acceptodds
Under review as a conference paper at ICLR 2027

Frequency-Guided Temporal Correction for Hallucination Suppression in Diffusion-Based Real-World Image Super-Resolution

Abstract

Diffusion models perform strongly in real-world image super-resolution (Real-ISR), yet their generative priors can introduce structural distortions and spurious details. Without ground truth (GT) at inference, distinguishing useful refinement from hallucinations remains challenging. Using GT only for offline analysis, we examine how deviations from a selected early prediction and recent prediction history relate to reconstruction errors within each frequency band. We find a frequency-dependent reference preference: the selected early prediction provides stronger error-related cues in the low-frequency band, whereas recent prediction history is more informative in the high-frequency band. Based on this finding, we propose Frequency-Guided Temporal Correction (FGTC), a training-free and plug-and-play framework for multi-step diffusion-based Real-ISR. FGTC selects a structurally reliable early low-frequency reference and maintains recent high-frequency prediction history, then performs frequency-decoupled correction with deviation-adaptive strength before feeding the corrected prediction back into the original reverse diffusion process. Experiments on three benchmarks with StableSR and SeeSR show consistent improvements in reconstruction fidelity, including PSNR gains of 0.92–1.31 dB, while reducing structural distortions and spurious textures without backbone retraining or GT at inference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Real-World Image Super-Resolution, Hallucination Suppression, Diffusion Models

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_V2NaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces FGTC, a training-free correction for multi-step diffusion-based super-resolution. It selects an early low-frequency prediction according to structural similarity with the upsampled LR input and uses recent prediction history as the high-frequency reference. Subsequent predictions are blended with these references using deviation-dependent weights, then re-encoded for the original sampler. Experiments on StableSR and SeeSR across three datasets show consistent improvements in PSNR, with perceptual improvements in most settings. The paper makes a modest contribution to super-resolution. However, the comparisons with relevant state-of-the-art methods are insufficient to establish its incremental value, and I find the overall novelty limited.

Critical strengths
  1. The paper addresses a practical problem: improving reconstruction fidelity without retraining a diffusion backbone or requiring GT at inference. Using different temporal references for structure and detail is an intuitive design, and FGTC is demonstrated on two existing models.

  2. The reported gains over the original backbones are consistent. Table 1 shows PSNR improvements of 0.92-1.31 dB across all six backbone/dataset combinations. StableSR also improves LPIPS and DISTS on all three datasets, and the visual examples illustrate cleaner boundaries and reduced spurious textures.

  3. The ablations help assess the method beyond its overall performance. Table 3 compares all four early/history reference assignments, while Table 4 separates the proposed correction from decode-encode and wavelet round trips. In particular, decode-encode alone improves PSNR by about 0.18 dB on SeeSR/DRealSR, compared with about 1.07 dB for full FGTC. This supports a contribution from the correction itself.

Critical weaknesses
  1. My main reservation concerns the incremental contribution relative to StructSR, which is discussed in the related work but not included in the experiments. StructSR already selects an early prediction using structural similarity to the LR input and uses that prediction to guide subsequent restoration without additional training. FGTC differs by selecting the reference in the low-frequency band, using recent history for high frequencies, and adapting correction strength to reference deviations. These are meaningful distinctions to investigate, but comparisons with unmodified backbones do not establish how much they add over the existing structural-reference approach. A matched comparison with StructSR would be more informative for assessing the contribution than further comparisons with general-purpose SR models. The absence of this comparison does not imply that FGTC would perform worse; it leaves the size and nature of the advance unresolved.

  2. The internal ablations support the design, although they do not replace that external comparison. On SeeSR/DRealSR, Table 3 reports PSNR of 27.82 for the proposed Early/History assignment versus 27.70 for Early/Early, together with improvements in SSIM, LPIPS, and DISTS. This suggests a benefit from the high-frequency history branch. Its consistency across images or sampling seeds would help judge the strength of this particular contribution. Importantly, the Early/Early variant is an FGTC ablation, not an implementation of StructSR. MASF is also relevant because it aggregates predictions by frequency; a short account of the differences, and an adapted comparison if feasible, would further clarify FGTC's position.

