acceptodds
Under review as a conference paper at ICLR 2027

PhoenixSR: Generative Heterogeneous Distillation Unleashes Efficient Models for Real-World Super-Resolution

Abstract

Real-world image super-resolution (SR) requires recovering perceptually realistic high-resolution images from complex low-resolution observations while preserving faithful content. Diffusion-based SR benefits from strong generative priors but incurs substantial computational overhead, whereas feed-forward CNN and Transformer SR models are efficient yet often struggle to recover realistic high-frequency details. This motivates a natural question: can diffusion priors be transferred to existing diffusion-free SR networks without introducing diffusion components at inference time? To this end, we propose PhoenixSR, a generative heterogeneous distillation framework that transfers diffusion priors to independently designed feed-forward SR networks through score-based distribution matching. Rather than aligning heterogeneous features or imitating sampled diffusion outputs, PhoenixSR uses the pretrained diffusion model as distribution-level supervision, while paired SR supervision preserves reconstruction fidelity. To make distribution matching effective for fidelity-sensitive SR, we introduce Heterogeneous Distribution Adaptation, which adapts the target score to the SR domain, improves tracking of the evolving student distribution, and anchors training with paired supervision. We further employ Directional Reliability Weighting, a lightweight residual-consistency-based reweighting strategy that reduces unstable distributional guidance. All diffusion-related components are removed after training, leaving the original student architecture and inference cost unchanged. Experiments on three SR benchmarks and six feed-forward backbones, including SwinIR, HAT, Real-ESRGAN, and SeeMoRe, show consistent perceptual improvements with largely preserved reconstruction fidelity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Real World Super-resolution, Heterogeneous distillation, Distribution Matching Distillation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aMRpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhoenixSR, a framework for transferring pretrained diffusion-model priors into existing feed-forward super-resolution networks without modifying their inference architectures. The method applies distribution matching distillation in the latent space of Stable Diffusion 2.1, using a real-score model and an online fake-score model to supervise student outputs. Also, the paper proposes two main components: 1) HDA: The real-score model is adapted to the target HR distribution using LoRA, while the fake-score model is regularized through representation alignment with DINOv3 features. 2) DRW: The distribution-matching objective is reweighted according to the directional consistency of denoising residuals around generated and HR latents.

Critical strengths
  1. Practical motivation: Transferring diffusion-based supervision into existing feed-forward SR networks is a practically relevant problem. PhoenixSR does not require changes to the student architecture or additional diffusion computation at inference time. Its applicability to CNN, Transformer, and lightweight SR backbones is an important strength.

  2. Broad empirical evaluation: The experiments cover six student networks, three benchmarks, and multiple reconstruction and perceptual metrics. The supplementary material further includes three-seed experiments for two students and a user study with statistical analysis. This provides reasonably broad evidence of the method's effectiveness, although the robustness experiments cover only a subset of the reported settings.

  3. Controlled comparison with alternative transfer strategies: Table 4 compares PhoenixSR against task-only fine-tuning, output-level knowledge distillation, SDS, and DMD2-based transfer under a common initialization, dataset, and student-update budget. Including a strong task-only fine-tuning baseline is particularly valuable because it helps distinguish the effects of diffusion-based supervision from ordinary continued training.

Critical weaknesses
  1. Novelty: The framework combines established components: DMD-style real/fake score matching, LoRA adaptation, representation alignment, paired reconstruction losses, timestep restriction, and sample-wise reweighting. The individual technical components are largely adaptations of existing approaches, and the paper offers limited analysis of why their combination is particularly effective.

  2. Effects of each component: The main results in Tables 1–3 compare PhoenixSR against the original student checkpoints. However, these comparisons do not establish how much improvement comes specifically from diffusion-prior transfer rather than additional paired SR training. Table 4 includes an equal fine-tuning baseline, but only for three students on RealSR under an 80K-iteration setting. The authors explicitly note that these results are not directly comparable to the full-schedule experiments.

