- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
To address the challenge of obtaining spatiotemporally aligned training data for arbitrary-scale super-resolution (SR) tasks in real-world scenarios, this paper proposes a "degradation-aware" synthetic pipeline. This pipeline incorporates a degradation feature extractor—trained via a ranking-based approach—and a scale-adaptive kernel generator, which together enable the transfer of degradation characteristics from reference low-resolution (LR) images to new high-resolution (HR) contexts. Furthermore, we introduce the DP-ArbSR model, which integrates degradation embeddings, pre-trained diffusion-based texture priors, and a local dual-attention feature fusion mechanism. We evaluate the synthesized data across various SR architectures and compare the performance of DP-ArbSR against existing methods using both real-world and synthetic benchmark datasets.
Critical strengths
The data synthesis task aims to address a significant practical limitation: extending real-world degradation processes to scenarios where obtaining "aligned multi-scale image pairs" is difficult. By conducting experiments on the same underlying scenes, researchers can isolate the effects of "degradation synthesis" from those of the "scene content" itself. In settings based on real-world captured scenes, Table 3 shows that, compared to DegAE, the proposed method achieves average performance gains of 0.354 dB, 0.310 dB, and 0.588 dB on the COZ ×4, COZ ×6, and RealArbiSR ×4 datasets, respectively. The performance improvements observed across various super-resolution (SR) architectures strongly demonstrate that the value of this dataset extends beyond any single, specific network model. 2. The experiments cover several complementary experimental settings. The supplementary material includes detailed results for each model, comparisons with InterFlow, SSIM/LPIPS metric measurements, architectural ablation studies, inference time analysis, and two user studies. DP-ArbSR consistently achieves improvements across various reconstruction metrics; its performance in terms of LPIPS and its qualitative results further indicate that the method excels in enhancing perceptual quality. Furthermore, Table 9 shows that, despite having a larger parameter count, DP-ArbSR is not necessarily slower in terms of inference speed than all the other implicit super-resolution models included in the comparison.
Critical weaknesses
1、The authors report that DegFlow—Continuous Degradation Modeling via Latent Flow Matching for Real-World Super-Resolution (Kim et al., AAAI 2026)—already synthesizes degradations at continuous scales from a single HR image at inference and evaluates the resulting data for arbitrary-scale SR. This directly overlaps with the paper’s central motivation and is absent from the discussion and comparisons. The contribution should therefore be distinguished more precisely in terms of reference-conditioned degradation transfer, native-resolution synthesis, and dataset coverage. A controlled comparison would help establish the practical value of these differences.
2、 Appendix D.3 attributes much of InterFlow’s disadvantage to misregistered endpoint images. This supports robustness to imperfect alignment, but does not isolate the claimed superiority in modeling nonlinear degradation trajectories. Please add an evaluation on an accurately aligned subset to separate these explanations. 3、 The current evidence does not sufficiently establish what the degradation embedding represents. Equation (2) supervises a scalar ordering based on image resolution. This can encourage scale discrimination, but does not by itself establish that the retained 512-dimensional embedding captures transferable degradation characteristics independently of scene content. In particular, lower cosine similarity between adjacent-scale embeddings, as reported in Table 10, is not sufficient evidence of better degradation modeling: greater separation can also reflect sensitivity to resampling or texture differences.
4、Though the downstream results are encouraging, the ranking-specific ablation primarily reports embedding similarity. I suggest adding a comparison using the same extractor architecture with and without ranking, followed by identical synthesis and SR training, to better establish the mechanism.
5 、The quantitative evaluation does not prove the fine-grained, arbitrary-scale claim. Though the paper aims to realize continuous scales, the quantitative SR results use integer magnifications. Suggset add quantitative evaluation at real fractional scales, such as ×2.5 and ×3.5 in RealArbiSR. Furthermore, results on synthetic test measure generalization to new content under the same synthesis process; they cannot independently validate realism. The real benchmark results and user studies provide additional evidence, but direct evaluation of synthesized LR images against held-out real observations would strengthen the claim that the generator faithfully reproduces scale-dependent degradation. 6、The architectural contribution is not clearly separated from the pretrained diffusion prior. DP-ArbSR uses a substantial pretrained component, and Table 9 reports 181.2M parameters, compared with 3.15M for EDSR-DDIR. This does not invalidate the comparison, particularly given the reported inference speed, but makes attribution important. 7、The current ablation descriptions are ambiguous: variant (a) removes the plug-and-play module in Section 4.4, whereas the qualitative discussion describes it as removing the texture prior. Since texture features also enter DAFF, these are different interventions. An explicit ablation removing the diffusion prior from every pathway, together with a simple fusion baseline using the same pretrained features, would clarify the contribution of the proposed integration mechanisms.
