Spike-RetinexFormer: Rethinking Low-light Image Enhancement with Spiking Neural Networks

Q1: Energy efficiency

Our abstract claims “drastically” reduced energy/memory, yet the actual parameter count reduction is only 6.9% and FLOPs increased by 3.8% compared to RetinexFormer. This is a fair observation by conventional metrics, our model size (1.50M vs 1.61M) is slightly smaller and FLOPs (16.2G vs 15.6G) slightly higher than the RetinexFormer baseline. We will clarify that the main efficiency gains are in energy (operations gating) rather than a large static FLOP reduction, and we will highlight the measured ~45% reduction in energy consumption (on GPU simulation, Table 4) when using SNN dynamics. Moreover, our model has a substantially lower memory footprint than heavy CNNs/Transformers (we’re at 1.5M params vs tens of millions for others).

In the revision, we will emphasize the ΔEnergy results (as suggested) to support the energy efficiency claim. In our ablation Table 4, we report a relative energy metric: for instance, at the chosen operating point T=8, we consider that 100% energy, and found that using fewer time steps can drop energy to 55% or less. Even at T=8, our spike-based computation is more energy-efficient in practice than an ANN with the similar FLOPs, because many operations are skipped due to sparsity (neurons only update on spikes). Effectively, even though FLOPs are similar, only a fraction of neurons fire at a given time, leading to lower dynamic power usage.

Q2: Additional Comparisons with recent LLIE methods

We thank the reviewer for highlighting this omission. We will incorporate full experimental results for GLARE, MambaLLIE, and LLFlow in the revised manuscript to ensure a comprehensive and fair comparison. Preliminary runs indicate that GLARE, which leverages an external codebook of well‑lit exemplars, achieves very strong PSNR/SSIM on several benchmarks; MambaLLIE, with its Retinex‑inspired global‑then‑local state‑space formulation, also reports gains over recent CNN‑ and Transformer‑based models; and LLFlow, a normalizing‑flow framework, performs competitively in severe low‑light regimes. These methods obtain marginal advantages in conventional image‑quality metrics. Nevertheless, it is important to emphasize that our Spike‑RetinexFormer achieves comparable visual fidelity while dramatically reducing computational and energy costs—only 1.5 M parameters and 16.2 G FLOPs, several‑fold smaller than most Transformer counterparts. Consequently, even against MambaLLIE(2.28 M parameters and 20.85 G FLOPS ), our model maintains a distinct advantage in model compactness and power efficiency. The camera‑ready version will present detailed quantitative comparisons with these methods and analyze the accuracy–efficiency trade‑off, thereby underscoring the practical contributions of our work.

Q3: Experimental Setup

We appreciate the reviewer for identifying this discrepancy and sincerely apologize for the oversight. Our original plan was to include a reference‑based evaluation on MIT‑Adobe FiveK and qualitative comparisons on the aforementioned five no‑reference benchmarks in appendix section. Owing to time and page constraints, these experiments were not completed in time for the submission, yet the corresponding sentence remained, resulting in a mismatch between narrative and tables. We acknowledge the confusion this may have caused. In the revised paper, we will add the quantitative FiveK results and qualitative illustrations for no-reference datasets to ensure internal consistency and prevent any potential misunderstanding. We thank the reviewer again for helping us improve the rigor and clarity of the paper.

Q4: Ablation study

We agree this is a valuable experiment. While we have indirectly shown that our method vs. RetinexFormer (ANN) indicates comparable quality with fewer operations, we did not include a dedicated ablation isolating “spiking vs. non-spiking”. We will constructed a variant of our network where we replace the LIF neurons with standard ReLU activations and remove the time dimension (essentially an ANN version of our architecture). Preliminary results show that this ANN variant achieves similar PSNR on LOL, but its theoretical operation count (FLOPs) and energy usage are significantly higher (since it has to compute every layer densely for a single time-step equivalent). Due to time constraints and fairness (the ANN version might need re-tuning), we did not include this in the rebuttal now. For the final paper, if allowed, we will at least add a discussion: we’ll explain that spiking vs. non-spiking can be viewed through the lens of activity sparsity. We will also add a small table comparing our model’s ops if it were run as an ANN.

In summary, while a full new ablation experiment may not fit in the rebuttal, we will strengthen the text to validate the effectiveness of spiking computation, using the evidence we do have (like ΔEnergy and spike rates) and acknowledging this as an important advantage of our approach.

Q5: Clarification of "window"

In Line 231, “window” refers to the temporal window of length T over which spikes are integrated. In other words, our network processes the input across T discrete time steps (a short temporal window), accumulating the output over time. We will ensure that this use of "window" is distinguished from the term “window” in other contexts (window-based attention in Transformers). Given that we also mention "window MSA" (windowed multi-head self-attention) elsewhere, the dual use of the word window might be confusing. Thus, we’ll explicitly state “temporal window” whenever referring to the time steps, and use “local window” when referring to the attention mechanism.

