acceptodds
Under review as a conference paper at ICLR 2027

Lightweight but Global: Event-Guided Broad-Range Brightness Restoration

Abstract

Broad-range brightness restoration maps RGB frames captured under low, normal, or high illumination to a target exposure. Events provide complementary measurements when the frames are noisy or saturated. Balancing global spatial context, temporal information, and reconstruction detail within a compact model remains challenging. To this end, we present MambaSEE, a compact spatiotemporal architecture that assigns these functions to different operators. First, fixed-width additive fusion and attentive state-space modules model each frame–event pair with global context. Next, a weight-and-activation 8-bit quantization-aware ConvGRU aggregates three first-in, first-out pairs through backward-then-forward propagation, without observations later than the prediction time. Finally, a prompt-conditioned depthwise decoder reconstructs fine image details. On the SDE and SEE-600K benchmarks, MambaSEE achieves the highest reported average PSNR with the lowest FP32-equivalent weight storage among the compared methods. On SEE-600K, its PSNR values are 2.01, 1.39, and 5.32 dB higher than the published SEENet results for low-, high-, and normal-light inputs, respectively. MambaSEE uses mixed-precision weight storage equivalent to 0.48M FP32 parameters, about one quarter of SEENet's.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Event-based Vision, State Space Models, Mamba, Quantization-Aware Training, Efficient Architecture

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_4dp8OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes MambaSEE, a compact event-guided framework for broad-range brightness restoration. The architecture combines fixed-width image-event fusion, MambaIRv2-inspired global state-space blocks, a quantization-aware FIFO ConvGRU for short-term temporal aggregation, and a prompt-conditioned lightweight decoder. The model reports strong results on SDE and SEE-600K with a relatively small parameter/storage footprint. Overall, the paper is technically polished and substantially stronger than a routine architecture submission, but I am not yet convinced that the methodological novelty and the current experimental comparisons are sufficient for acceptance.

Critical strengths
  • The empirical results are strong. MambaSEE consistently improves over the reported baselines on both SDE and SEE-600K, including sizeable gains over SEENet, while maintaining a compact model. The evaluation also covers real event-camera datasets rather than relying on synthetic data alone.
  • The paper is generally well organized and technically detailed. The authors clearly document the input construction, temporal aggregation, quantization scheme, complexity accounting, and evaluation protocol, and the supplementary material is substantially more complete than what is typical for many architecture papers.

Overall, I appreciate that this submission does not read like a fully automated “auto-research” paper. The figures and overall presentation appear to have received genuine manual attention, which I value given the current trend toward heavily automated submissions. I am still not convinced that the paper reaches the acceptance bar in its current form, but I would be open to raising my score if the rebuttal addresses the main concerns convincingly and the paper remains competitive within my review batch.

Critical weaknesses
  • The technical novelty and task-specific motivation are not yet sufficiently convincing. The dominant spatial component, GSSB, closely follows the Attentive State-Space Module of MambaIRv2, while the remaining ingredients—additive fusion, ConvGRU, quantization-aware training, and depthwise decoding—are individually standard. The paper states that the novelty lies in their task-specific organization, but it does not clearly explain why the MambaIRv2-style semantic routing and state-space formulation are particularly suitable for event-guided brightness restoration, beyond providing generic global context. This is especially important because the largest ablation gain (4.54 dB) comes from replacing Swin with GSSB, whereas the proposed temporal module contributes only about 0.33 dB. The method section therefore reads more as a sequence of architectural design choices than as a mechanism driven by a clear insight into the task.

  • The experimental comparison is generally reasonable as an end-to-end benchmark evaluation, but several issues limit the interpretation of the claimed accuracy-efficiency advantage. First, recent directly relevant methods are missing, including BiEvLight (CVPR 2026) and EIC-LIE (CVPR 2026), both of which evaluate event-guided low-light enhancement on SDE. Second, MambaSEE uses three causal frame–event pairs, whereas some prior methods, such as SEENet, use a single frame–event pair. This is a legitimate architectural choice, but it makes it difficult to determine how much of the gain over prior work comes from the proposed architecture versus the additional temporal evidence. The efficiency comparison is also based on heterogeneous accounting: MambaSEE reports a cached steady-state MAC estimate and mixed-precision FP32-equivalent storage, while baseline FLOPs and parameter counts are quoted under their original FP32 settings, and the MambaSEE MAC estimate excludes routing and selective-scan operations. The accounting is clearly disclosed, but the current results support a promising accuracy-efficiency trade-off more strongly than a fully controlled efficiency superiority claim.

Questions for authors
  • Beyond inserting MambaIRv2-style GSSBs after image-event fusion, what task-specific modification is made to the state-space/routing mechanism itself? Can the authors explain what property of event-guided exposure restoration specifically motivates semantic token routing, rather than simply replacing Swin with a stronger generic restoration block?

  • Could the authors provide a more controlled comparison with MambaIRv2-style or other global restoration backbones under matched width/parameter budget? The current Swin-to-GSSB ablation shows a large gain, but does not clarify how much comes from the proposed task-specific design versus adopting a stronger pretrained/restoration-oriented spatial operator.

  • Why are BiEvLight (CVPR 2026) and EIC-LIE (CVPR 2026) absent from the SDE comparison? Both are recent event-guided LLIE methods evaluated on the same benchmark.

