acceptodds
Withdrawn from ICLR 2027

Faithful or Hallucinated? A Fidelity Metric and Benchmark for Real-World Super-Resolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world image super-resolution aims to restore high-resolution images from low-resolution inputs under unknown degradations. Reliable SR should be both perceptually appealing and faithful to the LR observation, meaning that restored content should not contradict the visual evidence in the input. Automatically assessing such fidelity is challenging without paired HR references. Existing metrics either ignore LR consistency or rely on degraded LR proxies that may penalize valid restoration changes. Human judgments can identify input-inconsistent content but are costly to collect at scale. These challenges motivate an HR-reference-free metric for LR–SR fidelity assessment. To this end, we propose FAIM, a Fidelity-Aware Input-consistency Metric for real-world super-resolution. FAIM leverages frozen multi-layer DINOv3 features to assess LR–SR fidelity by jointly modeling spatial discrepancies and region-level contextual inconsistencies. We further introduce FAIM-Bench, a human-annotated benchmark containing 3,720 LR–SR pairs constructed from 465 real-world LR images and eight diffusion-based SR models. It includes separate ratings of LR quality, SR perceptual quality, and LR–SR fidelity, together with 13,020 unique pairwise fidelity comparisons. Experiments show that FAIM agrees more strongly with human fidelity judgments than existing metrics, including on unseen SR methods. In addition, FAIM supports zero-shot artifact localization without additional localization supervision.

Primary area:
datasets and benchmarks
Keywords:
Real-World Image Super-Resolution, Fidelity Assessment, LR-SR fidelity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gXaROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces FAIM, a super-resolution fidelity metric that does not need an HR reference. It judges whether the SR output contradicts its LR input. First, the LR image is upsampled with bicubic interpolation. Then, a frozen DINOv3 extracts multi-layer features from both the upsampled LR and the SR images, and small learned modules score the differences between them through two branches (a local branch and a region-level branch). The final score is the negative sum of both branches, and their maps can be used for artifact localization. The authors also propose a benchmark, FAIM-Bench, with 465 LR images from five real-world SR datasets, each super-resolved with eight diffusion-based models (3,720 LR–SR pairs). Twelve annotators rated LR quality, SR quality, and LR–SR fidelity, with at least eight judgments per pair.

Critical strengths
  • Well-posed problem.
  • The benchmark, although small, could be useful for the community.
Critical weaknesses
  1. Formulation of the problem. The paper is motivated by the lack of GT data in real-world SR. However, FAIM is trained on the labels of its own benchmark and all fidelity results are reported on it, where FAIM is the only method trained in-domain. Holding out SR methods helps, but the annotators and the protocol remain the same. Evaluating FAIM on independently collected human judgments, and on non-diffusion SR methods, would show whether it generalizes beyond this benchmark.
  2. Missing related work and baselines. The authors miss two important works that, like FAIM, do not assume that the reference is perfect: A-FINE [1] and RQI [2]. I think both should be evaluated on FAIM-Bench. In addition, the Hallucination Score [3], although a preprint, also targets hallucinations in generative SR and should be discussed in the related work.
  3. Benchmark size. The benchmark is a central part of the paper, as FAIM is both trained and evaluated on it. However, it is small, 465 LR images (only 149 used for training), 12 annotators, and at least eight judgments per pair, below the 15 observers usually recommended for subjective studies (ITU-R BT.500). Together with the low agreement between annotators (Krippendorff's alpha of 0.307), I am not sure that solid conclusions can be drawn from it.
  4. Effectiveness. FAIM reaches an average SRCC of 0.458. In Table 2, several components barely help (e.g., RCCB alone obtains 0.452 versus 0.458 for the full model). A noise ceiling for the human scores would help to judge whether 0.458 is a good result.

[1] D. Chen et al. Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption. CVPR 2025.

[2] S. Su. et al. Bridging the Perception Gap in Image Super-Resolution Evaluation. CVPR 2026.

