acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Multi-Exposure Regularization For Dynamic Hdr Scene Reconstruction

Abstract

Dynamic HDR reconstruction with Gaussian Splatting relies on neural tone mappers to model the nonlinear relationship between reconstructed HDR radiance and multi-exposure observations. Under varying exposure conditions, the pre-activation distribution within the tone mapper can shift substantially, altering its activation behavior and limiting the nonlinear responses effectively involved in reconstruction. To address this issue, we propose Adaptive Multi-Exposure Regularization (\method), which calibrates the tone mapper's pre-activation distribution across exposure conditions using exposure-conditioned residual offsets. Extensive experiments across multiple dynamic HDR Gaussian Splatting pipelines demonstrate consistent improvements in HDR reconstruction quality, supporting the general applicability of our method.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
HDR novel view synthesis, dynamic scene reconstruction, 3D Gaussian Splatting, regularization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RobjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the exposure-dependent shift in the operating range of tone mappers for dynamic HDR reconstruction and proposes AME-REG. The method calibrates the pre-activation distributions before ReLU using exposure-stop-conditioned biases and quantile-range constraints. Experiments across multiple dynamic HDR frameworks and both real and synthetic scenes demonstrate improved HDR reconstruction quality.

Critical strengths

The paper analyzes the internal activation behavior and shows that exposure variations systematically shift the pre-activation distributions of the tone mapper, causing some units to remain inactive or operate within a single linear regime for extended periods. This perspective provides a clear and interpretable mechanism for understanding how multi-exposure supervision affects HDR scene optimization.

The paper has a simple design with relatively low parameter overhead, while consistently improving HDR PSNR across multiple dynamic HDR reconstruction frameworks.

Critical weaknesses

I have two primary concerns.

First, the current method is largely designed around ReLU. The paper treats zero as the only nonlinear transition point and defines the feasible range using quantiles on either side of zero. While this formulation is reasonable for ReLU, it does not directly generalize to GELU, SiLU, or more sophisticated learnable activation functions. Although the paper compares against ReActNet and ACON-C, these methods are evaluated only as alternative activation functions, and the experiments do not examine whether AME-REG can be combined with them. Therefore, the paper’s claims regarding the generality of the method appear somewhat stronger than what is supported by the current evidence.

Second, the calibration bias depends only on discrete exposure levels, namely low, medium, and high EV, without explicitly accounting for scene content, spatial regions, or temporal variations. Even under the same EV setting, different scenes, or highlight and shadow regions within the same scene, may exhibit substantially different pre-activation distributions, yet they still share the same set of biases. Although the training-time quantiles indirectly capture scene-level statistics, the biases remain fixed during inference and cannot be dynamically adapted to the current frame.

Questions for authors

Can AME-REG be generalized to activation functions other than ReLU, such as GELU, SiLU, or learnable activation functions?

How robust is the fixed exposure-conditioned bias when the pre-activation distribution varies significantly across scenes, spatial regions, or consecutive frames?

4 comments

Q1: Can AME-REG be generalized to activation functions other than ReLU, such as GELU, SiLU, or learnable activation functions?

W1: First, the current method is largely designed around ReLU. The paper treats zero as the only nonlinear transition point and defines the feasible range using quantiles on either side of zero. While this formulation is reasonable for ReLU, it does not directly generalize to GELU, SiLU, or more sophisticated learnable activation functions. Although the paper compares against ReActNet and ACON-C, these methods are evaluated only as alternative activation functions, and the experiments do not examine whether AME-REG can be combined with them. Therefore, the paper’s claims regarding the generality of the method appear somewhat stronger than what is supported by the current evidence.

Re: We thank the reviewer for raising this concern. To investigate whether AME-Reg is restricted to the specific activation behavior of ReLU, we conducted additional experiments with seven activation functions: ReLU, Leaky-ReLU, PReLU, SiLU, GELU, ACON-C, and RPReLU (an activation function used in ReActNet).

These functions cover diverse activation response structures, including fixed and learnable piecewise-linear activations, smooth nonlinear activations, and learnable smooth activations. In particular, ACON-C and RPReLU allow us to examine the compatibility of AME-Reg with more sophisticated learnable activation functions.

Experimental results. We combine AME-Reg with each activation function and report the reconstruction results below.

Activation FunctionHDR PSNRHDR SSIMLDR PSNRLDR SSIM
SiLU23.740.94422.820.857
SiLU + AME-Reg31.840.94826.290.877
PReLU24.190.94622.940.853
PReLU + AME-Reg32.750.95226.800.882
GELU24.270.94623.000.854
GELU + AME-Reg31.640.94526.530.875
Leaky-ReLU24.400.94822.940.853
Leaky-ReLU + AME-Reg33.260.95227.080.885
ReLU24.230.94722.980.856
ReLU + AME-Reg33.510.95327.400.887
ACON-C24.170.94623.210.857
ACON-C + AME-Reg32.520.95026.000.875
RPReLU24.500.94823.390.861
RPReLU + AME-Reg32.590.95226.310.877

AME-Reg consistently improves HDR PSNR by 7.37–9.28 dB and LDR PSNR by 2.79–4.42 dB across all seven activation functions. SSIM is also largely maintained or improved, with only a marginal HDR SSIM decrease of 0.001 for GELU.