  • Could the authors clarify how much of the gain over SEENet is attributable to the architectural changes versus the additional temporal context? A matched comparison between SEENet and a K=1 MambaSEE variant, followed by K=1 versus K=3 within MambaSEE, would help disentangle these factors.

  • For SDE, how much performance remains if the oracle target-brightness prompt is replaced with a fixed or input-predicted value? Since the current prompt is computed from the ground-truth normal-exposure reference, this seems important when comparing against methods that do not receive equivalent target information.

  • For the “lightweight” claim, could the authors clarify the practical deployment advantage beyond FP32-equivalent storage? The current W8A8 result is based on fake quantization and theoretical bit-width accounting, while the reported MAC estimate excludes some state-space/routing operations. A unified runtime or memory comparison under the same hardware/backend would make this claim substantially stronger.

12 comments
AuthorsOpenReview ↗

W1 — Technical novelty and task-specific motivation

Thank you for asking us to clarify the task-specific contribution. MambaSEE combines global spatial modeling, bounded temporal evidence, and explicit target-brightness conditioning within a compact restoration framework; its contribution is not limited to the GSSB replacement. We attribute GSSB's underlying state-space/routing mechanism to MambaIRv2 [3], rather than claiming a new operator.

Table K. SDE accuracy, cost, and input settings (selected rows from manuscript Table 2).

MethodFLOPs (G)Eq. Storage (M)EventsTemporalBrightness PromptIndoor PSNR ↑Indoor SSIM ↑Outdoor PSNR ↑Outdoor SSIM ↑Average PSNR ↑Average SSIM ↑
SEENet405.721.90✓×✓22.540.775624.600.769223.570.7724
EvLight++225.9126.21✓✓×22.670.779123.340.767623.010.7733
BiEvLight61.592.471✓××22.870.775024.360.745123.610.7601
EIC-LIE70.952.13✓××23.330.757324.220.820023.780.7887
MambaSEE (Ours)45.19†0.48✓✓✓25.010.788326.400.779825.710.7841

The columns and values follow manuscript Table 2; only rows are omitted. Average equally weights Indoor and Outdoor. Temporal denotes multi-frame/recurrent processing, not matched history lengths; MambaSEE uses three causal pairs. Eq. Storage denotes theoretical FP32-equivalent weight storage. †MambaSEE counts one operation per MAC with cached steady-state feature reuse; accuracy uses full-window recomputation. Baseline FLOPs retain their source conventions.

  1. Brightness conditioning and long-context modeling. Our problem definition in Section 3, Eqs. (1)–(2) highlights the importance of the brightness prompt: it specifies the target exposure to be restored. We use Mamba's long-range spatial context to better model illumination across the image, rather than relying only on local cues.

Table 3. Progressive module-level ablation on SEE-600K (reproduced from the manuscript).

CaseSpatial BlockDecoderTemporalEq. Storage (M)Low Light PSNR ↑Low Light SSIM ↑Low Light L1 ↓High Light PSNR ↑High Light SSIM ↑High Light L1 ↓Normal Light PSNR ↑Normal Light SSIM ↑Normal Light L1 ↓
1SwinR1--0.9316.370.59710.128817.020.60440.111118.030.72120.1030
2GSSBR1--0.2119.350.63940.096920.400.67780.073325.290.85520.0391
3GSSBDWConv--0.4019.980.66500.088520.550.69520.071126.230.87360.0343
4GSSBDWConvFP32 GRU0.7320.690.69220.080620.670.69750.069726.400.87710.0334
5GSSBDWConvW8A8 GRU0.4820.780.69610.079420.600.69810.070126.240.87610.0338

Case numbers, configurations, storage values, and per-condition scores follow manuscript Table 3. R1 denotes the original SEENet decoder, DWConv the depthwise-convolutional decoder, and “--” disables temporal aggregation. Cases 4 and 5 use the same architecture with selected GRU layers in FP32 and W8A8, respectively. Bold marks the best score in each metric column.

  1. Why GSSB is important but not the whole explanation. Table 3 shows the largest gain from the spatial replacement (Cases 1–2), while the decoder (Cases 2–3) and temporal module (Cases 3–4) add 0.57 dB and 0.33 dB mean PSNR with GSSB retained. These means equally weight the three lighting conditions. The temporal gain reaches 0.71 dB under low light. Quantization serves a different purpose: reducing theoretical weight storage from 0.73M to 0.48M while retaining comparable reported accuracy.

  2. Simple components with a favorable complete-model trade-off. Table K, excerpted from manuscript Table 2, reports 25.71 dB for MambaSEE, exceeding SEENet and EIC-LIE by 2.14 and 1.93 dB, respectively, with 0.48M FP32-equivalent weight storage—about one quarter of SEENet's 1.90M. The design allocates capacity to global spatial reasoning while keeping fusion, history handling, and reconstruction compact, rather than uniformly shrinking a large fusion backbone. These are end-to-end accuracy–storage results; component attribution comes from the ablations, not the benchmark ranking alone.

Reference: manuscript Section 3 (Problem Formulation and spatial/temporal definitions), Tables 2–3, and Section 4.1.

