Spik-NeRF: Spiking Neural Networks for Neural Radiance Fields

Thanks for your efforts in reviewing our paper and your recognition of our novel method, clear presentation, and open source. The response to your questions is given piece by piece as follows.

---

W1: All experiments were reported at a timestep of T=2. Since SNN behaviour often varies substantially with temporal resolution, it would be beneficial to evaluate the model using additional timesteps.

A1: Thanks for the advice. We add the experiment results for 4 timesteps here. We did the experiments for Chair, Lego, and Horns, considering the limited time. It can be seen that the performance of Spik-NeRF with 4 timesteps could be improved further and is very close to the NeRF.

---

W2: The comparison between models appears to be unfair. While the baseline NeRF model was trained for 200,000 iterations, the proposed model was trained for 300,000 iterations. This discrepancy in training duration could influence performance and may partially account for the observed improvements, rather than reflecting the superiority of the proposed method.

A2: Thanks for the advice. Since SNN is more difficult to train, we use more training iterations as is commonly done in this field. Here, we also train the NeRF with 300,000 iterations for Chair, Lego, and Horns, considering the limited time. It can be seen that the performance remains almost the same as the iteration increases for the ANN-based NeRF. For a fair comparison, we choose their converged performance, and thus different iterations for ANN and SNN, respectively.

---

Q1: Could the authors extend their experiments to include additional timesteps to better assess the temporal generalization of the proposed approach?

A1: Thanks for the question. Please see our response to W1.

---

Q2: What is the rationale for training the proposed model for 300,000 iterations, especially given that the baseline NeRF model was trained for only 200,000? Could this discrepancy have influenced the reported performance gains?

A2: Thanks for the question. Please see our response to W2.

Thanks for your advice again. If you have any other questions, please feel free to let us know.

Thanks for your efforts in reviewing our paper and your recognition of our novel method and clear presentation. The response to your questions is given piece by piece as follows.

---

W1: The experiments conducted in the paper are only for a few timesteps. It would be good to test the proposed method with a broader range of timesteps to understand its performance across different configurations.

A1: Thanks for the advice. We add the experiment results for 4 timesteps here. We did the experiments for Chair, Lego, and Horns, considering the limited time. It can be seen that the performance of Spik-NeRF with 4 timesteps could be improved further and is very close to the NeRF.

---

W2: It does not compare its results to other works that have applied SNNs to NeRF. A comparison with existing studies in this area would help contextualize the significance and impact of this contribution.

A2: Thanks for the advice. There are 4 recent related SNN-NeRF variants. However, we feel only [1] is suitable to compare with us. In [1], an ANN-SNN model is proposed to develop energy-efficient spiking neural rendering using the ANN-to-SNN conversion approach, which requires many timesteps. We here report the detailed comparison below on both the synthetic dataset and the realistic dataset. It can be seen that ANN-SNN NeRF is much worse than our method. With only 2 timesteps, our method obtains similar performance to ANN-SNN NeRF, even with 256 timesteps.

Other studies differ from our approach. For instance, hybrid ANN-SNN models were proposed in [2] and [3]. These works employ non-linear, non-spike functions to post-process the density-related outputs of the original ANN-based NeRF models. At the same time, these work did not report their PSNR. Thus, we can not compare our work with theirs. Another relevant study [4] presents a spiking neural radiance field model based on the DVGO and TensoRF frameworks, which are much different from our work. Additionally, it uses different datasets from our work. Thus, it is not suitable to compare our work with theirs either.

[1] Spiking-NeRF: Spiking neural network for energy-efficient neural rendering

[2] Spiking nerf: Representing the real-world geometry by a discontinuous representation

[3] Sharpening Your Density Fields: Spiking Neuron Aided Fast Geometry Learning

[4] SpikingNeRF: Making Bio-inspired Neural Networks See through the Real World

---

Q1: Could the authors extend their experimental evaluation to include other timestep values (e.g., 4) to assess how the ternary spike neurons perform across various temporal resolutions?

A1: Thanks for the question. Please see our response to W1.

Thanks for your advice again. If you have any other questions, please feel free to let us know.

