acceptodds
Under review as a conference paper at ICLR 2027

DS-LIF: Novel Spiking Neuron for Reliable Low-Light Image Enhancement

Abstract

The event-triggered mechanism and rich dynamic characteristics of spiking neural networks (SNNs) are highly compatible with the dynamic adaptive repair requirements of low-level visual restoration tasks. However, SNNs adopted for low-light image enhancement (LLIE) fail to accurately distinguish variations in structural information within under-illuminated regions and produce ambiguous spiking responses triggered by image edges, textures, and noise. To address this issue, we propose a reliability-aware spiking neuron named Dynamic-Sleep Leaky Integrate-and-Fire (DS-LIF). It scales the input current prior to membrane potential integration according to local illumination and color contrast, and dynamically masks weakly effective feature channels for image restoration. Subsequently, it combines the mean spike response across all time steps with the local noise, chromatic stability, and underexposure level of the generated image to produce a reliability-guided map. Based on DS-LIF, we construct a dual-stream HVI enhancement framework, SpikeRelight, which can effectively improve the quality of LLIE. To better facilitate cross-stream information interaction, we design a Reliability-Guided Dual-Frequency Interaction (RGDFI) module in this architecture, guided by the reliability-guided map. It decomposes features in the illumination flow and chrominance flow into low-frequency and high-frequency parts, and performs asymmetric cross-stream interaction. Extensive experiments on both paired and unpaired benchmarks, along with ablation studies, demonstrate that SpikeRelight achieves competitive performance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Spiking Neural Network (SNN), Low-Light Image Enhancement (LLIE)

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZJJPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SpikeRelight, a low-light image enhancement framework built around a new spiking unit called DS-LIF. DS-LIF modulates the input current using illumination and local structure cues, suppresses less active channels, and uses temporal spike statistics together with image priors to predict a reliability map. The proposed RGDFI module then uses this map to guide cross-stream interaction between intensity and chrominance features in the HVI space.

Critical strengths
  1. The paper is reasonably well organized and the proposed framework is presented in a modular way. The roles of DS-LIF and RGDFI are separated clearly, which makes the overall pipeline relatively easy to follow.
  2. The paper includes experiments on multiple LLIE benchmarks and several component-level ablations. The evaluation covers both paired and unpaired datasets, providing a reasonably broad empirical assessment of the proposed framework.
Critical weaknesses
  1. The central motivation and claimed novelty of DS-LIF are not sufficiently justified. The paper motivates DS-LIF by arguing that temporal spiking dynamics can help distinguish meaningful structures from noise in low-light regions. However, the same projected current is simply repeated across all timesteps, i.e., X_t=X, so it is unclear what genuinely new information is introduced by the temporal dimension. Moreover, the membrane update itself remains essentially standard LIF, while most of the additional design comes from handcrafted illumination/structure modulation, channel gating, and a reliability prediction head.

  2. The experimental evidence does not substantiate the paper’s key claims about spiking computation, reliability, efficiency, or superiority over relevant prior work. The paper does not directly show that the predicted reliability map actually corresponds to a meaningful notion of restoration reliability. It also motivates SNNs through sparse/event-driven computation, yet provides no latency, energy, or hardware-level efficiency results; the reported dense MACs/FLOPs do not indicate a clear computational advantage. In addition, the most closely related SNN-based LLIE method, Spike-RetinexFormer, is discussed in the related work but omitted from the main quantitative comparison.

(PS: this paper should be desk rejected since the authors do not provide the AI USE STATEMENT required by ICLR.)

Questions for authors
  1. Could the authors provide the exact definitions of the hard channel mask M, the quantities N_c, N_I, C_s, and E_n used by the reliability head, and the full formulation of L_{\mathrm{cal}}? It is also unclear why Eq. (4) introduces an input feature F while the equation itself appears to use only the intensity map I.

  2. How is the proposed reliability map actually validated? The paper repeatedly interprets R as restoration reliability, but there is no direct evaluation showing that it correlates with reconstruction error, noise level, structural correctness, or another measurable notion of reliability. The appendix also mentions “ground-truth reliability calibration,” but does not clearly explain how this ground truth is constructed.

  3. What is the actual efficiency benefit of using an SNN here? The paper motivates the design partly through event-driven sparse computation, but no latency, energy, or sparse execution measurements are reported. The appendix states that spike-rate and channel-sleep statistics are analyzed separately, but such an analysis is not clearly presented in the submission. Could the authors provide direct efficiency comparisons against matched ANN and SNN baselines?