Questions for authors
- How are reference LR images selected for each synthesized scene and target scale? Does the reference scale need to match the requested scale, and how is synthesis handled when they differ? Please specify the reference pool and its train/test separation.
- Please clarify the DFE implementation. Section 2.1 and Appendix C.2 describe a ResNet-50-based extractor, whereas Appendix D.1 refers to frozen pretrained SRGAN weights. Which architecture produces the final embedding, and what are its held-out SROCC results?
- Do the improvements persist at held-out fractional scales? Please also specify the scale ranges used by every component and dataset in the mixed-training experiments, so that the meaning of “out-of-scale” can be assessed for the complete pipeline.
- What exactly is removed in variant (a)? Are texture features still supplied to DAFF? A precise description and results for a complete removal of the diffusion prior would materially affect my assessment of the network contribution.
4 comments
Response to Reviewer M421 (2/2)
(Part 2 of 2: Weaknesses 6–7 and Questions 1–4.)
Weakness 6: separating the architecture from the pretrained diffusion prior.
The parameter count of DP-ArbSR in Appendix Table 12 includes the frozen pretrained diffusion prior; the caption now states this. Without any diffusion prior (DP-ArbSR-T3), DP-ArbSR obtains 28.92 / 26.83 / 25.40 dB at ×2 / ×3 / ×4, still above every EDSR-based ASSR baseline in Table 2 (best of these at each scale: 28.87 / 26.75 / 25.33 dB). The frozen prior adds a further 0.12–0.15 dB on top of this. Table 2 now groups methods by encoder and by whether a pretrained prior is used, and lists DP-ArbSR-T3 in the EDSR group.
Weakness 7 and Question 4: ablation descriptions.
Variant (a) removes the entire Plug-and-Play module, which includes both the texture-prior injection and the degradation-prior injection; DAFF still receives texture features. Sec. 4.4 now states this explicitly. Three new variants on COZ (×2 / ×3 / ×4 PSNR, Table 3(b)) remove the texture prior from Plug-and-Play (T1: 28.96 / 26.89 / 25.44 dB), from DAFF (T2: 29.00 / 26.91 / 25.46 dB), and from every pathway (T3: 28.92 / 26.83 / 25.40 dB). The full model obtains 29.05 / 26.98 / 25.52 dB. A simple feature-fusion baseline using the same pretrained features is not included in this revision.
Question 1: reference selection.
Synthesis takes an HR image, a target scale s, and the degradation embedding of one real reference LR image; it does not require an aligned LR-HR pair or any real capture of the target scene. The reference scale need not equal the target scale. The reference pool consists of DFE embeddings of the real LR images in the COZ training split, grouped by capture scale into 50 scale intervals. For each HR image and target scale s, two embeddings are drawn at random from the interval containing s and mixed with a random weight α ∈ [0.2, 0.8], so the reference scale matches the target scale up to the interval width. No image from the COZ or RealArbiSR test sets is used as a reference. During training of the synthesis network, the reference is the real LR image of the same scene. This procedure is now described in Sec. 2.4.
Question 2: DFE implementation.
Thank you for catching this inconsistency. The DFE used for data synthesis is the ResNet-50-based extractor described in Sec. 2.1 and Appendix C.2 (ResNet-50 with a 512-d projection, trained with the ranking objective). The sentence on frozen SRGAN weights in Appendix D.1 describes the DegAE-RL variant (a frozen pretrained SRGAN extractor with only its downsampling layers retrained) and was misplaced; it is now moved to the DegAE-RL description in Appendix D.4. The degradation prior inside DP-ArbSR uses the DegAE-RL encoder, i.e., the SRGAN-based extractor retrained with our ranking objective; Sec. 3.2 now states this. On held-out COZ scenes, which are used neither to train the DFE nor as references, the DFE orders the real captures of each scene by degradation level with SROCC = 0.88 across the ×2–×4 captures (Appendix C.2).
Question 3: scale ranges and out-of-scale.
The training scales of all three sources lie in ×1.0–×4.0 (new Appendix Table 10; Appendix Fig. 7). Our synthesized data sample 100 scales per scene in this range; COZ provides continuous scales with more samples at low magnifications; RealArbiSR has six fixed scales (×1.5–×4.0). The DFE and the synthesis network are trained only on the COZ V2 training split (264 scenes; 90% of them for the DFE). "In-scale" (×2–×4) and "out-of-scale" (×5–×6) refer to test scales inside and outside this ×1.0–×4.0 training range.
Thank you for the detailed response, which addresses most of my previous concerns. However, I share the other reviewer’s concern about the missing experimental comparison with DegFlow. A useful comparison would generate equally sized datasets from the same HR images and scale factors, then train an identical SR backbone with the same training budget and evaluate on the same real-world test sets. Differences in generator training data and preprocessing should be clearly reported. This comparison is important for establishing the benefit of the proposed synthesis method. Given this remaining gap, I maintain my score.
