acceptodds
Under review as a conference paper at ICLR 2027

BRACE: Adversarial and Contrastive Endpoint Regularization for Schrödinger-Bridge Image Enhancement

Abstract

Unpaired image enhancement can be viewed as transporting a degraded image distribution toward a clean one under a strong constraint of source-scene fidelity. Unlike broad domain translation, the output must undergo substantial appearance correction while remaining closely aligned with the input structure. Schrödinger-bridge methods provide a principled stochastic transport framework for this setting, but existing bridge-based objectives can be overly conservative, underconstrain target realism, and require multiple refinement steps at inference time. We propose Bridge Regularization with Adversarial and Contrastive Endpoints (BRACE), a Schrödinger-bridge method that strengthens endpoint supervision with explicit constraints for target realism and local structural consistency. The resulting translator preserves the bridge construction while producing more realistic and faithful outputs with very small inference budget. Across low-light enhancement, dehazing, and underwater enhancement benchmarks, BRACE consistently improves over a Schrödinger-bridge baseline and remains competitive with dedicated unpaired baselines. Our results indicate that explicit endpoint supervision enables effective bridge-based transport for image enhancement with few inference steps.

Primary area:
generative models
Keywords:
unpaired image enhancement, Schrödinger bridges, stochastic transport, image-to-image translation

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_3JPDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BRACE, a Schrödinger-bridge method for unpaired image enhancement. It combines source-endpoint regression, adversarial supervision, and PatchNCE to improve target-domain appearance while preserving input content. The authors evaluate the method on low-light enhancement, dehazing, underwater enhancement, and unpaired image translation. They report improvements over bridge baselines and one-step inference for the enhancement tasks.

Critical strengths

The method has a clear design goal. It combines target-domain supervision with input-based constraints to address both visual quality and content preservation.

The experiments cover several enhancement tasks. The loss ablations and inference-step comparisons provide useful evidence for the effects of the added supervision.

Critical weaknesses

Technical Contribution. Adversarial supervision and patch-based contrastive losses have already been used together in bridge-based translation. UNSB [1] uses adversarial learning and patch-wise contrastive regularization. DehazeSB [2] combines adversarial supervision with PatchNCE for unpaired dehazing. BRACE uses different endpoint predictions and bidirectional training, but the paper does not clearly explain the benefit of these changes over prior work. A direct comparison of their objectives and training procedures is needed.

Experimental Results and Metrics. On LOL-Real, BRACE improves LPIPS and FID over SBFlow, but PSNR drops from 18.25 to 14.85 dB and SSIM drops from 0.393 to 0.191. These results show a clear trade-off between perceptual quality and reconstruction accuracy. PSNR and SSIM may not fully reflect perceptual quality, but such large drops still raise concerns about faithful restoration. The visual comparisons also do not clearly show whether the gains in image quality outweigh these losses. The paper should examine these cases and provide stronger evidence for content preservation.

Low-Step Inference. SBFlow also uses one-step inference on the enhancement tasks. CycleGAN and CUT require only one forward pass. A lower number of function evaluations (NFE) does not necessarily mean faster inference. The authors should report inference time and memory use at the same image sizes, together with the corresponding image quality.

[1] Unpaired Image-to-Image Translation via Neural Schrödinger Bridge. ICLR 2024.

[2] When Schrödinger Bridge Meets Real-World Image Dehazing with Unpaired Training. ICCV 2025.

Questions for authors

What does bidirectional training motivate to the such vision enhancement task? A comparison with one-direction training or discriminative models would help explain whether the reverse direction is necessary.

8 comments
AuthorsOpenReview ↗

Response to Reviewer 3JPD(1/2)

We thank the reviewer for the thoughtful review and for pointing out that NFE alone does not determine inference cost. Following this suggestion, we now report inference time and peak memory at equal resolution together with exposure-aligned quality metrics (Table R1). We also clarify the role of bidirectional training and how BRACE differs from prior bridge-based methods.

All new tables and the corrected figures are also collected at https://anonymous.4open.science/r/BRACE_ICLR_Rebutal-14E0/. The tables relevant to this review are included below as well.