  4. Why is Spike-RetinexFormer not included in the main comparison? This appears to be one of the most directly related prior works, since it also studies spiking neural networks for low-light enhancement, and it is already cited in the related work section. A direct comparison under the same benchmark and evaluation protocol seems important for assessing the contribution of the proposed method.

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

This paper proposes the DS‑LIF spiking neuron, which adaptively modulates input current according to local illumination structures, dynamically masks invalid channels, and outputs a pixel‑wise reliability guidance map. The SpikeRelight two‑stream framework spatially separates illumination and chrominance based on HVI. RGDFI dual‑frequency interaction: the low‑frequency branch preserves global color and illumination continuity, while the high‑frequency branch suppresses noise interference using the reliability map to realize asymmetric cross‑stream information fusion. What's more, DS‑LIF can be plug‑and‑play migrated to other restoration networks.

Critical strengths

Different from conventional LIF neurons that adopt fixed integration‑and‑firing rules, DS‑LIF introduces local illumination conditions and structural responses into the input‑current modulation process, and meanwhile designs a channel‑wise dynamic dormancy mechanism.

Critical weaknesses
  1. The references are relatively outdated, with few publications from the past two years. Although Spike‑retinexformer is cited, it is not included for comparison in the experimental section. The appendix is formatted rather crudely.
  2. Although Table 4 compares DS‑LIF, standard LIF and the continuous‑activation variant, it lacks rigorous comparisons against non‑spiking networks equipped with identical gating structures. The existing experiments cannot rule out the possibility that performance gains mainly stem from the additional attention mechanisms and image priors rather than the spiking firing mechanism itself.
  3. The paper leverages the event‑driven computation and dynamic channel dormancy mechanism of SNNs to illustrate the potential high‑efficiency computing capability of the proposed method. However, sufficient practical runtime data are currently missing to support this claim. Table 11 in the appendix reports the number of parameters and computational complexity. SpikeRelight has approximately 2.58 M parameters and 9.78 G MACs, while CIDNet achieves 1.98 M parameters and 7.98 G MACs. That is, according to the reported dense computation metrics, SpikeRelight is not more lightweight than this baseline.
Questions for authors
  1. As stated in Weakness 1, what exactly are the practical advantages of DS‑LIF compared with existing SNN‑based image enhancement methods?
  2. Can DS‑LIF be applied to other low‑level vision tasks, such as rain removal and dehazing?
  3. Several prior works have already designed illumination‑aware spiking attention modules, such as [1]Spike-HTR, [2]V-LIF.

[1] Spike-HTR: Spiking Neural Transformer for Handwritten Text Recognition. ICML2026. [2] Exploring the Potentials of Spiking Neural Networks for Image Deraining. AAAI2026.

Reviewer_HzN4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DS-LIF, a reliability-aware spiking neuron for low-light image enhancement. DS-LIF introduces illumination-dependent current modulation and channel-wise activity suppression, and combines temporal spike statistics with image-derived cues to estimate restoration reliability. Building upon DS-LIF, the authors develop SpikeRelight, an HVI-based dual-stream framework equipped with a Reliability-Guided Dual-Frequency Interaction (RGDFI) module. Experiments on paired and unpaired benchmarks demonstrate competitive enhancement performance.

Critical strengths
  1. The paper addresses a relevant challenge in low-light image enhancement by investigating adaptive neuronal responses and reliability-guided feature processing.

  2. DS-LIF combines illumination-aware current modulation with channel-wise activity suppression, providing a structured extension of conventional LIF neurons.

  3. RGDFI integrates reliability guidance with frequency-selective cross-stream interaction, allowing intensity and chrominance features to exchange information at different frequency scales.

Critical weaknesses
  1. In Eq. (4), how does the intermediate feature FF contribute to the projected input current? Is the use of the intensity map II intentional? 
  2. How does DS-LIF distinguish meaningful structural responses from noise when both can produce large local contrast values in Eq. (2)? 
  3. Is the channel activity mask recomputed during temporal integration or fixed across all timesteps? What specific mechanism makes the proposed channel sleeping dynamic? 
  4. How are the noise confidence, chrominance stability, and exposure-need priors constructed? How is the reliability map calibrated, and what supervision is used on datasets where reliability calibration is disabled? 
  5. Can the authors clarify the individual contributions of the low-frequency and high-frequency interaction paths in RGDFI? 
  6. How does the proposed spiking mechanism improve computational efficiency when its reported dense FLOPs exceed those of CIDNet? Are measured inference latency, spike activity, or energy consumption results available?
Questions for authors

Please see the weaknesses above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.