Comment 1.1:

Relevant baselines are missing. At line 117, paper refers to 3 works [25,30,9] which work in this field. However, none of them are present as a baseline in Table 1. The work does mention that those methods reconstruct only neural activity, but this work also reconstructs entire volumetric structures. However, the baselines mentioned in Table 1 were also not developed for XLFM 3D reconstruction task, but the authors did the work to adapt them to this task. So, I donot see why such an effort was not made for any of these three works, given the fact that they are much closer to the domain than the used baselines.

Summary. While Refs. [9, 25, 30] are closer to our imaging domain, they fundamentally assume sparse, static settings or require impractical memory, making them inapplicable to our dense, dynamic full-volume reconstruction task. In contrast, our selected model-agnostic baselines enable fair, scalable, and reproducible comparisons.

We appreciate the reviewer’s thoughtful suggestion to include prior domain-specific methods as baselines. After thorough analysis, we found that these methods—despite their relevance—cannot be meaningfully adapted to our setting due to the following limitations:

• Refs. 25 & 30 focus exclusively on sparse neural signal extraction under the assumption of fully immobilized specimens. Both adopt the SLNet architecture to extract sparse signals prior to reconstruction and are fundamentally incompatible with our goal of reconstructing dense volumetric tissue. Moreover, they process temporally fused multi-frame inputs, which cannot be applied to our single-frame, dynamic zebrafish recordings.
• Ref. 9 (FourierNet) performs all computation in the Fourier domain and demands excessive GPU memory (>80GB) even for modest output sizes. In our case (300×600×600 volumes), FourierNet is not practically reproducible or scalable, even with aggressive approximations.

Comment 1.2:

Also, since richardson-lucy is used for 3D reconstruction in light field microscopy, it is expected that that should be present as a baseline, arguably along with a quantitative estimate of its compute requirements (the relevant ones from time, gpu, cpu or RAM).

Authors’ response:

Summary. RLD achieves higher PSNR but is prohibitively slow and memory-intensive; our method offers a practical tradeoff, enabling real-time, scalable 3D reconstruction essential for biological applications.

We appreciate the reviewer’s suggestion to compare against classical iterative deconvolution algorithms. As shown below, we benchmarked RLD at different iteration counts (20/30/40) on the challenging H2B-Nemos dataset. While (Richardson-Lucy Deconvolution)RLD can achieve high PSNR/SSIM given sufficient iterations, it incurs orders of magnitude higher memory usage and lower speed, making it impractical for real-time or high-throughput applications.

Comment 2:

The authors themselves mention that only one dataset, that has been generated by them, was used for the experiments. The authors should explain the reason: Are there no other datasets which could be adapted to this task? In such a case, even simulated data will hold value, in my opinion.

Authors’ response:

We appreciate the reviewer’s suggestion about including additional datasets, including simulated ones, to strengthen the evaluation. Since the submission, we have extended our experiments to a newly collected dataset—H2B-Nemos, which involves a different zebrafish line under distinct experimental conditions. This allows us to directly test the generalizability and robustness of our method.

As shown below, we re-run all methods under identical settings using ResNet-101 as a common baseline, and observe consistent performance advantages of our XLFM-Former. Notably, our method achieves +0.92 dB PSNR gain over ResNet-101 in the supervised setting, and a +2.29 dB gain in the zero-shot setting, further demonstrating its cross-domain generalization.

**Comparison on the H2B-Nemos Dataset (baseline: ResNet-101) **

We will include this dataset and evaluation in the final version of the paper and commit to releasing both datasets (G8S and Nemos) publicly under the supervision of the AC.

Comment 3.1:

The loss function for XLFM reconstruction tasks (equation 7) arguably has redundancies. It has a loss component using PSNR as well as using MSE. Since PSNR internally uses MSE, it is not clear why both are needed. Indeed the analysis done in the supplement shows that the PSNR loss component provides limited benefit.

Authors’ response:

Summary. Despite being related, PSNR and MSE offer complementary benefits—our ablation shows that including both improves fidelity and contrast, supporting their joint use for high-quality reconstruction.

We thank the reviewer for the thoughtful comment. While it is true that PSNR is derived from MSE, we argue that including both explicit PSNR and MSE losses encourages the model to balance overall numerical accuracy (via MSE) with perceptual contrast and dynamic range (via PSNR).

