PG-ControlNet), particularly in handling partially observable image components?

  2. Could the authors clarify the practical significance of the bound in Eq. (19)? Does it provide any meaningful guarantee on the physical consistency of generated details, beyond bounding the increase in forward-model error?

  3. How much benefit does explicit observability conditioning provide over an otherwise identical learned gate using only uncertainty, confidence, and forward disagreement? This comparison would help isolate the contribution of observability from that of the learned gating mechanism itself.

  4. Could the authors provide a more intuitive explanation of how Eqs. (13)–(15) and (19) relate to the overall framework, particularly the role of the physical inconsistency bound?

  5. What are the main failure modes under inaccurate optical calibration or severely underconstrained image regions?

5 comments
AuthorsOpenReview ↗

Clarifications on Observability-Guided Residual Allocation, Physical Consistency, and Reconstruction Reliability

We thank the reviewers for their thorough comments. In response to the issues raised, we further clarify the distinctions between PCAR-Flow and existing methods, the independent contribution of observability in residual gating, and the scope of application for physical consistency constraints. We also present supplementary experiments to analyze local reconstruction fidelity and reliability under calibration mismatches.

Q1. Distinction from Existing Null-Space, Physics-Conditioned, and Wavelet-Domain Generative Methods

The key distinction of PCAR-Flow is its explicit treatment of partially observable components, rather than the use of wavelet decomposition or flow matching itself.

Null-space methods such as NPN [1] and GSNR [2] construct learned priors over components in the null space of the sensing operator. NPN learns low-dimensional null-space projections, while GSNR introduces graph-smooth representations to exploit informative structures within the null space. PCAR-Flow addresses a complementary problem: image components with nonzero optical responses may nevertheless remain poorly recoverable under signal-dependent noise. Instead of a strict observable/null-space partition, effective observability Ωs\Omega_s (Eqs. 4–7) quantifies the spatially and spectrally varying degree of measurement support, including components with weak but nonzero responses.

PG-ControlNet [3] incorporates spatially varying blur information into a ControlNet-conditioned diffusion model for image deblurring, while FlowLet [4] employs conditional flow matching in an invertible 3D wavelet domain for brain MRI synthesis. PCAR-Flow differs in how these generative corrections are incorporated into reconstruction: a deterministic, measurement-supported estimate is first obtained, conditional flow matching then generates candidate high-frequency residuals, and a separate reliability gate controls their admission into the final reconstruction.

Specifically, the gate jointly uses Ωs\Omega_s, baseline uncertainty UbU_b, confidence CbC_b, and forward disagreement rbr_b (Eq. 16) to regulate residual injection. Eq. (19) further bounds the additional forward-model discrepancy introduced by admitted residuals. Thus, unlike approaches that primarily use physical or measurement information to condition generation, PCAR-Flow explicitly separates residual generation from measurement-dependent residual admission. This distinction is reflected in Table 4: compared with full-image conditional flow, PCAR-Flow reduces EphysE_{\mathrm{phys}} from 0.0748 to 0.0398 (46.8%), with an LPIPS difference of only 0.0041 (0.1289 vs. 0.1248). These results support the benefit of observability-guided residual allocation in balancing perceptual quality and physical consistency.

Reference

[1] Roman Jacome, Romario Gualdrón-Hurtado, León Suárez-Rodríguez, and Henry Arguello. NPN: Non-Linear Projections of the Null-Space for Imaging Inverse Problems. Advances in Neural Information Processing Systems 38, 2025.

[2] Romario Gualdrón-Hurtado, Roman Jacome, Rafael S. Suárez, and Henry Arguello. GSNR: Graph Smooth Null-Space Representation for Inverse Problems. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 12469–12479.

[3] Hakki Motorcu and Mujdat Cetin. PG-ControlNet: A Physics-Guided ControlNet for Generative Spatially Varying Image Deblurring. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2026 (NTIRE Workshop). arXiv:2511.21043.