The role of zero. In AME-Reg, zero serves as a reference point for pre-activation distribution calibration, rather than being assumed to be a universal nonlinear transition point. For ReLU, Leaky-ReLU, and PReLU, zero is the boundary where the response slope changes. For SiLU, GELU, and ACON-C, the response passes smoothly through zero. RPReLU has a learnable transition location, which is close to zero in our experiments. These differences highlight that zero can serve as a common calibration reference without requiring different activation functions to share the same nonlinear transition behavior.

Joint bias–quantile calibration. AME-Reg jointly uses exposure-conditioned biases and quantile-based regularization to calibrate pre-activation distributions. The biases introduce exposure-specific shifts, while the quantile constraint guides their calibration by constraining the bias-adjusted pre-activations according to their empirical distribution. This joint process defines a data-dependent operating range, rather than imposing arbitrary fixed thresholds or a prescribed activation ratio. It compensates for exposure-induced distribution shifts and prevents calibrated features from remaining concentrated in locally near-linear response regions. Gradient isolation further preserves this calibration objective by preventing reconstruction gradients from overriding it.

Overall, these results demonstrate that AME-Reg is not restricted to ReLU or a specific activation response structure, but remains effective across the seven evaluated activation functions, including learnable variants.

Q2: How robust is the fixed exposure-conditioned bias when the pre-activation distribution varies significantly across scenes, spatial regions, or consecutive frames?

W2: Second, the calibration bias depends only on discrete exposure levels, namely low, medium, and high EV, without explicitly accounting for scene content, spatial regions, or temporal variations. Even under the same EV setting, different scenes, or highlight and shadow regions within the same scene, may exhibit substantially different pre-activation distributions, yet they still share the same set of biases. Although the training-time quantiles indirectly capture scene-level statistics, the biases remain fixed during inference and cannot be dynamically adapted to the current frame.

Re: We appreciate the reviewer’s concern regarding the robustness of fixed exposure-conditioned biases. We evaluate their effectiveness on HDR-SplineGS using 123 paired test frames from four real scenes (Bed, Excavator, Tank, and Toys), as detailed in Appendix C.3.

Spatial robustness. We evaluate Dark, Mid-tone, and Bright regions using scene-wide one-third and two-thirds quantiles of GT log-luminance. Both methods use identical GT-derived masks without region-specific normalization. As reported in Table 4 and Fig. 5 in Appendix C.3, AME-Reg achieves scene-macro mean PSNR gains of +9.24, +10.07, and +9.90 dB, respectively. All 369 frame-region cases show positive gains, with the weakest improvement being +0.54 dB.

Temporal robustness. We separately evaluate Low-, Mid-, and High-EV subsequences. As shown in Fig. 6 and Table 4 in Appendix C.3, the corresponding scene-macro mean gains are +9.86, +9.38, and +9.18 dB. All 123 matched frames show positive gains, with the weakest improvement being +2.31 dB.

Together, these results show that fixed offsets shared across pixels and frames within each exposure bracket remain beneficial across the evaluated scenes, luminance regions, and temporal variations. Although they do not dynamically adapt to individual regions or frames, their benefits persist under these variations.

I thank the authors for the detailed response and additional experiments. The new results across seven activation functions substantially alleviate my concern that AME-Reg is intrinsically restricted to ReLU. In particular, the consistent improvements obtained with SiLU, GELU, ACON-C, and RPReLU provide useful evidence that zero can serve as a practical calibration reference beyond piecewise-linear activations. The additional spatial- and temporal-robustness evaluations also show that the fixed exposure-conditioned biases remain beneficial across the evaluated luminance regions, scenes, exposure subsequences, and frames.

Nevertheless, I do not believe these additions fully resolve my original concerns. For non-ReLU activations, the justification for using zero remains primarily empirical. The current formulation still lacks a principled characterization of when zero is an appropriate calibration reference, particularly for activations with learnable transition locations or substantially different response geometries. Moreover, while the new robustness results demonstrate that fixed biases work well on the evaluated four scenes and 123 frames, they do not directly address the method’s inability to adapt its calibration to scene content, spatially varying statistics, or individual frames. The evidence supports the robustness of the current design within the tested setting, but not the stronger conclusion that explicit content- or frame-dependent adaptation is unnecessary more generally.

Overall, the rebuttal has meaningfully strengthened the empirical support for the method and reduced the severity of my concerns. However, I still view the limitations in the formulation and the scope of the robustness evidence as significant enough to prevent me from recommending acceptance. I therefore maintain my original rating of 2: Weak Rejection.

Re: We sincerely thank the reviewer for recognizing that our additional experiments substantially alleviate the concern regarding the generality of AME-Reg. We further clarify the two remaining points concerning the calibration reference and the adaptation of the exposure-conditioned biases.