  3. The practical trade-off deserves a clearer account. For SeeSR on DIV2K-Val, PSNR improves from 22.36 to 23.38, while LPIPS changes from 0.3194 to 0.3321 and DISTS from 0.1968 to 0.2111. This is an exception to otherwise broadly positive results, rather than evidence by itself that FGTC over-smooths images. Discussing representative cases would help explain when the correction preserves useful detail and when it may trade perceptual similarity for fidelity. Runtime is also not reported, despite the additional decoding and encoding during sampling, so the cost of the gains is difficult to assess.

Questions for authors
  1. Could the authors compare FGTC with StructSR using the same backbone, sampler, step count, and evaluation protocol? A focused experiment on the existing SeeSR/DRealSR setting would already help. Please also explain the differences in reference selection and how the selected reference enters the sampling update.

  2. How consistent is the Early/History improvement over Early/Early across test images or sampling seeds? A paired summary on the existing evaluation would help establish the contribution of using recent high-frequency history.

  3. For the Table 5 analysis, which images and timesteps are used? Could the authors clarify whether the reference preference remains within comparable sampling stages? Shared temporal trends are a possible alternative explanation for part of the association, rather than an established problem with the analysis.

  4. What are the runtime and peak memory at matched sampling settings, and what kinds of images account for the SeeSR/DIV2K-Val perceptual-metric degradation?

4 comments
AuthorsOpenReview ↗

Response to Reviewer V2Na (part 1)

Thank you for your detailed review and suggestions. We hope our responses address your questions and concerns. We also hope the added comparisons and analyses support a more positive assessment of our work.

Reply to W1, W2 & Q1: Comparison with StructSR

To compare FGTC with StructSR, we ran quantitative tests with SeeSR and StableSR on DRealSR, RealSR, and DIV2K-Val, using the same backbone and sampling budget for each comparison. The results show that FGTC achieves better reconstruction fidelity in all six settings and also improves perceptual metrics in most settings. The table below includes the original models and both added methods, with the best results in each group in bold.

DatasetBackboneMethodPSNR ↑SSIM ↑MS-SSIM ↑LPIPS ↓DISTS ↓
DRealSRSeeSRBaseline26.750.73730.86250.31740.2315
DRealSRSeeSR+StructSR27.170.77470.86960.30290.2229
DRealSRSeeSR+FGTC27.820.78320.87380.29240.2073
DRealSRStableSRBaseline26.790.71900.85610.32680.2248
DRealSRStableSR+StructSR27.530.76160.87160.30540.2175
DRealSRStableSR+FGTC27.930.78440.87880.28160.2005
RealSRSeeSRBaseline23.830.69040.86100.30070.2224
RealSRSeeSR+StructSR24.310.71930.86220.29870.2189
RealSRSeeSR+FGTC24.750.72570.86630.26560.2001
RealSRStableSRBaseline23.320.67510.85230.30560.2164
RealSRStableSR+StructSR23.970.71750.86350.29070.2091
RealSRStableSR+FGTC24.410.73510.87290.26010.1966
DIV2K-ValSeeSRBaseline22.360.55920.79720.31940.1968
DIV2K-ValSeeSR+StructSR23.190.59610.80930.31670.2109
DIV2K-ValSeeSR+FGTC23.380.60560.81190.33210.2111
DIV2K-ValStableSRBaseline21.940.52430.78280.31190.2052
DIV2K-ValStableSR+StructSR22.530.55540.79730.30450.2028
DIV2K-ValStableSR+FGTC23.250.59910.81540.28010.1911

Reply to Q2: PSNR improvements across images

With the selected early LL reference unchanged, we compared Early/History with Early/Early using SeeSR on 93 DRealSR images. PSNR improves on 89 images (95.70%), and the 95% confidence interval for the mean paired gain is above zero, as estimated by paired image bootstrapping. These results show that recent HF prediction history benefits most images.

LL ReferenceHF ReferencePSNR ↑ΔPSNR ↑95% CIImproved Images
EarlyEarly27.70———
EarlyHistory27.82+0.12[0.046, 0.253]89/93 (95.7%)
AuthorsOpenReview ↗

Response to Reviewer V2Na (part 2)

Reply to Q3: Reference preference at different sampling stages

The analysis in Table 5 uses 93 DRealSR images and decoded clean predictions along their StableSR sampling trajectories, with GT used to compute reconstruction errors. We exclude the initial t=200t=200 snapshot. Both Early and History references are evaluated on the same steps, 194194–00, with steps 199199–195195 used to initialize the history buffer with K=5K=5.