W1. BRACE learns the bridge by IMF regression in continuous time instead of solving a sequence of adversarial problems on a fixed discrete grid, as UNSB and DehazeSB do. The network regresses real images from Brownian-bridge states, so structure is supervised by real data rather than by proximity to the generator's own output, which ties the output more directly to the source scene. Bidirectional training is what makes this possible: the reverse direction generates the endpoint that defines the bridge for the enhancement direction, so the enhancement output is regressed toward real clean images. The transport cost comes from the Brownian reference instead of an explicit loss with a learned entropy estimator, which removes a source of estimation error. The number of steps is not tied to a training grid and is chosen at test time, so one step suffices. In BRACE, the adversarial loss is not the transport mechanism but a stabilizer of IMF, and this is our specific novelty: regression carries the transport, and the discriminator only corrects the domain. In online IMF, the generated endpoints define the next bridges, yet no loss term keeps them in the target domain, so the procedure drifts, as SBFlow shows on LOL (Table 1). BRACE anchors each predicted endpoint to the target domain, and removing this anchor reproduces the failure (Table 5). It does so with one task-agnostic objective, without the dehazing-specific priors of DehazeSB, and outperforms UNSB on 20 of 24 metrics with the same backbone.

W2. By source-scene fidelity we mean preserving the input's geometry, edges and local detail, not the exact pixel values of a reference the model never sees. BRACE enforces it through bridge regression and PatchNCE, while the adversarial term moves the output toward the target domain (see W1). Without paired supervision, the input does not determine one correct exposure, and PSNR and SSIM penalize a global brightness shift as heavily as a structural error [1]. To separate the two, we re-evaluated all methods after scaling each output to the mean intensity of its reference (GT-mean protocol), which removes exposure differences only (Table R1 - part 2/2). With exposure aligned, the LOL-Real drop disappears: BRACE outperforms SBFlow in LPIPS (0.209 vs 0.279), PSNR (23.53 vs 23.13) and SSIM (0.692 vs 0.668). The original gap therefore reflects exposure mismatch, not lost structure. On LOL, BRACE achieves the best LPIPS and PSNR among all methods. We will add this analysis to the paper.

[1] The Perception-Distortion Tradeoff. CVPR 2018.

AuthorsOpenReview ↗

Response to Reviewer 3JPD (2/2)

W3. We agree that NFE alone does not determine speed, and we report inference time and peak memory at native resolution (Table R1). BRACE runs in 44/34 ms per image on LOL/LOL-Real. This is 5–6× faster than UNSB (≈206 ms, NFE 5) and over 350× faster than AGLLDiff (≈16 s), with 3–4× less memory. GANs remain the fastest (6–9 ms), but BRACE is better on LOL in all quality metrics and on LOL-Real in LPIPS and SSIM. SBFlow has a comparable cost, since both run one evaluation of the same network, and at this cost BRACE improves every quality metric. Our claim is therefore not that BRACE enables one-step inference, but that it makes one-step bridge inference competitive in quality. We will state this precisely in contribution (ii).

Table R1.

LOLLOL-Real
MethodNFELPIPS↓PSNR↑SSIM↑ms↓MB↓LPIPS↓PSNR↑SSIM↑ms↓MB↓
CycleGAN10.18624.44‾\underline{24.44}0.6978.8‾\underline{8.8}1235‾\underline{1235}0.29523.57‾\underline{23.57}0.6458.2‾\underline{8.2}1235‾\underline{1235}
CUT10.21823.290.6529.011690.30722.700.6148.31169
EnlightenGAN10.21620.450.6527.113360.21121.430.6495.61336
GDP10000.31719.470.4987.2×10⁵34690.33119.360.5447.2×10⁵3469
AGLLDiff100.182‾\underline{0.182}24.170.7551.6×10⁴69090.17326.740.7941.6×10⁴6909
ASBM40.41419.500.534134.329740.47118.540.492129.72974
DDSB*50.46619.410.447206.515730.49519.750.435208.01573
UNSB50.24722.860.578206.216630.29323.110.581205.61663
SBFlow10.37117.420.45737.426440.27923.130.66833.32658
BRACE (ours)10.17425.370.732‾\underline{0.732}44.416550.209‾\underline{0.209}23.530.692‾\underline{0.692}33.52434

GT-mean evaluation: each output is scaled to the mean intensity of its reference before computing metrics. Time [ms/img] and peak GPU memory [MB] at native resolution, batch 1, on an NVIDIA A100 (40 GB). Best in bold, second underlined.