  3. Performance: The controlled results show substantial, but smaller, gains over task-only fine-tuning. For example, PhoenixSR improves MUSIQ from 56.0 to 63.4 for SwinIR-S and from 53.7 to 61.5 for SeeMoRe. These are meaningful improvements, but they should be distinguished from the larger gains reported relative to the original checkpoints. The reconstruction results are also mixed. In Table 4, PhoenixSR slightly improves PSNR over task-only fine-tuning for SwinIR-L and SwinIR-S, but reduces it for SeeMoRe. Table 5 further shows that SFA-only training has almost no effect on SwinIR-L PSNR in the controlled setting, although this does not establish its contribution over the full schedule. More discussion on this issue would be beneficial.

  4. Non-reference-based metrics-oriented: Most of the reported perceptual improvements are measured using no-reference image-quality metrics, including MANIQA, MUSIQ, CLIP-IQA, LIQE, ARNIQA, TOPIQ, and NIQE. These metrics are useful for evaluating perceptual quality but do not directly establish whether generated details are faithful to the input scene. The diffusion teacher is also conditioned on a fixed quality-oriented prompt, potentially encouraging visually appealing high-frequency details. This makes the distinction between faithful restoration and plausible texture generation particularly important. There are also exceptions to the reported fidelity and perceptual trends, including a PSNR decrease for HAT on DIV2K-Val and small regressions in certain no-reference metrics for Real-ESRGAN.

  5. Effects of DRW: The authors motivate DRW as a mechanism for identifying reliable distribution-matching guidance. Tables 9 and 10 show that the proposed residual-direction consistency score correlates with guidance stability under independently sampled diffusion perturbations. This supports its use as a proxy for directional stability. However, the reliability and the stability are both based on related directional properties of the score-model guidance. Their correlation does not directly establish that samples receiving higher weights produce more beneficial student updates or improve reconstruction fidelity.

Questions for authors
  1. Perceptual fidelity: Can the authors provide additional reference-based perceptual evaluation, such as DISTS, together with an assessment of hallucinated details or faithfulness to the LR input?
5 comments
AuthorsOpenReview ↗

Q1. Insufficient analysis of the effectiveness of combining existing components

Using existing techniques does not imply a lack of methodological innovation. What matters is how these techniques address a previously unresolved transfer problem. To our knowledge, PhoenixSR is the first framework to use DMD to transfer pretrained diffusion priors to independently designed non-diffusion SR students that inherit neither the diffusion teacher's denoising architecture nor its pretrained weights, while preserving the students' original inference architectures. It is also effective on SeeMoRe with only 0.97M parameters and SwinIR-S with 1.05M parameters, requiring distributional priors to translate into actual reconstruction capabilities within the students' limited capacity and original inference costs.

These components act at different stages of distributional supervision. DMD constrains the overall output distribution, but distributional proximity does not automatically determine whether each output corresponds to its own LR input. SFA therefore maintains input correspondence through pixel and paired perceptual supervision, and supplements texture realism with an adversarial term. Real-score LoRA addresses the difference between generic diffusion priors and the HR target domain. The fake-score model is fully updated online to track the continuously changing student distribution, while REPA constrains the representations of this online estimator. The representation constraint is applied to the fake-score model that provides supervision, and the original SR student retains its own reconstruction structure.

Timestep truncation and DRW further regulate how distributional guidance is used. The former selects noise states that still retain more image structure, while the latter adjusts each sample's contribution to supervision according to the directional consistency of denoising residuals near the generated image and its paired HR image. Paired constraints preserve the content to be restored, score adaptation improves distributional supervision, and noise scheduling and sample weighting regulate how supervision is applied.