Importantly, our ablation in Table 4 shows that removing the PSNR loss results in a noticeable drop in PSNR (from 54.04 to 53.07 dB) and SSIM (from 0.9944 to 0.9935), which demonstrates non-negligible improvement.

Comment 3.2:

The other issue with having five loss components is that one has 5 hyper-parameters, and then it is not clear whether the same hyperparameter set can be used for other datasets. Since there are no other datasets on which the same hyperparameter set has been tested, this becomes a bit more problematic.

Authors’ response:

Summary. Our loss configuration generalizes well across datasets with only minimal tuning, as confirmed by strong performance on H2B-Nemos in new ablation results.

We thank the reviewer for raising this important concern regarding the generalizability of our loss configuration across datasets.

To verify this, we conducted additional experiments on the H2B-Nemos dataset, using the same loss configuration as in our main experiments on the original dataset. We only slightly adjusted the ORC loss weight (from 1e-4 to 3e-5) to account for minor noise differences; all other hyperparameters were kept unchanged.

As shown in the table, the results remain consistently strong:

**Ablation Study on the H2B-Nemos Dataset. **

Comment 4:

At Line 230, the authors mention that the batch size is set to 1 and mention that it aligns with the requirements of the volumetric task. How does batch size of 1 align with the requirements of the volumetric task?

Authors’ response:

Thank you for pointing this out. Due to the large size of each reconstructed 3D volume (up to 300×600×600) and the complexity of our Swin-Transformer-based architecture, the peak memory usage approaches 79GB on an A100 GPU, leaving little room for increasing the batch size. Therefore, we set the batch size to 1 to ensure stable training and inference under these memory constraints.

Comment 5:

At line 66, the authors say that Richardson-Lucy deconvolution should not be used because of being computationally intensive. It will be helpful to know quantitatively how much the computational requirements are. For example, a statement like ‘using it on a XxYxZ sized data takes A amount of time with B GPU/CPU specs’ can be helpful.

Authors’ response:

We appreciate the reviewer’s suggestion and now provide quantitative justification. Applying Richardson-Lucy deconvolution (RLD) on a single 600×600×300 volume takes over 42 seconds for 60 iterations and consumes more than 20 GiB of GPU memory, even on an A100 GPU (80GB).

Comment 1:

Generalisability beyond zebrafish Could you add—or at least simulate—experiments on a different specimen (e.g., mouse brain slice or a thicker adult zebrafish) to show that the gains hold outside the current benchmark? Score up if cross-domain PSNR/SSIM or qualitative slices confirm similar improvements; score down if performance drops sharply without a clear explanation.

Authors’ response:

Summary. We validated generalization using a new dataset from a distinct zebrafish strain (H2B-NemoS); our method showed strong cross-strain and OOD performance, confirming robustness to physiological and imaging variations.

We appreciate the reviewer’s suggestion to evaluate the generalizability of our method. While publicly available XLFM datasets beyond zebrafish are currently lacking, we took an important step by introducing a new dataset—XLFM-H2B-NemoS—captured using a genetically distinct zebrafish strain and under varied stimulation protocols.

Specifically, our original dataset is based on the H2B-G8S strain [1], which expresses a fast-responding GCaMP calcium indicator optimized for high signal-to-noise imaging in freely behaving animals. In contrast, the new H2B-NemoS dataset is based on an indicator engineered for larger dynamic range and higher sensitivity to weak signals [2], introducing more sparse and heterogeneous neural activity along with higher background noise. This substantial difference in imaging properties and biological characteristics makes the two datasets effectively distinct in terms of signal distribution, sample noise, and dynamic response profiles.

To simulate domain shift, we collected 4 zebrafish samples from H2B-NemoS under four levels of electrical stimulation (5V, 10V, 15V, 20V), each producing ~70 volumetric frames at 10 fps. The 10V and 15V data were used for training and in-distribution validation, while 5V and 20V data were held out for out-of-distribution (OOD) testing. Apart from a minor increase in PSF regularization weight (from 1e-4 to 5e-3 to adapt to the brighter dynamics), all network and training settings were kept unchanged.

Our method demonstrated strong generalization across this challenging domain shift. As shown in Tables 1 and 2, it outperformed all baselines with clear margins on PSNR and SSIM. Moreover, the zero-shot version of XLFM-Former (trained only on the G8S and tested directly on H2B-NemoS) still delivered 53.72 dB PSNR, indicating impressive robustness to cross-strain variation.

