acceptodds
Under review as a conference paper at ICLR 2027

Decouple to Reconcile: Explicit Inter- and Intra-Exposure Motion Modeling for Night Image Restoration

Abstract

Photon scarcity in low-light environments makes high-quality night image restoration from a single exposure inherently challenging. Dual-exposure imaging offers an effective solution by exploiting the complementary characteristics of short- and long-exposure images. However, motion-induced inter-frame misalignment and intra-exposure blur hinder their effective collaborative restoration. To this end, we propose an explicit inter- and intra-exposure motion decoupling framework that separately models cross-exposure displacement and within-exposure motion blur, enabling accurate alignment and collaborative restoration of complementary dual-exposure information. Specifically, I2EM follows a two-stage decouple-to-reconcile paradigm. The decoupling stage separately models cross-exposure displacement and intra-exposure blur using bidirectional optical flow and exposure-specific motion-conditioned kernels. The reconciliation stage restores and aligns long-exposure features against the sharper short-exposure reference, then integrates complementary information through confidence-guided multi-scale fusion and residual denoising. On D2, I2EM achieves state-of-the-art performance among the compared methods, demonstrating the effectiveness of explicit inter- and intra-exposure motion modeling for dual-exposure night-image restoration.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dual-exposure night image restoration, low-light imaging, motion–blur decoupling, bidirectional optical flow, spatially varying blur, confidence-aware fusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aU3mOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes I²EM, a dual-exposure night image restoration framework that explicitly separates inter-exposure displacement and intra-exposure motion blur using optical flow and spatially varying kernels. The method achieves strong quantitative results on the D2 benchmark. However, the methodological novelty appears relatively limited compared with existing multi-exposure restoration approaches, and the effectiveness of the proposed framework on real captured dual-exposure images requires more convincing validation.

Critical strengths
  1. Well-motivated design. Explicitly distinguishing inter-exposure displacement from intra-exposure blur provides a reasonable approach to improving cross-exposure alignment and information fusion.

  2. Strong quantitative performance. The method achieves 40.78 dB PSNR on the D2 benchmark, outperforming the directly comparable D2HNet baseline by 2.83 dB.

  3. Extensive ablation studies. The paper evaluates several key components, including bidirectional flow estimation, kernel modeling, reconstruction constraints, and confidence-aware fusion.

Critical weaknesses
  1. Insufficient dual-exposure comparisons. Although Table 1 includes numerous baselines, D2HNet is the only directly comparable method under the same dual-exposure setting. Additional comparisons with relevant methods, such as LSD₂, LSFNet, and DeLECA, would strengthen the evaluation.

  2. Unclear baseline training protocols. The paper does not clearly specify whether the compared methods were retrained on the D2 dataset or evaluated using pretrained weights. The authors should clarify the training settings, checkpoint sources, and input adaptations to ensure fair comparisons.

Questions for authors

Please refer to the weakness.

3 comments

Q1:Insufficient dual-exposure comparisons. Although Table 1 includes numerous baselines, D2HNet is the only directly comparable method under the same dual-exposure setting. Additional comparisons with relevant methods, such as LSD₂, LSFNet, and DeLECA, would strengthen the evaluation.

In section 4.2(manuscript lines 342–378), we have compared with D2HNet on the same pair and have also evaluated methods that use multiple observations, such as CDVD-TSP and ESTRNN, on D2. To extend the direct comparison, the revised Table 2 evaluates LSD₂, LSFNet-RGB, and DeLECA using the same long–short pair as I²EM (manuscript lines 363–377). We use the released checkpoint for LSD₂ and adapt and train the other two implementations on the official D2 training split (Section 4.1, manuscript lines 331–336).

Under the shared protocol, these three methods reach 38.42, 38.15, and 39.15 dB, respectively, while I²EM reaches 40.78 dB. The comparison therefore shows that the PSNR advantage persists across the additional methods with the same exposure inputs.