[4] Danilo Danese, Angela Lombardi, Matteo Attimonelli, Giuseppe Fasano, and Tommaso Di Noia. FlowLet: Conditional 3D brain MRI synthesis using wavelet flow matching. Medical Image Analysis, 113 (2026), 104161.

Q2. Practical Meaning and Scope of the Physical-Consistency Bound in Eq. (19)

Eq. (19) provides a quantitative upper bound on the forward-model discrepancy introduced by admitted residuals:

\leq \mathcal E_{\mathrm{phys}}(\mathbf x_b) + \sum_{s\in\mathcal S_{\mathrm{hf}}} K_s \|\mathbf g_s\odot\hat{\mathbf z}_{1,s}\|_1.$$ Here, $\mathcal E_{\mathrm{phys}}(\mathbf x)$ measures the mean noise-whitened forward disagreement over the valid optical support, and $K_s$ is the induced $L_1$ operator norm of the calibrated, noise-whitened forward operator restricted to wavelet band $s$. It therefore quantifies how strongly an admitted residual in each band can affect the measured-data discrepancy. The bound separates the discrepancy already present in the deterministic reconstruction from the additional contribution that may arise through residual admission. For a fixed residual proposal, reducing its admitted magnitude tightens this upper bound.

Importantly, the guarantee concerns measurement consistency under the calibrated forward operator, not the ground-truth correctness of generated details. In weakly observable directions, different image structures may produce similarly small forward disagreements; consequently, satisfying the bound cannot establish that a generated texture or edge is authentic. Moreover, a tighter upper bound does not imply a monotonically lower actual forward error, since an admitted residual may also correct part of the discrepancy already present in the base reconstruction. The bound is conditional on the accuracy of the calibrated operator and noise covariance, and is not a hard feasibility constraint enforced at inference. Instead, Eq. (22) incorporates the differentiable physical-consistency term into gate training, encouraging residual admission to account for its effect on the measured data.

The practical relevance of this allocation is reflected in Table 4. Full-image conditional flow achieves a slightly lower LPIPS of 0.1248, but its EphysE_{\mathrm{phys}} reaches 0.0748. PCAR-Flow obtains an LPIPS of 0.1289 while reducing EphysE_{\mathrm{phys}} to 0.0398, a 46.8% reduction in forward discrepancy at an LPIPS difference of only 0.0041. Compared with the deterministic base, PCAR-Flow also improves LPIPS from 0.1568 to 0.1289, while increasing EphysE_{\mathrm{phys}} by only 0.0029 (0.0369 to 0.0398). These results illustrate the intended balance: generated residuals improve perceptual reconstruction without allowing the forward inconsistency to increase substantially.

Q3. The Independent Contribution of Observability to Learning-Based Gating

The accuracy of observability estimates is not directly equivalent to their contribution to residual gating; therefore, it is necessary to validate the two separately. Table 3 shows that, compared to using only optical transmission information, the combined use of noise and local image statistical features reduces the calibration MAE of the observability estimate from 0.118 to 0.047 and increases the Spearman correlation coefficient rsr_s from 0.69 to 0.92. Table 5 validates the complementary information provided by uncertainty, confidence, and forward disagreement. However, these results have not yet directly isolated the independent role of Ωs\Omega_s as a gating input.

To this end, we conducted further controlled ablation experiments: while keeping the deterministic baseline, residual flow, remaining reliability features, gating structure, and training protocol consistent, we varied only whether the gating network received Ωs\Omega_s. In both control groups, the monotonicity regularization term Lmono\mathcal L_{\mathrm{mono}}, which depends on Ωs\Omega_s, was removed to rule out its influence. As show in Table L1, under conditions where neither set uses the monotonicity regularization term, introducing Ωs\Omega_s causes LPIPS to decrease from 0.1392 to 0.1327 and EphysE_{\mathrm{phys}} to decrease from 0.0501 to 0.0439, representing reductions of approximately 4.7% and 12.4%, respectively. This indicates that even though the gated network already utilizes uncertainty, confidence, and forward disagreement, explicit observability still provides additional information for residual control. Upon further incorporating a monotonicity regularization term, the LPIPS and EphysE_{\mathrm{phys}} of the complete model decreased to 0.1289 and 0.0398, respectively, indicating that the observability-conditioned input and the corresponding regularization constraint have complementary effects.