Q1. In BRACE, the reverse direction is not an extra task: it builds the bridge for the enhancement direction. Each bridge connects a real image and an endpoint generated by the opposite direction, and the MSE always regresses the real image (Algorithm 1). Regression toward real clean images therefore requires the reverse direction. A clean image is degraded by the model, and the network recovers it from any state on the resulting bridge. This cycle (clean → degraded → bridge state → clean) acts as a cycle loss that trains a continuous bridge with real clean targets. With the enhancement direction alone, the clean output would be supervised only by adversarial and PatchNCE losses, as in CUT. As requested, we trained a one-direction, one-step translator on LOL with the same U-Net, discriminator and PatchNCE (adversarial + PatchNCE). It reaches higher PSNR/SSIM (21.39/0.584 vs 20.19/0.470) but clearly worse FID (92.2 vs 71.2) and LPIPS (0.202 vs 0.193), trading realism for lower distortion. The reverse direction adds no inference cost: it is used only in training, shares one network, and BRACE runs with NFE = 1.

“Source-scene fidelity” is not a widely accepted concept in low-level vision. LOL is used to evaluate low-light enhancement. Although low-light enhancement, like many image restoration tasks, does not have a strictly defined ground truth, lighting and color are themselves part of the restoration target. The reported results do not show consistent improvements in PSNR, SSIM, and perceptual metrics.

BRACE also does not offer a substantial advance in model efficiency. The authors should report parameter, computational cost, and inference latency to reflect method complexity. The improvements over compared methods reported in Table R1 are only incremental.

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

This paper investigates unpaired image enhancement (low‑light enhancement, dehazing, underwater restoration) using the Schrödinger‑Bridge (SB) framework. Schrödinger‑Bridge learns a stochastic SDE transport process that maps the source degraded‑image distribution towards the clean target‑image distribution, trained only by two disjoint image collections without paired input‑ground‑truth samples. The proposed BRACE introduces endpoint‑oriented regularizers: it applies adversarial loss on the predicted target endpoint to enforce target‑domain visual realism, and PatchNCE contrastive loss on the predicted source endpoint to preserve local patch‑level scene correspondence. Empirically, BRACE outperforms earlier Schrödinger‑Bridge baselines across three enhancement benchmarks and supports very‑low‑NFE inference.

Critical strengths
  1. It identifies a concrete practical limitation of existing Schrödinger‑Bridge image translation: weak endpoint supervision causes overly‑conservative enhancement outputs. The proposed endpoint‑regularization strategy alleviates this symptom within the SB training pipeline.
  2. Consistent quantitative and qualitative improvements are observed compared against prior Schrödinger‑Bridge family methods (SBFlow, UNSB, ASBM, DDSB) on low‑light enhancement, dehazing and underwater restoration tasks.
  3. The method enables fast inference with very small NFE budgets (NFE=1), which is practically attractive for image‑enhancement deployment. Ablations verify the individual contribution of adversarial loss and PatchNCE loss.
  4. Beyond image enhancement, the approach is validated on a general unpaired image‑to‑image translation benchmark AFHQ, demonstrating that the endpoint‑regularized SB formulation is not task‑specific.
Critical weaknesses
  1. Adversarial loss for domain alignment and PatchNCE for patch‑level correspondence are well‑established from prior unpaired translation work (e.g. CUT). The novelty lies in placing them as endpoint regularizers inside the SB training loop, rather than inventing new loss functions.
  2. A time‑conditioned vanilla U‑Net can be trained with exactly the same set of losses (MSE + adversarial + PatchNCE) under the unpaired setting, without any Schrödinger‑Bridge SDE / IPF machinery. Without this ablation, we cannot tell how much performance gain originates from the loss combination itself versus SB‑specific training dynamics with intermediate sampled bridge states.
  3. BRACE achieves clear gains over older SB methods, but it does not consistently outperform CycleGAN / CUT across all benchmarks and metrics.
  4. For many practical image‑enhancement tasks such as dehazing, synthesizing degraded inputs to build paired supervised training datasets is relatively straightforward. This raises a natural question, and the paper should further clarify the practical motivation for pursuing unpaired training in these settings.
  5. Optimal loss weights λadv\lambda_{adv} and λnce\lambda_{nce} differ substantially across different benchmarks. It remains unclear how well these hyper‑parameters transfer to unseen real‑world mixed‑degradation inputs.
Questions for authors

Please address the concerns in the above weaknesses section.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Zrbh (1/2)

We thank the reviewer for the constructive and precise feedback. The suggested ablation with the same losses but no bridge states was particularly valuable: it isolates the contribution of the bridge training dynamics, and we now report it together with a direct-translator control. We also appreciate the questions on the comparison with GANs, the motivation for unpaired training and the choice of loss weights, which we answer below.