The preference for temporal references in each frequency band remains consistent both after timestep control and within comparable sampling stages: Early for LL and History for HF. Using the original analysis settings, we calculated partial Spearman correlations for each image by removing the effect of timestep ranks from the ranks of reference deviations and reconstruction errors measured against GT, then correlating the residuals. The table compares these correlations with the original results as mean ± standard deviation across images. Higher means are bold.

FrequencyMetricEarly referenceRecent history
LLSpearman ρ0.796 ± 0.2600.206 ± 0.240
LLPartial Spearman ρ0.742 ± 0.2010.167 ± 0.367
LLHigh-error AUC0.820 ± 0.2560.717 ± 0.223
HFSpearman ρ0.776 ± 0.2910.825 ± 0.198
HFPartial Spearman ρ0.623 ± 0.2350.726 ± 0.207
HFHigh-error AUC0.909 ± 0.1790.943 ± 0.093

To further check whether the preference holds within comparable sampling stages, we compare Early and History (Hist.) on the same 93 images and the same steps, dividing 194194–00 into three stages of 65 steps each: early (194194–130130), middle (129129–6565), and late (6464–00). We calculate ordinary Spearman correlations for each image within each stage, then average across images. Early has a higher LL correlation and History has a higher HF correlation in all three stages. The second table reports the mean correlations for each stage, with the higher mean for each frequency band in bold.

Sampling StageReferenceLL Spearman ρHF Spearman ρ
Early (194–130)Early0.60210.2934
Early (194–130)Hist.-0.01340.6401
Middle (129–65)Early0.71480.7388
Middle (129–65)Hist.0.16950.7612
Late (64–0)Early0.82420.8022
Late (64–0)Hist.0.50610.9098
AuthorsOpenReview ↗

Response to Reviewer V2Na (part 3)

Reply to W3 & Q4: Image quality, runtime, and peak memory

The perceptual tradeoff on SeeSR/DIV2K-Val is closely related to the recoverability of fine structures under synthetic degradation. On DIV2K-Val, severe degradation can obscure fine lines and densely repeated structures. SeeSR uses semantic guidance and pretrained diffusion priors to generate realistic details, but insufficient structural information in the LR input may lead to local structural distortions and textures inconsistent with GT. FGTC corrects these deviations using a selected early LL reference and recent HF prediction history. For severely degraded, densely structured regions, improving structural fidelity may also change the appearance of fine lines and local texture contrast, resulting in higher reconstruction accuracy at some perceptual cost. This tradeoff is specific to SeeSR/DIV2K-Val: FGTC improves both LPIPS and DISTS with SeeSR on RealSR and DRealSR, as well as with StableSR on DIV2K-Val. The two backbones also differ in their training datasets (LSDIR and FFHQ10K for SeeSR versus DF2K and OST for StableSR), which may contribute to their different perceptual responses.

To further illustrate this tradeoff across different image contents, we provide two representative examples in Section 4.3 (Figure 5 and Table 3) of the revised manuscript. For the building facade, severe degradation obscures the densely repeated window grids, making their geometry difficult to recover. FGTC brings these structures closer to GT, improving structural fidelity, while changes in the appearance of fine lines come at a certain perceptual cost. In contrast, the stained glass window retains relatively clear main contours. FGTC preserves these structures while removing textures inconsistent with GT, improving all five metrics. These examples demonstrate that FGTC can simultaneously improve fidelity and perceptual similarity when the main structures remain sufficiently recoverable. When severe degradation blurs densely repeated structures, FGTC can improve their alignment with GT, while changes in fine-line appearance may incur a certain perceptual cost. The table below reports all five metrics for these two examples, with PSNR in dB and the better result for each image in bold.

ImageMethodPSNR ↑SSIM ↑MS-SSIM ↑LPIPS ↓DISTS ↓
0821_pch_00015SeeSR19.350.53210.89070.23960.2480
0821_pch_00015SeeSR+FGTC20.240.65870.91920.17630.2120
0845_pch_00025SeeSR14.980.40820.71520.21720.1957
0845_pch_00025SeeSR+FGTC16.290.47970.76010.25770.2217

Runtime and peak memory.