AuthorsOpenReview ↗

W2 - Recent methods, temporal inputs, and efficiency

Thank you for highlighting these comparison and efficiency concerns. We distinguish broader benchmark coverage from controlled attribution and practical deployment measurements. These three issues now have separate responses:

  1. Recent methods. BiEvLight and EIC-LIE have been added to Related Work and the SDE comparison [6,7]. EvLight++ is also included as a published video-method comparison [2]. Q3 provides the added accuracy values and their provenance.

  2. Temporal evidence. The single-frame Case 3 and new window/direction controls show that multi-frame input is not required for all of the reported improvement. Q4 separates that observation from the still-unavailable unified SEENet rerun.

  3. Efficiency scope. The current evidence supports an accuracy–weight-storage trade-off. It does not establish a common-backend runtime advantage from heterogeneous FLOP/MAC counts or fake quantization. Q6 states that boundary explicitly.

Evidence: Tables 1–4 (pp. 7–10); Appendix A.3 and B.4–B.5.

AuthorsOpenReview ↗

Q1 - What changes in the state-space or routing mechanism?

Thank you for requesting a more precise description of the task-specific changes. Our task-specific adaptation lies in the input representations and the architectural organization surrounding the inherited state-space operator.

  1. What is inherited. The attentive state-space/routing mechanism follows MambaIRv2 [3], including the original-order prompt choice explained by its authors [4].

  2. What is task-specific. GSSBs receive fused features from separately adapted RGB and event inputs. Spatial processing operates within each observation, the FIFO module aggregates across observation times, and the decoder incorporates the target-brightness prompt.

  3. Why use global context? When local RGB evidence is unreliable, information from spatially separated related regions can be useful. Semantic routing is one way to provide that interaction. This motivates our choice, but it is not evidence that the operator is inherently better for events than every other global backbone.

The scope of novelty is therefore the compact restoration system and its division of roles, not a renamed or newly derived MambaIRv2 operator.

Evidence: Related Work and Method (pp. 3–6); Appendix B.2 (p. 19).

Reviewer_DW4UOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MambaSEE, a compact architecture for event-guided brightness restoration across low-, normal-, and high-light conditions. The model combines fixed-width additive image–event fusion, Global State-Space Blocks adapted from MambaIRv2, backward-then-forward ConvGRU propagation over a three-observation causal window, and an exposure-conditioned depthwise convolutional decoder. Selected temporal convolutions use W8A8 quantization-aware training. The model contains 0.73M trainable parameters, corresponding to approximately 0.48M FP32-equivalent weight storage under the stated mixed-precision accounting.

Critical strengths
  1. Restoring images across both underexposed and overexposed conditions is practically important. Exploiting complementary event measurements is well motivated, particularly when RGB observations suffer from noise or saturation.
  2. Separating spatial modeling, temporal aggregation, and reconstruction is a sensible design. Fixed-width fusion and depthwise reconstruction keep the model compact, while the bounded temporal window supports causal prediction without requiring observations beyond the output time.
  3. On SEE-600K, the reported PSNR improvements over SEENet are 2.01, 1.39, and 5.32 dB for low-, high-, and normal-light inputs, respectively. The qualitative examples also demonstrate substantial exposure correction and recovery of visible scene structure.
Critical weaknesses
  1. The introduction would benefit from a more rigorous and convincing motivation. The identified limitations of existing methods appear overly tailored to justify the proposed architecture, rather than being established through a balanced analysis of prior work. In particular, the discussion of these limitations focuses predominantly on a single method, SEENet, providing insufficient evidence that the stated bottlenecks are broadly representative of existing approaches. A broader discussion of relevant methods, together with a clearer explanation of their shared limitations and unresolved challenges, would strengthen the motivation and better establish the significance of the proposed contribution.
  2. Appendix B.3 states that MambaSEE is evaluated on the top-left (256x256) region of SDE images, whereas the original resolution is (260x346). This retains approximately 73% of the pixels and changes both the evaluated content and the reference-derived brightness prompt. Since baseline scores are taken from previous publications, the paper needs to establish that they use identical spatial regions, test samples, and scoring implementations.
  3. The contribution mainly combines established components, but the ablations do not match parameter or computation budgets. The FP32/W8A8 comparison explicitly does not isolate quantization, and the necessity of backward-then-forward propagation remains unverified.
  4. No retrained image-only counterpart is provided, making it difficult to separate event benefits from backbone improvements and additional RGB observations. Single-frame, repeated-current-frame, and forward-only temporal baselines would strengthen the analysis.
  5. Appendix B.2 states that token features are reordered while their prompts are not. The authors should clarify the resulting correspondence and explain how it supports the proposed semantic interpretation.
Questions for authors

Please refer to the concerns raised in the Weaknesses section.

8 comments
AuthorsOpenReview ↗

W1 - Broader and more rigorous task motivation / W2 - SDE crop, samples, and scoring protocol

W1 - Broader and more rigorous task motivation

Thank you for highlighting the need for a more balanced analysis of prior work. We have focused the revision on the Introduction's prior-work paragraph and its transition to our design goal, while retaining the original organization and main narrative.

  1. Broader prior-work context. The paragraph now acknowledges signal-to-noise-guided fusion in EvLight [5] and existing temporal modeling in EvLowLight and EvLight++ [2], before discussing SEENet's broad-range brightness adjustment [1]. This avoids using a single method's configuration to characterize the entire field. More detailed coverage, including BiEvLight and EIC-LIE, remains in the Event-Guided Methods discussion in Related Work.