All new tables and the corrected figures are also collected at https://anonymous.4open.science/r/BRACE_ICLR_Rebutal-14E0/. The tables relevant to this review are included below as well.

Q1. The novelty lies in using these losses to stabilize the IMF procedure, not as a standalone objective as in CUT. BRACE learns the translation by bridge regression to real images and applies the adversarial loss to the endpoint predicted from each bridge state. This keeps the generated endpoints, which define the next bridges, in the target domain. SBFlow has no such term and drifts in practice (Table 1). BRACE stabilizes IMF, and removing the anchor reproduces the failure (Table 5). PatchNCE reinforces local structure on top of the regression. The gains therefore depend on the bridge: the same losses without bridge states perform much worse (see our answer to Q2).

Q2. We ran exactly this control on LOL. The network is BRACE's time- and direction-conditioned U-Net, trained with the same losses, weights, discriminators and alternating loop. Only the bridge is removed. For a batch A → B with source a, the network translates directly from the endpoint, ŷ = G(a, t_a, B), and the adversarial and PatchNCE losses act on ŷ, as in BRACE. The MSE becomes a stop-gradient cycle loss ‖G(sg[ŷ], t_B, A) − a‖². This is BRACE's regression on the same synthetic pair, evaluated only at the endpoint rather than at bridge states x_t, t ~ U(0,1). The pair comes from one forward pass instead of 5-step SDE sampling. The results are shown in Table R5 under the name “Same losses, no bridge”. We also include a direct one-step translator (adversarial + PatchNCE), requested by another reviewer (R8J3).

Table R5.

LOLBridgeMSEAdvNCELPIPS↓PSNR↑SSIM↑FID↓
Direct translator––✓✓0.202‾\underline{0.202}21.390.58492.2
Same losses, no bridge–✓✓✓0.39615.350.333126.1
BRACE w/o Adv✓✓–✓0.4758.430.054146.7
BRACE w/o NCE✓✓✓–0.21318.820.45585.9‾\underline{85.9}
BRACE✓✓✓✓0.19320.19‾\underline{20.19}0.470‾\underline{0.470}71.2

All rows share the U-Net, discriminator, PatchNCE, optimizer, data, steps, seed and checkpoint selection; NFE = 1. Best in bold, second underlined.

With the loss combination held fixed, removing the bridge states degrades every metric (FID 71.2 → 126.1, PSNR 20.19 → 15.35). At the endpoint, the MSE reduces to a cycle-consistency constraint. On the bridge, it becomes a regression from every intermediate state, and this is what makes the loss combination effective. The gains therefore come from the SB training dynamics, not from the losses alone. The direct translator (row 1) reaches higher PSNR and SSIM, but neither control matches BRACE's LPIPS and FID.

AuthorsOpenReview ↗

Response to Reviewer Zrbh (2/2)

Q3. We agree and do not claim uniform superiority over GANs; our aim is to make bridge-based enhancement competitive with them. Against the stronger of CycleGAN and CUT on each metric, BRACE is better in 14 of 24 cases across the six enhancement benchmarks (Tables 1–3), including LPIPS on 5/6 and FID on 4/6. The GANs keep an advantage mainly in PSNR, which partly reflects exposure rather than structure. After scaling outputs to the reference brightness (Table R1), BRACE outperforms both GANs on LOL in LPIPS, PSNR and SSIM, and on LOL-Real in LPIPS and SSIM, with PSNR on par (23.53 vs 23.57). UIEB is the clear exception, as noted in the paper.

Table R1.

LOLLOL‑Real
MethodNFELPIPS↓PSNR↑SSIM↑Time [ms]↓Memory [MB]↓LPIPS↓PSNR↑SSIM↑Time [ms]↓Memory [MB]↓
CycleGAN10.18624.44‾\underline{24.44}0.6978.8‾\underline{8.8}1235‾\underline{1235}0.29523.57‾\underline{23.57}0.6458.2‾\underline{8.2}1235‾\underline{1235}
CUT10.21823.290.6529.011690.30722.700.6148.31169
EnlightenGAN10.21620.450.6527.113360.21121.430.6495.61336
GDP10000.31719.470.4987.2×10⁵34690.33119.360.5447.2×10⁵3469
AGLLDiff100.182‾\underline{0.182}24.170.7551.6×10⁴69090.17326.740.7941.6×10⁴6909
ASBM40.41419.500.534134.329740.47118.540.492129.72974
DDSB*50.46619.410.447206.515730.49519.750.435208.01573
UNSB50.24722.860.578206.216630.29323.110.581205.61663
SBFlow10.37117.420.45737.426440.27923.130.66833.32658
BRACE (ours)10.17425.370.732‾\underline{0.732}44.416550.209‾\underline{0.209}23.530.692‾\underline{0.692}33.52434