Response to Reviewer M421
Thank you for the follow-up. We did not include this comparison earlier because the two generators are trained on different data: the released DegFlow generator is trained on registered pairs from a different camera (RealSR-V2), whereas our generator is trained on COZ, so a direct comparison could not distinguish whether a difference comes from the training data of the two generators or from the methods themselves. The required SR training cannot be completed within the remaining discussion period. We will compare three generators: the released DegFlow, DegFlow retrained on the same COZ data as ours, and ours. Each synthesizes an equally sized dataset from the same HR images. On each dataset we train the same standard arbitrary-scale SR backbone that is not part of our method (LIIF), with the same training budget, and evaluate it on real-world images. The results will be added to the supplementary material, with a statement in the main text.
Response to Reviewer M421 (1/2)
Thank you for the detailed review. (Part 1 of 2: Weaknesses 1–5.)
Weakness 1: DegFlow.
We now discuss DegFlow in Sec. 1 and in the related-work section (Appendix B), and distinguish our contribution in three respects. (i) Native-resolution synthesis: our degradation is applied at the target resolution and supervised by real captures at that resolution, whereas DegFlow generates LR images at the HR resolution and then applies bicubic downsampling. (ii) Reference-conditioned transfer: the degradation is specified by the embedding of an unpaired real reference LR image, rather than by a trajectory learned from one camera's registered HR-LR pairs. (iii) Supervision density: our DFE and synthesis network are supervised by the continuous-zoom real captures of COZ, which cover many scales in ×1.0–×4.0 within each scene, while DegFlow is supervised only at the ×1, ×2 and ×4 captures of its training pairs. We make no quantitative claim against DegFlow; our response to Reviewer UZ7r explains what a matched comparison would require.
Weakness 2: InterFlow.
Appendix Table 11 compares the two synthesis methods on RealArbiSR. As you note, part of the roughly 0.9 dB gap can come from imperfectly aligned ×2 and ×4 endpoint pairs, so we do not use this comparison to claim better modeling of the degradation trajectory. The difference we rely on is structural: InterFlow interpolates linearly in latent space between two aligned captures of the same scene, whereas our synthesis needs one unpaired real reference per scale and covers scales continuously, so it also applies to scenes without aligned endpoints. We have revised the corresponding statements in Sec. 1 and Appendix D.3.
Weakness 3: what the embedding represents.
New Appendix C.3 adds two measurements. First, cross-reference consistency: in the controlled cross-content experiment described in our response to Reviewer UZ7r, changing the reference changes the synthesized images less than changing the scale by 0.5 (cosine similarity 0.978 vs. 0.952). Second, the embedding encodes more than resolution: at identical image size, real and bicubic-downsampled LR images of the same scene are separated by a linear probe with 85% accuracy on held-out scenes (COZ test set). Sec. 4.4 now states that the adjacent-scale similarity is only a diagnostic of how the embedding responds to scale, not evidence of better degradation modeling.
Weakness 4: ranking-specific ablation.
Appendix C.2 reports the downstream effect of the embedding. When the DFE is trained with coarser scale sampling (35, 25, and 10 scales per scene) and then used for synthesis, DP-ArbSR on COZ drops by 0.07 / 0.08 / 0.10 dB (35 scales), 0.19 / 0.23 / 0.16 dB (25 scales), and 0.33 / 0.42 / 0.41 dB (10 scales) at ×2 / ×3 / ×4. These drops follow the same order as the embedding similarity (0.79, 0.86, and 0.92 for 35, 25, and 10 scales, against 0.73 for the final DFE). Appendix Table 13 now lists similarity and downstream PSNR together.
Weakness 5: fractional scales and realism.
On COZ we evaluate at non-integer scales that are real optical-zoom pairs rather than resampled images. At ×2.5, ×3.5, and ×5.5, DP-ArbSR obtains 27.64, 25.81, and 24.05 dB, compared with 27.53, 25.72, and 23.94 dB for LMI-RDN and 27.54, 25.65, and 23.88 dB for DDIR-RDN, i.e., 0.10, 0.09, and 0.11 dB above the strongest baseline. These values fall between the neighbouring integer-scale results (new Appendix Table 9).
On realism, the synthetic test set measures generalization to new content rather than realism. Our evidence for realism is downstream and on real captures: across five SR architectures (LIIF, LTE, LMI, DDIR, and DP-ArbSR), models trained on our synthesized data outperform those trained on bicubic-downsampled data from the same scenes by 0.47, 0.45, and 0.69 dB, and those trained on DegAE data by 0.35, 0.31, and 0.59 dB in mean PSNR on the real COZ ×4, COZ ×6, and RealArbiSR ×4 test sets (Table 3(a)). Both user studies also use real-world inputs. Sec. 4.2 now states this explicitly.