  2. A more precise design goal. The revised transition identifies our focus as retaining global spatial context and a bounded history within a compact, target-exposure-conditioned model. We motivate the design through the trade-off between contextual information and model capacity/computation, rather than presenting SEENet's attention cost or frame-wise processing as limitations shared by all prior methods.

Reference: Introduction, primarily the third paragraph and its transition to the method overview. Supporting prior-work coverage: Section 2, Event-Guided Methods.

W2 - SDE crop, samples, and scoring protocol

Thank you for raising this concern about evaluation consistency. The 256×256 crop was chosen to follow earlier crop-based evaluation practice and accommodate resolution-constrained comparison implementations, rather than because MambaSEE requires a fixed input size. For example, the released SEE implementation [1] of Liu et al.'s Low-Light Video Enhancement with Synthetic Event Guidance explicitly requires both input height and width to be multiples of 4. Since 346 is not divisible by 4, native 260×346 frames cannot be processed directly by this unmodified implementation. Its normalization and linear layers are also tied to configured spatial dimensions.

Table F. SDE resolution comparison (manuscript Table 2 baseline and additional tests).

Resolution (H×W)Indoor PSNR / SSIMOutdoor PSNR / SSIMEqual-subset average
256×25625.01 / 0.788326.40 / 0.779825.71 / 0.7841
260×34624.99 / 0.797026.51 / 0.782525.75 / 0.7897

The comparison of these two experimental results demonstrates that the performance difference at the 260×346 resolution is negligible, and it still achieves state-of-the-art (SOTA) accuracy.

AuthorsOpenReview ↗

W3 - Matched budgets, temporal design, and quantization

Thank you for highlighting the need to justify these design choices with quantitative evidence. The ablations provide separate evidence for spatial replacement, temporal propagation, and GRU quantization, but they do not establish equal-compute superiority for every component. We clarify the scope of each comparison.

  1. Complete-model accuracy and weight storage. Our main result is the accuracy–storage trade-off of the complete MambaSEE model, not the size of an isolated spatial block. MambaSEE achieves 22.54 dB PSNR / 0.7568 SSIM with 0.725986M trainable parameters. Under the stated mixed-precision accounting, its theoretical weight storage is 0.484018M FP32-equivalent weights, a 33.33% reduction relative to an all-FP32 representation of the same parameters. The spatial variants serve as intermediate ablation references:

Table G. Complete-model accuracy and weight storage (selected entries from main-text Table 3 and appendix Table 8, with derived accuracy averages).

ConfigurationFrame–event pairsParameters (M)FP32-equivalent weights (M)Average PSNR / SSIM
MambaSEE, complete model30.7259860.48401822.54 / 0.7568
Swin + R1 decoder, Case 110.9280640.92806417.14 / 0.6409
GSSB + R1 decoder, Case 210.2117750.21177521.68 / 0.7241

The Swin-to-GSSB comparison provides local evidence for the spatial replacement: at the same 64-channel width and with the same R1 decoder, average PSNR improves by 4.54 dB with 77.18% fewer parameters. The complete MambaSEE model additionally includes the DWConv decoder and temporal aggregation, achieving higher reported accuracy while retaining a smaller parameter count and theoretical weight storage than the Swin-based reference. We distinguish this complete-model trade-off from the local spatial ablation; neither is presented as an equal-computation comparison.

Accuracy values are retained from the manuscript; the MambaSEE parameter/storage counts refer to the current implementation. FP32-equivalent weights describe theoretical weight storage, not the trainable parameter count, serialized model size, or runtime GPU memory.

  1. Forward-only versus bidirectional propagation. We compare the forward-only variant with MambaSEE using the same K=3 observation window. Each cell reports PSNR / SSIM / L1:

Table H. Propagation-direction comparison on SEE-600K (Cases 2 and 4 excerpted from manuscript Table 4).

Temporal configurationLow lightHigh lightNormal light
Forward-only, K=320.49 / 0.6881 / 0.083120.49 / 0.6880 / 0.073226.34 / 0.8741 / 0.0337
MambaSEE, K3 bidirectional20.78 / 0.6961 / 0.079420.60 / 0.6981 / 0.070126.24 / 0.8761 / 0.0338

MambaSEE raises average PSNR from 22.44 to 22.54 dB (+0.10 dB) and average SSIM from 0.7501 to 0.7568 (+0.0067), while lowering average L1 from 0.0633 to 0.0611. Low- and high-light PSNR improve by 0.29 and 0.11 dB, respectively, whereas normal-light PSNR decreases by 0.10 dB. These results support a modest overall accuracy advantage for the bidirectional configuration, especially under low light, rather than a uniform gain in every condition. The comparison fixes the observation count, not the parameter or computation budget.

Result sources: MambaSEE from Table 3, Case 5; forward-only from Table 4. Averages equally weight the three lighting conditions.

  1. Quantization. In the original ablation, Case 4 and Case 5 have the same model architecture and differ in the quantization settings of selected GRU layers. Their averages are 22.59/22.54 dB PSNR and 0.7556/0.7568 SSIM, respectively. Thus, the W8A8 configuration retains comparable reported accuracy, rather than being the source of the large spatial-block gain.