1. On the choice of zero as the calibration reference.

We clarify that zero is not intended as a universal nonlinear transition point for all activation functions. Our initial analysis focuses on ReLU because ReLU is the standard Stage-I activation shared by the tone-mapper designs commonly used in existing HDR Gaussian Splatting methods. In AME-Reg, zero instead serves as a reference coordinate for pre-activation calibration.

For ReLU, Leaky-ReLU, and PReLU, zero is the point where the response slope changes. For smooth activations such as SiLU and GELU, zero is not a transition point, but remains a natural reference coordinate of the activation input. For RPReLU, whose transition location is learnable, AME-Reg does not require this location to be fixed at zero; in our experiment, the learned transition remains close to zero while AME-Reg still substantially improves reconstruction.

More generally, the absolute location of an activation transition is not separately identifiable from the bias of the preceding layer. A horizontal shift of the activation can be absorbed into the bias, e.g., ϕ(z−c+b)=ϕ(z+(b−c))\phi(z-c+b)=\phi(z+(b-c)). Thus, choosing zero as the reference coordinate does not restrict the expressive power of a biased layer; another fixed reference would simply correspond to a different parameterization of the bias. The key quantity for AME-Reg is therefore the relative alignment between the pre-activation distribution and the activation response. Exposure-conditioned biases and quantile regularization jointly calibrate this alignment according to the empirical input distribution, rather than requiring all activation functions to share the same transition behavior.

2. On the adaptation of exposure-conditioned biases.

We would also like to clarify that our reconstruction follows a per-scene optimization setting: each scene is optimized independently and learns its own exposure-conditioned biases. Therefore, no single set of biases is shared across different scenes. Within each scene, the biases are conditioned on exposure rather than scene content, spatial location, or individual frames. Their purpose is to compensate for the systematic pre-activation shifts introduced by known exposure variations, and the learned Low/Mid/High-EV biases are shared across pixels and frames within their respective exposure brackets.

Importantly, different spatial regions may have different feature statistics, but this does not necessarily imply that they require different exposure calibrations. What AME-Reg addresses is the systematic distribution shift associated with exposure. The spatial and temporal experiments in Appendix C.3 directly examine whether this exposure-conditioned calibration remains effective across different regions and frames within each scene. Across the four independently optimized scenes, we observe positive gains in all 369/369 spatial frame-region cases and all 123/123 temporal frames.

Thus, content- or frame-dependent calibration is not a requirement of the current formulation, nor is its absence part of our claim. Such finer-grained adaptation could be explored as a possible extension to provide additional flexibility beyond the exposure-induced distribution shift targeted by AME-Reg. Our current results instead show that, within the per-scene reconstruction setting, exposure-conditioned calibration provides consistent improvements without requiring explicit pixel- or frame-dependent biases.

Overall, AME-Reg uses zero as a reference coordinate rather than imposing a universal nonlinear transition at zero, and employs per-scene exposure-conditioned calibration to address systematic shifts caused by different exposures. The additional experiments demonstrate that this calibration remains effective across diverse activation functions, luminance regions, and temporal variations.

Reviewer_n6vYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Adaptive Multi-Exposure Regularization (AME-REG) to improve dynamic HDR reconstruction with Gaussian Splatting. The method introduces exposure-conditioned residual offsets to calibrate tone-mapper pre-activation distributions across different exposures. Experiments on multiple HDR Gaussian Splatting frameworks demonstrate improvements in reconstruction quality.

Critical strengths
  1. The paper is well structured, with a clear presentation of its core concepts and methodology.
  2. The authors conduct extensive experiments across multiple reconstruction frameworks, covering both real-world and synthetic datasets.
Critical weaknesses
  1. As discussed in the section “Tone-mapping and Radiometric Regularization,” existing HDR Gaussian Splatting (HDR-GS) methods employ different tone-mapping strategies. However, the authors restrict their analysis to the tone mapper used in HDR-GS in Figure 1. To better support the generality of the proposed method, the analysis should be extended to other tone mappers, accompanied by corresponding ablation studies.

  2. The authors analyze the activation behavior of the HDR-GS tone mapper in Figure 1, but conduct the ablation studies on HDR-SplineGS. This makes the connection between the activation analysis and the ablation findings less direct: it remains unclear whether the mechanisms identified through the HDR-SplineGS ablations also explain the improvements observed in the analyzed HDR-GS setting. The authors should clarify the rationale for this experimental design and provide key ablations in the HDR-GS setting to strengthen this connection.

  3. The current cross-framework comparisons do not convincingly establish that the reported improvements can be attributed specifically to AME-REG or that the proposed method generalizes consistently across different frameworks. For example, in Table 1, the HDR-4DGS baseline results on real-world scenes are taken from the original paper, whereas the integration of AME-REG into GaussHDR involves replacing the original tone-mapping-specific supervision. These differences raise concerns about the fairness of the comparisons and make it difficult to isolate the contribution of AME-REG. Moreover, the paper does not clearly explain how AME-REG is integrated into each backbone, particularly whether the original tone-mapping architectures and training objectives are preserved or modified. Since cross-framework applicability is a central claim of the paper, these ambiguities substantially weaken the experimental evidence supporting this contribution.