Table 2. Dual-exposure night-image restoration comparison on D2-2880p with a shared long–short input pair

MethodWeights/trainingPSNRSSIMLPIPSParams (M)FLOPs (G)Time (ms)
CDVD-TSPD2-adapted training37.000.96680.08716.688.523
ESTRNND2-adapted training37.100.96750.08820.578.525
LSD₂Released38.420.97410.07128.5142.818
HINetD2-adapted training37.220.96850.08388.7112.36
D2HNetReleased37.950.97230.07635.2165.312
LSFNet-RGBD2-adapted training38.150.97320.07332.5158.422
SelfHDRReleased38.850.97520.065118.2425.838
DeLECAD2-adapted training39.150.97580.06345.8198.525
BracketIREReleased39.250.97610.061128.5498.245
STGDNetReleased38.120.97290.07228.4142.615
I²EMD2 training40.780.97830.054133.79462.878

Q2:Unclear baseline training protocols. The paper does not clearly specify whether the compared methods were retrained on the D2 dataset or evaluated using pretrained weights. The authors should clarify the training settings, checkpoint sources, and input adaptations to ensure fair comparisons.

In the manuscript, we have specified the common D2-2880p scenes, full-resolution inference, and metric implementation. In the revision, Section 4.1 clarifies that we use released weights where an applicable checkpoint is available and otherwise adapt the implementation and train it on the official D2 split (manuscript lines 331–340).

The “Weights/training” columns in Tables 1 and 2 identify the protocol for each baseline (manuscript lines 345–377). DarkIR, NoiseDiff, CDVD-TSP, ESTRNN, HINet, LSFNet-RGB, and DeLECA use D2-adapted training. The other listed baselines use released checkpoints. The tables also distinguish long-only and short-only input from the common long–short pair used for the ordinary dual-exposure methods. Appendix E.5 specifies the crop setting, number of epochs, optimizers, and loss for the two capacity controls (manuscript lines 1410–1419).

Table 1. Single-exposure night-image restoration comparison on D2-2880p

MethodInputWeights/trainingPSNRSSIMLPIPSParams (M)FLOPs (G)
DeepDeblurL-onlyReleased36.550.96360.09542.3156.8
DeblurGAN-v2L-onlyReleased36.820.96550.09060.9182.3
MPRNetL-onlyReleased37.080.96840.08623.598.2
SwinIRL-onlyReleased37.700.97120.07912.162.3
MIMO-UNet+L-onlyReleased37.310.96900.08316.142.1
MPRNet-DNS-onlyReleased37.350.95800.08515.7492.3
NAFNetL-onlyReleased37.700.97100.07923.492.4
RestormerL-onlyReleased37.870.97170.07726.1105.7
UformerL-onlyReleased37.550.97020.08128.3118.4
HATL-onlyReleased37.850.97150.07812.498.3
PromptIRL-onlyReleased37.450.96950.08287.2138.6
UFPNetL-onlyReleased37.720.97120.07980.3290.4
EVSSML-onlyReleased38.090.97270.07217.1126.0
DeblurDiffL-onlyReleased37.550.97030.050386.22890.0
DarkIRS-onlyD2-adapted training37.650.97050.08035.2128.5
NSPS-onlyReleased37.250.96850.08528.5115.2
NoiseDiffS-onlyD2-adapted training37.950.97180.075120.5285.3
I²EMLong–short pairD2 training40.780.97830.054133.79462.8

Table 2. Dual-exposure night-image restoration comparison on D2-2880p with a shared long–short input pair

MethodWeights/trainingPSNRSSIMLPIPSParams (M)FLOPs (G)Time (ms)
CDVD-TSPD2-adapted training37.000.96680.08716.688.523
ESTRNND2-adapted training37.100.96750.08820.578.525
LSD₂Released38.420.97410.07128.5142.818
HINetD2-adapted training37.220.96850.08388.7112.36
D2HNetReleased37.950.97230.07635.2165.312
LSFNet-RGBD2-adapted training38.150.97320.07332.5158.422
SelfHDRReleased38.850.97520.065118.2425.838
DeLECAD2-adapted training39.150.97580.06345.8198.525
BracketIREReleased39.250.97610.061128.5498.245
STGDNetReleased38.120.97290.07228.4142.615
I²EMD2 training40.780.97830.054133.79462.878

I maintain my weak reject decision, as I believe the experimental comparisons are unfair.

Reviewer_UqcYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes I2EM for dual-exposure nighttime image restoration, which models the displacement between different exposures and the motion blur occurring within each exposure separately. In the decoupling stage, bidirectional optical flow and exposure-specific motion kernels are used to estimate these two types of motion. In the reconciliation stage, long-exposure features are restored and aligned to the short-exposure coordinate system; the final image is then generated through confidence-guided fusion and residual denoising. The method achieves higher PSNR and SSIM scores than existing methods on the D2 benchmark.