An additional parameter-/compute-matched alternative backbone has not been evaluated. The existing controls should therefore be interpreted at their stated scopes, rather than as a factorial demonstration of component synergy.

Evidence: manuscript Table 3 (p. 8), Table 4 (p. 10), and Section 4.1 (pp. 8–10); weight-storage accounting in manuscript Table 8 (p. 21).

AuthorsOpenReview ↗

W4 - Retrained no-event and temporal-input controls

Thank you for suggesting these controls. The zero-event control shows that our network remains effective without event measurements, while the single-frame, repeated-observation, and forward-only baselines support additional gains from our temporal design. The zero-event and repeated-observation models were retrained and evaluated with their respective input modifications, rather than changing the inputs only at inference.

Table I. Retrained event and temporal-input controls on SEE-600K (selected rows of manuscript Table 4, with equal-condition averages).

ConfigurationLow: PSNR / SSIM / L1High: PSNR / SSIM / L1Normal: PSNR / SSIM / L1Average: PSNR / SSIM / L1
MambaSEE, K3 bidirectional20.78 / 0.6961 / 0.079420.60 / 0.6981 / 0.070126.24 / 0.8761 / 0.033822.54 / 0.7568 / 0.0611
Zero events, retrained K319.86 / 0.6443 / 0.091320.71 / 0.6973 / 0.069126.24 / 0.8740 / 0.034122.27 / 0.7385 / 0.0648
Single-frame, K120.43 / 0.6778 / 0.082420.61 / 0.6955 / 0.070426.20 / 0.8738 / 0.034322.41 / 0.7490 / 0.0624
Repeated current RGB + event observation, K320.37 / 0.6757 / 0.083320.56 / 0.6949 / 0.070626.22 / 0.8728 / 0.034122.38 / 0.7478 / 0.0627
Forward-only, K320.49 / 0.6881 / 0.083120.49 / 0.6880 / 0.073226.34 / 0.8741 / 0.033722.44 / 0.7501 / 0.0633
  1. Event information and network capability. Setting the event input to zero reduces mean PSNR from 22.54 to 22.27 dB (−0.27 dB), but the network retains substantial restoration quality, including 26.24 dB under normal light. This indicates that the network can effectively restore RGB inputs without event measurements. Events nevertheless provide valuable additional information: under low light, MambaSEE improves PSNR from 19.86 to 20.78 dB (+0.92 dB) and SSIM from 0.6443 to 0.6961 (+0.0518). Thus, the network's restoration capability is not solely dependent on events, while event information further improves accuracy, particularly under low light.

  2. Temporal aggregation. The single-frame, repeated-current-observation, and forward-only baselines examine complementary aspects of the temporal design. K1 retains the temporal module but receives only one observation, unlike Case 3, which removes the module. The repeated-observation control retains K3 but feeds the same current RGB frame and its K1-window event representation into all three slots; unlike RGB-only repetition, both modalities are repeated. The forward-only baseline uses three distinct observations but omits backward propagation. Compared with these three baselines, MambaSEE improves mean PSNR by 0.13, 0.16, and 0.10 dB, respectively, and low-light PSNR by 0.35, 0.41, and 0.29 dB. Together, these results support the benefit of aggregating distinct historical observations and using bidirectional temporal propagation, with higher average accuracy and the strongest gains under low light.

The zero-event model retains the network and brightness-conditioning pathway; it is not a separately optimized image-only network with the event branch removed. Averages equally weight the three lighting conditions.

Reference: manuscript Table 4 (p. 10), excerpted in Table I.

Reviewer_WagwOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MambaSEE, a framework designed for event-guided brightness restoration. The architecture combines image–event fusion, global state-space blocks, and a depthwise convolutional decoder. The paper reports improvements on SEE-600K and SDE datasets.

Critical strengths
  1. The framework is well motivated by the specific task. Assigning global spatial reasoning, short-term temporal aggregation, and local reconstruction to separate modules is reasonable.

  2. The reported experimental gains are substantial. On SEE-600K, average PSNR increases from 19.63 dB for SEENet to 22.54 dB. On SDE, the reported average PSNR improvement over SEENet is 2.14 dB.

Critical weaknesses
  1. Insufficient comparisons: Historical inputs, brightness conditioning are not consistently matched across methods, making architectural improvements difficult to isolate.

  2. Limited ablations: The largest improvement comes from a state-space operator. Progressive ablations do not verify whether the gains are from temporal design, quantization, or the proposed combination.

Questions for authors
  1. Can SEENet and a strong video baseline be rerun using matched histories, brightness conditioning, image regions, and evaluation scripts?

  2. Can you add analysis about quantization and temporal design, forward-only/bidirectional, and different-window-length experiments?

  3. How are state prompts aligned with reordered tokens? Please provide an index-level explanation and a consistently permuted-prompt comparison.

9 comments
AuthorsOpenReview ↗

W1 - Historical inputs and brightness conditioning

We thank you for this important comment. We have added two columns, Temporal and Brightness Prompt, to Table 2 to clarify the input settings of all compared methods, as summarized below.

Table B. Input-setting summary (selected columns of manuscript Table 2, with methods grouped by configuration).