  4. In Table 3(a), applying L2L_2 regularization with a fixed 50% activation target already improves HDR PSNR from 24.23 dB to 33.32 dB, whereas AME-REG achieves 33.51 dB, with a slightly lower SSIM. The relatively small improvement of 0.19 dB raises questions about whether the proposed adaptive mechanism provides a meaningful advantage over a simpler fixed-target regularization strategy. Additional experiments are needed to demonstrate the necessity and practical benefits of the adaptive design.

  5. The individual contributions of the proposed components are not adequately isolated. For instance, the authors could evaluate a variant in which the exposure-conditioned bias is optimized solely through the reconstruction loss, as well as a variant that removes the gradient isolation mechanism. Furthermore, given the marginal improvement over L2L_2 regularization with a fixed 50% activation target, experiments across multiple random seeds would be valuable to determine whether the reported gains are consistent and statistically meaningful.

Questions for authors
  1. For each evaluated backbone framework, which original tone-mapping components and loss functions are retained, modified, or removed when integrating AME-REG? In particular, please specify which supervision terms in GaussHDR are replaced. Additionally, are the HDR-4DGS baseline and its AME-REG-enhanced counterpart evaluated under identical preprocessing pipelines, training configurations, and experimental settings?
6 comments

Q1: For each evaluated backbone framework, which original tone-mapping components and loss functions are retained, modified, or removed when integrating AME-REG? In particular, please specify which supervision terms in GaussHDR are replaced. Additionally, are the HDR-4DGS baseline and its AME-REG-enhanced counterpart evaluated under identical preprocessing pipelines, training configurations, and experimental settings?

W3: The current cross-framework comparisons do not convincingly establish that the reported improvements can be attributed specifically to AME-REG or that the proposed method generalizes consistently across different frameworks. For example, in Table 1, the HDR-4DGS baseline results on real-world scenes are taken from the original paper, whereas the integration of AME-REG into GaussHDR involves replacing the original tone-mapping-specific supervision. These differences raise concerns about the fairness of the comparisons and make it difficult to isolate the contribution of AME-REG. Moreover, the paper does not clearly explain how AME-REG is integrated into each backbone, particularly whether the original tone-mapping architectures and training objectives are preserved or modified. Since cross-framework applicability is a central claim of the paper, these ambiguities substantially weaken the experimental evidence supporting this contribution.

Re: Thank you for raising this important question. We clarify the backbone-specific modifications to the tone-mapping components and supervision terms below.

GaussHDR: We retain the global and local tone-mapping branches and the uncertainty branch, together with the original 2D/3D local reconstruction, uncertainty-fusion, Gaussian regularization, and HDR supervision terms. We remove the auxiliary global-CRF supervision and the global-to-local residual composition, while applying AME-Reg to the local tone-mapping branch.

HDR-4DGS: Its original tone-mapping network and all original reconstruction and regularization terms are retained, with AME-Reg added at the first stage as the only methodological modification. We integrate AME-Reg into the authors' official implementation, following the released experimental configurations and evaluation protocol.

HDR-GS: Its original tone-mapping network and all original reconstruction and regularization terms are retained. AME-Reg is added at the first stage without removing any original loss.

HDR-HexPlane: We regard its exposure-conditioning branch as part of the tone-mapping module. We replace the unconstrained per-image exposure embedding with known physical exposure values, retain the subsequent exposure MLP and ToneMapper, and insert AME-Reg at the first stage of the exposure-conditioning MLP. All original reconstruction and regularization losses are retained.

W1: As discussed in the section “Tone-mapping and Radiometric Regularization,” existing HDR Gaussian Splatting (HDR-GS) methods employ different tone-mapping strategies. However, the authors restrict their analysis to the tone mapper used in HDR-GS in Figure 1. To better support the generality of the proposed method, the analysis should be extended to other tone mappers, accompanied by corresponding ablation studies.

Re: Thank you for this helpful comment. Although existing HDR reconstruction methods employ different overall tone-mapping strategies, the tone-mapping branches considered in our work share the same core architecture. As clarified in Sec. 3.1, our functional decomposition provides a common perspective for analyzing this component, including in more elaborate tone mappers with additional conditioning or local mapping branches. Fig. 1 uses the tone mapper adopted from HDR-GS as a representative example (rather than the entire HDR-GS framework) to investigate its exposure-dependent behavior.

W2: The authors analyze the activation behavior of the HDR-GS tone mapper in Figure 1, but conduct the ablation studies on HDR-SplineGS. This makes the connection between the activation analysis and the ablation findings less direct: it remains unclear whether the mechanisms identified through the HDR-SplineGS ablations also explain the improvements observed in the analyzed HDR-GS setting. The authors should clarify the rationale for this experimental design and provide key ablations in the HDR-GS setting to strengthen this connection.