These results validate that our method generalizes well to specimens with distinct physiological properties and imaging conditions, and support its broader applicability to future XLFM data beyond the current benchmark.

Table 1: Ablation Study on the H2B-Nemos Dataset

Table 2: Comparison with Existing Methods on the H2B-Nemos Dataset (baseline: ResNet-101)

References: [1]Zhang, Y., Rózsa, M., Liang, Y., Bushey, D., Wei, Z., Zheng, J., ... & Looger, L. L. (2023). Fast and sensitive GCaMP calcium indicators for imaging neural populations. Nature, 615(7954), 884-891. [2]Li, J., Shang, Z., Chen, J. H., Gu, W., Yao, L., Yang, X., ... & Wang, Y. (2023). Engineering of NEMO as calcium indicators with large dynamics and high sensitivity. Nature methods, 20(6), 918-924.

Comment 2:

Runtime, throughput, and memory footprint Please report FPS and peak GPU memory for one full-resolution volume on a single A100 and compare with the strongest baseline. Showing that XLFM-Former meets real-time (>30 fps) or near-real-time speeds with ≤80 GB will raise confidence; opaque or unfavourable numbers will lower it.

Authors’ response:

We appreciate the reviewer’s concern regarding the practical efficiency of our method. To address this, we benchmarked XLFM-Former against four widely used baselines on a single A100 GPU (80 GB), reporting inference speed (FPS), peak memory usage, and reconstruction accuracy (PSNR).

As shown below, XLFM-Former achieves 48.44 FPS, well above the 30 FPS real-time threshold, while maintaining competitive memory usage (2631 MiB) and the highest reconstruction fidelity (PSNR 52.34 dB) among all compared methods. These results demonstrate that our approach is not only accurate but also efficient enough for real-time 3D microscopy applications.

Inference Efficiency and Reconstruction Accuracy on H2B-Nemos Dataset (Single A100 GPU)

Comment 3:

Masked-view vs. standard MAE baseline Include a control experiment where an off-the-shelf pixel-patch MAE (ViT-MAE) is pre-trained on the same data and fine-tuned identically. If MVM-LF still wins by a clear margin, originality concerns are alleviated; if the gap is marginal, the novelty claim weakens.

Authors’ response:

We thank the reviewer for the valuable suggestion. To assess the effectiveness and originality of our proposed Masked View Modeling for Light Fields (MVM-LF), we conducted a control experiment using a standard ViT-based pixel-patch MAE [He et al., 2022] pre-trained on the same dataset (∼20k unlabelled frames) and fine-tuned with the same downstream setup.

The results are as follows:

ViT-MAE: PSNR = 46.55, SSIM = 0.9752

MVM-LF (Ours): PSNR = 54.04, SSIM = 0.9944

This substantial performance gap highlights that our MVM-LF design, which incorporates prior knowledge of the multi-view structure of XLFM (e.g., angular correlations), is significantly more effective than standard MAE when operating on limited-scale microscopy data. We believe this validates both the originality and practical value of MVM-LF in the context of structured light field inputs.

Comment 4:

Robustness to PSF mis-calibration Demonstrate reconstruction quality when the forward PSF used in ORC is intentionally perturbed (e.g., ±10 % axial FWHM, mild aberration). Stable performance would validate the physics-loss design; sensitivity would suggest the need for adaptive PSF estimation.

Authors’ response:

These results suggest that the reconstruction performance of our ORC loss is largely stable under ±10% perturbations in the axial full-width at half-maximum (FWHM) of the PSF. Specifically, the PSNR fluctuates within ±0.12 dB and SSIM remains within ±0.0002, indicating that the ORC loss is not overly sensitive to small mismatches in PSF calibration.

This robustness arises from the fact that the forward projection consistency enforced by the ORC loss penalizes global mismatches between reconstructed and measured views, while allowing for small inaccuracies in the PSF shape. We believe this property is crucial for real-world applicability, where slight PSF variations are inevitable due to experimental imperfections.

Moreover, these findings suggest that our method does not require exact PSF estimation and can tolerate mild miscalibration, reducing the burden on the calibration process in practical deployments.

Robustness of ORC Loss to PSF Perturbation on H2B-Nemos Dataset.