MethodsTemporalBrightness Prompt
SNR, UFormer, LLFlow, Retinexformer, ELIE, Wang et al., eSL-Net, Liu et al., EvLight, BiEvLight, EIC-LIE✗✗
E2VID+, HDRev, EvLowLight, EvLight++✓✗
SEENet✗✓
MambaSEE (Ours)✓✓
  1. Why this setting? The setting follows the target-conditioned restoration problem in Section 3, Eqs. (1)–(2). Eq. (1) defines Ipred=fsee({Ik,Ek}k=i−K+1i,p)I_{\mathrm{pred}}=f_{see}(\{I_k,E_k\}_{k=i-K+1}^{i},p): the model reconstructs the normally exposed target at time ii from K=3K=3 causal RGB–event pairs and a target-brightness prompt. The history ends at ii, with no access to future observations, and allows preceding observations to contribute restoration evidence. Following the SEE task protocol [1], Eq. (2) defines pp as the global RGB mean of the target reference. This scalar specifies the desired output exposure but supplies no spatial structure. Thus, historical observations provide reconstruction evidence, whereas pp specifies the requested exposure; our evaluation addresses restoration to a known target exposure, not automatic exposure selection. Table B explicitly discloses these input conditions rather than assuming that they are matched across methods.

Table C. Architectural and input-ablation summary on SEE-600K (derived from selected results in manuscript Tables 3 and 4). Average PSNR is computed across low-, high-, and normal-light conditions. Detailed configurations and per-condition PSNR/SSIM/L1 results are provided in Tables 3 and 4 of the revised manuscript.

AblationComparisonAvg. PSNR (dB)Improvement
Spatial ModelingSwin → GSSB17.14 → 21.68+4.54 dB
Lightweight DecoderR1 → DWConv21.68 → 22.25+0.57 dB
Temporal AggregationNo GRU → FP32 GRU22.25 → 22.59+0.33 dB
Temporal WindowK1 → K322.41 → 22.54+0.13 dB
Historical ObservationsRepeated Current → Original K322.38 → 22.54+0.16 dB
  1. Comparison fairness and ablations. Table 2 presents an end-to-end comparison of complete methods under their respective settings, rather than a strictly matched-input comparison. To further distinguish architectural improvements from additional temporal information, we conduct extensive ablation studies, summarized in Table C. Importantly, Table 3 (Cases 1–3) uses identical single-frame inputs and brightness conditioning, demonstrating that the major architectural improvements are obtained without additional historical observations. Table 4 further examines the contribution of temporal context through window-length and repeated-observation experiments. These results provide complementary evidence for the architectural design and the proposed input setting.
AuthorsOpenReview ↗

W2 - Spatial, temporal, and quantization contributions

Thank you for raising this distinction. The state-space block provides the largest gain, but not the entire gain: the decoder and temporal module further improve restoration, while quantization targets lower weight storage. Table D presents all results from manuscript Table 3, retaining its configurations, storage column, and separate results for each lighting condition, with additional differences from Case 1.

Table D. Progressive module-level ablation with Case 1 differences (manuscript Table 3 augmented for this rebuttal).

CaseSpatial BlockDecoderTemporalEq. Storage (M)Low Light PSNR ↑Low Light SSIM ↑Low Light L1 ↓High Light PSNR ↑High Light SSIM ↑High Light L1 ↓Normal Light PSNR ↑Normal Light SSIM ↑Normal Light L1 ↓
1SwinR1--0.9316.370.59710.128817.020.60440.111118.030.72120.1030
2GSSBR1--0.2119.35 (+2.98)0.6394 (+0.0423)0.0969 (-0.0319)20.40 (+3.38)0.6778 (+0.0734)0.0733 (-0.0378)25.29 (+7.26)0.8552 (+0.1340)0.0391 (-0.0639)
3GSSBDWConv--0.4019.98 (+3.61)0.6650 (+0.0679)0.0885 (-0.0403)20.55 (+3.53)0.6952 (+0.0908)0.0711 (-0.0400)26.23 (+8.20)0.8736 (+0.1524)0.0343 (-0.0687)
4GSSBDWConvFP32 GRU0.7320.69 (+4.32)0.6922 (+0.0951)0.0806 (-0.0482)20.67 (+3.65)0.6975 (+0.0931)0.0697 (-0.0414)26.40 (+8.37)0.8771 (+0.1559)0.0334 (-0.0696)
5GSSBDWConvW8A8 GRU0.4820.78 (+4.41)0.6961 (+0.0990)0.0794 (-0.0494)20.60 (+3.58)0.6981 (+0.0937)0.0701 (-0.0410)26.24 (+8.21)0.8761 (+0.1549)0.0338 (-0.0692)

Parentheses in Cases 2–5 give the signed difference from Case 1 in the same column, not a percentage: PSNR differences are in dB, and SSIM/L1 differences retain their original units. Negative L1 differences indicate lower error. R1 is the original SEENet decoder, DWConv is the depthwise-convolutional decoder, and “--” disables temporal aggregation. Eq. Storage denotes theoretical FP32-equivalent weight storage. Bold marks the best value in each metric column.