Re: Thank you for raising this concern. HDR-SplineGS is constructed by integrating the HDR-GS tone mapper into the SplineGS framework. Therefore, both methods share the same tone-mapping architecture and exposure-conditioned input formulation, while differing in their underlying scene representations. Since AME-Reg operates directly on the tone mapper rather than the scene representation, the ablations conducted on HDR-SplineGS are intended to investigate the contributions of individual components within the same HDR-GS tone-mapping design.

W4. In Table 3(a), applying regularization with a fixed 50% activation target already improves HDR PSNR from 24.23 dB to 33.32 dB, whereas AME-REG achieves 33.51 dB, with a slightly lower SSIM. The relatively small improvement of 0.19 dB raises questions about whether the proposed adaptive mechanism provides a meaningful advantage over a simpler fixed-target regularization strategy. Additional experiments are needed to demonstrate the necessity and practical benefits of the adaptive design.

W5. The individual contributions of the proposed components are not adequately isolated. For instance, the authors could evaluate a variant in which the exposure-conditioned bias is optimized solely through the reconstruction loss, as well as a variant that removes the gradient isolation mechanism. Furthermore, given the marginal improvement over regularization with a fixed 50% activation target, experiments across multiple random seeds would be valuable to determine whether the reported gains are consistent and statistically meaningful.

Re: Thank you for these helpful comments.AME-Reg addresses exposure-induced feature distribution shifts through a coupled calibration process. Specifically, exposure-conditioned biases first compensate for exposure-dependent shifts in Stage-I pre-activation distributions, bringing features from different exposure brackets into a more aligned operating range. However, distribution alignment alone does not guarantee effective nonlinear feature utilization, as the calibrated features may still concentrate in locally near-linear response regions. Quantile-based regularization therefore further constrains the bias-adjusted pre-activations within an effective nonlinear operating range, adapting the calibration to the empirical input distribution. Gradient isolation prevents reconstruction gradients from overriding this distributional calibration.

Component contributions (W5). We isolate the contributions of exposure-conditioned bias, quantile regularization, and gradient isolation. The baseline corresponds to HDR-SplineGS without AME-Reg.

MethodHDR PSNRHDR SSIM
Baseline24.230.947
Exposure-conditioned Bias (reconstruction loss only)31.910.954
Exposure-conditioned Bias + Quantile Regularization (without gradient isolation)32.980.955
AME-Reg (with gradient isolation)33.510.953

The results confirm that exposure-conditioned bias calibration provides the primary performance gain, while quantile regularization and gradient isolation bring further improvements. This supports their complementary roles: bias calibration aligns exposure-conditioned feature distributions, while quantile regularization guides them toward an effective nonlinear operating range, with gradient isolation preserving the distributional calibration objective.

Fixed-target L2 (W4). On HDR-SplineGS, fixed 50% L2 regularization achieves 33.32 dB (0.954 SSIM), compared with 33.51 dB (0.953 SSIM) for AME-Reg. To clarify its contribution, we conducted the following ablation:

MethodHDR PSNRHDR SSIM
Fixed-50% L2 without Exposure-conditioned Bias24.470.949
Fixed-50% L2 with Exposure-conditioned Bias33.320.954

Without exposure-conditioned bias calibration, Fixed-50% L2 achieves only 24.47 dB, a marginal improvement of 0.24 dB over the baseline without AME-Reg (24.23 dB). This demonstrates that the strong performance of fixed-target regularization largely relies on bias calibration, which first compensates for exposure-induced distribution shifts and establishes a more consistent operating range for subsequent regularization.

To further investigate the advantage of adaptive regularization, we rerun and compared Fixed-50% L2 and quantile-based regularization on four real-world scenes using HDR-4DGS. Exposure-conditioned bias calibration was retained in both variants to isolate the effect of the regularization design.

SceneFixed 50% L2 (HDR PSNR )Quantile-based Regularization (HDR PSNR )
Tank35.9331.40
Toys30.7535.36
Bed31.9630.82
Excavator28.3332.05
Average31.7432.41

Although quantile-based regularization does not outperform the fixed target on every individual scene, it achieves a higher average HDR PSNR (32.41 vs. 31.74 dB). This suggests that a fixed activation ratio may not be equally suitable across different scenes and exposure configurations, which exhibit different pre-activation distributions. In contrast, our quantile-based design determines the feasible operating range according to the empirical input distribution, allowing the calibration to adapt to scene-dependent variations.

Overall, these results demonstrate that exposure-conditioned bias calibration and quantile regularization play complementary roles. Together with gradient isolation, these components constitute our proposed AME-Reg.

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

This paper explores the exposure-dependent operating range of the ReLU stage in an existing HDR tone mapper, and proposes AME-REG, which adds bracket-conditioned residual offsets to Stage-I pre-activations and constrains these offsets using a quantile-derived feasible interval. The experiments on real and synthetic dynamic HDR benchmarks provide initial evidence for the effectiveness of the proposed method.

