acceptodds
Under review as a conference paper at ICLR 2027

When Capture Is Not Enough: Synthetic Data for Real-World Arbitrary-Scale Image Super-Resolution

Abstract

Achieving high-quality arbitrary-scale super-resolution (ASSR) results is crucial for real-world multimedia applications. However, existing real-world SR datasets are primarily constructed through multi-camera synchronization with post-alignment, beam splitters for direct paired capture, or zoom lenses for sequential imaging. Yet these approaches either fail to provide fine-grained scaling factors required for arbitrary SR or struggle to preserve degradation consistency in dynamic scenes due to temporal misalignment. Therefore, these methods are not suited for capturing real-world datasets for arbitrary-scale super-resolution. Since capturing genuine arbitrary-scale datasets is infeasible, we construct a real-world SR dataset with truly arbitrary-scales through a degradation-aware synthesis approach. Specifically, we extract degradation characteristics from real low-resolution (LR) images and employ them to synthesize LR images at arbitrary scales, ensuring that synthesized data faithfully reflects real-world degradation patterns. We further propose the DP-ArbSR, a Dual-Prior Arbitrary-Scale Super-Resolution Network, which explicitly incorporates both degradation priors and texture priors to enhance feature representation and local feature fusion. Extensive experiments demonstrate that our synthesized dataset enables models to generalize robustly to real-world degradations, and DP-ArbSR achieves competitive performance. The dataset and the implementation code used in this work will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
arbitrary-scale super-resolution, real-world super-resolution, synthetic data, degradation modeling, dual-prior network

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_M421OpenReview ↗
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
  1. 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.
  2. 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?
  3. 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.
  4. 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
AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

Reviewer_UZ7rOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the limited availability of practical training data for real-world arbitrary-scale image super-resolution. The authors propose a degradation-aware synthesis framework that learns fine-grained degradation embeddings using rank-based supervision and uses these embeddings to generate scale-conditioned adaptive kernels for realistic LR image synthesis. They also introduce DP-ArbSR, which combines degradation and diffusion-based texture priors for feature extraction and local fusion. Experiments on COZ, RealArbiSR, and synthesized test scenes show consistent improvements over existing ASSR methods and degradation-synthesis baselines, suggesting better generalization to real-world degradations.

Critical strengths
  1. The paper addresses a practical limitation of constructing real-world datasets for arbitrary-scale super-resolution. Capture-based datasets make it difficult to obtain densely sampled scales while maintaining spatial and temporal consistency, especially for dynamic scenes. The proposed framework offers a practical alternative by synthesizing finely sampled training data while retaining degradation characteristics learned from real LR images. It can also extend the training data to scenes that are difficult to collect using conventional multi-scale capture.

  2. The experimental results provide convincing evidence that the synthesized data itself is useful. The authors evaluate the proposed synthesis method with several representative arbitrary-scale SR models, rather than restricting the evaluation to DP-ArbSR. Across different SR architectures, training with the proposed synthetic data consistently outperforms alternative data-generation methods, including bicubic downsampling, DegAE, and InterFlow. This makes it more likely that the improvements come from the proposed synthesis framework rather than from a specific reconstruction architecture.

Critical weaknesses
  1. Content dependence of the degradation representation.

The synthesis framework assumes that a degradation embedding extracted from a reference LR image can be transferred to a different HR image while preserving the intended real-world degradation. It is not clear, however, how well the embedding separates degradation characteristics from the content of the reference image. Real-world degradation can depend on local signal properties such as brightness and texture, whereas the DFE extracts a single degradation embedding from the reference image and uses it to generate adaptive degradation kernels.

A controlled cross-content experiment would make this assumption easier to assess. For example, the authors could fix the source HR image and target scale while varying the reference LR image across substantially different brightness and texture conditions. The resulting images could then be examined to determine whether the intended degradation remains consistent or changes with the reference content. This would provide clearer evidence of whether the learned embedding captures transferable degradation characteristics or retains scene-specific cues.

Questions for authors
  1. How robust is the degradation representation when the reference LR image and target HR image differ substantially in scene characteristics, such as brightness or texture? Could the authors provide a controlled cross-content analysis to examine whether the degradation embedding remains consistent across different reference contents?

  2. DegFlow (AAAI 2026) also studies continuous real-world degradation synthesis for arbitrary-scale super-resolution. It learns degradation trajectories from real HR-LR observations and generates LR images at continuous scales from a single HR input. Given the closely related problem setting, could the authors provide a direct comparison with DegFlow using a matched training and evaluation protocol?