Table 5 of the paper has already conducted the above ablations under the same setting:

ConfigurationPSNR ↑MUSIQ ↑
Without SFA, retaining score adaptation and DRW20.710740.7941
SFA only26.312456.8245
SFA + score-model adaptation, without DRW26.158763.5729
Full PhoenixSR26.440964.3102

Paired supervision provides a fidelity constraint for distributional transfer. Adding distributional supervision and score adaptation substantially improves MUSIQ. Further adding DRW improves both PSNR and MUSIQ. Table 11 also separately examines fake-score REPA and real-score LoRA, with the two adaptations increasing PSNR from 25.9778 to 26.4409 dB. Table 12 shows that removing timestep truncation reduces PSNR to 25.6564 dB. This evidence corresponds to the roles of input fidelity, score estimation, and guidance conditions, respectively. The contribution of this paper is to establish and validate this distributional transfer method for independent non-diffusion students, enabling existing SR students, including networks with several million to over ten million parameters, to use diffusion priors.

AuthorsOpenReview ↗

Q2. Full-training experiments do not distinguish the contributions of diffusion-prior transfer and additional paired training

The total gains relative to the original checkpoint and the incremental gains after controlling for additional training should be addressed using different references. Tables 1–3 show the changes in students after full training relative to their original checkpoints. Table 4 compares PhoenixSR with task-only fine-tuning under the controlled 80K setting on RealSR, using the same initialization, training data, paired task supervision, and student-update budget. The latter has already demonstrated perceptual gains beyond ordinary continued training on SwinIR-L, SwinIR-S, and SeeMoRe.

Following your suggestion, we add a task-only fine-tuning comparison under the full training schedule. Each student starts from the same checkpoint and uses the same training data, paired task supervision, number of student updates, and model-selection rule, and is compared with the corresponding PhoenixSR result on RealSR.

StudentMethod and data sourcePSNR ↑LPIPS ↓MUSIQ ↑
SwinIR-LTask-only fine-tuning (additional comparison)26.070.28258.24
SwinIR-LPhoenixSR (Table 1 of the paper)26.740.25061.36
SwinIR-STask-only fine-tuning (additional comparison)25.070.34260.40
SwinIR-SPhoenixSR (Table 1 of the paper)25.250.30365.34
SeeMoReTask-only fine-tuning (additional comparison)24.980.30957.83
SeeMoRePhoenixSR (Table 1 of the paper)25.930.26662.41

With the full training budget matched, PhoenixSR achieves higher PSNR and MUSIQ and lower LPIPS on SwinIR-L, SwinIR-S, and SeeMoRe. In particular, SeeMoRe's PSNR also increases from 24.98 to 25.93 dB, together with improvements in perceptual metrics. This shows that, in the evaluated setting, PhoenixSR's gains cannot be attributed solely to additional paired SR training. With distributional-prior transfer added, students obtain further reconstruction and perceptual gains, and this effect is also observed on lightweight students with approximately 1M parameters. This full-training comparison and the controlled comparison in Table 4 under the 80K setting complement each other, with each quantifying the incremental gains beyond ordinary continued training under the corresponding training budget.

AuthorsOpenReview ↗

Q3. Controlled gains and overall gains should be distinguished, and the inconsistency of reconstruction performance improvements should be discussed

The paper has separately reported the incremental gains beyond ordinary continued training in Table 4 and explicitly distinguished them from the overall changes relative to the original checkpoint. Smaller controlled gains than overall changes do not contradict the method's effectiveness. This is precisely the purpose of including a continued-training baseline and separating the sources of gains. Under the same student-update budget, PhoenixSR's results relative to task-only fine-tuning are as follows:

StudentPSNR changeLPIPS changeMUSIQ change
SwinIR-L+0.03−0.013+4.6
SwinIR-S+0.08−0.014+7.4
SeeMoRe−0.07−0.010+7.8

LPIPS and MUSIQ improve for all three students, while PSNR also improves for SwinIR-L and SwinIR-S. SeeMoRe's PSNR decreases by 0.07 dB, while LPIPS decreases by 0.010 and MUSIQ increases by 7.8, reflecting this student's reconstruction-accuracy and perceptual-quality trade-off under the 80K setting. This local trade-off cannot directly represent performance after full training. In the added comparison with matched full training budgets, PSNR, LPIPS, and MUSIQ all improve for SwinIR-L, SwinIR-S, and SeeMoRe, with SeeMoRe's PSNR increasing from 24.98 to 25.93 dB. The full-budget comparison therefore directly answers this question: these gains are not brought about solely by additional paired training, and all three evaluated students obtain further reconstruction and perceptual improvements.