Rebuttal Table L1. Controlled ablation of observability conditioning within the admission gate.

Gate configurationΩ inputMonotonicity lossLPIPS ↓Ephys ↓
Full PCAR-FlowYesYes0.12890.0398
Gate with Ω, without LmonoYesNo0.13270.0439
Gate without Ω, without LmonoNoNo0.13920.0501

Q4. The Relationship Between Residual Flow, Reliability Gating, and Physical Consistency

Eqs. (13–15) describe the process by which the conditional flow model generates high-frequency residuals, with the goal of generating candidate details based on a deterministic reconstruction. Eq. (14) introduces calibrated physical descriptors and reliability information into the conditional inputs of the flow model. However, the generated residuals are not directly superimposed onto the final reconstruction.

The actual injection of residuals is controlled by an independent gate defined in Eq. (16). The gate combines effective observability Ωs\Omega_s, baseline uncertainty UbU_b, confidence CbC_b, and forward disagreement rbr_b to yield effective corrections for each frequency band: Ws−1(gs⊙z^1,s).W_s^{-1}(g_s\odot\hat z_{1,s}). The residual flow is responsible for generating candidate high-frequency details, the gating determines the magnitude of their injection, and Eq. (19) bounds the forward measurement errors that these corrections may introduce.

Sec. 3.3 employs sequential training: the deterministic baseline and residual flow are fixed during gate optimization, allowing the gate to focus on selecting existing residual candidates rather than adapting to the gate loss by altering the generative model. This not only limits unconstrained injection of generated residuals but also enables the independent analysis of the impact of residual corrections on physical consistency.

Q5. Local Reconstruction Fidelity and Reliability of Generated Details

We agree that the compact layout of Fig. 3 makes fine-grained comparisons, particularly between DMDiff and PCAR-Flow, difficult. In the revision, we will include enlarged crops of representative regions and use a shared, explicitly labeled color scale for the absolute error maps. Quantitatively, Table 2 shows that PCAR-Flow improves over DMDiff in PSNR (30.62 vs. 30.31 dB), SSIM (0.8841 vs. 0.8731), LPIPS (0.1289 vs. 0.1705), DISTS (0.1218 vs. 0.1499), and MUSIQ (52.42 vs. 44.48). Supplementary Table 16 further shows that in the optically degraded edge region, PCAR-Flow improves PSNR by 0.87 dB (29.18 vs. 28.31) and reduces LPIPS by 16.1% (0.1624 vs. 0.1936).

To examine whether these perceptual gains also translate into improved local reference fidelity, we conducted a patch-level analysis. Test patches are stratified into low-, medium-, and high-observability groups, with PSNR and LPIPS evaluated on the same patches and reference images. We additionally report the degradation rate, defined as the proportion of patches in which residual injection increases reference MSE relative to the deterministic base. As shown in Table L2, in low-observability regions, always-on injection improves LPIPS over DMDiff (0.1943 vs. 0.2231), but reduces PSNR (28.11 vs. 28.38 dB). PCAR-Flow instead achieves both higher PSNR (28.94 dB) and lower LPIPS (0.1619), while reducing the degradation rate from 27.1% to 18.7%. The same trend holds at medium and high observability: compared with always-on injection, PCAR-Flow improves both PSNR and LPIPS and reduces the degradation rate by 7.8 and 7.4 percentage points, respectively.

Rebuttal Table L2. Local reconstruction quality across observability levels.