To compare FGTC's inference cost with the original models and StructSR, we measured runtime and peak GPU memory on SeeSR and StableSR under the same settings for each backbone. All tests use an NVIDIA GeForce RTX 3090 for 4× super-resolution with 512×512 outputs. StableSR uses 200 sampling steps and SeeSR uses 50.

FGTC introduces no trainable parameters and keeps the original sampling step count. Its extra runtime over the original models comes from reference computation, decoding, encoding the predictions again, and frequency correction during sampling. The table below reports runtime and peak GPU memory. Lower is better for both measures.

BackboneMethodPeak GPU memory (GiB) ↓Runtime (s/image) ↓
SeeSRBaseline6.71537.32
SeeSR+StructSR6.715314.84
SeeSR+FGTC (Ours)6.720311.77
StableSRBaseline7.074619.53
StableSR+StructSR7.112041.78
StableSR+FGTC (Ours)7.112030.07
AuthorsOpenReview ↗

Updated Manuscript and Summary of Revisions

Dear Reviewer V2Na,

Thank you again for your detailed review and suggestions. We have uploaded a revised manuscript incorporating the following additions:

  • Direct quantitative and visual comparisons with StructSR are provided in Section 4.2, Table 2 and Figure 5.
  • Paired PSNR results across images in Appendix B.2, Table 10.
  • Analysis settings, results after timestep control, and comparisons within sampling stages in Appendix A, Tables 7 and 8.
  • Representative cases showing detail preservation and the perceptual tradeoff in Section 4.3, Figure 6 and Table 3.
  • Runtime and peak GPU memory measurements in Appendix D, Table 11.

All changes are marked in blue. We hope this additional evidence and our responses address your concerns and support a more positive assessment of our work. We would appreciate your feedback and would be happy to answer any remaining questions.

Best regards,

The Authors

Reviewer_8UpBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Frequency-Guided Temporal Correction (FGTC), a training-free inference-time module for multi-step diffusion-based real-world image super-resolution. Its key contribution is identifying frequency-dependent preferences for temporal references and leveraging this insight to improve structural fidelity and detail reconstruction during diffusion sampling. Experiments on three benchmarks with two diffusion-based backbones demonstrate consistent improvements in reconstruction quality and hallucination suppression, supported by comprehensive ablation studies validating the key design choices.

Critical strengths
  1. FGTC provides a simple and practical inference-time design. It requires neither backbone retraining nor ground truth at inference. Its frequency-specific reference strategy is well motivated and readily integrated into existing diffusion-based backbones.
  2. The empirical results are consistent across two backbones and three benchmarks. Direct backbone-versus-backbone-plus-FGTC comparisons show substantial fidelity gains, while most perceptual metrics also improve. The ablation study supports the complementary roles of the LL and HF branches.
Critical weaknesses
  1. The per-image Spearman correlations and high-error AUCs provide useful evidence for the proposed frequency-dependent reference preference. However, since both the deviation and GT-error sequences are evaluated along the same diffusion trajectory, part of their association may reflect a shared dependence on the sampling timestep. A partial Spearman analysis controlling for timestep, or an equivalent detrended analysis, would make the interpretation more convincing.
  2. While the first two examples in Figure 4 show clear structural improvements, the third appears to retain fragmented high-frequency patterns that are not clearly supported by the GT. It would be helpful if the authors could clarify the underlying reason for this behavior and provide further analysis of the LL/HF errors and additional examples to better characterize the effectiveness and limitations of the proposed hallucination-suppression strategy.
Questions for authors
  1. Table 5 correlates reference deviation and GT-based error over 200 consecutive sampling steps, although both quantities may vary systematically with timestep. Could the authors report the per-image partial Spearman correlation between dtd_t and ete_t conditioned on t/Tt/T, for each frequency band and reference? Please clarify whether Early remains preferable for LL and History for HF after removing this shared temporal trend.
  2. What are the wall-clock runtime and peak GPU memory of StableSR and SeeSR with and without FGTC under the same resolution, hardware, and sampling settings?
  3. In the third example of Figure 4, FGTC removes some surface artifacts but appears to retain fragmented high-frequency patterns relative to the GT. Could the authors discuss whether hallucinations that persist for several steps can enter the recent-history buffer and escape correction?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 8UpB

Thank you for your helpful comments and suggestions. We hope our responses address your questions and concerns.