2 comments
AuthorsOpenReview ↗

Response to Reviewer UZ7r

Thank you for the review and for the concrete suggestion of a cross-content experiment.

Weakness 1 and Question 1: content dependence of the degradation representation.
We ran this experiment (new Appendix C.3 and Fig. 8). We fix the HR image and the target scale, and vary the reference across four real LR captures at the same scale that span dark/bright and flat/textured content (20 HR images, ×2.0–×3.9). We re-embed the synthesized images with the DFE. Images synthesized from different references are closer to each other (cosine similarity 0.978) than images synthesized from the same reference at scales 0.5 apart (0.952).

We also take references from different regions of the same capture, which holds the real degradation fixed. Reference brightness has little effect on the synthesized degradation (about 0.2 dB in PSNR relative to bicubic). Reference texture density has a moderate effect (about 1 dB). We list this texture dependence as a limitation in Sec. 5. The global tone of the output follows the HR image, because the per-channel mean and variance of the synthesized LR image are aligned to the downsampled HR image; we now state this step in Sec. 2.4. When the dataset is constructed, each sample mixes the embeddings of two randomly drawn references (Sec. 2.4), so this dependence appears as random variation across samples rather than as a fixed offset.

Question 2: comparison with DegFlow.
DegFlow and our method differ in three respects: native-resolution synthesis (our degradation is applied at the target resolution and supervised by real captures at that resolution, whereas DegFlow generates at HR resolution and then downsamples), reference-conditioned transfer (our degradation is specified by an unpaired real reference, whereas DegFlow learns a trajectory from one camera's registered pairs), and supervision density (continuous-zoom real captures covering many scales in ×1.0–×4.0 per scene, versus the ×1, ×2 and ×4 captures of DegFlow's training pairs). A fair quantitative comparison would require training DegFlow on the same COZ captures and then retraining the SR models on data synthesized by each method. The released DegFlow model was trained on registered pairs from a different camera and covers scales up to ×4, so it does not provide such a matched setting. We therefore make no quantitative claim against DegFlow. DegFlow is now cited and discussed in Sec. 1 and Appendix B.

Thank you for the detailed response. My initial concerns have been resolved, and I believe the paper retains its strengths. I would also suggest that a comparison with DegFlow using the same training dataset could be valuable for future readers. Therefore, I maintain my original score.

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

This paper tackles the problem os synthesizing real-world super-resolution (SR) dataset on arbitrary scales. A degradation extractor trained with RankIQA-style ranking on real zoom, which removes the need for a paired HR LR data to train the degradation model. They show the effectiveness of their synthesized SR dataset and further proposes an SR method that utilizes the degradation prior and the texture prior for effective real-world SR.

Critical strengths
  1. They tackle a new frontier problem on the SR degradation modeling domain, where they aim to train an arbitrary scale degradation synthesizer from an unpaired LR dataset, which can by highly challenging. This was achieved by applying the ranking loss that was previously used in RankIQA, which seems quite novel.

  2. The generated dataset and proposed SR method gives consistent performance boost on various settings.

  3. The paper contains detailed ablative studies and additional information about the dataset and the method as well as comparison with the closest method as InterFlow.

Critical weaknesses
  1. Only PSNR values are reported to show reconstruction accuracy. At least MS-SSIM and LPIPS need to be reported so that the readers have a sense of the nature of the output results that the trained SR models give.

  2. Some recent related works [1, 2] are not referenced and not compared with.

  3. Real-world degradation models are often not deterministic, rather stochastic. However, the authors don't provide any justification on the deterministic design of their LR generation method.

References:

[1] Kim, Hyeonjae, et al. "Continuous degradation modeling via latent flow matching for real-world super-resolution." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 7. 2026.

[2] Peng, Long, et al. "Towards realistic data generation for real-world super-resolution." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer AUgz

Thank you for the careful review.

Weakness 1: only PSNR is reported.
SSIM and LPIPS for all eight compared methods, on COZ and on our synthesized test set, were already reported in the appendix (now Appendix Table 5), and DP-ArbSR is best on both metrics at every scale. The key columns are now also in the main paper (new Table 4). MS-SSIM is a multi-scale variant of SSIM: the structural aspect it measures is covered by the reported SSIM, and the perceptual aspect by LPIPS.

Weakness 2: DegFlow and Peng et al. are not discussed.
We now cite and discuss both works in Sec. 1 and in the related-work section (Appendix B).