Critical strengths

None

Critical weaknesses
  • The practical motivation for the dual-exposure capture setup is insufficient. The proposed method assumes a scenario where short-exposure and long-exposure images are captured sequentially with the same camera. However, this capture method requires changing exposure settings and additional capture time, potentially leading to motion and occlusion issues due to the time gap between the two shots. It is unclear what substantial advantages this complex capture setup offers compared to single-exposure restoration or short-exposure burst capture. In particular, as the primary quantitative evaluation focuses on the D2 benchmark, it is difficult to assess the method's practical applicability.

  • The computational cost of the proposed method is excessive, and the necessity of its complex structure has not been sufficiently validated. I2EM requires 133.79M parameters and an inference time of 78ms, which is significantly higher than D2HNet's 35.2M parameters and 12ms. However, it remains unverified whether this performance gap persists when the model capacity of conventional methods is increased or when they are extended to a similar dual-branch structure. Consequently, it is difficult to distinguish whether the reported performance gains stem from the structural advantages of the proposed motion modeling or simply from the significantly larger model capacity.

Questions for authors
  • A capacity-matched comparison with existing methods is required. Please conduct a comparison under conditions where the number of parameters and FLOPs are equivalent to those of I2EM. For instance, increasing the model capacity of D2HNet or by extending existing restoration methods (such as Restormer) to a dual-branch structure that takes short and long exposures as input should be included. This is necessary to disentangle the effects of the proposed architecture from the effects of increased model capacity.

  • There is a lack of direct evaluation regarding how accurately the proposed bidirectional optical flow estimates the actual inter-exposure displacement. While training using flow cycle consistency and reconstruction loss have been demonstrated, quantitative results verifying the accuracy of the predicted flow are missing. Please evaluate the EPE or warping error in a synthetic setting where ground-truth correspondences are established.

  • A direct analysis is needed to determine whether the restored and aligned long-exposure features actually provide meaningful information during the reconciliation stage. Although ablation studies concerning gating and certain reconciliation components are presented in Table 2 and Appendix Table 14, it remains unclear exactly what information is being preserved or selected through feature restoration and confidence-guided fusion. Please visualize the features (or alignment results) and spatial fusion gates before and after reconciliation, and analyze how information from each exposure is utilized in areas affected by motion blur or occlusion.

7 comments

Q1:The practical motivation for the dual-exposure capture setup is insufficient. The proposed method assumes a scenario where short-exposure and long-exposure images are captured sequentially with the same camera. However, this capture method requires changing exposure settings and additional capture time, potentially leading to motion and occlusion issues due to the time gap between the two shots. It is unclear what substantial advantages this complex capture setup offers compared to single-exposure restoration or short-exposure burst capture. In particular, as the primary quantitative evaluation focuses on the D2 benchmark, it is difficult to assess the method's practical applicability.

In fact, we build our dual-exposure capture framework on the successive long–short acquisition used by D2HNet (ECCV 2022, manuscript lines 1078–1079 and 785–789). The paper has discussed related acquisition designs, including exposure bracketing (BracketIRE, ICLR 2025) and dual-exposure Quad Bayer sensing (IEEE Transactions on Image Processing, 2025), and the revision clarifies how their timing and sampling differ (Section 2.1, manuscript lines 119–136). The long exposure collects more light, while the short exposure limits motion blur and preserves sharper structures. A single exposure must trade off these two benefits. A short-exposure burst can reduce noise by combining several frames. The successive pair instead combines the light collected over a longer interval with the sharper structure of one short frame.

To improve restoration when the capture gap introduces displacement or unreliable correspondence, we propose a framework that models motion between exposures, estimates blur within each exposure, and uses confidence to control fusion (manuscript lines 87–106). We have evaluated acquisition robustness on MiDual, our 1,192-scene captured dataset (manuscript lines 1078–1113). The supplementary results compare sensor noise, exposure ratios, resolution, repeatability, and illumination (Tables 7–11, manuscript lines 1243–1315). The MiDual tests show a consistent PSNR gain over D2HNet across day and night scenes. We have also evaluated transfer using GoPro, HIDE, RealBlur, Köhler, and REDS (Tables 12–13, manuscript lines 1318–1380). Those tests supply the same blurred frame to both branches to assess transfer without complementary inputs. I²EM improves on D2HNet in this setting, with smaller gains than on D2.