[3] W. Ren et al. Hallucination Score: Towards Mitigating Hallucinations in Generative Image Super-Resolution. Arxiv 2025.

Questions for authors
  1. Why do 6 different ablation variants nearly identical numbers? Given Table 2, do most of the modules actually contribute to the final performance?
  2. How is the localization of FAIM evaluated? Fig. 5 is only qualitative, FAIM-Bench has no artifact annotations, and on SR-Prominence FAIM is below the baselines in pixel-level metrics.
8 comments
AuthorsOpenReview ↗

Response to the Review (1/4)

We thank the reviewer for the constructive feedback. In response, we have added an evaluation using newly collected human judgments on images outside FAIM-Bench, including non-diffusion SR methods (Appendix G.1); expanded the discussion of A-FINE, RQI, and Hallucination Score and added the requested A-FINE and RQI comparisons (Section 2.2, Section 5.2, and Appendix G.2); expanded the annotation-agreement and split-half reproducibility analyses (Appendix B.4); and expanded the quantitative artifact-localization evaluation and the analysis of architectural contributions (Appendices G.5–G.6). All revisions to the manuscript are highlighted in blue. Below, we provide point-by-point responses to the weaknesses and questions raised in your review.

W1: In-domain training, independent human evaluation, and non-diffusion generalization

R: To assess generalization beyond diffusion-based SR and ×4 upscaling, we evaluated FAIM on 20 RealLQ250 LR images disjoint from FAIM-Bench. This evaluation covered six methods from three SR families: regression-based (RCAN), GAN-based (MambaIR_real, BSRGAN, and FeMaSR), and diffusion-based (DreamClear and SeeSR). We tested ×4 and ×16, with seven annotators evaluating 120 LR–SR pairs per scale (840 fidelity ratings). Human fidelity scores combine absolute ratings (Stage I) and relative rankings (Stage II). Full details are provided in Appendix G.1.

We compare FAIM with 11 IQA metrics (four FR-IQA and seven NR-IQA) and four additional methods:

  • RefReward-SR (ECCV 2026) [2]: Fine-tunes Qwen3-VL-8B on human SR preferences to assess overall SR quality.
  • A-FINE (CVPR 2025) [3]: Adaptively combines fidelity and naturalness to assess image quality.
  • RQI (CVPR 2026) [4]: Learns relative quality differences to align SR evaluation with human perceptual preferences.
  • Hallucination Score (arXiv 2026) [5]: Uses multimodal large language models to assess hallucinations in SR outputs. We adopt its GPT-NR variant using gpt-6.1-sol.

(1) Overall performance across scales.

Table 1. SRCC/PLCC-log comparison of ILR–SR proxy-reference metrics (Appendix G.1, Tables 18–19).

ScalePSNRSSIMLPIPSDISTSRQIA-FINERefReward-SRHallucination ScoreFAIM (Ours)
×40.422/0.4490.407/0.4610.478/0.5310.443/0.479-0.295/0.3990.304/0.4410.468/0.6680.484/0.4970.638/0.664
×160.642/0.6360.607/0.6200.575/0.6010.608/0.658-0.578/0.6720.633/0.6600.290/0.3290.658/0.6660.757/0.763

Table 2. SRCC/PLCC-log comparison of NR-IQA metrics (Appendix G.1, Tables 18–19).

ScaleNIQEMANIQACLIP-IQAMUSIQQ-AlignTOPIQ-NRLIQE
×4-0.292/0.400-0.173/0.303-0.136/0.287-0.325/0.466-0.065/0.202-0.302/0.521-0.371/0.474
×16-0.561/0.644-0.270/0.327-0.484/0.590-0.446/0.479-0.189/0.262-0.551/0.686-0.577/0.641

As shown in Tables 1 and 2, at ×4, FAIM achieves the highest SRCC of 0.638, exceeding the strongest baseline, Hallucination Score (0.484), by 0.154. Its PLCC-log of 0.664 closely matches the best result of 0.668 from RefReward-SR. At ×16, FAIM ranks first in both SRCC (0.757) and PLCC-log (0.763), outperforming the strongest baselines by 0.099 and 0.077, respectively. These results demonstrate strong overall agreement with human fidelity judgments across both upscaling factors.