Critical strengths
  • The paper addresses a meaningful and challenging problem in dynamic HDR scene reconstruction under alternating exposure.
  • The effectiveness of the proposed AME-REG is initially validated on both real and synthetic dynamic HDR datasets.
  • The paper provides a clear analysis of exposure-dependent pre-activation shifts and detailed implementation of the proposed regularization.
Critical weaknesses
  • The novelty of the proposed method appears reasonable but relatively incremental. The method primarily augments an existing tone-mapping pipeline with exposure-conditioned bias calibration and a quantile-based regularization objective. As stated in Sec. 3.1, the Stage-I/Stage-II formulation is introduced as a functional decomposition of the adopted tone mapper and does not alter its underlying architecture. Moreover, as shown in Eqs. (4)-(5), the exposure-dependent shift in pre-activations follows directly from the log-domain exposure formulation and the affine transformation in the first layer. While the proposed formulation is technically well motivated, Table 3(a) shows that AME-REG improves HDR PSNR by 0.19 dB over the Fixed-50% L2 baseline (33.51 versus 33.32 dB). This result suggests that the empirical advantage of the adaptive quantile formulation over a strong fixed-target alternative is modest in the reported setting.

  • The experimental evaluation remains incomplete in several respects. 1) Although Sec. 4 reports results on four real and eight synthetic scenes using multiple reconstruction backbones and HDR/LDR metrics, the current ablations do not sufficiently isolate the contribution of each component. As shown in Table 3, it remains unclear how much of the observed gain is attributable to the bracket-conditioned offsets, the quantile-derived feasible range, and the gradient-isolation design, respectively. In particular, Appendix A.2 states that the calibration offsets are detached from the reconstruction objective and optimized only by the range regularization; however, the paper does not provide a direct comparison with a jointly optimized alternative. 2) Moreover, for GaussHDR, AME-REG replaces the original tone-mapping-specific supervision rather than being evaluated as an additive component under the same objective. 3) The qualitative evaluation is similarly limited to module-level comparisons. Figs. 3 and 4 compare each backbone with and without AME-REG, but do not provide direct qualitative comparisons with competing complete HDR reconstruction methods under the same settings. A more controlled component-wise analysis and comprehensive qualitative comparison would better support the claimed effectiveness and general applicability of the proposed method.

  • Several mechanism-related claims of the proposed method appear stronger than the current evidence can fully support. As explained in the introduction and Sec. 3, the paper establishes a clear connection between exposure variation and shifts in the pre-activation distribution, which motivates the proposed operating-range constraint. Fig. 1 further shows that AME-REG is associated with both a broader activation-frequency distribution and improved reconstruction quality. However, these observations demonstrate correlation rather than directly establishing a causal relationship between the changed activation behavior and the performance gains. Moreover, as described in Appendix A.2, the calibration offsets are not directly optimized by the reconstruction objective. Therefore, the term “adaptive calibration” may somewhat overstate the extent to which these offsets are learned to optimize reconstruction quality.

Questions for authors
  • Table 3(d) reports the best PSNR at q=0.15, while q=0.05 is used as the default setting. The rationale for this choice should be clarified.
  • The related-work discussion could be updated with recent HDR Gaussian Splatting literature, such as PhysHDR-GS, which also studies exposure-dependent optimization challenges in HDR reconstruction.
  • The authors should update the reference information where appropriate. For example, HDR-4DGS, one of the main comparison methods, has been published at ICLR 2026 rather than remaining only an arXiv preprint.
  • Lack of dynamic visual results. Video-based evaluation is particularly important for dynamic 3D/4D scene reconstruction, as static-frame results cannot adequately reveal temporal consistency, motion quality, or the stability of HDR reconstruction across a sequence. However, the paper does not provide video results or temporal visualizations for the reconstructed dynamic HDR scenes. As a result, it is difficult to assess whether the proposed method improves the perceptual quality and temporal coherence of the full 4D representation, rather than only improving frame-wise quantitative metrics or selected static examples.
8 comments

Q1: Table 3(d) reports the best PSNR at q=0.15, while q=0.05 is used as the default setting. The rationale for this choice should be clarified.

Re: Thank you for this helpful question. We initially set (q=0.05) empirically during the early development of AME-Reg and kept this setting throughout the subsequent experiments.

In our ablation study on HDR-SplineGS, increasing qq consistently improves HDR PSNR, with q=0.15q=0.15 achieving the best result. To investigate whether this trend generalizes across reconstruction frameworks, we further evaluated different quantile settings on four real scenes using HDR-HexPlane, HDR-4DGS, HDRGS, and GaussHDR. The results are summarized below (all values are mean HDR PSNR in dB).