GT-mean evaluation: each output is scaled to the mean intensity of its reference before computing metrics. Time [ms/img] and peak GPU memory [MB] at native resolution, batch 1, on an NVIDIA A100 (40 GB). Best in bold, second underlined.

Q4. Synthetic pairs are easy to generate but hard to make realistic, and models trained on them inherit this gap:

  • Dehazing. Synthesis relies on the atmospheric scattering model, with homogeneous haze, global airlight and known depth. Real haze is non-homogeneous and wavelength-dependent, and models trained on synthetic RESIDE pairs degrade on real hazy images [1].
  • Low light. Real low-light images carry sensor noise and camera-specific ISP processing (tone mapping, denoising, white balance), which brightness or gamma synthesis does not reproduce. Even real paired data are tied to the sensor: [2] trains a separate model per camera.
  • Underwater. Clean references do not exist. UIEB references are selected by human voting among outputs of existing enhancement methods [3], so a supervised model is bounded by those methods.

Unpaired training learns directly from real degraded and real clean images. We use paired benchmarks only for evaluation.

Q5. The weights are not tuned per benchmark, but rather they maintain a certain ratio. We set λ_adv and λ_nce so that, averaged over training, the weighted terms contribute about 75% (MSE), 12.5% (adversarial) and 12.5% (PatchNCE) of the total loss. The raw magnitudes of the adversarial and PatchNCE losses depend on the data, so the same proportions yield different λ values per benchmark (Table 10). For a new dataset, including mixed-degradation data, the weights follow from the raw loss magnitudes measured in a short initial run, without a search. We will state the weighting rule in Appendix D.

[1] Benchmarking Single-Image Dehazing and Beyond (RESIDE). IEEE TIP 2019.

[2] Learning to See in the Dark. CVPR 2018.

[3] An Underwater Image Enhancement Benchmark Dataset and Beyond (UIEB). IEEE TIP 2019.

We have uploaded a revised paper that addresses the points raised in this review. The bridge-free controls, including the same-losses control without bridge states, are now Table 8 and are discussed in Sec. 5.3, with details in Appendix D. The GT-mean evaluation and inference cost referenced in our answer to Q3 are now Tables 2 and 3. Appendix E.2 now states the rule for the endpoint loss weights, and the per-benchmark values are in Table 13. Because of the new tables, the ablation (Table 5 in our responses) is now Table 7. We are happy to discuss any further points.

Reviewer_R8J3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Bridge Regularization with Adversarial and Contrastive Endpoints (BRACE), a Schrödinger-bridge method for unpaired image enhancement that strengthens endpoint supervision with explicit constraints for target realism and local structural consistency. It augments an SBFlow-based objective with adversarial supervision at the translated endpoint and a PatchNCE constraint against the source image. Experiments cover low-light enhancement, dehazing, and underwater enhancement.