(2) Generalization across SR families.

We further examine FAIM's performance across different SR families and training exposure. Correlations are computed across 20 images per method and averaged within each group. Brackets show 95% CIs from 5,000 bootstrap resamples of LR images.

Table 3. FAIM correlations across SR families (Appendix G.1, Table 20).

ScaleMetricRegressionGAN-basedDiffusion-basedDreamClear (unseen)SeeSR (seen)Five unseen methods
×4SRCC0.341[-0.179,0.738]0.592[0.303,0.780]0.104[-0.301,0.505]0.033[-0.464,0.522]0.174[-0.300,0.627]0.430[0.233,0.587]
PLCC-log0.488[0.260,0.805]0.719[0.589,0.890]0.547[0.355,0.789]0.501[0.201,0.813]0.592[0.344,0.868]0.629[0.544,0.793]
×16SRCC-0.132[-0.549,0.338]0.552[0.253,0.747]-0.208[-0.522,0.167]-0.368[-0.773,0.119]-0.047[-0.463,0.436]0.231[-0.025,0.452]
PLCC-log0.180[0.123,0.726]0.666[0.524,0.847]0.465[0.353,0.788]0.519[0.282,0.845]0.411[0.221,0.835]0.539[0.492,0.745]

FAIM achieves positive correlations on unseen GAN-based methods at both scales (95% CIs above zero). Across all five unseen methods, average correlations are positive in both metrics at ×4 and in PLCC-log at ×16. Together, these findings support FAIM’s applicability to unseen SR methods beyond diffusion-based SR.

In summary, FAIM leads in overall SRCC at both scales and PLCC-log at ×16, with positive correlations on unseen GAN-based methods supporting generalization beyond diffusion-based SR. Details are provided in Appendix G.1, Tables 18–20.

AuthorsOpenReview ↗

Response to the Review (2/4)

W2: Missing Related Work and Baseline Comparisons

R: We have expanded Section 2.2 (Image Quality Assessment) to discuss A-FINE, RQI, and Hallucination Score, and additionally included RefReward-SR in our related work and baseline comparisons.

Specifically, A-FINE, RQI, and RefReward-SR have been incorporated into the FAIM-Bench evaluation under the same protocols as other metrics. Hallucination Score is evaluated on the separate 20-image RealLQ250 evaluation set in Appendix G.1 using gpt-6.1-sol, rather than on the full FAIM-Bench, due to the substantial computational cost of MLLM-based evaluation.

The corresponding revisions are located at:

  • Related Work: Section 2.2 (Image Quality Assessment).
  • Main Experiments: Sections 5.1–5.2 and Table 1.
  • Additional Analyses: Appendix G.1, Tables 18–19 (external evaluation); Appendix G.2, Table 21 (known-method evaluation); Appendix G.3, Table 22 (ground-truth HR assessment); and Appendix G.8, Table 37 (computational costs).

W3: Benchmark Size and Annotation Reliability

R: We acknowledge that FAIM-Bench has a relatively limited number of LR images and annotators. While ITU-R BT.500 generally recommends at least 15 observers, our study involves 12 annotators, with at least eight valid judgments per LR–SR pair. We recognize this as a limitation. To further assess the reliability of the collected annotations, we conducted additional agreement and reproducibility analyses.

Importantly, the originally reported Krippendorff's α = 0.307 reflects agreement on individual absolute ratings rather than the final aggregated fidelity labels. As described in Sec. 3.2 and Appendix B.3, our two-stage annotation protocol combines five-level absolute fidelity ratings (Stage I) with relative rankings of eight SR outputs sharing the same LR input (Stage II). Following Ye and Doermann [5], both sources are jointly aggregated into the final latent fidelity scores. We therefore assess annotation reliability at three levels:

  • For absolute ratings, we identified and corrected a calculation error that excluded the lowest valid rating category. The corrected interval Krippendorff's α\alpha is 0.359, compared with the previously reported 0.307.

