- 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
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.
| LOL | LOL-Real | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Method | NFE | LPIPS↓ | PSNR↑ | SSIM↑ | ms↓ | MB↓ | LPIPS↓ | PSNR↑ | SSIM↑ | ms↓ | MB↓ |
| CycleGAN | 1 | 0.186 | 0.697 | 0.295 | 0.645 | ||||||
| CUT | 1 | 0.218 | 23.29 | 0.652 | 9.0 | 1169 | 0.307 | 22.70 | 0.614 | 8.3 | 1169 |
| EnlightenGAN | 1 | 0.216 | 20.45 | 0.652 | 7.1 | 1336 | 0.211 | 21.43 | 0.649 | 5.6 | 1336 |
| GDP | 1000 | 0.317 | 19.47 | 0.498 | 7.2×10⁵ | 3469 | 0.331 | 19.36 | 0.544 | 7.2×10⁵ | 3469 |
| AGLLDiff | 10 | 24.17 | 0.755 | 1.6×10⁴ | 6909 | 0.173 | 26.74 | 0.794 | 1.6×10⁴ | 6909 | |
| ASBM | 4 | 0.414 | 19.50 | 0.534 | 134.3 | 2974 | 0.471 | 18.54 | 0.492 | 129.7 | 2974 |
| DDSB* | 5 | 0.466 | 19.41 | 0.447 | 206.5 | 1573 | 0.495 | 19.75 | 0.435 | 208.0 | 1573 |
| UNSB | 5 | 0.247 | 22.86 | 0.578 | 206.2 | 1663 | 0.293 | 23.11 | 0.581 | 205.6 | 1663 |
| SBFlow | 1 | 0.371 | 17.42 | 0.457 | 37.4 | 2644 | 0.279 | 23.13 | 0.668 | 33.3 | 2658 |
| BRACE (ours) | 1 | 0.174 | 25.37 | 44.4 | 1655 | 23.53 | 33.5 | 2434 |
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.
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.