Reply to W1 & Q1: Reference preference after timestep control

To check whether the reference preference stays the same after controlling for sampling timestep, we calculated partial Spearman correlations for each image using the same settings as in the original analysis. We control for normalized timestep t/Tt/T by using regression to remove the effect of timestep ranks from the ranks of reference deviations and reconstruction errors measured against GT, then correlate the remaining values. The results show that the preference for temporal references in each frequency band remains consistent after controlling for the sampling timestep: Early has a higher correlation with LL reconstruction errors, while History has a higher correlation with HF reconstruction errors. The table gives mean ± standard deviation across images, together with the original results. Higher means are bold.

FrequencyMetricEarly referenceRecent history
LLSpearman ρ0.796 ± 0.2600.206 ± 0.240
LLPartial Spearman ρ0.742 ± 0.2010.167 ± 0.367
LLHigh-error AUC0.820 ± 0.2560.717 ± 0.223
HFSpearman ρ0.776 ± 0.2910.825 ± 0.198
HFPartial Spearman ρ0.623 ± 0.2350.726 ± 0.207
HFHigh-error AUC0.909 ± 0.1790.943 ± 0.093

Reply to Q2: Runtime and peak GPU memory

To compare FGTC's inference cost with the original models and StructSR, we measured runtime and peak GPU memory on SeeSR and StableSR under the same settings for each backbone. All tests use an NVIDIA GeForce RTX 3090 for 4× super-resolution with 512×512 outputs. StableSR uses 200 sampling steps and SeeSR uses 50.

FGTC introduces no trainable parameters and keeps the original sampling step count. Its extra runtime over the original models comes from reference computation, decoding, encoding the predictions again, and frequency correction during sampling. The table below reports runtime and peak GPU memory. Lower is better for both measures.

BackboneMethodPeak GPU memory (GiB) ↓Runtime (s/image) ↓
SeeSRBaseline6.71537.32
SeeSR+StructSR6.715314.84
SeeSR+FGTC (Ours)6.720311.77
StableSRBaseline7.074619.53
StableSR+StructSR7.112041.78
StableSR+FGTC (Ours)7.112030.07

Reply to W2 & Q3: Residual high-Frequency artifacts

In the third example of Figure 4, severe synthetic degradation obscures fine textures and object boundaries, making accurate reconstruction challenging for existing SR methods, some of which introduce noticeable artifacts or structural distortions in the highlighted region. FGTC uses the selected early LL reference to correct structural deviations and recent HF prediction history to suppress unstable details. Compared with other methods in Figure 4, FGTC removes much of the irregular surface texture and preserves a more continuous object boundary, achieving a closer visual match to GT. The few remaining fragmented patterns may arise when persistent artifacts enter the history buffer and exhibit small deviations from the recent historical reference, resulting in milder correction. Overall, FGTC effectively reduces spurious textures while improving structural consistency.

AuthorsOpenReview ↗

Updated Manuscript and Summary of Revisions

Dear Reviewer 8UpB,

Thank you again for your helpful comments and suggestions. We have uploaded a revised manuscript with the following additions:

  • The analysis controlling for sampling timestep is introduced in Section 3.1, with the settings and full results provided in Appendix A, Table 7. The results show that the frequency-dependent reference preference remains consistent after timestep control.
  • Runtime and peak GPU memory measurements are provided in Appendix D, Table 11.
  • Additional visual analysis and the corresponding image metrics are presented in Section 4.3, Figure 6 and Table 3.

All changes are marked in blue. We hope these additions, together with our explanation of the remaining textures in the challenging example, can address your concerns. We would appreciate your feedback and would be happy to provide further clarification.

Best regards,

The Authors

Thank you for providing the timestep-controlled analysis, the efficiency comparisons with StructSR, and the clarification of the artifact case. This addresses my remaining questions about the temporal-reference preference and the practical inference cost. The added evidence strengthens the empirical support for FGTC. Good luck!