  • For relative rankings, Stage II achieves a Kendall's τ-b of 0.312, measuring rank-order consistency between annotators, and a pairwise ordering agreement of 65.6%, indicating that annotators agree on which SR output is better in 65.6% of pairwise comparisons.

  • For the final aggregated fidelity labels, we conducted a split-half analysis across all 465 LR images. We independently reconstructed the hybrid fidelity scores from two non-overlapping groups of six annotators across 184 valid splits. The median macro-averaged SRCC between the two groups' scores is 0.571. Across all 13,020 within-LR output pairs, the median ordering agreement reaches 77.1%, exceeding the 63.0% obtained using absolute ratings alone under identical splits.

These results demonstrate that, despite variability in individual judgments, the final aggregated fidelity labels retain reproducible ranking information across annotator groups. We have corrected Table 4 and expanded Appendix B.4 to report the detailed agreement and reproducibility analyses.

AuthorsOpenReview ↗

Response to the Review (3/4)

W4&Q1: Architectural Contributions and Human-Score Noise Ceiling

R: We provide additional statistical analyses to clarify the contributions of individual components and examine the human-score noise ceiling to help interpret FAIM's performance.

(1) Architectural Contributions.

FAIM integrates multi-level features, spatial discrepancies, and region-level contextual reasoning. The ablations show that different components contribute to different extents.

  • Component effectiveness. We ran paired bootstrap analyses with 9,999 resamples using the predictions underlying main-text Table 2 and applied Holm correction to the resulting p-values. Cross-Attention significantly improves SRCC (Δ = +0.0687, p = 0.0014), while multi-layer DINOv3 features significantly improve PLCC-log (Δ = +0.0625, p = 0.0221). Additional Appendix G.6 ablations under the known-method protocol support Spatial Discrepancy Gate, Geometry, and Contextual Contradiction (ΔSRCC = +0.0605, +0.0308, +0.0168; adjusted p = 0.014 for each).

  • Overall architectural benefit. Across three test datasets (RealDeg-test, RealSR, and RealPhoto60) and two evaluation protocols (leave-one-SR-method-out and known-method LR-held-out), Full outperforms the stronger single-branch variant by up to 0.035 SRCC, while never underperforming it by more than 0.010 SRCC. In contrast, RCCB-only and SCB-only exhibit larger performance gaps of up to 0.051 and 0.113 SRCC, respectively. These results highlight the complementary strengths of SCB and RCCB, with their combination providing more consistent performance across datasets and evaluation settings.

  • Computational efficiency. Table 4 compares the performance and computational cost of the architectural variants under the known-method LR-held-out protocol. Full FAIM adds at most 0.023 MB of model storage over the ablated variants, with latency increases of 0.025–0.328 s per image pair.

Table 4. Accuracy and computational cost of architectural ablations. The parameter and storage columns exclude the shared frozen DINOv3 backbone (Appendix G.6, Table 34).

VariantAvg. SRCC / PLCC-log ↑FAIM parameters (M) ↓FAIM model size (MB) ↓GMACs ↓Latency (s) ↓Peak GPU memory (GB) ↓
Full FAIM0.719 / 0.7660.0460.1751410.2820.9431.520
SCB-only0.699 / 0.7460.0460.1751399.4030.6151.520
RCCB-only0.684 / 0.7320.0460.1751410.2630.9181.520
DINOv3 layer 24 only0.564 / 0.6160.0400.1521409.8110.8561.464
w/o Cross-Attention0.697 / 0.7540.0460.1751410.2820.8521.520
w/o Spatial Discrepancy Gate0.658 / 0.7130.0460.1751410.2820.8251.520
w/o Geometry0.688 / 0.7300.0460.1751410.2820.8861.520
w/o Contextual Contradiction0.702 / 0.7440.0460.1751399.4080.6161.520

Together, these results support the full architecture as a balanced design that combines effective components with consistent performance across evaluation settings.