  1. Spatial modeling and local reconstruction (Cases 1 → 2 → 3). With R1 retained, replacing Swin with GSSB raises PSNR from 16.37 to 19.35 dB under low light, 17.02 to 20.40 dB under high light, and 18.03 to 25.29 dB under normal light. SSIM increases and L1 decreases in every condition, while Eq. Storage falls from 0.93M to 0.21M. Replacing R1 with DWConv then improves all nine quality entries: for example, low-light PSNR/SSIM/L1 changes from 19.35 / 0.6394 / 0.0969 to 19.98 / 0.6650 / 0.0885, with Eq. Storage increasing from 0.21M to 0.40M.

  2. Temporal aggregation (Cases 3 → 4). Adding the FP32 FIFO module improves all three metrics in each lighting condition. The largest PSNR gain occurs under low light, where PSNR/SSIM/L1 changes from 19.98 / 0.6650 / 0.0885 to 20.69 / 0.6922 / 0.0806. High- and normal-light PSNR also increase from 20.55 to 20.67 dB and 26.23 to 26.40 dB, respectively. These improvements accompany an increase in Eq. Storage from 0.40M to 0.73M. Q2 provides the window-length and propagation-direction controls.

  3. Quantization and the final combination (Cases 4 → 5). These cases use the same architecture, with the selected GRU layers in FP32 and W8A8, respectively. Eq. Storage decreases from 0.73M to 0.48M, with mixed accuracy changes: low-light PSNR/SSIM/L1 improves from 20.69 / 0.6922 / 0.0806 to 20.78 / 0.6961 / 0.0794, whereas high- and normal-light PSNR decrease slightly from 20.67 to 20.60 dB and 26.40 to 26.24 dB. Quantization therefore provides a storage trade-off rather than a uniform accuracy gain. The final Case 5 still has higher PSNR/SSIM and lower L1 than the no-temporal Case 3 in all three lighting conditions.

These results support distinct restoration and storage benefits, rather than attributing the accuracy gain to quantization or establishing component synergy through a factorial experiment.

Reference: Table 3 (p. 8); Section 4.1 (pp. 8–10); Appendix B.5, Table 8 (p. 21).

AuthorsOpenReview ↗

Q1 - SEENet and a strong video baseline

Thank you for suggesting a controlled comparison with SEENet and a strong video baseline. For a fair and consistent benchmark comparison, we retain each method's reported configuration and explicitly document its input settings in Table 2. The Events, Temporal, and Brightness Prompt columns identify the use of event observations, multi-frame/recurrent aggregation, and explicit target-brightness conditioning for every method. We apply the same reporting rule across the comparison while keeping these method-specific settings visible.

Table 2 includes both SEENet [1] and the event-guided video method EvLight++ [2]. SEENet is also included in the SEE-600K comparison in Table 1. Their SDE results are reproduced below:

Table E. SDE comparison with SEENet and EvLight++ (selected rows and accuracy columns from manuscript Table 2).

SDE methodIndoor PSNR / SSIMOutdoor PSNR / SSIMAverage PSNR / SSIM
SEENet [1]22.54 / 0.775624.60 / 0.769223.57 / 0.7724
EvLight++ [2], video + events22.67 / 0.779123.34 / 0.767623.01 / 0.7733
MambaSEE25.01 / 0.788326.40 / 0.779825.71 / 0.7841
  1. Comparison scope and evidence. These baseline scores are quoted from the published reports; we have not performed a new rerun that matches histories, brightness conditioning, image regions, and evaluation scripts across methods. Table 2 documents the input categories, while Appendix A.3 explains the numerical sources and protocol boundaries. Thus, consistency refers to the comparison and reporting policy, not identical information budgets. For component-level evidence, W1 discusses the module ablations in Table 3, and Q2 examines the temporal-input controls in Table 4.

Revision: Table 2 (p. 7), Section 4 (Compared Methods and Protocol), and Appendix A.3 (pp. 17–18).

Comments on the paper

4 comments
AuthorsOpenReview ↗

Acknowledgment of Reviewers' Positive Feedback

We sincerely thank all reviewers for their constructive feedback and encouraging assessments. We appreciate their recognition of task-specific motivation, modular architecture, strong empirical performance, real-world evaluation, and technical presentation. These positive assessments highlight both the practical significance and the effectiveness of our approach. We summarize the reviewers' original comments in Table A.

Table A. Summary of positive feedback from all three reviewers.

No.Recognized StrengthOriginal Reviewer Comments (Verbatim)
1Task-specific Motivation & Practical RelevanceWagw.S1: “The framework is well motivated by the specific task.” <br>DW4U.S1: “Restoring images across both underexposed and overexposed conditions is practically important. Exploiting complementary event measurements is well motivated, particularly when RGB observations suffer from noise or saturation.”
2Modular Architecture & Efficient DesignWagw.S1: “Assigning global spatial reasoning, short-term temporal aggregation, and local reconstruction to separate modules is reasonable.” <br>DW4U.S2: “Separating spatial modeling, temporal aggregation, and reconstruction is a sensible design. Fixed-width fusion and depthwise reconstruction keep the model compact, while the bounded temporal window supports causal prediction without requiring observations beyond the output time.”
3Strong Quantitative ImprovementsWagw.S2: “The reported experimental gains are substantial. On SEE-600K, average PSNR increases from 19.63 dB for SEENet to 22.54 dB. On SDE, the reported average PSNR improvement over SEENet is 2.14 dB.” <br>DW4U.S3: “On SEE-600K, the reported PSNR improvements over SEENet are 2.01, 1.39, and 5.32 dB for low-, high-, and normal-light inputs, respectively.” <br>4dp8.S1: “The empirical results are strong. MambaSEE consistently improves over the reported baselines on both SDE and SEE-600K, including sizeable gains over SEENet, while maintaining a compact model.”
4Qualitative Restoration QualityDW4U.S3: “The qualitative examples also demonstrate substantial exposure correction and recovery of visible scene structure.”
5Real-World Event-Camera Evaluation4dp8.S1: “The evaluation also covers real event-camera datasets rather than relying on synthetic data alone.”
6Technical Details & Documentation4dp8.S2: “The paper is generally well organized and technically detailed. The authors clearly document the input construction, temporal aggregation, quantization scheme, complexity accounting, and evaluation protocol, and the supplementary material is substantially more complete than what is typical for many architecture papers.”
7Careful Presentation4dp8.S3: “Overall, I appreciate that this submission does not read like a fully automated “auto-research” paper. The figures and overall presentation appear to have received genuine manual attention, which I value given the current trend toward heavily automated submissions.”
AuthorsOpenReview ↗