Thank you for your helpful feedback and support. We sincerely appreciate your time and effort.

Reviewer_hn1NOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FGTC introduces a training-free approach to reduce hallucinations in diffusion-based super-resolution by combining early low-frequency references with recent high-frequency prediction histories. Extensive ablations demonstrate consistent reconstruction fidelity gains and validate the proposed temporal correction design.

Critical strengths

(1) It works as a plug-and-play addition without needing any backbone retraining or ground-truth images during inference.

(2) The choice of temporal references is well justified, using early predictions for low-frequency structures and recent histories for high-frequency details.

(3) It brings solid, consistent PSNR gains of 0.92–1.31 dB across benchmarks, with control experiments showing these improvements really come from the correction mechanism.

Critical weaknesses

(1) Key related baselines like StructSR and MASF are missing from the quantitative comparisons, making it hard to tell FGTC's exact advantage over existing methods.

(2) There is a perceptual trade-off, where pairing SeeSR with FGTC on DIV2K-Val improves PSNR but actually worsens LPIPS and DISTS metrics.

(3) The extra computational overhead from intermediate decoding, re-encoding, and frequency processing is left unmeasured in terms of runtime and memory.

(4) Looking at the visual results, SeeSR+FGTC appears blurrier in Figure 4 (bottom), both StableSR+FGTC and SeeSR+FGTC lose sharpness in Figure 6 (middle), and SeeSR+FGTC still leaves leaf contour discrepancies in Figure 7 (top).

Questions for authors

(1) How does FGTC compare quantitatively against StructSR and MASF under identical backbones and sampling budgets?

(2) What causes the perceptual metric degradation (LPIPS/DISTS) on DIV2K-Val when paired with SeeSR?

(3) What are the exact inference latency and peak memory overheads relative to the base models?

(4) Could you explain the loss of sharpness in SeeSR+FGTC in Figure 4 (bottom) and both models in Figure 6 (middle), and why leaf contour discrepancies persist in SeeSR+FGTC in Figure 7 (top)?

5 comments
AuthorsOpenReview ↗

Response to Reviewer hn1N (part 1)

Thank you for your careful review and useful comments. We hope our revisions and additional experiments address your concerns and support a more favorable assessment of our work.

Reply to W1 & Q1: Comparison with StructSR and MASF

To compare FGTC with StructSR, we ran quantitative tests with SeeSR and StableSR on DRealSR, RealSR, and DIV2K-Val, using the same backbone and sampling budget for each comparison. The results show that FGTC achieves better reconstruction fidelity in all settings and also improves perceptual metrics in most settings. The best scores in each group in the table are displayed in bold.

DatasetBackboneMethodPSNR ↑SSIM ↑MS-SSIM ↑LPIPS ↓DISTS ↓
DRealSRSeeSRBaseline26.750.73730.86250.31740.2315
DRealSRSeeSR+StructSR27.170.77470.86960.30290.2229
DRealSRSeeSR+FGTC27.820.78320.87380.29240.2073
DRealSRStableSRBaseline26.790.71900.85610.32680.2248
DRealSRStableSR+StructSR27.530.76160.87160.30540.2175
DRealSRStableSR+FGTC27.930.78440.87880.28160.2005
RealSRSeeSRBaseline23.830.69040.86100.30070.2224
RealSRSeeSR+StructSR24.310.71930.86220.29870.2189
RealSRSeeSR+FGTC24.750.72570.86630.26560.2001
RealSRStableSRBaseline23.320.67510.85230.30560.2164
RealSRStableSR+StructSR23.970.71750.86350.29070.2091
RealSRStableSR+FGTC24.410.73510.87290.26010.1966
DIV2K-ValSeeSRBaseline22.360.55920.79720.31940.1968
DIV2K-ValSeeSR+StructSR23.190.59610.80930.31670.2109
DIV2K-ValSeeSR+FGTC23.380.60560.81190.33210.2111
DIV2K-ValStableSRBaseline21.940.52430.78280.31190.2052
DIV2K-ValStableSR+StructSR22.530.55540.79730.30450.2028
DIV2K-ValStableSR+FGTC23.250.59910.81540.28010.1911

MASF is not an image super-resolution method. It is designed to stabilize diffusion sampling. Therefore, a direct and fair comparison with FGTC on SeeSR and StableSR is not feasible. We cite MASF in the related work to discuss prior studies of diffusion sampling that uses frequency information. MASF stabilizes sampling with weighted moving averages computed separately for each frequency, while FGTC uses a selected early LL reference and recent HF prediction history, adapting correction to reference deviations to address structural distortions and spurious details in Real-ISR.