Comment 5:

Reproducibility details Provide (in supplementary material or repo) the exact decoder architecture diagram, LR schedule, and data-normalisation script that yield the 54 dB result.

Authors’ response:

Thank you for raising this important point. Due to the NeurIPS 2025 rebuttal policy, we are unable to include external links in our response. However, we will submit an anonymized link containing the full training code (including the decoder architecture diagram, learning rate schedule, and data normalization pipeline) to the Area Chair. If you are interested, please feel free to request it via the AC. We apologize for this inconvenience and appreciate your understanding.

Comment 1:

The paper is technically reasonable in a general sense; however, the implementation seems suspicious, and the experimental results appear contradictory. The quantitative results in Table 1 are excessively high, showing a PSNR of 45–54 dB for all methods (while the maximum PSNR for 8-bit signals is 48.13), essentially making them “perfect” equally. Such high PSNR values do not match the poor visualization results in Figure 5.

Authors’ response:

Summary.

• Our PSNR values are computed on 32-bit floating-point data normalized to [0, 2000], not 8-bit integers, making values >48.13 dB valid and consistent with prior XLFM work.

We would like to respectfully but firmly correct a misunderstanding in this comment.

The claim that PSNR values “cannot exceed 48.13 dB” is based on an incorrect assumption that the evaluation is performed on 8-bit integer images in the [0, 255] range. This is not the case in our setting. In line with standard practice in XLFM and calcium imaging, we work with 32-bit floating-point volumetric data, and normalize signals to a biologically meaningful range of [0, 2000]. Under this dynamic range, PSNR values significantly above 48 dB are not only mathematically valid, but widely reported in prior XLFM literature. For example, recent works such as Fast light-field 3D microscopy with out-of-distribution detection and adaptation through Conditional Normalizing Flows (Fast et al. [1]) clearly report PSNR values exceeding 50 dB under the same normalization convention, as shown in their Figure 3.

We acknowledge the reviewer’s concern regarding the visual appearance of Figure 5, but would like to clarify that Figure 5 highlights a challenging anatomical region (the hindbrain) with sparse neural signals. In such cases, PSNR—which evaluates fidelity across the entire 3D volume—may remain high even when small, localized discrepancies are visible. This is a known property of PSNR and does not indicate implementation issues.

Due to this year’s NeurIPS policy prohibiting the release of anonymized links during the rebuttal phase, we are unable to share our implementation at this stage. However, we will additionally release our code upon acceptance, under the supervision of the Area Chair, to ensure full transparency and reproducibility.

Comment 2:

The paper does not present the technical details clearly. The design of the lightweight decoder is unclear, and the pipeline presented in Appendix A is poorly written. Additionally, the details of the data, such as image resolution, sub-aperture image resolution, and reconstruction volume size, are not specified.

Authors’ response:

We thank the reviewer for highlighting these points. Below, we provide the requested clarifications.

(1) Design of the Lightweight Decoder

Our decoder follows a simple yet effective architecture to upsample the Swin Transformer output to a full-resolution 3D volume. It consists of three key stages:

• Projection to depth: A 1×1 convolution maps the Swin features (channel dimension 96) to the target depth size (D=300).
• Spatial upsampling: Two successive transposed convolutions upscale the spatial resolution from 150×150 → 300×300 → 600×600.
• Final refinement: A 1×1 convolution is applied to preserve local semantics and reduce boundary artifacts.

The corresponding PyTorch implementation is provided below for clarity:

class Lightweight3DDecoder(nn.Module):
    def __init__(self):
        super().__init__()
        self.project_to_depth = nn.Conv2d(96, 300, kernel_size=1)
        self.upsample_1 = nn.ConvTranspose2d(300, 300, kernel_size=4, stride=2, padding=1)
        self.upsample_2 = nn.ConvTranspose2d(300, 300, kernel_size=4, stride=2, padding=1)
        self.refine_output = nn.Conv2d(300, 300, kernel_size=1)

    def forward(self, x):
        x = self.project_to_depth(x)       # [B, 96, 150, 150] → [B, 300, 150, 150]
        x = F.relu(self.upsample_1(x))     # → [B, 300, 300, 300]
        x = F.relu(self.upsample_2(x))     # → [B, 300, 600, 600]
        x = self.refine_output(x)          # → [B, 300, 600, 600]
        return x

(2) Clarification of Dataset Specifications

• Raw image resolution: 2048 × 2048, captured by an XLFM system with a microlens array.
• Sub-aperture view resolution: Each of the 27 angular views is aligned and cropped to 600 × 600 pixels.
• Reconstructed 3D volume size: 600 × 600 × 300 voxels.