DegFlow (Kim et al., AAAI 2026). Both methods synthesize real-world LR images at continuous scales from HR images, and they differ in three ways. (i) Native-resolution synthesis: DegFlow generates LR images at the HR resolution and applies bicubic downsampling afterwards, which its authors note may introduce artificial degradation. Our degradation is applied after resizing to the target resolution and is supervised by real LR captures at that resolution. (ii) Reference-conditioned transfer: DegFlow learns one latent trajectory from registered HR-LR pairs at ×1, ×2 and ×4 of a single camera. Our generator is conditioned on the embedding of an unpaired real reference LR image, so changing the reference changes the synthesized degradation without retraining. (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, whereas DegFlow is supervised only at the ×1, ×2 and ×4 captures of its training pairs.

A fair quantitative comparison would require training DegFlow on the same COZ captures and then retraining the SR models on data synthesized by each method. The released DegFlow model was trained on registered pairs from a different camera and covers scales up to ×4, so it does not provide such a matched setting. We therefore make no quantitative claim against DegFlow.

Peng et al. (RealDGen, ICLR 2025). This work also conditions LR generation on real LR references, but it targets fixed-scale SR (×4) and has no scale control. Its degradation extractor is pre-trained with contrastive pairs built from a synthetic (Real-ESRGAN) degradation model, whereas our ranking supervision comes directly from the relative degradation order of real multi-scale captures. Its diffusion generator is stochastic and relies on post-hoc sample filtering; ours is a lightweight deterministic network that outputs LR images directly at the target resolution. Because RealDGen is fixed-scale, a matched arbitrary-scale comparison is not straightforward.

Weakness 3: the deterministic design is not justified.
The synthesis network is a deterministic mapping of (HR image, target scale, reference embedding). The variability of the dataset is stochastic: for each HR image and target scale, two reference embeddings are drawn at random from real LR observations at that scale and mixed with a random weight (Sec. 2.4). The variation in synthesized degradations therefore follows the empirical distribution of real captures rather than a hand-designed noise model. Appendix Fig. 8 shows the same HR image and scale synthesized with different references. DegFlow also uses a deterministic bridge. Random noise inside a single image is not modeled explicitly; we now state this as a limitation in Sec. 5.

Thank you for your response. W1 is resolved and W2 and W3 are partly resolved. The strengths of this paper still remain, and I will stick to my initial rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank the reviewers and the AC for their feedback. We have uploaded a revised PDF with the following main changes.

Related work (all reviewers). DegFlow (Kim et al., AAAI 2026) and RealDGen (Peng et al., ICLR 2025) are now cited and discussed in Sec. 1 and Appendix B. The statements on InterFlow in Sec. 1, Appendix B and Appendix D.3 are rewritten as factual descriptions.

Synthesis procedure (Sec. 2.4). We now describe how references are selected (DFE embeddings of real LR images from the COZ training split, 50 scale intervals, two embeddings mixed with α ∈ [0.2, 0.8], no test images used as references) and the per-channel mean-variance alignment of the output to the downsampled HR image.

Cross-reference consistency (new Appendix C.3, Fig. 8). A controlled experiment that fixes the HR image and scale and varies the reference in brightness and texture, and a linear probe that separates real from bicubic LR images.

Degradation encoders (Sec. 3.2, Appendix C.2, D.1, D.4). The ResNet-50 DFE is used for data synthesis; the degradation prior inside DP-ArbSR uses the DegAE-RL encoder. The misplaced sentence on frozen SRGAN weights in Appendix D.1 is corrected, and the SROCC of the DFE on held-out COZ scenes (0.88) is reported.

Attribution of the diffusion prior (Table 2, Table 3(b), Sec. 4.4, Appendix Table 12). Table 2 now groups methods by encoder and by whether a pretrained prior is used, and includes DP-ArbSR without the texture prior. Three new ablation variants remove the texture prior from the Plug-and-Play module, from DAFF, and from every pathway. The definition of variant (a) is stated explicitly.

Additional results. SSIM/LPIPS key columns in the main text (new Table 4; all methods in Appendix Table 5); downstream PSNR of the DFE sampling-granularity variants next to the embedding similarity (Appendix Table 13); real non-integer scales on COZ (Appendix Table 9); training scale ranges of all data sources (Appendix Table 10).

Appendix table numbers have shifted (e.g., the InterFlow comparison is now Table 11); our replies use the revised numbering.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.