Reply to W2 & Q2: Changes in LPIPS/DISTS on DIV2K-Val

The perceptual tradeoff on SeeSR/DIV2K-Val is closely related to the recoverability of fine structures under synthetic degradation. On DIV2K-Val, severe degradation can obscure fine lines and densely repeated structures. SeeSR uses semantic guidance and pretrained diffusion priors to generate realistic details, but insufficient structural information in the LR input may lead to local structural distortions and textures inconsistent with GT. FGTC corrects these deviations using a selected early LL reference and recent HF prediction history. For severely degraded, densely structured regions, improving structural fidelity may also change the appearance of fine lines and local texture contrast, resulting in higher reconstruction accuracy at some perceptual cost. This tradeoff is specific to SeeSR/DIV2K-Val: FGTC improves both LPIPS and DISTS with SeeSR on RealSR and DRealSR, as well as with StableSR on DIV2K-Val. The two backbones also differ in their training datasets (LSDIR and FFHQ10K for SeeSR versus DF2K and OST for StableSR), which may contribute to their different perceptual responses.

To further illustrate this behavior, we provide two representative examples in Section 4.3 of the revised manuscript. For the building facade, severe degradation obscures the densely repeated window grids, making their geometry difficult to recover. FGTC brings these structures closer to GT, but changes in fine-line appearance incur a perceptual cost. In contrast, the stained glass window retains relatively clear main contours. FGTC preserves these structures while removing textures inconsistent with GT, improving all five metrics. These examples demonstrate that FGTC can simultaneously improve fidelity and perceptual similarity when the main structures remain sufficiently recoverable. When severe degradation blurs densely repeated structures, FGTC can improve their alignment with GT, while changes in fine-line appearance may incur a certain perceptual cost.

AuthorsOpenReview ↗

Response to Reviewer hn1N (part 2)

Reply to W3 & Q3: Runtime and memory

To compare FGTC's inference cost with the original models and StructSR, we measured runtime and peak GPU memory on SeeSR and StableSR under the same settings for each backbone. All tests use an NVIDIA GeForce RTX 3090 for 4× super-resolution with 512×512 outputs. StableSR uses 200 sampling steps and SeeSR uses 50.

FGTC introduces no trainable parameters and keeps the original sampling step count. Its extra runtime over the original models comes from reference computation, decoding, encoding the predictions again, and frequency correction during sampling. The table below reports runtime and peak GPU memory. Lower is better for both measures.

BackboneMethodPeak GPU memory (GiB) ↓Runtime (s/image) ↓
SeeSRBaseline6.71537.32
SeeSR+StructSR6.715314.84
SeeSR+FGTC (Ours)6.720311.77
StableSRBaseline7.074619.53
StableSR+StructSR7.112041.78
StableSR+FGTC (Ours)7.112030.07

Reply to W4 & Q4: Sharpness and edge differences

These visual differences can be understood by considering the characteristics of the image content and the temporal correction that FGTC applies separately to LL and HF.

For Fig. 4 (bottom), FGTC keeps the structure stable and limits structural drift, although some edges appear slightly softer. FGTC uses a selected early LL reference to preserve reliable structural information and mitigate structural drift during subsequent sampling. Since the reference is selected based on its structural consistency with the LR input rather than its sharpness, some selected predictions may retain relatively coarse edge information. When the current prediction differs substantially from this reference, the adaptive LL correction increases the reference contribution to maintain structural consistency. In such cases, this constraint may also moderately limit subsequent edge refinement, resulting in a slightly softer appearance.