Comment 3:

The work proposes both a unique dataset and a method specifically tailored for the task of XLFM reconstruction. Although it is rooted in masked pretraining and the Swin Transformer architecture, it introduces a fair rendering consistency loss to inject domain-specific prior knowledge.

Authors’ response:

We thank the reviewer for recognizing the novelty of our dataset and the design of our domain-specific loss function, as well as its conceptual connection to masked pretraining.

While our pretraining strategy is indeed inspired by masked autoencoders (e.g., MAE), our core contribution lies in a fundamentally different view-level masking mechanism tailored to light field microscopy. Instead of masking random image patches, we mask out entire angular views from the 27-view XLFM input and train the model to reconstruct the missing views from the remaining ones.

This design forces the network to learn explicit spatial-angular dependencies across views—an essential property for XLFM systems where angular consistency and optical geometry play a critical role. Unlike standard MAE, which operates on spatial patches, our approach requires the model to reason about structured angular correlations inherent in the light field representation.

Furthermore, as shown in Table 1 of our manuscript, our view-masked pretraining strategy significantly outperforms random-patch masking (MAE-style) approaches, especially in low-supervision settings and when generalizing to new zebrafish strains. These results validate the advantages of our domain-aware masking paradigm and its alignment with the underlying physics of the imaging system.

Comment 4:

Implementation of XLFM-Transformer: It is unclear what does the “cropping operation” does. According to appendix A, it raises the image to a 3D representation of shape H x W x D x S ( the channel S =1). I wonder in what sequency does the sub-aperture images are stacked in the D dimension, how to ensure the 2D neighboring information of sub-images is reflected in this one dimensional stacking? Moreover, what is changed in the positional embedding for this added dimension? The equation 13 and 14 contradicts each other, and it seems duplicated illustration in the last part in Appendix A.

Authors’ response:

We appreciate the reviewer’s close reading and raise the following clarifications:

• (1) “Cropping operation” explanation We extract 27 sub-aperture views of size 600×600 from the 2048×2048 input image based on known microlens coordinates. This operation corresponds to spatially aligning and cropping the raw XLFM measurement into angular sub-images.

• (2) Stacking order and spatial adjacency The 27 views are stacked into a tensor following their spatial layout on the microlens array (i.e., consistent with optical geometry). While we maintain this order for reproducibility, we note that our 3D Swin Transformer possesses a large receptive field, making the model robust to minor changes in view order.

• (3) Positional embedding for the D dimension We do not alter the positional embedding mechanism of the Swin Transformer. During pretraining, we randomly zero out a subset of views (angular-wise masking), which does not require architectural modification. The model learns to complete missing views from visible ones, guided by inherent spatial-angular dependencies.

• (4) Appendix A redundancy and equations We acknowledge that Equation (13) in Appendix A is redundant and may have caused confusion. The decoder’s structure is accurately described by Equation (14)。

Comment 5:

The idea of masked pretraining doesn’t seems to fit quite well with such 3D Swin Transformer architecture. For masked pretraining in [1,2], each sub-image is treated as a token and are assigned a unique positional embedding, which is clearly not the case for this implementation of “3D Swin Transformer”. Does the pretraining implemented in this paper simply set the masked sub-image to zero? A more detailed implementation should be included, and the rationality to do that should be discussed.

Authors’ response:

We respectfully disagree with the reviewer’s concern, and appreciate the opportunity to clarify our design.

While our MVM-LF framework may superficially resemble pixel-level masked autoencoders such as MAE, its inductive bias, masking granularity, and objective are fundamentally different and tailored to XLFM.

• Masking strategy: Instead of masking image patches, we mask entire sub-aperture views (i.e., angular slices from XLFM’s 27-view input). This aligns with the physical imaging model of light field microscopy, where each view encodes angular information.
• Network compatibility: During pretraining, we randomly zero out a subset of views (i.e., set entire channels to zero) and train the model to reconstruct the missing views using the remaining ones.
• No change to positional embedding: Since Swin Transformer already uses relative positional encoding within each view, and our masking is done at the view level, we do not modify the embedding scheme. The model naturally learns to associate angular correlations across channels.
• Empirical evidence: Our ablation study (Table 2) shows that MVM-LF pretraining achieves 54.04 dB PSNR, outperforming random pixel masking (52.97 dB) by a clear margin (+1.07 dB). Given the high baseline, this is a significant improvement in 3D volume reconstruction tasks.

