- 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
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., . 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.
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.
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.