The incremental gains from continued training with SFA alone should likewise be distinguished from its constraining role in distributional distillation. In Table 5, the PSNR values of SFA-only and the base SwinIR-L are 26.3124 and 26.3071 dB, respectively. However, removing SFA while retaining distributional distillation and the other components reduces PSNR from 26.4409 for the full method to 20.7107 dB. This is because the two types of supervision serve different purposes. Distribution matching introduces distributional priors from HR images, while SFA constrains the paired correspondence between the content of each output and its corresponding input. Limited gains from continuing to optimize the reconstruction objective alone for a student that has already undergone paired training do not imply that this constraint can be omitted when new distributional supervision is introduced. The existing ablation directly supports SFA's role in preserving fidelity during joint training.

The existing evidence shows both improvements in outputs relative to the original models and additional gains from PhoenixSR after controlling for the continued-training budget. The 80K comparison illustrates the specific trade-offs under that setting, while the full-budget comparison further verifies simultaneous reconstruction and perceptual improvements for all three students. The method's contribution should be assessed through these comparisons under matched conditions, and should not be generalized solely from a decline in a single metric under a particular setting.

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

This paper proposed PhoenixSR to enhance the existing pixel-wise feed-forward SR model for nowadays practical applications by employing latent distribution matching to share the diffusion model's prior. Similar to one-step diffusion SR model, it combines distribution matching with paired reconstruction supervision, adapts the real and fake score models, and introduces a reliability-based weighting strategy called DRW. Experiments cover several student models and show improved perceptual quality with a reasonable fidelity trade-off, while retaining the original students’ deployment costs.

Critical strengths
  • The idea of employing distribution matching to these traditional feed-forward SR models is intuitive and effective, bringing a general improvement without prominent fidelity loss.
Critical weaknesses
  • Despite its practical value, the methodological contribution appears incremental. DMD and REPA are established techniques, while DRW follows prior work on direction-based weighting of distillation updates. The core contribution seems to extend these new techniques to old models.
  • The paper omits closely related work that is important for assessing its novelty. GenDR [1] already combines HR-adapted score networks with adversarial learning and representation alignment, and also uses an improved distribution matching for one-step SR. GenVC [2] uses directional alignment to reweight per-sample DMD updates. The authors should clarify what HDA and DRW add beyond these precedents and substantiate the benefits of those differences.

[1] GenDR: Lighten Generative Detail Restoration. ICLR 2026.

[2] Generative Video Compression with Adaptive Score Distillation.

Questions for authors

See Weaknesses.

7 comments
AuthorsOpenReview ↗

Q1. Limited methodological novelty, mainly extending existing techniques to conventional SR models

Using existing techniques does not amount to a lack of methodological innovation. The core innovation of PhoenixSR is to explore and achieve the transfer of the generative prior of a pretrained diffusion model through DMD to independent non-diffusion SR networks that inherit neither its denoising architecture nor its pretrained weights. To our knowledge, PhoenixSR is the first framework to achieve this distribution-matching transfer while preserving the student's original inference architecture.

It is particularly worth emphasizing that this transfer remains effective for non-diffusion students with only approximately 1M parameters: SeeMoRe has only 0.97M parameters, SwinIR-S has only 1.05M, and even SwinIR-L, the largest student used, has only 28.01M (Table 1). These students do not inherit the diffusion teacher's architecture or pretrained weights, and they must also balance detail realism and input fidelity within their limited model capacity. This makes the problem subject to the simultaneous constraints of architectural differences, student capacity, and paired reconstruction, requiring us to convert diffusion priors into training supervision that small reconstruction networks can effectively use without enlarging the students and while maintaining their original inference cost.