For Fig. 6 (middle), FGTC preserves useful fine details, with SeeSR+FGTC showing clearer textures resembling fine hairs than SeeSR even where local edges appear softer. The image contains densely distributed fine structures and textures with high contrast. FGTC employs recent HF prediction history to reduce unstable texture variations while retaining temporally consistent details. This correction may moderate certain local HF contrast, making some edges appear less sharp. However, SeeSR+FGTC also exhibits more discernible textures resembling fine hairs than SeeSR in this example. This suggests that the perceived softness does not necessarily correspond to a uniform reduction in fine details, but may reflect a different balance between local contrast and texture preservation.

For Fig. 7 (top), FGTC maintains the overall leaf structure and limits structural drift, with some local contour differences remaining in regions with closely spaced boundaries. FGTC anchors subsequent predictions to the selected early LL reference, preserving structural consistency and limiting the accumulation of structural drift during sampling. The overlapping leaves contain closely spaced boundaries whose exact geometry is difficult to recover from the degraded LR input. FGTC provides stable guidance for the overall leaf structure, while precise boundary placement also depends on the fine details reconstructed by the backbone. FGTC thus keeps the overall leaf structure anchored during sampling, while the residual differences from GT concern the precise placement of closely spaced boundaries.

Overall, these examples illustrate how FGTC's structural preservation and texture correction can produce different visual effects depending on the image characteristics. We will clarify these behaviors in the revised manuscript.

AuthorsOpenReview ↗

Updated Manuscript and Summary of Revisions

Dear Reviewer hn1N,

Thank you again for your careful review and helpful comments. We have uploaded a revised manuscript incorporating the following additions:

  • Direct quantitative and visual comparisons with StructSR are provided in Section 4.2, Table 2 and Figure 5.
  • Two representative visual examples and their metrics explaining the perceptual tradeoff on SeeSR/DIV2K-Val in Section 4.3, Figure 6 and Table 3.
  • Runtime and peak GPU memory measurements in Appendix D, Table 11.

All changes are marked in blue. We hope these additions and our responses address your concerns and support a more positive assessment of our work. We would appreciate your feedback and would be happy to clarify any remaining questions.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Manuscript Revisions

Dear Area Chair,

Thank you for overseeing the review of our paper. FGTC is a training-free framework that uses frequency-specific temporal references to improve reconstruction fidelity in diffusion-based Real-ISR. In response to the reviewers' comments, we have revised the manuscript and added several experiments and analyses to clarify its contributions.

  1. Direct comparison with StructSR. Under matched backbones and sampling budgets, FGTC achieves higher PSNR, SSIM, and MS-SSIM in all six settings, with PSNR improvements of 0.19–0.72 dB over StructSR. LPIPS and DISTS also improve in five settings. Quantitative and visual comparisons are provided in Section 4.2, Table 2, and Figure 5.

  2. Reference preference after timestep control. After controlling for the sampling timestep, partial Spearman correlations still favor the early reference for LL (0.742 vs. 0.167) and recent history for HF (0.726 vs. 0.623). Analyses within three sampling stages show the same preference, providing further evidence for FGTC's frequency-specific reference design (Section 3.1 and Appendix A, Tables 7–8).

  3. Image-level consistency. With the LL reference fixed, using recent HF history instead of the early HF reference improves PSNR on 89 of 93 DRealSR images with SeeSR. The mean gain is 0.12 dB, with a 95% bootstrap confidence interval of [0.046, 0.253] dB, indicating that the improvement is shared across most images (Appendix B.2, Table 10).

  4. Perceptual trade-off analysis. We added representative examples to explain the different perceptual responses of SeeSR+FGTC on DIV2K-Val. The stained-glass example improves all five metrics, while the building example demonstrates improved structural fidelity with some perceptual cost when fine structures are severely degraded (Section 4.3, Figure 6, and Table 3).

  5. Computational efficiency. FGTC requires less runtime than StructSR on both backbones and introduces minimal additional peak GPU memory. We also report its runtime overhead relative to the original backbones (Appendix D, Table 11).

Together, these additions provide stronger empirical support for the frequency-specific temporal reference design, demonstrate consistent reconstruction improvements, and clarify the perceptual and computational trade-offs. All revisions have been incorporated into the updated manuscript and marked in blue.

We sincerely appreciate the reviewers' constructive feedback and your time and consideration. We hope these additional analyses and revisions help clarify the main concerns raised during the review.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.