Thanks for your efforts in reviewing our paper and your recognition of our well-motivated and novel method, solid theoretical justification, good results, and clear presentation. The response to your questions is given piece by piece as follows.

---

W1: The model is evaluated only at T=2. Since the energy-performance trade-off is central to SNN research, a more thorough analysis across different timesteps would be valuable.

A1: Thanks for the advice. We add the experiment results for 4 timesteps here. We did the experiments for Chair, Lego, and Horns, considering the limited time. It can be seen that the performance of Spik-NeRF with 4 timesteps could be improved further and is very close to the NeRF.

---

W2: While comparisons with Spiking-NeRF (Li et al., 2025) and ANN-based NeRF are provided, more detailed ablations against other recent SNN-NeRF variants would strengthen the empirical contributions.

A2: Thanks for the advice. There are 4 recent related SNN-NeRF variants. However, we feel only [1] is suitable to compare with us. In [1], an ANN-SNN model is proposed to develop energy-efficient spiking neural rendering using the ANN-to-SNN conversion approach, which requires many timesteps. We here report the detailed comparison below on both the synthetic dataset and the realistic dataset. It can be seen that ANN-SNN NeRF is much worse than our method. With only 2 timesteps, our method obtains similar performance to ANN-SNN NeRF, even with 256 timesteps.

Other studies differ from our approach. For instance, hybrid ANN-SNN models were proposed in [2] and [3]. These works employ non-linear, non-spike functions to post-process the density-related outputs of the original ANN-based NeRF models. At the same time, these work did not report their PSNR. Thus, we can not compare our work with theirs. Another relevant study [4] presents a spiking neural radiance field model based on the DVGO and TensoRF frameworks, which are much different from our work. Additionally, it uses different datasets from our work. Thus, it is not suitable to compare our work with theirs either.

[1] Spiking-NeRF: Spiking neural network for energy-efficient neural rendering

[2] Spiking nerf: Representing the real-world geometry by a discontinuous representation

[3] Sharpening Your Density Fields: Spiking Neuron Aided Fast Geometry Learning

[4] SpikingNeRF: Making Bio-inspired Neural Networks See through the Real World

---

W3: The paper emphasizes energy efficiency, but no actual energy consumption or MAC operation estimates are reported. Including such results would better quantify the benefits of Spik-NeRF.

A3: Thanks for the advice. Here, we add the energy comparison of NeRF and Spik-NeRF below. Following the practice in [Enabling Spike-Based Backpropagation for Training Deep Neural Network Architectures] and [Spiking-NeRF: Spiking Neural Network for Energy-Efficient Neural Rendering], we utilize 3.2pJ/MAC and 0.1pJ/AC as the energy consumption baseline. In NeRF, the number of MAC operations required for a single forward propagation process is 0.56M. In Spik-NeRF, we first calculate the mean sparsity of activation of Spik-NeRF on the whole synthetic dataset using 2 timesteps, and it is 22.72%. Thus, the number of AC operations required for a single forward propagation process is 0.25M. It can be seen that the energy consumption of Spik-NeRF is much less than that of NeRF.

---

Q1: Why was only T=2 tested? How does Spik-NeRF quantitatively compare against other SNN-NeRF models?

A1: Thanks for the question. Please see our response to W1 and W2.

Thanks for your advice again. If you have any other questions, please feel free to let us know.

Thanks for your efforts in reviewing our paper and your recognition of our clear presentation, well-motivated and simple method, and solid theoretical analysis. The response to your questions is given piece by piece as follows.

---

W1: Although the authors claim in P3, L122, that the number of timesteps directly impacts the model’s energy efficiency, the experiments do not include any energy analysis. Compared to vanilla NeRF—which is now considered a relatively weak baseline—the approach may indeed consume less energy during inference, but the training overhead could be substantial due to the added complexity of ANN-NeRF distillation.

