- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
To address the limited reconstruction fidelity of VAE-based generative methods, this paper adapts a pretrained pixel-space generative model for one-step real-world image super-resolution. It introduces a Cross-Scale Wavelet Discriminator (CSWD) and Mixup Contrastive Self-Distillation (MixCSD) to mitigate patch artifacts, and Selective LoRA Pruning (LoRA-SP) to improve the fidelity–perception trade-off. It also introduces Dual-Q-Bench, a paired real-world super-resolution benchmark with sharp, texture-rich reference images designed to enable more reliable evaluation of both reconstruction fidelity and perceptual quality.
Critical strengths
[1] Adapting a large pixel-space generative prior to faithful, one-step restoration has practical value.
[2] The paper presents detailed ablation studies, evaluates the method across a broad range of datasets, and reports a comprehensive set of evaluation metrics.
Critical weaknesses
[1] The methodological contribution lies primarily in the engineering integration of existing techniques, including pixel-space generation, feature-based discrimination, contrastive learning, and LoRA.
[2] The method relies on several heuristic design choices, while key implementation and experimental details remain underspecified, raising concerns about reproducibility (see Questions for Authors [1] and [2]).
Questions for authors
[1] Section 3.4 describes LoRA-SP as a training-time mechanism that progresses from temporary windowed freezing to permanent freezing. Could the authors specify when probing begins, how frequently the CKA statistics and layer selections are updated, how long temporary freezing lasts? What operational criterion determines that the redundancy signal has stabilized sufficiently for permanent freezing? Please also clarify how the soft-decay fallback is implemented when the pruning budget is saturated. An explicit algorithm or pseudocode with these scheduling parameters would improve reproducibility.
[2] The InfoNCE formulation is explicit, but claims about “boundary-consistency” and “preventing representational collapse” require stronger evidence. Moreover, the specific settings for the activation step , and target weight $ \lambda_{\mathrm{InfoNCE}}^{\max}$ are not reported.
[3] What motivates the choice of the fixed timestep , and how should this time condition be interpreted after replacing Gaussian noise with degraded images? Please provide an ablation across different fixed timesteps, using matched training and inference settings, to assess their effects on reconstruction fidelity, perceptual quality, and patch-boundary artifacts.
[4] The connection between patch-level InfoNCE discrimination and cross-patch boundary continuity remains unclear. The objective in Eq. (5) directly encourages feature matching at corresponding spatial locations and discriminability between different patches, without explicitly enforcing continuity across adjacent patch boundaries.
[5] Could the authors compare MixCSD against simpler spatial strategies, such as overlapping tiles with spatial offsets and output blending? Such a comparison, including boundary-artifact measurements, texture preservation, and inference overhead, would help establish whether MixCSD provides benefits beyond straightforward spatial averaging.
3 comments
I sincerely thank the authors for their detailed response and additional experiments. Most of my questions have been answered, and I appreciate the practical value of this work. I will therefore raise my rating from 2 to 3.
Response to Reviewer ph1a’s Updated Assessment
Thank you for your positive feedback. We are glad that our response has addressed your concerns, and we sincerely appreciate you raising your review score. Thank you again for your time and thoughtful evaluation.
Response to Reviewer ph1a
Thank you for recognizing the practical significance of our work and comprehensiveness of our evaluations. Revisions to paper are highlighted in blue.
Q1:Method contribution. We respectfully clarify that HiFiPix is not a simple combination of existing techniques, but a set of tailored designs for distinct challenges in pixel-space Real-ISR: CSWD goes beyond discrimination by explicitly focusing on multi-scale high-frequency details; MixCSD goes beyond contrastive learning by using asymmetric distillation to improve patch-wise reconstruction; and LoRA-SP goes beyond LoRA fine-tuning by adaptively tuning layer updates to retain fidelity and generation. Their complementary roles are validated by ablations.
Q2:Details We will release codes to improve reproducibility.
LoRA-SP: Detection starts at iteration 200 to avoid early pruning instability. CKA statistics and layer selection are updated every 100 iterations, with temporary freezes lasting 400 iterations. Redundancy signals stabilize at 80%–90% of training, as indicated by minimal changes in selected layers. We thus apply freezing at 85% to ensure consistency between training and inference. Once pruning budget K is exhausted, new candidates undergo soft decay rather than freezing: lora_B is scaled by 0.7, while lora_A remains unchanged, reducing their contribution while retaining parameter updates. These hyperparameters were empirically selected. We have added Algorithm 1 of LoRA-SP in revised paper.
InfoNCE: InfoNCE promotes boundary consistency by learning token representations from the LQ-GT mixed teacher branch, which provides more accurate high-frequency features for each token and reduces high-frequency discontinuities across tokens in student. We measure gradient strength along patch boundaries on 50 test images. In below table, InfoNCE significantly improve boundary consistency.
Stop-gradient in InfoNCE prevents representational collapse by excluding collapse modes in which teacher and student branches jointly converge to same representation. Without it, we find average cosine similarity between teacher and student tokens increases from 0.81 to 0.97, indicating a loss of feature diversity.
tstart is 500 iters, λInfoNCEmax is 0.05. We have added these in Tab.19 of revised paper.
Q3:Fixed t=0.5. We choose a fixed mid-range timestep t=0.5 to balance fidelity and generative capacity. In pre-trained HiDream's flow-matching, xt=t×GT+(1−t)×Noise, so t=0.5 gives equal contributions from input signal and noise, providing a balance between content preservation and generation that suits SR.
When Gaussian noise is replaced with degraded images, t no longer represents degradation or noise level. Instead, it serves as a conditioning signal that determines model's operating mode. Specifically, t=0.5 indicates a balanced treatment of fidelity and generation.
Ablations on 50 images further confirm that t=0.5 achieves an optimal balance between fidelity and generation. t doesn't affect patch artifacts. We have added these in Tab.17 of paper.
Q4:Patch Boundary Continuity in InfoNCE. Although InfoNCE explicitly matches spatially aligned tokens, it indirectly improves cross-patch boundary consistency by improving each token's high-frequency recovery. Specifically, LQ-GT mixed teacher inputs provide more reliable high-frequency features, guiding student tokens to recover GT-consistent details and reduce high-frequency gaps across neighboring patches. Fig.3(a) of main text show inputs with richer high-frequency details yield better cross-patch consistency. We have tried directly enforcing pixel smoothness at adjacent patch boundaries but saw limited gains, possibly because local constraints cannot capture wider high-frequency links across patches. Thus, our token-level supervision better promotes cross-patch consistency. Table in Q2-2 also confirms the effectiveness of InfoNCE.
Q5:Compare with overlapping tiles. We perform two forward passes per image: one on original tile grid and the other on a grid shifted by (16, 16) pixels, using reflection padding at boundaries. Predictions from two passes are averaged pixel-wise in overlapping regions.
Table shows that overlapping slightly improves PSNR/LPIPS and patch boundary gradient, but averaging-induced smoothing severely harms gerenation metrics and doubles inference. In contrast, MixCSD better balances fidelity, generation, and speed.