This problem touches on a key direction in diffusion distillation: can independently designed deterministic reconstruction networks acquire diffusion priors through distributional supervision and convert them into reconstruction capabilities faithful to the input? Our study provides a concrete method and validates this objective across multiple SR architectures. Therefore, evaluating the novelty of this paper requires examining this heterogeneous transfer problem and its solution, rather than reducing it to "new techniques for old models" solely because the underlying techniques used already exist. DMD, REPA, and directional weighting each provide basic tools, but their supervision objectives, targets of application, and reliability criteria all require specific designs for the problem considered in this paper, as reflected in the following three aspects.

First, directly using the distributional objective of DMD cannot provide the per-input correspondence required by SR. DMD constrains the overall outputs to approach the target distribution, but closeness of the overall distributions does not mean that each output is faithful to its own LR input. Merely improving representations or reallocating the weights of distributional updates cannot automatically provide paired reconstruction constraints either. Therefore, we use the pixel and perceptual supervision in SFA to maintain input correspondence, constraining the generative prior by specific reconstruction targets. Table 5 in the paper provides direct evidence: after removing SFA, PSNR still drops to 20.7107 dB even with score adaptation and DRW retained. This shows that, in our setting, existing distributional supervision and its weighting cannot replace paired fidelity constraints.

At the same time, paired losses cannot correct the domain bias and tracking errors of the score models themselves. The pretrained real-score model targets a general natural-image distribution, while the fake-score model needs to estimate the continuously changing student output distribution; their estimates jointly determine the distributional guidance received by the student. HDA therefore adopts different adaptation approaches: real-score adapts to the HR target domain through LoRA, while fake-score is fully updated online and receives representation regularization. This step addresses whether the distributional supervision itself is suitable for the current SR student, while SFA addresses whether the reconstruction corresponds to the current input; the two cannot replace each other.

AuthorsOpenReview ↗

Q1. Limited methodological novelty, mainly extending existing techniques to conventional SR models

Continued.

Second, directly adding existing REPA to the generator cannot directly constrain the independent fake-score model responsible for providing supervision. GenDR, mentioned by the reviewer, incorporates REPA into the training objective of the diffusion generator; PhoenixSR applies REPA to the online fake-score model. GenDR §3.3 This adjustment targets a different learning problem: our student already has its own SR representations, whereas the fitting target of fake-score continuously changes as the student is updated, and fitting only the denoising error may make it sensitive to transient artifacts. Therefore, we align the intermediate representations of fake-score with frozen DINOv3 representations of the same student-generated image, providing a visual representation constraint for online distribution estimation. What changes here is the target of REPA and the training stage it improves: regularizing the estimator that provides distributional supervision so that the student receives more stable guidance. In Table 11, with real-score, SFA, and DRW fixed, adding fake-score REPA increases PSNR from 25.9778 to 26.3012 dB and MANIQA from 0.5422 to 0.5787, supporting the role of this design in our setting.

Third, directional weighting is only an operational form; what matters is whether the compared directions reflect the supervision problem that this paper needs to address. GenVC, the concurrent work mentioned by the reviewer, compares the DMD update with the reconstruction direction toward GT to determine whether the distributional guidance points toward the true target. GenVC §IV-C DRW focuses on another question: whether the online fake-score model exhibits consistent local denoising behavior near generated samples and paired HR samples. To this end, we let the two branches share the timestep and Gaussian noise, compute denoising residuals using the same fake-score model, and then compare the residual directions. Shared perturbations control the comparison conditions, while subtracting the known noise reduces the dominance of the common noise component in the directional comparison. This signal is used to adjust each sample's contribution to distributional supervision. Existing GT-direction gating does not compute this local denoising consistency, so the two cannot be regarded as the same mechanism merely because both use directional weighting. Table 10 shows that the Spearman rank correlation between this residual-direction signal and DMD guidance stability under independent perturbations is 0.59, higher than the 0.41 obtained by directly comparing noise-prediction directions, supporting the connection between the chosen criterion and the supervision problem considered in this paper.

AuthorsOpenReview ↗

Q1. Limited methodological novelty, mainly extending existing techniques to conventional SR models

Continued.