Methodq=0.05q=0.05q=0.10q=0.10q=0.15q=0.15
HDR-SplineGS33.5133.6933.76
HDR-HexPlane25.8424.8123.90
HDR-4DGS32.2832.4032.25
HDRGS30.3730.3729.40
GaussHDR31.0931.0931.09

The results reveal different sensitivity patterns across reconstruction frameworks. While increasing qq consistently benefits HDR-SplineGS, it leads to lower PSNR on HDR-HexPlane and HDRGS. HDR-4DGS achieves its best result at q=0.10q=0.10, while GaussHDR remains nearly unchanged across the three settings. These observations indicate that the effect of quantile selection is method-dependent, and a larger quantile does not necessarily yield better reconstruction.

We therefore retain q=0.05q=0.05 as a common empirical default across different baselines.

Q2: The related-work discussion could be updated with recent HDR Gaussian Splatting literature, such as PhysHDR-GS, which also studies exposure-dependent optimization challenges in HDR reconstruction.

Q3: The authors should update the reference information where appropriate. For example, HDR-4DGS, one of the main comparison methods, has been published at ICLR 2026 rather than remaining only an arXiv preprint.

Re: Thank you for these helpful suggestions. We have added PhysHDR-GS (Zeng et al., 2026) to the Related Work section to include this recent development in HDR Gaussian Splatting. We have also reviewed and updated the bibliographic information throughout the manuscript, including updating HDR-4DGS (Zhang et al., 2026) from its arXiv preprint to the published ICLR 2026 version, as well as correcting the publication information of other relevant references where applicable.

These revisions improve the completeness of our literature review and the overall manuscript. We sincerely appreciate the reviewer’s suggestions.

Q4: Lack of dynamic visual results. Video-based evaluation is particularly important for dynamic 3D/4D scene reconstruction, as static-frame results cannot adequately reveal temporal consistency, motion quality, or the stability of HDR reconstruction across a sequence. However, the paper does not provide video results or temporal visualizations for the reconstructed dynamic HDR scenes. As a result, it is difficult to assess whether the proposed method improves the perceptual quality and temporal coherence of the full 4D representation, rather than only improving frame-wise quantitative metrics or selected static examples.

Re: Thank you for this helpful suggestion. We have addressed this concern in two ways. First, we have uploaded supplementary videos based on HDR-4DGS, showing dynamic HDR reconstruction results before and after applying AME-Reg, allowing direct visual comparison of reconstruction quality and temporal behavior over complete sequences. Second, we have added consecutive-frame visual comparisons of four real scenes in Appendix D. Specifically, Fig. 7 presents the Bed and Excavator scenes using HDR-HexPlane, while Fig. 8 presents the Tank and Toys scenes using HDR-SplineGS. Each figure compares the ground truth, the baseline reconstruction, and the corresponding reconstruction with AME-Reg across six consecutive frames.

These additions complement the frame-wise quantitative evaluations and static visual comparisons, providing a more comprehensive assessment of dynamic HDR reconstruction. We hope these additional visual results make the temporal behavior of the reconstructed HDR scenes easier to assess.

W1: The novelty of the proposed method appears reasonable but relatively incremental. The method primarily augments an existing tone-mapping pipeline with exposure-conditioned bias calibration and a quantile-based regularization objective. As stated in Sec. 3.1, the Stage-I/Stage-II formulation is introduced as a functional decomposition of the adopted tone mapper and does not alter its underlying architecture. Moreover, as shown in Eqs. (4)-(5), the exposure-dependent shift in pre-activations follows directly from the log-domain exposure formulation and the affine transformation in the first layer. While the proposed formulation is technically well motivated, Table 3(a) shows that AME-REG improves HDR PSNR by 0.19 dB over the Fixed-50% L2 baseline (33.51 versus 33.32 dB). This result suggests that the empirical advantage of the adaptive quantile formulation over a strong fixed-target alternative is modest in the reported setting.

Re: We appreciate the reviewer's concern regarding the novelty of our method. The Stage-I/Stage-II decomposition is used simply as a functional organization to describe and analyze the adopted tone mapper, while exposure-induced pre-activation shifts follow from the log-domain formulation. Our contributions go beyond these observations, focusing on the implications of distribution shifts for nonlinear representation and the resulting calibration strategy.

From observation to calibration. Our analysis reveals that exposure-induced distribution shifts affect how inputs utilize the nonlinear responses of the tone mapper. Replacing ReLU with alternative activations, such as ACON-C, can improve neuron activation rates, but does not fully resolve the reconstruction issue. Further analysis reveals that inputs can still concentrate in locally near-linear response regions, limiting effective nonlinear representation. This led us to shift the focus from activation-function selection to the relationship between input distributions and nonlinear response regions. We first introduce exposure-conditioned biases to compensate for distribution shifts. Recognizing that bias adjustment alone lacks an explicit objective for controlling the calibrated operating range, we further introduce quantile-based regularization, with gradient isolation to separate calibration and reconstruction objectives.

Component contributions. Stepwise ablations on HDR-SplineGS demonstrate the complementary roles of these components.