A1: Thanks for the advice. Here, we add the energy comparison of NeRF and Spik-NeRF below. Following the practice in [Enabling Spike-Based Backpropagation for Training Deep Neural Network Architectures] and [Spiking-NeRF: Spiking Neural Network for Energy-Efficient Neural Rendering], we utilize 3.2pJ/MAC and 0.1pJ/AC as the energy consumption baseline. In NeRF, the number of MAC operations required for a single forward propagation process is 0.56M. In Spik-NeRF, we first calculate the mean sparsity of activation of Spik-NeRF on the whole synthetic dataset using 2 timesteps, and it is 22.72%. Thus, the number of AC operations required for a single forward propagation process is 0.25M. We also add the experiments with more timesteps for Spik-NeRF, When T = 4, the mean sparsity of activation of Spik-NeRF becomes 19.08% and the number of AC operations becomes 0.43M. It can be seen that the energy consumption of Spik-NeRF is much less than that of NeRF. But as the timestep increases, the energy consumption of Spik-NeRF increases too.

For the training overhead, we agree that training overhead is a common concern in the SNN field. However, this is a challenge in this field, not only specific to our method. Due to the inherent requirement of multiple timesteps in SNNs, their training typically incurs heavier burdens compared to ANNs, which is a well-recognized consensus in the community. To illustrate this, we have provided statistics on the training time for 100 iterations among NeRF, Spik-NeRF, and Spik-NeRF distilled from NeRF. These results clearly show that the training time of NeRF is significantly shorter than that of Spik-NeRF. Even with ANN-NeRF distillation, our method only increases 16% training time compared to the baseline. For the memory, it exhibits a similar phenomenon.

While we acknowledge that our approach does have a higher training burden compared to vanilla ANNs, we would like to emphasize its superiority over existing spiking NeRF methods in terms of training efficiency. Most current spiking NeRF methods require at least 10 timesteps to achieve satisfactory performance. In contrast, our method only needs 2 timesteps. We have also included the training burden of Spiking NeRF with 10 timesteps for comparison, which clearly demonstrates the substantial advantage of our approach in reducing training overhead. Thus, our method effectively alleviates the training burden commonly associated with SNNs, making it more practical for real-world applications.

---

W2: The experiments do not report training time or inference time. Another concern is whether the proposed methods are compatible with faster NeRF variants such as InstantNGP, Nerfacto, or TensoRF. Focusing exclusively on vanilla NeRF may limit the practical relevance of the work for SNN applications in neural rendering or NVS.

A2: Thanks for the advice. Here, we add statistics on the training time for 100 iterations and inference time for one image of Chair, Lego, and Horns, among NeRF, Spik-NeRF, and Spik-NeRF distilled from NeRF. We will add this to the final paper version. It should be noted that the training time and inference time of SNNs are worse than ANNs. This is an inherent problem in the field of SNN.

In addition, our method is also compatible with faster NeRF variants. We would like to address it from two aspects:

First, the core of Spik-NeRF lies in improving the neural network components within NeRF—specifically, optimizing the feedforward neural networks used for predicting properties like density and color. This design philosophy is inherently transferable to other NeRF variants, since all these variants fundamentally rely on neural networks to model and predict scene-related information. Thus, the idea of enhancing neural network components with spiking mechanisms, as in Spik-NeRF, can be naturally extended to these variants.

Second, to validate this transferability, we have applied Spik-NeRF’s core idea to TensoRF with 2 timesteps for the TensoRF-CP-384 backbone. We did the experiments for Chair and Lego, considering the limited time. The results demonstrate that our method retains its effectiveness in this setting.

---

Q1: It is unclear whether the capacity can be further increased. For example, could quaternary spike neurons be used? Is there an upper bound on the number of discrete states that remain compatible with efficient SNN operation?

A1: Thanks for the advice. The capacity can be further increased. We add the experiment results for quaternary spike neurons here. We did the experiments for Chair and Lego, considering the limited time. It can be seen that the performance of Spik-NeRF with quaternary spike is almost the same as NeRF. However, to retain the event-driven, multiplication-free advantages typical of SNNs, 3 is the upper bound on the number of discrete states, since only {-1,0,1} could be chosen for spike to keep multiplication-free advantages.

Thanks for your advice again. If you have any other questions, please feel free to let us know.