ObservabilityMethodPSNR ↑LPIPS ↓Degradation Rate ↓
Low ΩDMDiff28.380.2231--
Always-on28.110.194327.1%
PCAR-Flow28.940.161918.7%
Medium ΩDMDiff30.570.1661--
Always-on30.420.139820.0%
PCAR-Flow30.750.124412.2%
High ΩDMDiff32.890.1267--
Always-on32.720.106114.7%
PCAR-Flow32.960.10487.3%
Reviewer_RqjqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PCAR-Flow, a reconstruction framework for calibrated computational imaging, particularly metalens image reconstruction. It combines calibrated deterministic inversion, a signal-dependent noise model, and local observability analysis to produce a base reconstruction and reliability cues, then uses flow matching to generate only high-frequency residuals, with reliability-gated admission controlled by observability, uncertainty, confidence, and forward disagreement. Training is staged over the base estimator, residual flow, and admission gate, and a forward-consistency bound is provided.

Its contributions are decoupling calibrated physical inversion from generative priors and restricting generation to high-frequency content not sufficiently resolved by the measurement; defining effective observability via local Bayesian uncertainty reduction and using it to guide residual generation and admission; and evaluating cross-metalens calibration transfer by training on M1, freezing parameters, and using only the M2 calibration package at inference.

Critical strengths

A substantive strength is the explicit coupling of calibrated forward modeling with generative residual modeling: the deterministic base addresses resolvable structure, flow matching is restricted to high-frequency residuals, and admission is conditioned on observability, uncertainty, confidence, and forward disagreement, with a forward-consistency bound. This is consequential because it offers a physically interpretable mechanism for constraining prior-driven detail rather than applying a generative prior to the full image.

A further strength is the cross-metalens evaluation, in which M1-trained parameters are frozen and only the M2 calibration package is supplied at inference, together with ablations over observability components, prior allocation, reliability cues, and calibration mismatch. This is consequential because it examines whether the learned allocation mechanism can operate under a different calibrated forward model without target-domain retraining, which is relevant to the paper’s claim regarding calibration-dependent generalization.

Critical weaknesses

A limitation concerns the scope of evidence for calibration transfer. All learned components are trained only on M1; the cross-metalens evaluation freezes these parameters and supplies only the M2 calibration package at inference. This tests transfer to one secondary system, and the mismatch analysis is limited to controlled perturbations such as joint moderate drift, physical remount, and thermal drift. The evidence therefore supports transfer under the evaluated conditions, but it is limited for broader conclusions across substantially different optical systems, sensor changes, and calibration regimes.

A further limitation concerns dependence on calibration and the local stationarity approximation. Effective observability inherits error from the forward operator, the signal-dependent noise estimate, and the image-statistics model, and the approximation may weaken near abrupt PSF-field changes. The forward-consistency bound constrains measurement inconsistency but does not identify a unique latent solution in weakly observed directions, which limits interpretation of recovered high-frequency detail and the practical scope of claims without system-specific calibration.

Questions for authors

In the M2 cross-metalens evaluation, what are the test-set size, paired capture protocol, and reference-image acquisition procedure, and is the M2 test set independent of the M1 test set? The answer would materially affect how strongly the transfer result supports generalization. In deployment, would each device require its own system-specific calibration package, and how often would recalibration be expected under remounting, thermal drift, or sensor replacement? This bears directly on whether the calibration-dependent design is practical outside controlled laboratory conditions. Future-work suggestion: Evaluate the method in a concrete application setting with end-to-end capture, calibration, and reconstruction, and report the cost and frequency of recalibration under realistic deployment conditions.

3 comments
AuthorsOpenReview ↗

Clarifications on M2 Evaluation, Calibration-Dependent Transfer, and Observability-Based Physical Consistency

We thank the reviewer for the constructive assessment. We clarify the cross-metalens evaluation protocol and transfer results, the requirements for device-specific calibration, and the theoretical basis and physical consistency of observability-guided reconstruction.