(2) Human-Score Noise Ceiling.

A noise ceiling refers to the highest agreement with human scores that a model can be expected to achieve given variability in human judgments. To examine this variability, we assessed the reproducibility of independently aggregated human scores through a split-half analysis (Appendix B.4).

We divided the 12 annotators into two non-overlapping groups of six and retained 184 valid splits, with at least four absolute ratings per LR–SR pair in each group. For each split, we independently reconstructed each group's hybrid fidelity scores from its absolute ratings and relative rankings over all 465 LR images. On the fixed test set of 279 LR images, we computed human–human SRCC separately for each dataset and SR method, then averaged equally over methods and datasets, following the aggregation protocol in main-text Table 2.

The median human–human SRCC across these splits was 0.575, providing an empirical reference for FAIM's average SRCC of 0.458. This result shows that the aggregated human scores retain reproducible fidelity information while remaining affected by annotation variability. However, 0.575 is not a strict upper bound on FAIM's performance: the split-half analysis compares two smaller annotator groups, whereas FAIM is evaluated against the full consensus.

Reviewer_q9vJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a new metric and benchmark to assess the quality of real-world super resolution (SR) predictions algorithms. In real-world SR, a high resolution image needs to be restored from a low resolution (LR) image, and the degradation is not known. A reliable SR should be both perceptually appealing while still be faithful to the LR observation. The paper suggests a new HR-reference-free metric FAIM, that also account for LR-SR faithfulness. Additionally, a new benchmark is introduced, where human annotators have benchmarked LR images as well as the results of 8 diffusion based SR methods, providing scores on LR quality, SR prediction quality, and LR-SR faithfulness. In a mostly rank-based analysis, the paper shows that FAIM is more aligned with human evaluation than other standard metrics.

My current recommendation is to reject the paper in its current form. My main concerns are centered around the fact that the method and its evaluation are based on results of 8 diffusion models only. There is no evaluation or indication that this metric can be generally used for real-SR results of any kind of SR methods, significantly limiting the contribution.

Critical strengths
  1. The paper is generally written understandably, with extensive details. Ablations are provided in the supplementary material.
  2. The human labelled dataset can be a valuable contribution to the community.
Critical weaknesses
  1. Limited scope: The paper focusses its evaluation on the results of 8 diffusion based methods only. While many state of the art methods fall in this category (i.e. due to their capability to create perceptually appealing results), the metric and results can be biased to these kind of methods. No analysis or results are shown on non-diffusion based methods. This should be clearly pointed out and the scope should be limited ("real world SR metric for diffusion based SR algorithms"). Alternatively, more resultss should be provided that show the generalization capability. In its current form, I don't see the metric as broadly usable for evaluation of real-world SR methods.

  2. Complexity of the network: The complexity of the provided network seems to rather high. The provided ablations in Table 2, also seem to indicate that even when ablating (significant) parts of the network, one can still achieve similar results. As pointed out in the limitations, patch-level Dinov3 comes with the limitation that pixel level artifacts are not considered in the evaluation.

  3. Evaluation: The quality scales of LR and SR are defined in different ways (i.e. a score 2 in the LR scale might not be equivalent to a score 2 in the SR scale).

  4. Related work: The single image super resolution paragraph is written too narrowly. If the paper is framed about diffusion based SR methods, fewer references would need to be added (such as SR3, SRDiff, etc.). If the intention is more generally about real-world SR, there is a whole line works regarding blind super resolution that is closely connected to real world super resolution that should also be discussed.

  5. The promoted zero-short artifact localization is hard to interpret as it is very noisy. I understand that these results in principle should also take into account the user study results. However, a detailed analysis on how the results would be more informative compared to a more simple approach (e.g. a simple difference image between LR and SR, or default dino features) is missing.