The paper also directly examines whether existing DMD transfer schemes are already sufficient. Table 4 compares DMD2 + SFA and PhoenixSR with the same student initialization, training data, and training budget: on SwinIR-L, the former achieves a PSNR of 25.55 dB and MUSIQ of 57.8, whereas PhoenixSR without DRW already reaches 26.15 dB and 63.5, and the full method further reaches 26.44 dB and 64.3. Even when an existing DMD scheme is equipped with reconstruction supervision, the transfer design in this paper still brings additional gains, and these gains already exist without directional weighting.

The most important innovation of this paper is to propose and validate a feasible path for transferring diffusion priors to independent non-diffusion SR networks through distribution matching distillation, and this transfer remains effective even for students with only approximately 1M parameters. In this framework, the diffusion model provides distributional supervision during training, the original SR student obtains reconstruction gains from the prior, and deployment retains its own network architecture and computational cost. The six student backbones covering CNNs, Transformers, and lightweight networks validate this contribution across multiple architectures. Specifically, in Table 1, the 0.97M SeeMoRe and 1.05M SwinIR-S improve MUSIQ on RealSR by 13.6747 and 11.1050, respectively, with PSNR changes of −0.1410 dB and −0.0457 dB, respectively, while the students' parameter counts and inference costs remain unchanged. These results show that the benefits of diffusion priors can be converted into practical reconstruction capabilities of non-diffusion networks with approximately one million parameters through our distillation method. The designs of HDA and DRW jointly serve this objective, addressing the adaptation of distributional supervision, paired fidelity constraints, and sample guidance. The names of existing components cannot substitute for a judgment of the methodological contribution; the fact that DMD, REPA, or directional weighting already exists is insufficient to negate the novelty of this paper in its research problem, heterogeneous transfer method, and systematic validation.

Reviewer_L7TsOpenReview ↗
Rating
3
Confidence
1(−2)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PhoenixSR, a framework for transferring pretrained diffusion priors into conventional feed-forward super-resolution networks through distribution matching. Heterogeneous Distribution Adaptation (HDA) combines LoRA-adapted real-score estimation, REPA-regularized fake-score learning, and paired fidelity supervision. Directional Reliability Weighting (DRW) uses denoising-residual consistency to adjust sample-wise distribution-matching guidance. Experiments across six SR backbones and three benchmarks demonstrate improved perceptual quality without modifying student inference architectures.

Critical strengths
  1. PhoenixSR transfers distribution-level supervision directly to conventional SR networks without requiring feature correspondence or imitation of diffusion-generated outputs. This addresses an important limitation of conventional distillation while preserving the students' original inference architectures. The approach is particularly relevant for improving existing efficient SR models rather than designing new diffusion-based generators.

  2. Directly applying DMD to paired SR can severely degrade reconstruction accuracy because distribution-level guidance does not enforce input-specific correspondence. PhoenixSR addresses this issue by combining paired fidelity supervision with adapted score estimation, while DRW further accounts for unreliable guidance from the evolving fake-score model. The experiments help clarify how generative supervision can be incorporated into fidelity-sensitive restoration while mitigating the risk of unfaithful reconstruction.

Critical weaknesses
  1. The individual contribution of score adaptation remains somewhat unclear. Table 11 shows that adding LoRA real-score adaptation to the REPA-regularized configuration improves PSNR only from 26.30 to 26.44 dB, while slightly reducing MANIQA. Meanwhile, Table 12 shows a larger impact from timestep truncation. Although the full framework is effective, these results leave some uncertainty about the relative importance of the proposed score adaptation and the training schedule.

  2. The effect of DRW on the fidelity–perception trade-off could be better characterized. Table 8 shows that DRW improves both reconstruction and perceptual metrics over uniform weighting. However, the low-reliability-only variant achieves higher MANIQA and MUSIQ at lower PSNR. It remains unclear whether DRW improves the attainable fidelity–perception trade-off or mainly selects a different operating point.

Questions for authors
  1. Table 4 controls the number of student updates. Are the score-model update budgets and timestep schedules also matched between PhoenixSR and DMD2 + SFA? This would help clarify the source of the performance difference.