Q1. M2 Evaluation Protocol and Its Relationship to M1

The M2 test set consists of 500 paired samples, with its scale and data partitioning summarized in Table 1. All models are trained exclusively on the M1 training set, with model selection performed using the M1 validation set. No M2 data are used for training, fine-tuning, or model selection. The M1 and M2 test sets are not strictly source-image-disjoint, with a small overlap involving less than 2% of the M2 test samples (fewer than 10 out of 500 samples). Importantly, the ground-truth reference images used to construct the M2 test pairs do not overlap with those in the M1 training or validation sets. Thus, the M2 evaluation is free from source-image overlap with the data used for model training and selection, allowing cross-metalens generalization to be assessed without target-domain adaptation.

The M2 acquisition and reference-image procedures are detailed in Appendix A.4, while the optical characteristics and evaluation specifications of M1 and M2 are summarized in Appendix Table 9. During acquisition, source images are displayed on an sRGB-calibrated LCD at a fixed distance of 0.4 m in a dark environment, with automatic exposure disabled to maintain consistency with the calibrated signal-dependent noise conditions. The imaging system is characterized through spatial PSF measurement, spectral response calibration, flat-field data, and Poisson-Gaussian noise fitting. Reference-image metrics are computed after geometric registration and radiometric alignment. Patches with an NCC value below 0.9 after registration are excluded to minimize the influence of spatial misalignment on evaluation.

Q2. Cross-Metalens Transfer under Device-Specific Calibration

The transference experiments from M1 to M2 were designed to verify whether PCAR-Flow can adapt to different optical systems without retraining the network. Appendix Table 9 shows that M1 has an aperture of 0.5 mm and a focal length of 1.0 mm, while M2 has an aperture of 10 mm and a focal length of 24.5 mm at a wavelength of 532 nm. There are significant differences in the optical responses and observability distributions between the two systems, with low-observability region ratios of 0.47 and 0.39, respectively.

In the M2 test, all network parameters were kept frozen, and the forward operator and observability descriptors were constructed using only the parameters independently calibrated for M2. As shown in Table 7 of the original paper, compared to the high-frequency residual flow, PCAR-Flow reduced the LPIPS from 0.1460 to 0.1321 on M2, representing a 9.6% improvement; while simultaneously reducing the physical consistency error EphysE_{\mathrm{phys}} from 0.0492 to 0.0410, a 16.7% reduction. This improvement in LPIPS is comparable to the 10.2% improvement observed on M1.

These results demonstrate that the residual-allocation mechanism remains effective when transferring the M1-trained model to the evaluated M2 optical system using its independently calibrated physical parameters, without target-domain retraining.

Q3. Device-Specific Calibration and Practical Deployment Requirements

PCAR-Flow requires a device-specific calibration package for each imaging system, comprising the spatially varying PSF bank, MTF, chromatic shift, flat-field correction, and signal-dependent noise model (Appendix Table 8). These calibration parameters define the forward operator and observability descriptors, while the learned network weights remain independent of the imaging device. Accordingly, when transferring a trained model to a new system, the corresponding calibration package must be established and used to reconstruct the forward operator and observability descriptors; network retraining is not required. This is demonstrated by the M1-to-M2 experiment, where the M1-trained parameters remain frozen and the M2-specific calibration is used at inference.

Table 6 evaluates the sensitivity to several calibration changes. Following physical remounting, PSNR decreases from 30.62 to 29.31 dB and LPIPS increases from 0.1289 to 0.1512. Under a thermal drift of +15°C, the corresponding values are 29.66 dB and 0.1438. Under joint moderate calibration drift, PSNR and LPIPS reach 28.94 dB and 0.1544, respectively, while mean gate activation decreases from 0.382 to 0.255. These results show that calibration changes can affect reconstruction quality, while the gate responds more conservatively under degraded calibration. However, admission control mitigates rather than eliminates the effects of calibration errors.