Critical strengths
  1. The paper is generally clearly written and well organized. The motivation of improving target-domain realism while maintaining source-scene fidelity is easy to follow, and the method description is supported by training and inference algorithms in the appendix.
  2. The evaluation covers several datasets across three enhancement tasks and includes generic translation methods, bridge-based methods, and task-specific baselines. BRACE achieves good results on LOL and EUVP.
Critical weaknesses
  1. The novelty of this paper is limited. The main change is to add adversarial and PatchNCE losses to an SBFlow-based translator. However, UNSB (Kim et al., ICLR 2024) already combines Schrödinger-bridge transport with adversarial learning and patchwise contrastive regularization. DehazeSB (Lan et al., ICCV 2025) also uses adversarial supervision and PatchNCE for unpaired image enhancement. Introducing similar constraints to the SBFlow-based methods might not be a novel idea, the specific difference and its benefit need stronger justification.
  2. The claim of improved source-scene fidelity needs a more careful analysis. On LOL-Real, BRACE improves LPIPS over SBFlow but reduces PSNR from 18.25 to 14.85 and SSIM from 0.393 to 0.191. These are substantial decreases in reference fidelity. They may reflect a trade-off between perceptual quality and distortion, but the paper does not adequately examine this trade-off despite emphasizing faithful enhancement.
  3. Some reported results and sampling definitions are inconsistent. For SBFlow on EUVP at NFE = 1, Table 3 reports PSNR = 19.68 and LPIPS = 0.223, whereas Figure 6 appears to show approximately 17.5 and 0.29, respectively. Please clarify whether the table and figure use different checkpoints, splits, or configurations.
  4. The ablations do not isolate the contribution of the bridge formulation. Table 5 shows that the additional losses improve the current framework, especially the adversarial term. It does not establish whether bridge training is necessary for these improvements. A direct image-to-image translator using the same U-Net, discriminator, and PatchNCE configuration would provide an important control.
Questions for authors
  1. How does BRACE differ from existing methods such as UNSB and DehazeSB, and what is its specific novelty?

  2. How does BRACE balance perceptual quality and source-scene fidelity, given the substantial drops in PSNR and SSIM on LOL-Real despite improved LPIPS?

  3. Why are the reported SBFlow results on EUVP at NFE = 1 inconsistent between Table 3 and Figure 6, and do they use different checkpoints, data splits, or configurations?

  4. Is bridge-based training necessary for the reported improvements, and how does BRACE compare with a direct image-to-image translator using the same U-Net, discriminator, and PatchNCE configuration?

3 comments
AuthorsOpenReview ↗

Response to Reviewer R8J3 (1/2)

We thank the reviewer for the careful reading of our paper, including the detailed check of the SBFlow values across tables and figures. The questions on the relation to UNSB and DehazeSB and on the role of the bridge prompted two new bridge-free controls and an exposure-aligned evaluation, which we believe make the paper's claims considerably sharper. We address each point below.

All new tables and the corrected figures are also collected at https://anonymous.4open.science/r/BRACE_ICLR_Rebutal-14E0/. The tables relevant to this review are included below as well.

Q1. BRACE learns the bridge by IMF regression in continuous time instead of solving a sequence of adversarial problems on a fixed discrete grid, as UNSB and DehazeSB do. The network regresses real images from Brownian-bridge states, so structure is supervised by real data rather than by proximity to the generator's own output, which ties the output more directly to the source scene. The transport cost comes from the Brownian reference instead of an explicit loss with a learned entropy estimator, which removes a source of estimation error. The number of steps is not tied to a training grid and is chosen at test time, so one step suffices. In BRACE, the adversarial loss is not the transport mechanism but a stabilizer of IMF, and this is our specific novelty: regression carries the transport, and the discriminator only corrects the domain. In online IMF, the generated endpoints define the next bridges, yet no loss term keeps them in the target domain, so the procedure drifts, as SBFlow shows on LOL (Table 1). BRACE anchors each predicted endpoint to the target domain, and removing this anchor reproduces the failure (Table 5). It does so with one task-agnostic objective, without the dehazing-specific priors of DehazeSB, and outperforms UNSB on 20 of 24 metrics with the same backbone.

Q2. By source-scene fidelity we mean preserving the input's geometry, edges and local detail, not the exact pixel values of a reference the model never sees. BRACE enforces it through bridge regression and PatchNCE, while the adversarial term moves the output toward the target domain (see Q1). Without paired supervision, the input does not determine one correct exposure, and PSNR and SSIM penalize a global brightness shift as heavily as a structural error [1]. To separate the two, we re-evaluated all methods after scaling each output to the mean intensity of its reference (GT-mean protocol), which removes exposure differences only. Table R1 shows the results. With exposure aligned, the LOL-Real drop disappears: BRACE outperforms SBFlow in LPIPS (0.209 vs 0.279), PSNR (23.53 vs 23.13) and SSIM (0.692 vs 0.668). The original gap therefore reflects exposure mismatch, not lost structure. On LOL, BRACE achieves the best LPIPS and PSNR among all methods. We will add this analysis to the paper.

Table R1.