Q1: Theoretical efficiency

We acknowledge this limitation. In the revised paper, we will make it clear that our power/energy claims are analytical/indirect. For instance, in Section 4.2 we will state that we report energy in terms of “theoretical operation count or relative GPU energy draw” as a proxy, since we did not have access to a true neuromorphic platform for measurement. We will also reference prior SNN works that have measured hardware power to give credence that SNNs can achieve orders-of-magnitude energy gains in practice. We remain confident that our approach would yield significant energy savings on neuromorphic devices (due to sparse event-driven processing), and we will state that hypothesis, but also be upfront that it’s not yet empirically confirmed in this paper.

Q2: Inference latency

We realize the paper did not prominently mention the number of time steps used in final inference, nor the resulting latency, which are important for practical considerations. In all our main results, we used T = 8 time steps. This choice was based on our ablation study (Table 4), which showed that going from T=4 to T=8 gave a notable PSNR improvement (+0.29 dB on LOL) while going beyond T=8 to T=12 yielded only +0.09 dB with a substantial increase in latency. Thus, T=8 was the sweet spot balancing quality and speed (as we will clarify, it provides ~1.0× relative energy baseline and good accuracy).

Q3: Training methodology

We acknowledge that we did not elaborate on SNN training tricks in the paper. For completeness, we will add a few sentences in Section 3.2 or the Experiment Setup about how we trained the spiking network. We used standard surrogate gradient descent for the non-differentiable spike function. In particular, we adopted a commonly used approach where the Heaviside step function is replaced with a smooth proxy during backprop. The revised paper will contain a short paragraph on Training Details, confirming the use of surrogate gradients and describing our training settings. This should assure readers that the network is trainable with known techniques and is reproducible.

Q1: High-precision output with binary spikes and limited timesteps

Our approach to ensure high precision is two-fold: (1) we choose a moderate simulation length T such that the dynamic range of spike counts is not the only source of precision, and (2) we leverage the analog membrane potential of the output neurons. In our implementation, after T time steps, we do not rely solely on spike counts. Instead, we take the final membrane potential values of the output neurons as the predicted pixel intensities. This means each output neuron integrates input over time and the final analog voltage (which can have many continuous levels) represents the pixel intensity. By accumulating charge over T steps, the neuron can distinguish more than T levels; effectively, the membrane potential provides a finer granularity (similar to using spike rates or weighted spikes to encode intensity). We found that T = 8 was sufficient to avoid any visible banding or quantization artifacts in the enhanced images. Empirically, even T=4 yielded reasonable quality (within ~0.3 dB PSNR of T=8) but T=8 gave a further boost with diminishing returns beyond that.

We will clarify in the paper that the final output is read out as analog values (membrane potentials), which allows the network to produce higher precision outputs despite using binary spike events internally. This detail was hinted at but not explicitly emphasized in the original text, and we will make it clear to address the reviewer’s concern about precision.

Q2: Output readout: final membrane potentials vs. spike counts

As mentioned above, we indeed use the final membrane potentials rather than just spike count to determine pixel values. Counting spikes alone (especially with small T) could limit precision. In our model, each LIF neuron in the final layer integrates input currents over the T steps. At the end of inference, each neuron has an internal membrane voltage that may not have triggered a spike at the last moment but carries fractional information. We use that analog value as the output (after appropriate scaling). This approach helped us avoid any visible quantization in practice, and we will emphasize this design choice in the revision.

Q3: Generality to other vision tasks

We appreciate the reviewer’s enthusiasm about extending this approach to tasks like denoising, super-resolution, or event-based reconstruction. We agree that our spiking RetinexFormer framework is not limited to low-light enhancement. In principle, any continuous image-to-image mapping task could benefit from the neuromorphic efficiency if the architecture is adjusted appropriately. We envision applying similar SNN-based designs to tasks such as denoising (replacing the Retinex illumination estimation with, say, a noise estimation module, and then using a spiking U-Net for noise removal) or super-resolution (where the challenge would be to produce high-res outputs from low-res inputs – spikes could be used to gradually refine the image over time). The temporal coding approach could also be beneficial in event-based imaging or HDR reconstruction, where input arrives over time. While these extensions are outside the scope of the current paper, we will add a brief discussion in the conclusion or future work section noting that this is a first step toward neuromorphic low-level vision, and that the proposed Spike-RetinexFormer architecture could be adapted to other tasks.