Minor:

  • Many of the figures contain tiny fonts and are only readable when zooming in significantly. I heavily suggest the authors to revise the figures regarding readability.
Questions for authors

See weaknesses

4 comments
AuthorsOpenReview ↗

Response to the Review (1/3)

We thank the reviewer for the detailed comments on the scope, design, and presentation of our work. In response, we have added an external evaluation across SR families and upscaling factors (Appendix G.1); expanded the architectural and computational analyses and added a controlled study of sensitivity to localized perturbations (Appendices G.4, G.6, and G.8); clarified the distinction between LR quality and SR perceptual quality and broadened the related work on blind and diffusion-based super-resolution (Appendix B.2 and Section 2.1); and added quantitative artifact-localization comparisons with RGB-L1 and frozen DINOv3-Cosine baselines (Appendix G.5). All revisions to the manuscript are highlighted in blue. Below, we provide point-by-point responses to the weaknesses and questions raised in your review.

W1: Generalization beyond diffusion-based SR methods

R: To assess generalization beyond diffusion-based SR and ×4 upscaling, we evaluated FAIM on 20 RealLQ250 LR images disjoint from FAIM-Bench. This evaluation covered six methods from three SR families: regression-based (RCAN), GAN-based (MambaIR_real, BSRGAN, and FeMaSR), and diffusion-based (DreamClear and SeeSR). We tested ×4 and ×16, with seven annotators evaluating 120 LR–SR pairs per scale (840 fidelity ratings). Human fidelity scores combine absolute ratings (Stage I) and relative rankings (Stage II). Full details are provided in Appendix G.1.

We compare FAIM with 11 IQA metrics (four FR-IQA and seven NR-IQA) and four additional methods:

  • RefReward-SR (ECCV 2026) [2]: Fine-tunes Qwen3-VL-8B on human SR preferences to assess overall SR quality.
  • A-FINE (CVPR 2025) [3]: Adaptively combines fidelity and naturalness to assess image quality.
  • RQI (CVPR 2026) [4]: Learns relative quality differences to align SR evaluation with human perceptual preferences.
  • Hallucination Score (arXiv 2026) [5]: Uses multimodal large language models to assess hallucinations in SR outputs. We adopt its GPT-NR variant using gpt-6.1-sol.

(1) Overall performance across scales.

Table 1. SRCC/PLCC-log comparison of ILR–SR proxy-reference metrics (Appendix G.1, Tables 18–19).

ScalePSNRSSIMLPIPSDISTSRQIA-FINERefReward-SRHallucination ScoreFAIM (Ours)
×40.422/0.4490.407/0.4610.478/0.5310.443/0.479-0.295/0.3990.304/0.4410.468/0.6680.484/0.4970.638/0.664
×160.642/0.6360.607/0.6200.575/0.6010.608/0.658-0.578/0.6720.633/0.6600.290/0.3290.658/0.6660.757/0.763

Table 2. SRCC/PLCC-log comparison of NR-IQA metrics (Appendix G.1, Tables 18–19).

ScaleNIQEMANIQACLIP-IQAMUSIQQ-AlignTOPIQ-NRLIQE
×4-0.292/0.400-0.173/0.303-0.136/0.287-0.325/0.466-0.065/0.202-0.302/0.521-0.371/0.474
×16-0.561/0.644-0.270/0.327-0.484/0.590-0.446/0.479-0.189/0.262-0.551/0.686-0.577/0.641

As shown in Tables 1 and 2, at ×4, FAIM achieves the highest SRCC of 0.638, exceeding the strongest baseline, Hallucination Score (0.484), by 0.154. Its PLCC-log of 0.664 closely matches the best result of 0.668 from RefReward-SR. At ×16, FAIM ranks first in both SRCC (0.757) and PLCC-log (0.763), outperforming the strongest baselines by 0.099 and 0.077, respectively. These results demonstrate strong overall agreement with human fidelity judgments across both upscaling factors.

(2) Generalization across SR families.