MethodHDR PSNRHDR SSIM
Baseline24.230.947
Exposure-conditioned Bias (reconstruction loss only)31.910.954
Bias + Quantile Regularization (without gradient isolation)32.980.955
AME-Reg (with gradient isolation)33.510.953

Exposure-conditioned bias calibration provides the primary performance gain, while quantile regularization and gradient isolation bring further improvements.

**Why quantile-based calibration?**The Fixed-50% L2 achieves 33.32 dB, close to AME-Reg (33.51 dB). To clarify the contribution of bias calibration to this comparison, we conducted an additional ablation:

MethodHDR PSNRHDR SSIM
Fixed-50% L2 without Exposure-conditioned Bias24.470.949
Fixed-50% L2 with Exposure-conditioned Bias33.320.954

Without bias calibration, Fixed-50% L2 achieves only 24.47 dB, a marginal improvement of 0.24 dB over the baseline (24.23 dB). Introducing bias calibration increases PSNR to 33.32 dB. Thus, the 0.19 dB difference between Fixed-50% L2 and AME-Reg reflects the additional benefit of adaptive quantile-based regularization under the same bias calibration, rather than the overall contribution of our method.

Furthermore, across four real HDR-4DGS scenes, quantile-based regularization achieves a higher average HDR PSNR than Fixed-50% L2 (32.41 vs. 31.74 dB), although the better-performing variant varies across individual scenes. This supports deriving the feasible interval from empirical input distributions rather than prescribing a fixed activation ratio.

Overall, our contributions lie in two aspects: (1) identifying the underutilization of nonlinear responses caused by exposure-induced distribution shifts, particularly the concentration of inputs in locally near-linear response regions; and (2) developing a calibration strategy from this finding. The latter involves a non-trivial progression from identifying the distribution-response mismatch to recognizing the need for explicit range constraints, leading to the combination of exposure-conditioned bias calibration and quantile-based regularization.

W2: The experimental evaluation remains incomplete in several respects. 1) Although Sec. 4 reports results on four real and eight synthetic scenes using multiple reconstruction backbones and HDR/LDR metrics, the current ablations do not sufficiently isolate the contribution of each component. As shown in Table 3, it remains unclear how much of the observed gain is attributable to the bracket-conditioned offsets, the quantile-derived feasible range, and the gradient-isolation design, respectively. In particular, Appendix A.2 states that the calibration offsets are detached from the reconstruction objective and optimized only by the range regularization; however, the paper does not provide a direct comparison with a jointly optimized alternative. 2) Moreover, for GaussHDR, AME-REG replaces the original tone-mapping-specific supervision rather than being evaluated as an additive component under the same objective. 3) The qualitative evaluation is similarly limited to module-level comparisons. Figs. 3 and 4 compare each backbone with and without AME-REG, but do not provide direct qualitative comparisons with competing complete HDR reconstruction methods under the same settings. A more controlled component-wise analysis and comprehensive qualitative comparison would better support the claimed effectiveness and general applicability of the proposed method.

Re: Thank you for pointing out these aspects of our experimental evaluation. We have made additional revisions to clarify the component contributions, framework-specific adaptations, and qualitative results.

Component-wise analysis. We added a stepwise ablation in Table 4 to isolate the contributions of exposure-conditioned bias calibration, quantile regularization, and gradient isolation. In particular, the intermediate variant without gradient isolation retains reconstruction gradients on the calibration offsets, providing a direct comparison with the full AME-Reg, where the two objectives are separated. The results demonstrate the complementary contributions of these components.

GaussHDR integration. We have clarified the GaussHDR adaptation in Appendix A.4 and marked it in the main comparison table. AME-Reg is applied to its local tone-mapping branch, while the global and local branches, uncertainty estimation and fusion, and Gaussian reconstruction components are retained. The auxiliary global-CRF supervision and global-to-local residual composition are removed, with AME-Reg replacing the original tone-mapping-specific supervision. We explicitly describe this as a framework-specific integration.

Qualitative evaluation. We have added consecutive-frame visualizations of four real scenes in Appendix D. Fig. 7 shows Bed and Excavator using HDR-HexPlane, while Fig. 8 shows Tank and Toys using HDR-SplineGS. Each sequence contains six consecutive frames comparing the HDR ground truth, baseline reconstruction, and reconstruction with AME-Reg. We also provide supplementary videos based on HDR-4DGS, comparing reconstruction results before and after applying AME-Reg.

Comments on the paper

1 comment

We sincerely thank all reviewers for their thoughtful and constructive feedback, which has helped us improve the clarity and presentation of our work.

For ease of reference, we have color-coded the revisions in the manuscript according to the reviewers' comments: green for Reviewer RXeX, red for Reviewer n6vY, and blue for Reviewer Robj. We have also uploaded a supplementary video to provide additional visual results of our dynamic HDR reconstruction.

We hope that our revisions and responses help clarify the motivation, methodology, and contributions of our work, and make it easier for the reviewers to better understand our approach. We would be happy to further clarify any remaining questions or concerns and welcome continued discussion.

Thank you again for your time and valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.