Table 7. Robustness to calibrated sensor-noise variation on MiDual.

Method3,2006,40012,80025,60051,200Δ(12.8k→51.2k)
MPRNet35.5235.4835.4433.8531.20-4.24
MIMO-UNet+35.7535.7135.6834.1031.55-4.13
Restormer36.3136.2736.2434.7232.30-3.94
D2HNet36.3836.3536.3335.1032.85-3.48
I²EM39.2139.1839.1738.2836.23-2.94

Table 8. Cross-resolution stability on MiDual under single-scale and multi-scale training.

MethodTraining2880p PSNR2880p SSIM1440p PSNR1440p SSIMChangevs. D2HNet
D2HNetsingle-scale 51236.350.970835.820.9635-0.53–
I²EMsingle-scale 51239.180.976838.640.9741-0.54+2.82
I²EMmulti-scale [512,640]39.290.977438.910.9755-0.38+3.09

Table 9. Sensitivity to the short-to-long exposure ratio on MiDual.

Method2×4×8×16×AdaptiveMean
MPRNet36.2835.8835.4834.6835.5835.58
Restormer37.0736.6736.2735.4736.3736.37
D2HNet36.0536.1736.3536.0536.1536.15
I²EM39.3839.2839.1838.7839.1339.15
ΔPSNR+3.33+3.11+2.83+2.73+2.98+3.00

Table 10. Repeated-run stability and paired scene-level comparison on MiDual.

MethodRun 1Run 2Run 3Mean ± std.ΔPSNRp-value
MIMO-UNet+35.6935.7335.7135.71±0.02+3.47<0.001
Restormer36.2536.2936.2736.27±0.02+2.91<0.001
D2HNet36.3336.3736.3536.35±0.02+2.83<0.001
I²EM39.1639.1939.2039.18±0.02––

Table 11. Illumination-stratified MiDual results.

MethodDay PSNRDay SSIMDay ΔPSNRNight PSNRNight SSIMNight ΔPSNR
D2HNet37.120.9735–35.580.9662–
I²EM39.980.9805+2.8638.380.9731+2.80

Table 12. Cross-dataset PSNR under equivalent single-image input.

TrainTestDeepMIMO+HINetD2HNetEVSSMI²EMΔPSNR
GoProGoPro29.0832.4532.7130.8533.8931.65+0.80
GoProHIDE25.7329.9830.5029.2032.7329.85+0.65
GoProRealBlur-J27.8728.1528.3027.7031.9028.10+0.40
GoProRealBlur-R33.9035.3035.5034.8039.6535.20+0.40
GoProK"ohler26.1027.9028.2027.3029.0027.75+0.45
REDSGoPro26.5028.8029.1028.1030.0028.60+0.50

Table 13. Same-distribution RealBlur evaluation under equivalent single-image input.

MethodRealBlur-J PSNRRealBlur-J SSIMRealBlur-J ΔPSNRRealBlur-R PSNRRealBlur-R SSIMRealBlur-R ΔPSNR
DeepDeblur28.950.8510–36.350.9301–
MPRNet31.200.9005–39.050.9662–
MIMO-UNet+31.350.9028–39.150.9670–
HINet31.500.9045–39.300.9682–
Restormer31.650.9060–39.420.9690–
D2HNet30.600.8952–38.450.9625–
EVSSM32.500.9280–40.100.9740–
I²EM31.100.8990+0.5038.900.9655+0.45

Q2:The computational cost of the proposed method is excessive, and the necessity of its complex structure has not been sufficiently validated. I2EM requires 133.79M parameters and an inference time of 78ms, which is significantly higher than D2HNet's 35.2M parameters and 12ms. However, it remains unverified whether this performance gap persists when the model capacity of conventional methods is increased or when they are extended to a similar dual-branch structure. Consequently, it is difficult to distinguish whether the reported performance gains stem from the structural advantages of the proposed motion modeling or simply from the significantly larger model capacity.

Our method is not simply based on larger model capacity. Table 1 has included DeblurDiff (2025), which has 386.2M parameters yet achieves lower PSNR than I²EM (manuscript lines 345–360). We have also compared a simpler three-level network with the full five-level design. The three-level model reaches 40.13 dB with 89.35M parameters, and the full design gains 0.65 dB (Table 19, manuscript lines 1516–1530). At fixed capacity, removing coarse observation supervision also reduces performance, supporting the use of the coarse levels for motion modeling and reconstruction.