We further examine FAIM's performance across different SR families and training exposure. Correlations are computed across 20 images per method and averaged within each group. Brackets show 95% CIs from 5,000 bootstrap resamples of LR images.

Table 3. FAIM correlations across SR families (Appendix G.1, Table 20).

ScaleMetricRegressionGAN-basedDiffusion-basedDreamClear (unseen)SeeSR (seen)Five unseen methods
×4SRCC0.341[-0.179,0.738]0.592[0.303,0.780]0.104[-0.301,0.505]0.033[-0.464,0.522]0.174[-0.300,0.627]0.430[0.233,0.587]
PLCC-log0.488[0.260,0.805]0.719[0.589,0.890]0.547[0.355,0.789]0.501[0.201,0.813]0.592[0.344,0.868]0.629[0.544,0.793]
×16SRCC-0.132[-0.549,0.338]0.552[0.253,0.747]-0.208[-0.522,0.167]-0.368[-0.773,0.119]-0.047[-0.463,0.436]0.231[-0.025,0.452]
PLCC-log0.180[0.123,0.726]0.666[0.524,0.847]0.465[0.353,0.788]0.519[0.282,0.845]0.411[0.221,0.835]0.539[0.492,0.745]

FAIM achieves positive correlations on unseen GAN-based methods at both scales (95% CIs above zero). Across all five unseen methods, average correlations are positive in both metrics at ×4 and in PLCC-log at ×16. Together, these findings support FAIM’s applicability to unseen SR methods beyond diffusion-based SR.

In summary, FAIM leads in overall SRCC at both scales and PLCC-log at ×16, with positive correlations on unseen GAN-based methods supporting generalization beyond diffusion-based SR. Details are provided in Appendix G.1, Tables 18–20.

AuthorsOpenReview ↗

Response to the Review (2/3)

W2: Network complexity, limited gains from individual components, and sensitivity to fine-grained artifacts

R: (1) Network Complexity.

FAIM integrates multi-level features, spatial discrepancies, and region-level contextual reasoning. The ablations show that different components contribute to different extents.

  • Component effectiveness. We ran paired bootstrap analyses with 9,999 resamples using the predictions underlying main-text Table 2 and applied Holm correction to the resulting p-values. Cross-Attention significantly improves SRCC (Δ = +0.0687, p = 0.0014), while multi-layer DINOv3 features significantly improve PLCC-log (Δ = +0.0625, p = 0.0221). Additional Appendix G.6 ablations under the known-method protocol support Spatial Discrepancy Gate, Geometry, and Contextual Contradiction (ΔSRCC = +0.0605, +0.0308, +0.0168; adjusted p = 0.014 for each).

  • Overall architectural benefit. Across three test datasets (RealDeg-test, RealSR, and RealPhoto60) and two evaluation protocols (leave-one-SR-method-out and known-method LR-held-out), Full outperforms the stronger single-branch variant by up to 0.035 SRCC, while never underperforming it by more than 0.010 SRCC. In contrast, RCCB-only and SCB-only exhibit larger performance gaps of up to 0.051 and 0.113 SRCC, respectively. These results highlight the complementary strengths of SCB and RCCB, with their combination providing more consistent performance across datasets and evaluation settings.

  • Computational efficiency. Table 4 compares the performance and computational cost of the architectural variants under the known-method LR-held-out protocol. Full FAIM adds at most 0.023 MB of model storage over the ablated variants, with latency increases of 0.025–0.328 s per image pair.

Table 4. Accuracy and computational cost of architectural ablations. The parameter and storage columns exclude the shared frozen DINOv3 backbone (Appendix G.6, Table 34).