  2. Could the authors compare DRW against uniform DMD weighting at matched PSNR, for example by adjusting the fidelity or distribution-matching loss weights?

5 comments
AuthorsOpenReview ↗

Q1. The relative contributions of score-model adaptation and the training schedule remain unclear

In Table 11, REPA and real-score LoRA together provide a PSNR improvement of 0.4631 dB, with LoRA still contributing approximately 0.14 dB on top of REPA. For SR students that already have strong reconstruction capabilities, these are substantial additional reconstruction gains, without increasing inference computation. These results come from the existing ablation in Appendix D.4: the three relevant configurations all retain SFA and DRW, follow the unified controlled setting in Appendix D, and vary only REPA and real-score LoRA.

PSNR is computed on a logarithmic scale from the mean squared error between the reconstructed image and GT, whereas MANIQA predicts no-reference perceptual quality. Their evaluation objectives, numerical scales, and sensitivities to different distortions differ, so the absolute magnitudes of their changes cannot be directly compared. For the paired reconstruction fidelity emphasized in this paper, the PSNR improvement provides more direct and clear positive evidence. MUSIQ and CLIP-IQA also improve, supporting improved reconstruction accuracy while maintaining overall perceptual quality. MANIQA decreases by only 0.0042, a very small change. As a reference for the magnitude of training variability, Table 7 reports a three-seed MANIQA standard deviation of 0.006 for the full SwinIR-L model, which is of the same order as this difference. This change is therefore very likely due to training randomness. Together with the improvements in PSNR, MUSIQ, and CLIP-IQA, this small change should not obscure the fidelity gains from score-model adaptation.

Eq. (2) shows that DMD guidance depends both on the noisy states at which it is computed and on the real–fake score difference at those states. The timestep schedule adjusts the noise range and sampling distribution used for supervision, whereas score-model adaptation improves estimation of the target HR distribution and the current student distribution. They act on the supervision conditions and guidance quality, respectively.

The timestep schedule controls the noise scales used for supervision. Appendix C.1 restricts timesteps to [0,200), allowing the noisy latent to retain more structure related to the current reconstruction; noise-shift then adjusts the sampling distribution within this range at different training stages. This corresponds to a core requirement of SR: introducing natural-image priors while maintaining the content determined by the input. Score-model warm-up and asynchronous updates in Appendix C.2 further coordinate the learning pace of the student and score models, reducing tracking lag in online distribution estimation.

Score-model adaptation in HDA acts directly on the two score estimates in Eq. (2). Eq. (3) trains real-score LoRA with a denoising objective on real HR images, updating only low-rank parameters while freezing the pretrained backbone, so that the general prior becomes closer to the SR target domain through adaptation with a restricted parameter space. Eq. (4) allows the fake-score model to track the continuously changing student outputs through full optimization, while constraining its intermediate features using representations extracted from student-generated images by frozen DINOv3, improving representation learning in the online estimator. The former handles adaptation to the target HR distribution, and the latter handles tracking of the current student distribution. Even with the same timesteps, both designs change the score difference provided to the student.

The training schedule and score-model adaptation therefore play complementary roles in the same supervision process: the schedule provides suitable noisy states and update conditions, while adaptation improves target-distribution estimation and student-distribution tracking under these conditions. SFA in Eq. (5) then maintains LR–HR correspondence through pixel and paired perceptual constraints, with an adversarial term supplementing texture realism. This reflects our starting point of using diffusion priors moderately under fidelity constraints: real-score LoRA restricts the parameter space of target-domain adaptation, low-noise supervision retains input-related structure, and SFA maintains paired reconstruction. The improved reconstruction accuracy and broadly maintained perceptual quality in Table 11 are consistent with this design objective, while the student retains its original deployment architecture.

AuthorsOpenReview ↗

Q1. The relative contributions of score-model adaptation and the training schedule remain unclear

Continued.