We sincerely thank the reviewer for their detailed and thoughtful feedback.

Comment 1:

Explanation of ORC Loss and presentation issues: While the authors have a promising method, the paper suffers from widespread issues in its presentation quality that limit the clarity of the paper. The most major issue with presentation is that one of the major contributions of the paper as claimed by the authors, the "ORC Loss," does not obviously appear in any explanation of the losses used for the reconstruction phase of the training process. The authors later introduce a "PSF Loss" which is not explained, which is likely the "ORC Loss," but this is not clear. Other issues with the presentation are listed in the minor issues section below.

Authors’ response:

"ORC" stands for Optical Reconstruction Consistency, which encourages consistency between the forward projection of the reconstructed volume and the measured snapshot image. Specifically, this is implemented by comparing the forward projections (using the PSF) of both the reconstructed volume and the ground-truth label, thereby enforcing alignment between the predicted and real-world optical measurements. Thus, we introduced it as "PSF Loss" in the implementation section, but this is indeed the same as the ORC Loss.

Comment 2.1:

Analysis of ORC Loss and hallucinations: Further, to substantiate the claim that the ORC/PSF Loss can reduce artifacts in the reconstructions and improve cross-sample generalization, the authors should include some examples of reconstructions from networks trained with and without the ORC Loss.

Authors’ response:

We appreciate the reviewer’s suggestion to provide visual examples comparing models trained with and without the ORC Loss. We have prepared visualization results similar to Figure 5 in the main paper. However, due to the NeurIPS 2025 policy which prohibits images and figures in the rebuttal, we are unable to include them here.

To address this concern within the policy constraints, we instead provide quantitative comparisons on multiple held-out zebrafish samples from the H2B-G8S and H2B-Nemos datasets. As shown in the table below, incorporating the ORC (PSF) Loss consistently improves both PSNR and SSIM across all test samples. These improvements demonstrate that the ORC Loss helps reduce artifacts and enhances cross-sample generalization, particularly in challenging anatomical regions such as the hindbrain.

Comment 2.2:

While it is apparent that penalizing the forward projections of the reconstruction and the ground truth can enforce some consistency between the reconstruction and the measured snapshot image, there can still be hallucinated cells/artifacts in the reconstruction. Indeed, some cells/background appear in the XLFM-Former in Figure 5 (hindbrain) which are not present in the ground truth.

Authors’ response:

We appreciate the reviewer’s observation. While some background or cellular structures appear in the XLFM-Former output but not in the ground truth, we believe these are not necessarily artifacts. Instead, they may reflect limitations of the imaging and labeling process.

Specifically, the ground truth may miss low-SNR or ambiguously localized structures—especially in dense regions like the hindbrain—where XLFM captures signals beyond what is confidently annotated. Since downstream analysis focuses on neuronal clusters rather than individual cells, such discrepancies are unlikely to affect practical utility.

Ultimately, these differences likely arise from imperfect ground-truth alignment, not reconstruction failure. Resolving this would require single-cell-resolution annotations, which XLFM currently lacks.

Comment 2.3:

Also, some explanation of why this ORC Loss does not use the measured image instead of a forward projection would be useful (e.g. whether this helps avoid reconstruction artifacts due to some background/dark frame in the measurement).

Authors’ response:

We appreciate the reviewer’s question. In the early stages of developing XLFM-Former, we did explore using the raw measured light field image directly in the ORC Loss. However, we found this approach to be significantly less stable during training due to high levels of noise and background artifacts inherent in the measured data (e.g., dark current, scattering, and sensor noise).

By contrast, using the forward projection of the ground-truth volume via the PSF offers a cleaner and more structured supervision signal, ensuring that the network focuses on learning the physical consistency between volume and measurement, without being distracted by low-level measurement noise.

Comment 3:

Appropriate baselines: The authors compare their reconstructions against reconstructions from a range of other neural network architectures. However, they do not compare their reconstructions using the methods described by prior work in snapshot imaging (Fourier light field or mask-based ) with deep learning, for example: Yanny et al.: "Deep learning for fast spatially varying deconvolution", Optica, 2022, https://doi.org/10.1364/OPTICA.442438 and references [25, 30, 9] from their paper. If memory is a concern, an explanation of why these other methods are not appropriate for these data would be useful.

