- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper presents SwiftLight, a one-step low-light enhancement framework combining staged diffusion distillation, trajectory-aware supervision, adversarial refinement, and global–local patch correction. Experiments on three LOL benchmarks demonstrate improved perceptual quality and reduced inference time.
Critical strengths
1.The method achieves the best LPIPS and DISTS among the compared methods on all three benchmarks. 2.The global–local correction addresses illumination inconsistency during patch-based inference, and the manuscript is generally clear.
Critical weaknesses
1.Limited methodological novelty. Segment-wise consistency distillation, teacher–student training, and adversarial refinement are established techniques. The additional contribution of trajectory-aware weighting is insufficiently isolated. 2. Limited fidelity advantage. The method achieves neither the best PSNR nor the best SSIM on any evaluated dataset. On LOL-v2-real, its PSNR is approximately 3.21 dB below HVI-CIDNet, indicating a substantial fidelity trade-off despite improved perceptual metrics. 3.Incomplete ablation controls. Table 3 uses different training schedules across variants and does not separately evaluate trajectory weighting and darkness weighting, making their individual benefits difficult to assess.
Questions for authors
- What distinguishes the proposed distillation strategy from existing consistency distillation beyond spatial loss reweighting?
- What explains the PSNR gap on LOL-v2-real, and how much of it results from adversarial refinement?
- Can the authors clarify the contributions of trajectory weighting and darkness weighting under matched training budgets? 4.Were runtime comparisons conducted using identical hardware, precision, and comparable patch settings? The exceptionally large MDMS runtime warrants clarification.