To further separate architecture from capacity, Appendix E.5 varies model size in both directions using the same long–short pair (Table 15, manuscript lines 1410–1452). The 34.8M I²EM variant achieves 38.85 dB, compared with 37.95 dB for the 35.2M D2HNet, while using fewer FLOPs. At the full model's capacity, I²EM retains a PSNR lead of 1.26 dB over the 132.5M D2HNet-Wide and 1.10 dB over the 134.2M Dual-Branch Restormer. The full model takes 78 ms versus 12 ms for the standard D2HNet (Table 14, manuscript lines 1383–1408). The smaller variant therefore provides a lower-cost option, while the full configuration prioritizes restoration quality.

Table 1. Single-exposure night-image restoration comparison on D2-2880p for long- and short-exposure inputs.

MethodInputWeights/trainingPSNRSSIMLPIPSParams (M)FLOPs (G)
DeepDeblurL-onlyReleased36.550.96360.09542.3156.8
DeblurGAN-v2L-onlyReleased36.820.96550.09060.9182.3
MPRNetL-onlyReleased37.080.96840.08623.598.2
SwinIRL-onlyReleased37.700.97120.07912.162.3
MIMO-UNet+L-onlyReleased37.310.96900.08316.142.1
MPRNet-DNS-onlyReleased37.350.95800.08515.7492.3
NAFNetL-onlyReleased37.700.97100.07923.492.4
RestormerL-onlyReleased37.870.97170.07726.1105.7
UformerL-onlyReleased37.550.97020.08128.3118.4
HATL-onlyReleased37.850.97150.07812.498.3
PromptIRL-onlyReleased37.450.96950.08287.2138.6
UFPNetL-onlyReleased37.720.97120.07980.3290.4
EVSSML-onlyReleased38.090.97270.07217.1126.0
DeblurDiffL-onlyReleased37.550.97030.050386.22890.0
DarkIRS-onlyD2-adapted training37.650.97050.08035.2128.5
NSPS-onlyReleased37.250.96850.08528.5115.2
NoiseDiffS-onlyD2-adapted training37.950.97180.075120.5285.3
I²EMLong–short pairD2 training40.780.97830.054133.79462.8

Table 14. Accuracy and efficiency comparison under the unified evaluation platform.

MethodInputPSNR (dB)Params (M)FLOPs (G)Time (ms)Mem. (MB)
DeepDeblursingle36.5542.3156.8141,245
DeblurGAN-v2single36.8260.9182.321,580
CDVD-TSPmulti37.0016.688.5231,120
HINetmulti37.2288.7112.361,980
MPRNetsingle37.0823.598.2101,450
MIMO-UNet+single37.3116.142.19980
Restormersingle37.8726.1105.7121,680
D2HNetdual37.9535.2165.3121,520
UFPNetsingle37.7280.3290.4292,350
EVSSMsingle38.0917.1126.031,100
DeblurDiffsingle37.55386.22,890.030412,580
STGDNetmulti-sensor38.1228.4142.6152,350
I²EMdual40.78133.79462.87810,284
Reviewer_4TEtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes I²EM, a dual-exposure restoration framework that models displacement between exposures with bidirectional optical flow and blur within each exposure with spatially varying kernels. It uses these estimates to align and fuse short- and long-exposure information for night image restoration.

Critical strengths
  1. The method models the relationship between short and long exposures more explicitly and achieves stronger restoration results on the reported benchmark.
  2. The paper includes extensive ablations, robustness tests, and comparisons across acquisition conditions.
Critical weaknesses
  1. Capacity and input confounding. I²EM is much larger than D2HNet and the single-image baselines; although the paper includes module ablations, it lacks a capacity-matched dual-exposure baseline to establish how much of the gain comes from the proposed motion modeling rather than model capacity and the input difference.
  2. Limited input-matched comparisons. D2HNet is the only baseline in the main table using the same short-long image pair, while most single-image methods receive only the long exposure and are evaluated against a short-reference target; a strong short-exposure denoising baseline would make the comparison more informative.
  3. Unverified motion disentanglement. The ablations measure output quality but do not directly show that the predicted flow and kernels capture inter-exposure displacement and intra-exposure blur, respectively.
Questions for authors
  1. How to verify that the predicted flow and kernels capture inter- and intra-exposure motion separately, rather than compensating for each other?
3 comments

Q1:Capacity and input confounding. I²EM is much larger than D2HNet and the single-image baselines; although the paper includes module ablations, it lacks a capacity-matched dual-exposure baseline to establish how much of the gain comes from the proposed motion modeling rather than model capacity and the input difference.