Authors’ response:

• (1) Incompatible forward model (Yanny et al. [Optica 2022]): Their method assumes a spatially variant convolutional model, which does not hold for XLFM. XLFM relies on complex wavefront propagation and angular multiplexing, violating the convolution assumption. Even recent XLFM-specific works (e.g., Ref. 25) are not compatible with Yanny et al.'s formulation.
• (2) Different reconstruction goals (Refs. 25, 30): These works focus only on sparse neural signals, ignoring structural context. In contrast, our method reconstructs full tissue volumes, critical for tasks like optogenetic targeting. Moreover, their approaches require immobilized samples (e.g., Ref. 25 fuses three frames), making them unsuitable for dynamic settings.
• (3) FourierNet (Ref. 9) is infeasible at our scale: Operating fully in the Fourier domain, it demands >80GB memory even for smaller outputs. Our volumes (up to 300×600×600) render this approach impractical. Instead, we use Swin Transformer for efficient global modeling with far lower memory cost.

Comment 3.1:

It would also be useful to compare against a reconstruction using an iterative deconvolution algorithm such as Richardson-Lucy.

Authors’ response:

We appreciate the reviewer’s suggestion to compare against classical iterative deconvolution algorithms. As shown below, we benchmarked RLD at different iteration counts (20/30/40) on the challenging H2B-Nemos dataset. While (Richardson-Lucy Deconvolution)RLD can achieve high PSNR/SSIM given sufficient iterations, it incurs orders of magnitude higher memory usage and lower speed, making it impractical for real-time or high-throughput applications.

We appreciate the reviewer’s suggestion to compare against classical iterative deconvolution algorithms. As shown below, we benchmarked RLD at different iteration counts (20/30/40) on the challenging H2B-Nemos dataset. While (Richardson-Lucy Deconvolution)RLD can achieve high PSNR/SSIM given sufficient iterations, it incurs orders of magnitude higher memory usage and lower speed, making it impractical for real-time or high-throughput applications.

Comment 4:

Ground truth in datasets: Another major issue is that the collected dataset as described, though it is valuable for the computational optics community, does not describe the ground truth data collection process or resolution, though some kind of volumetric ground truth must be present and is shown in Figure 5. The authors should clarify the process of collecting the ground truth for the dataset.

Authors’ response:

We appreciate the reviewer’s attention to dataset clarity. We follow the established practice in XLFM literature and generate ground truth volumetric labels using 60 iterations of Richardson-Lucy Deconvolution (RLD-60), a widely adopted baseline in previous works. This iterative deconvolution algorithm is considered the gold-standard for recovering high-resolution volumes from XLFM snapshot measurements, and serves as our supervision target.

Thank you for the clarifications. The training methodology and architecture are promising, especially due to the speed of the reconstructions which was not originally highlighted in the paper. I will raise my score, but I still think that there could be a more thorough evaluation against prior work. I could still further increase my score pending response from the authors.

Specifically, I am satisfied with the response to all but the following:

3. Appropriate baselines: While Yanny et al. allow for a spatially varying PSF in the forward, their reconstruction architecture is still applicable in the case that the number of varying regions is 1. The forward model for XLFM is still a convolution between the PSF and the sample, where the "angular multiplexing" is a result of different beams of the PSF convolving with the sample at different plane. Indeed, the authors' own Equation 1 confirms this. Also, the argument in this response is about the forward model rather than the reconstruction method (what the submission is about). Finally, I appreciate that for these large volumes, you would need multiple GPUs to implement some of the prior architectures. However, it would still be possible to place this architecture in the context of prior work in deep learning for light field reconstructions using artificially smaller/easier volumes.

4. Ground truth in datasets: Thank you for the clarification. Richardson-Lucy may be the standard for recovering volumes from XLFM images, but it does not guarantee that the reconstruction is free of artifacts (and the projections with a saturated color map shown in Figure 5 make it difficult to judge). One of the advantages of using deep learning for snapshot reconstructions is that the network can take into account the prior of the particular kind of sampled being imaged (in this case zebrafish) for improved reconstruction quality. This cannot be evaluated if the ground truth is itself a reconstruction.