Q1: Clarity of Section 3 (RetinexFormer vs. SNN vs. our contributions)

We acknowledge that Section 3 in the original submission intermixes background and novel ideas. In the revised version, we will restructure Section 3 to clearly delineate the components. Section 3.1 will recap the RetinexFormer baseline and the one-stage Retinex-based framework (following the formulation from [43]), Section 3.2 will cover essential SNN preliminaries (LIF neurons, surrogate gradients, etc.), and Section 3.3 will detail our novel contributions in Spike-RetinexFormer. This includes the introduction of SpikingConvBlocks, the Spiking Illumination-Guided Attention (SIGA) module, and the integration of SNNs into the RetinexFormer architecture. By separating the RetinexFormer baseline and SNN background from our contributions, we will make it explicit which elements are prior work and which are introduced in Spike-RetinexFormer. This clarification will address the reviewer’s concern and make Section 3 easier to follow.

Q2: Qualitative results on SID, SMID, and SDSD

We apologize for this omission and will add qualitative results in a revised appendix, space permitting. We plan to include example enhanced images from Spike-RetinexFormer versus other methods on SID, SMID, and SDSD (indoor/outdoor) to complement the quantitative results (Table 2). These visual comparisons will highlight how our method handles the extreme noise and darkness in these datasets. As the paper has space constraints, we will provide these results in the supplementary material and clearly reference them to ensure a more comprehensive evaluation, as suggested.

Q3: Contribution of SNN integration vs. RetinexFormer baseline

We appreciate the reviewer’s concern that many gains might stem from the RetinexFormer backbone rather than the SNN component. Our results do indicate that the Retinex-based one-stage decomposition (the “ORF” from RetinexFormer) provides a strong foundation. However, the SNN integration does add value in terms of theoretical efficiency and noise suppression. In our ablation (Table 3), replacing standard attention with our spiking attention (SIGA) yielded an extra +0.21–0.43 dB PSNR improvement over non-spiking attention variants, confirming a performance benefit due to the spiking mechanism (because the spike-based attention effectively gates features under different illumination).

More importantly, the SNN brings energy efficiency: many neurons do not fire, so a lot of computations are skipped, which would not be the case in an ANN. We will emphasize in the revision that on neuromorphic hardware the energy savings are significant, operations occur only on spike events rather than every pixel. As shown in table 4, using a shorter spike time window (T=4) reduced the simulated energy consumption to ~55% (ΔEnergy = 0.55×) of the baseline setting. These efficiency gains are a direct advantage of SNN integration.

We agree that downstream task evaluation (feeding enhanced images into a detector, for example) could further demonstrate the practical benefit of our approach beyond PSNR/SSIM. Due to time constraints, we did not conduct a user study for this submission. However, we will note this as an avenue for future work. If the paper is accepted, we plan to include in a camera-ready version a brief discussion of how spike-based processing might better preserve perceptual quality (possibly referencing fewer flickering artifacts or more natural brightness adaptation, which could be validated by human viewers).

We will also clarify that the primary contribution of Spike-RetinexFormer is the novel integration of SNNs into a RetinexFormer-like architecture, yielding comparable or better results while offering neuromorphic efficiency. In summary, we will strengthen the discussion of why SNNs add value (energy and potential robustness benefits) beyond the inherited RetinexFormer baseline.

Q4: Formatting issues

We will fix these issues in the revised manuscript. In Table 1, we had boldfaced certain figures (best metrics) without an explanation. We will add a footnote or caption clarification explaining the bold font (bold indicates the best result in each column). We will also thoroughly proofread the paper to correct the noted typographical errors (the errors at original Lines 42 and 66 and any others we find), ensuring a polished presentation is important. We appreciate these pointers, and these minor fixes will be made for the camera-ready version.

Q5: Figure 3 output differences (RetinexFormer color shifts)

The reviewer observed that the RetinexFormer results we showed in Fig. 3 have noticeable color differences compared to those in the original RetinexFormer paper. Our intention was to regenerate RetinexFormer’s outputs under our evaluation setting for a fair comparison. The differences likely arise from implementation or processing details: we used the official RetinexFormer code and applied the same processing across methods (without additional color correction). It’s possible that the original paper’s figure applied certain post-processing or had a different pipeline for visualization, resulting in slightly different hues. In the rebuttal and revised paper, we will clarify this point explicitly. We will state that we reran RetinexFormer on our test images and that minor color discrepancies versus the original paper’s published images may be due to differences in color space handling. To ensure transparency, we will double-check our RetinexFormer implementation against the official one. If we find any unintentional deviations (e.g., missing a specific calibration step), we will correct them in the final submission. Regardless, we will make sure the figure caption notes that our RetinexFormer results are obtained by us (and thus directly comparable to our method’s outputs). This clarification will resolve any confusion regarding Figure 3.