We have included DeblurDiff (Kong et al., NeurIPS 2025), a state-of-the-art method for real-world single-image deblurring, as a larger reference in Table 1 (manuscript lines 345–360). It has 386.2M parameters versus 133.79M for I²EM, yet scores 3.23 dB lower in PSNR. DeblurDiff obtains a lower LPIPS (0.050 vs. 0.054), so the PSNR result indicates that more parameters alone do not ensure higher pixel fidelity. The revised Table 2 also compares I²EM with multi-exposure methods of comparable size, including SelfHDR and BracketIRE, under the same long–short input protocol (manuscript lines 363–377).

To further validate the contribution of the proposed design under controlled inputs and comparable capacity, we evaluate D2HNet-Wide and Dual-Branch Restormer using the same D2-2880p pairs as I²EM in Appendix E.5 (Table 15, manuscript lines 1410–1452). I²EM gains 1.26 dB and 1.10 dB in PSNR over the two baselines, respectively, while using fewer FLOPs than either. The capacity sweep confirms that increasing model size improves restoration, but the advantage persists when both exposure inputs and parameter budgets are controlled.

Table 1. Single-exposure night-image restoration comparison on D2-2880p

MethodInputWeights/trainingPSNRSSIMLPIPSParams (M)FLOPs (G)
DeepDeblurL-onlyReleased36.550.96360.09542.3156.8
DeblurGAN-v2L-onlyReleased36.820.96550.09060.9182.3
MPRNetL-onlyReleased37.080.96840.08623.598.2
SwinIRL-onlyReleased37.700.97120.07912.162.3
MIMO-UNet+L-onlyReleased37.310.96900.08316.142.1
MPRNet-DNS-onlyReleased37.350.95800.08515.7492.3
NAFNetL-onlyReleased37.700.97100.07923.492.4
RestormerL-onlyReleased37.870.97170.07726.1105.7
UformerL-onlyReleased37.550.97020.08128.3118.4
HATL-onlyReleased37.850.97150.07812.498.3
PromptIRL-onlyReleased37.450.96950.08287.2138.6
UFPNetL-onlyReleased37.720.97120.07980.3290.4
EVSSML-onlyReleased38.090.97270.07217.1126.0
DeblurDiffL-onlyReleased37.550.97030.050386.22890.0
DarkIRS-onlyD2-adapted training37.650.97050.08035.2128.5
NSPS-onlyReleased37.250.96850.08528.5115.2
NoiseDiffS-onlyD2-adapted training37.950.97180.075120.5285.3
I²EMLong–short pairD2 training40.780.97830.054133.79462.8

Table 15. Capacity-matched dual-exposure comparison and I²EM capacity sweep on D2-2880p

MethodParams (M)FLOPs (G)PSNRSSIMLPIPSΔPSNR
D2HNet35.2165.337.950.97230.076--
D2HNet-Wide132.5485.239.520.97690.059-1.26
Dual-Branch Restormer134.2512.839.680.97720.057-1.10
I²EM (35M)34.8142.538.850.97510.065-1.93
I²EM (70M)69.5285.439.920.97710.058-0.86
I²EM (100M)99.2405.840.450.97790.055-0.33
I²EM133.79462.840.780.97830.0540.00

Q2:Limited input-matched comparisons. D2HNet is the only baseline in the main table using the same short-long image pair, while most single-image methods receive only the long exposure and are evaluated against a short-reference target; a strong short-exposure denoising baseline would make the comparison more informative.

We have also compared with methods that use multiple observations, such as CDVD-TSP and ESTRNN, on D2-2880p. We have already evaluated MPRNet-DN (2021) using only the short exposure, and its PSNR is 37.35 dB (Table 1, manuscript line 352).

To further test how I²EM compares with short-exposure restoration methods, we extend the comparison to DarkIR (2025), NSP (2026), and NoiseDiff (2026) under the same short-only protocol. NoiseDiff obtains the highest PSNR among the short-only baselines at 37.95 dB, while I²EM reaches 40.78 dB using both exposures (Table 1, manuscript lines 358–360). For a direct comparison using identical inputs, Table 2 also evaluates LSD₂, LSFNet-RGB, and DeLECA on the same long–short pair, where I²EM achieves the highest PSNR (manuscript lines 363–377).