The existing training-schedule ablation in Table 12 is consistent with this explanation:

Training schedulePSNR ↑MANIQA ↑MUSIQ ↑CLIP-IQA ↑
Without timestep truncation25.65640.543961.67310.4733
Without the noise-shift curriculum26.37970.550362.73150.4977
Without score-model warm-up26.40110.563764.11490.4893
With the full training schedule26.44090.574564.31020.5026

In the current ablation setting, removing timestep truncation causes a PSNR decrease of 0.7845 dB, showing that a suitable noise range is an important foundation for fidelity-sensitive supervision. On this foundation, REPA and real-score LoRA still provide an additional gain of 0.4631 dB. The importance of the foundational conditions does not imply that the other components are secondary: timestep truncation establishes suitable supervision conditions, and score-model adaptation further improves the quality of distributional guidance under those conditions. The components serve different roles, and their value lies both in maintaining effective training conditions and in further improving reconstruction quality. Their relative contributions should therefore be assessed together with their roles and incremental gains within the full framework, rather than solely by the magnitude of the performance decrease after removal.

AuthorsOpenReview ↗

Q2. Does DRW improve the fidelity–perceptual quality trade-off, or merely change the operating point?

In Table 8, with the same loss coefficients, DRW improves PSNR from 26.1587 to 26.4409 dB over uniform DMD weighting, a gain of 0.2822 dB, while also improving MANIQA, MUSIQ, and CLIP-IQA. This is consistent with our design objective: using diffusion priors to improve perceptual quality while maintaining reconstruction fidelity corresponding to the input.

In principle, the overall loss coefficient controls the relative strength of distribution-matching supervision and fidelity constraints, whereas DRW controls the relative contributions of different samples to distribution-matching updates. According to Eq. (5), adjusting the DMD coefficient uniformly scales the distributional guidance of all samples, changing its balance with SFA. According to Eq. (8), DRW assigns different weights to individual samples based on the behavior of the current score estimator. Let the DMD parameter gradient of sample i be gᵢ. The batch gradient is Σᵢgᵢ/N under uniform weighting and Σᵢωᵢgᵢ/N under DRW. Because the weights use stop-gradient, they do not change the guidance direction of an individual sample, but can change the batch update direction formed by combining different directions. Sample-wise weighting therefore generally cannot be replaced by a single global loss coefficient.

This design addresses sample differences in online distribution estimation. Even when HDA improves overall score estimation, as the SR student's output distribution continues to change, the local behavior of the fake-score model may still differ across samples. Eqs. (6)–(7) use the same fake-score model to compare denoising residual directions at the generated image and its corresponding HR image, with the same timestep and noise, measuring local denoising consistency at the two locations. DRW adjusts the contributions of distributional supervision based on this signal, while SFA continues to maintain paired reconstruction constraints. DRW therefore concerns how to use the distributional guidance provided by the current score model; its reliability score is not directly equivalent to image perceptual quality.

This also explains the low-reliability-only result in Table 8. Appendix D.2 shows that using only low-reliability samples favors perceptual quality, whereas using only high-reliability samples favors fidelity. Higher MANIQA and MUSIQ in the former do not imply that the other samples should be discarded. DRW retains the contributions of both types of samples and coordinates their relative strength through soft weighting, obtaining the highest PSNR in Table 8 while maintaining good perceptual quality. The reviewer's suggested comparison at matched PSNR further examines whether this sample allocation still provides perceptual benefits at the same reconstruction accuracy.

Following your suggestion, we added a comparison at matched PSNR on SwinIR-L / RealSR. Taking the DRW result in Table 4 as the reference, we keep student initialization, training data, SFA, score-model adaptation, training schedule, and update budget unchanged, and adjust the DMD loss coefficient for uniform weighting so that its PSNR approaches DRW's 26.44 dB as closely as possible (an exact match is difficult):

Comparison at matched PSNRPSNR ↑LPIPS ↓MUSIQ ↑
Uniform DMD weighting26.420.27161.65
DRW26.440.24464.3

At similar PSNR, DRW obtains lower LPIPS and higher MUSIQ. This shows that, at the reconstruction accuracy compared here, sample-wise allocation of distributional guidance provides additional perceptual benefits, consistent with the mechanism described above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.