LOLLOL-Real
MethodNFELPIPS↓PSNR↑SSIM↑ms↓MB↓LPIPS↓PSNR↑SSIM↑ms↓MB↓
CycleGAN10.18624.44‾\underline{24.44}0.6978.8‾\underline{8.8}1235‾\underline{1235}0.29523.57‾\underline{23.57}0.6458.2‾\underline{8.2}1235‾\underline{1235}
CUT10.21823.290.6529.011690.30722.700.6148.31169
EnlightenGAN10.21620.450.6527.113360.21121.430.6495.61336
GDP10000.31719.470.4987.2×10⁵34690.33119.360.5447.2×10⁵3469
AGLLDiff100.182‾\underline{0.182}24.170.7551.6×10⁴69090.17326.740.7941.6×10⁴6909
ASBM40.41419.500.534134.329740.47118.540.492129.72974
DDSB*50.46619.410.447206.515730.49519.750.435208.01573
UNSB50.24722.860.578206.216630.29323.110.581205.61663
SBFlow10.37117.420.45737.426440.27923.130.66833.32658
BRACE (ours)10.17425.370.732‾\underline{0.732}44.416550.209‾\underline{0.209}23.530.692‾\underline{0.692}33.52434

GT-mean evaluation: each output is scaled to the mean intensity of its reference before computing metrics. Time [ms/img] and peak GPU memory [MB] at native resolution, batch 1, on an NVIDIA A100 (40 GB). Best in bold, second underlined.

[1] The Perception-Distortion Tradeoff. CVPR 2018.

AuthorsOpenReview ↗

Response to Reviewer R8J3 (2/2)

Q3. We thank the reviewer for spotting this. The discrepancy comes from EMA usage rather than data splits or configuration: both use the same training run, test split and sampler. Table 3 reports SBFlow with EMA weights, whereas the NFE analysis in Appendix F was generated from the non-averaged weights of the same checkpoint. BRACE is trained without EMA (Table 7), so its values match across tables and figures. The same issue affects the SBFlow curves in Figs. 4 and 5. We have regenerated Figs. 4–6 with EMA weights for SBFlow. At NFE = 1, the corrected values coincide with Tables 1–3, and BRACE outperforms SBFlow at every NFE on LOL, SOTS-Indoor and EUVP. The corrected figures, together with the new tables from this rebuttal, are available at https://anonymous.4open.science/r/BRACE_ICLR_Rebutal-14E0/. We will also state the EMA setting of each baseline in Appendix D.

Q4. We trained the requested control on LOL. It is the BRACE U-Net used as a one-step translator G(x), without conditioning on timestep or direction, with the same discriminator and PatchNCE configuration. A direct unpaired translator has no regression target, so it uses adversarial + PatchNCE only. The results are in the Table R5 under the name “Direct translator”. We also report a second control, requested by another reviewer (Zrbh): the same network trained with BRACE's full loss set but without bridge states.

Table R5.

LOLBridgeMSEAdvNCELPIPS↓PSNR↑SSIM↑FID↓
Direct translator––✓✓0.202‾\underline{0.202}21.390.58492.2
Same losses, no bridge–✓✓✓0.39615.350.333126.1
BRACE w/o Adv✓✓–✓0.4758.430.054146.7
BRACE w/o NCE✓✓✓–0.21318.820.45585.9‾\underline{85.9}
BRACE✓✓✓✓0.19320.19‾\underline{20.19}0.470‾\underline{0.470}71.2

All rows share the U-Net, discriminator, PatchNCE, optimizer, data, steps, seed and checkpoint selection; NFE = 1. Best in bold, second underlined.

The direct translator achieves higher PSNR and SSIM, but its FID is about 30% higher (92.2 vs 71.2) and its LPIPS is also worse. It trades realism of the clean domain for lower distortion, which is the perception–distortion trade-off. Giving the bridge-free network BRACE's full loss set does not close this gap; it degrades every metric (row 2). The adversarial and PatchNCE terms alone therefore do not explain BRACE's LPIPS and FID. These gains require the bridge.

We have uploaded a revised paper that addresses the points raised in this review. The GT-mean evaluation on LOL and LOL-Real is now Table 2, and the inference cost is Table 3, both discussed in Sec. 5.2. The bridge-free controls, including the direct translator, are now Table 8 and are discussed in Sec. 5.3, with details in Appendix D. Figs. 4–6 are regenerated with EMA weights for SBFlow, and their values at NFE = 1 match Tables 1, 4 and 5. Appendix D now states the EMA setting. Because of the new tables, the ablation (Table 5 in our responses) is now Table 7, and the EUVP results (Table 3 in our responses) are now Table 5. We hope the revision addresses the reviewer's concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.