In practical deployment, recalibration should be triggered by changes affecting the calibrated optical or sensor response. Optical remounting calls for remeasurement of the spatial PSF field and the associated optical characteristics. Sensor replacement requires recalibration of the spectral response, flat-field correction, and signal-dependent noise parameters. Thermal drift may also require updating the affected calibration components when measurable changes occur. We do not prescribe a universal recalibration interval, as its frequency depends on hardware stability and application-specific fidelity requirements; the present study has not quantified long-term recalibration frequency or calibration acquisition costs. Once calibration is available, the reconstruction inference time is 0.98 s per 400 × 400 image at 12 NFE on an RTX A5000 (Supplementary Table 12). End-to-end deployment studies are needed to further quantify calibration overhead and maintenance requirements under sustained operation.

Q4. Observability Assumptions and Physical Consistency of Generated Details

PCAR-Flow's effective observability integrates the optical transfer function, signal-to-noise ratio, and local image statistics to quantify the degree to which measurements constrain different spatial locations and frequency components. Under the local Gaussian model described in Appendix A.3, the normalized posterior uncertainty satisfies Vs∣μ,c/Ps,c=1−ΩsV_{s|\mu,c}/P_{s,c}=1-\Omega_s. Its practical effectiveness is verified by Table 3 in the original paper: compared to using only the optical transfer function, the complete model reduces the calibration MAE from 0.118 to 0.047 and increases the Spearman correlation coefficient from 0.69 to 0.92.

Considering the non-Gaussian characteristics of natural images and the applicability of the local stationarity assumption, Eq. (16) does not rely solely on Ωs\Omega_s, but further incorporates reconstruction uncertainty UbU_b, confidence level CbC_b, and forward residual rbr_b to modulate the admission of generated residuals. Table 5 shows that after removing these three pieces of information individually, EphysE_{\mathrm{phys}} increased from 0.0398 to 0.0460, 0.0452, and 0.0493, respectively, indicating that they play complementary roles in residual allocation.

Eq. (19) bounds the additional forward-model discrepancy introduced by admitted residuals under the calibrated operator. It therefore provides a quantitative constraint on measurement inconsistency, but does not guarantee the semantic correctness of generated details or identify a unique solution in weakly observable directions. Multiple high-frequency structures may remain compatible with the same measurement. Accordingly, we interpret the bound as controlling the physical impact of generative corrections, rather than certifying their ground-truth accuracy. Together, the observability calibration in Table 3, the reliability-cue ablations in Table 5, and the calibration-mismatch results in Table 6 support the use of measurement-dependent gating to improve perceptual reconstruction while limiting additional forward-model inconsistency.

Dear Reviewer Rqjq,

Thank you again for your thoughtful and constructive review. As the discussion period is nearing its deadline, we would greatly appreciate it if you could revisit our detailed responses and let us know whether they have adequately addressed your questions and concerns.

We sincerely hope our clarifications and additional evidence have strengthened your confidence in our work, and we would be grateful if you could reconsider your overall assessment accordingly.

Thank you for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Responses and Manuscript Revisions

Dear Area Chair and Reviewers,

We sincerely thank you for your thoughtful comments and constructive feedback. We have carefully considered the concerns raised during the review process and made every effort to address them through additional experiments, theoretical clarifications, and manuscript revisions.

In particular, we have strengthened the analysis of observability and reliability gating, examined alternative training objectives and local reconstruction fidelity, and conducted further evaluations under calibration perturbations and cross-metalens settings. The additional experiments and analyses presented in our rebuttal have now been incorporated into the revised manuscript and supplementary material, with the relevant changes highlighted for ease of review. We have also clarified the theoretical assumptions, experimental protocols, and limitations of our approach.

This work represents more than six months of dedicated research and continuous refinement. We have devoted considerable effort to developing the method, investigating its limitations, and strengthening the experimental evidence. We sincerely hope that these efforts have addressed your concerns and provided a clearer understanding of the contributions and reliability of our work.

If our responses and revisions have satisfactorily resolved your concerns, we would be very grateful if you could kindly reconsider your assessments and, where appropriate, consider a more favorable rating. We fully respect your academic judgment and deeply appreciate the time and care you have devoted to evaluating our work.

Thank you again for your valuable feedback 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.