Summary of Manuscript Revisions

Thank you for the constructive comments and valuable suggestions. We have carefully revised the manuscript to strengthen the motivation, expand experimental validation, clarify evaluation protocols, and improve technical presentation. All major revisions are highlighted in blue in the updated PDF and summarized in this Table.

Table. Summary of manuscript revisions and corresponding reviewer concerns.

No.Revisions in the ManuscriptCorresponding Reviewer Concerns
1Motivation and Related Work: Expanded the discussion of existing event-guided and temporal methods and clarified the task-specific motivation (Introduction and Section 2, pp. 2–3).DW4U.W1; 4dp8.W1, Q1
2Expanded SOTA Comparisons: Added BiEvLight and EIC-LIE, with their reported performance and numerical sources (Table 2; Appendix A.3, Table 6, pp. 7, 17–18).Wagw.Q1; 4dp8.W2, Q3
3Comparison Settings and Protocols: Added Temporal and Brightness Prompt columns and clarified method-specific inputs and evaluation conditions (Table 2; Appendix A.3, Table 6, pp. 7, 17–18).Wagw.W1, Q1; DW4U.W2; 4dp8.W2, Q4
4Architectural Ablations: Clarified the contributions of spatial modeling, lightweight decoding, temporal aggregation, and quantization (Section 4.1, Table 3, Figure 4, pp. 8–9).Wagw.W2, Q2; DW4U.W3; 4dp8.W1, Q2, Q4
5Temporal and Event Ablations: Added K1/K3/K5, forward-only, zero-event, and repeated-observation experiments, including qualitative comparisons (Table 4, Figure 5, p. 10; Appendix B.5, p. 20).Wagw.W1, W2, Q2; DW4U.W3, W4; 4dp8.W2, Q4
6SDE Resolution Evaluation: Added full-resolution (260×346) evaluation to investigate the influence of spatial cropping (Appendix B.3, pp. 19–20).DW4U.W2
7State-Prompt Correspondence: Clarified token–prompt indexing and added synchronized-prompt experiments (Appendix B.2, Eq. 15, p. 19; Table 7, p. 20).Wagw.Q3; DW4U.W5
8Parameter and Quantization Analysis: Clarified trainable parameters, FP32-equivalent storage, and FP32/W8A8 comparisons (Tables 3 and 8; Appendices B.4–B.5, pp. 9, 20–21).Wagw.W2, Q2; DW4U.W3; 4dp8.W2, Q6
9Runtime and Memory Evaluation: Added RTX 5090 inference-time and GPU-memory measurements, distinguishing fake quantization from native INT8 execution (Appendix B.6, Table 9, p. 21).4dp8.W2, Q6

The detailed explanations and experimental evidence are provided in our point-by-point responses to each reviewer.

Dear Reviewers,

As the discussion phase is coming to an end, could you please take a moment to review the authors' responses and finalize your ratings?

If you have any remaining concerns or questions, please feel free to raise them with the authors before the discussion closes.

Thank you very much for your time and support!

Best regards,

Xiaojuan

AuthorsOpenReview ↗

Discussion Summary and Sincere Appreciation

Dear Area Chair, Program Chairs, and Reviewers,

We would like to express our sincerest gratitude to the Area Chair, Program Chairs, and all three reviewers for your valuable time, careful evaluation, constructive feedback, and thoughtful consideration throughout the review and discussion process.

We are especially grateful that our rebuttal has satisfactorily addressed the reviewers' main concerns through additional comparisons, extensive ablation studies, clarifications of evaluation protocols and technical designs, and further runtime and memory analyses. We are truly encouraged that all three reviewers acknowledged our efforts and kindly raised their ratings.

Your insightful comments and constructive suggestions have been invaluable in helping us strengthen our experimental validation, clarify our contributions, and substantially improve the quality of our paper. We have carefully incorporated the corresponding revisions into the updated manuscript and will ensure that all additional results and clarifications are reflected in the final version.

We are also committed to publicly releasing our code and trained models to support reproducibility and benefit the broader research community.

Once again, we sincerely thank the Area Chair, Program Chairs, and all reviewers for your dedication, encouragement, and invaluable contributions to improving our work. We deeply appreciate your time and consideration.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.