Table 2. Dual-exposure night-image restoration comparison on D2-2880p with a shared long–short input pair

MethodWeights/trainingPSNRSSIMLPIPSParams (M)FLOPs (G)Time (ms)
CDVD-TSPD2-adapted training37.000.96680.08716.688.523
ESTRNND2-adapted training37.100.96750.08820.578.525
LSD₂Released38.420.97410.07128.5142.818
HINetD2-adapted training37.220.96850.08388.7112.36
D2HNetReleased37.950.97230.07635.2165.312
LSFNet-RGBD2-adapted training38.150.97320.07332.5158.422
SelfHDRReleased38.850.97520.065118.2425.838
DeLECAD2-adapted training39.150.97580.06345.8198.525
BracketIREReleased39.250.97610.061128.5498.245
STGDNetReleased38.120.97290.07228.4142.615
I²EMD2 training40.780.97830.054133.79462.878

Q3:Unverified motion disentanglement. The ablations measure output quality but do not directly show that the predicted flow and kernels capture inter-exposure displacement and intra-exposure blur, respectively.

Thanks for your helpful comments. We agree that the restoration ablations alone do not directly establish the roles of the two motion representations. We have clarified the distinction between displacement across exposure centers and blur accumulated within each exposure in the revised analysis (manuscript lines 478–485). The paper has tested the effect of constraining kernel displacement in its design ablations (Table 3, manuscript lines 433–449).

To provide direct evidence, the revision compares the estimated correspondence and kernels with their diagnostic references in Figure 4 (manuscript lines 486–512). The correspondence views preserve the main scene structures and show where alignment errors remain. The long-exposure kernels show broader directional spread than the short-exposure kernels. These observations are consistent with the intended roles of the two representations.

Table 3. Design ablations for factorization, decoupling, and reconciliation on D2-2880p

MechanismInterventionParams (M)PSNRΔPSNRSSIMLPIPS
FactorizationOne-way flow133.7940.20-0.580.97560.062
FactorizationSequential flow–kernel133.7940.36-0.420.97630.059
FactorizationImplicit joint motion68.4239.29-1.490.96890.077
Decoupling constraintsNo kernel centering133.7940.60-0.180.97610.057
Decoupling constraintsShort-to-long blur loss133.7939.54-1.240.97120.070
Decoupling constraintsBlur before warp133.7939.99-0.790.97430.065
ReconciliationLong-exposure residual133.7939.00-1.780.96460.084
ReconciliationNo flow-conditioned kernels133.7940.32-0.460.97610.060
ReconciliationNo residual denoising121.6440.09-0.690.97500.066
ReconciliationConcatenation instead of gating134.5140.47-0.310.97720.057
Complete designFull model133.7940.78--0.97830.054

Q4:How to verify that the predicted flow and kernels capture inter- and intra-exposure motion separately, rather than compensating for each other?

We have tested whether one representation can absorb part of the motion assigned to the other in the ablations reported in Table 3 (manuscript lines 433–449). In the revision, we explain more clearly that removing kernel centering allows the kernel to absorb part of the cross-exposure translation and reduces restoration PSNR by 0.18 dB.

To further test whether the two representations provide distinct information, we replace each internal variable with its reference while holding the other prediction fixed. In Table 4 (manuscript lines 459–477), replacing the flow improves PSNR by 0.47 dB, while replacing the kernel improves it by 0.30 dB. The gains from the two interventions support distinct contributions to restoration under the same inputs.

Table 4. Counterfactual motion and coordination controls on D2-2880p

ConditionFlowKernelPSNRSSIMLPIPSΔPSNR
Complete modelPredictedPredicted40.780.97830.0540.00
GT flowGround truthPredicted41.250.97920.051+0.47
GT kernelPredictedGround truth41.080.97880.052+0.30
No flow confidencePredictedPredicted39.850.97620.062-0.93
No kernel confidencePredictedPredicted40.120.97680.059-0.66
Uniform 0.5 gatePredictedPredicted40.250.97710.058-0.53
Swapped long–short featuresSwappedSwapped36.520.96250.095-4.26

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.