VariantAvg. SRCC / PLCC-log ↑FAIM parameters (M) ↓FAIM model size (MB) ↓GMACs ↓Latency (s) ↓Peak GPU memory (GB) ↓
Full FAIM0.719 / 0.7660.0460.1751410.2820.9431.520
SCB-only0.699 / 0.7460.0460.1751399.4030.6151.520
RCCB-only0.684 / 0.7320.0460.1751410.2630.9181.520
DINOv3 layer 24 only0.564 / 0.6160.0400.1521409.8110.8561.464
w/o Cross-Attention0.697 / 0.7540.0460.1751410.2820.8521.520
w/o Spatial Discrepancy Gate0.658 / 0.7130.0460.1751410.2820.8251.520
w/o Geometry0.688 / 0.7300.0460.1751410.2820.8861.520
w/o Contextual Contradiction0.702 / 0.7440.0460.1751399.4080.6161.520

Together, these results support the full architecture as a balanced design that combines effective components with consistent performance across evaluation settings.

(2) Pixel-Level Artifacts.

FAIM uses DINOv3-Large to represent images with feature tokens corresponding to 16×16 patches. This patch-level representation may reduce sensitivity to small, localized artifacts, which is the limitation acknowledged in Sec. 6.

To quantify this limitation, we added a controlled sensitivity study in Appendix G.4. Using FAIM without retraining, we introduced LR-inconsistent perturbations of 4×4 to 64×64 pixels into SR candidates from 24 scenes while keeping the LR inputs fixed. We measured the change in image-level fidelity scores as Δs=Sclean−Sperturbed\Delta s = S_{clean} − S_{perturbed}.

As reported in Table 23 (Appendix G.4), perturbations of 4×4, 8×8, 16×16, 32×32, and 64×64 pixels decrease the fidelity score in 50.0%, 70.8%, 87.5%, 91.7%, and 95.8% of the scenes, respectively. The corresponding mean score decreases are 0.0020, 0.0373, 0.1514, 0.3765, and 0.6251.

These results show that FAIM can respond to localized inconsistencies, including some perturbations smaller than a single patch. Responses become more consistent as perturbation size increases, while sensitivity to very small artifacts remains limited.

W3: Comparability of LR and SR quality scales

R: As defined in Appendix B.2, LR quality measures how much visual evidence is available in the LR input itself. Annotators judge the recognizability and informativeness of the LR image without considering any SR result. In contrast, SR perceptual quality measures the visual quality of the restored image. Annotators evaluate whether the SR image looks clear, natural, and artifact-free, regardless of whether it is faithful to the LR input. Although both aspects are rated on five-level discrete scales, their rating criteria are defined separately, and identical numerical ratings do not imply equivalent quality across the two dimensions.

AuthorsOpenReview ↗

Response to the Review (3/3)

W4: Insufficient coverage of diffusion-based and blind super-resolution literature

R: We agree that the original Section 2.1 (Single Image Super-Resolution) focused too narrowly on recent diffusion-based SR methods and did not adequately reflect the broader development of single-image super-resolution. We have therefore revised and reorganized Section 2.1 to provide a more comprehensive overview, progressing from early regression-based methods to GAN-based approaches, followed by blind SR and diffusion-based methods for real-world restoration.

W5: Localization interpretability and comparisons with LR–SR differences and DINO features

R: We clarify that FAIM's zero-shot artifact localization arises from its spatial and region-level inconsistency modeling (SCB and RCCB), rather than localization-specific supervision. While human judgments supervise image-level fidelity assessment, artifact localization is neither an explicit training objective nor a prerequisite for fidelity assessment.

To assess these maps against human annotations, we added comparisons on the official OpenImages and Urban100 subsets of SR-Prominence [5], which provide human-annotated artifact regions and artifact-prominence ratings. FAIM is evaluated without training or adaptation on this benchmark. The experimental protocol and baseline definitions are provided in Appendix G.5.

We compare FAIM with two simple discrepancy maps using the same bicubic-upsampled LR (ILR) and SR image pairs:

  • RGB-L1: The mean absolute difference across the three RGB channels at each pixel, providing a direct pixel-level ILR–SR difference map.
  • DINOv3-Cosine: The cosine distance (1 − cosine similarity) between corresponding final-layer patch features extracted from frozen DINOv3 ViT-L/16, providing a feature-level ILR–SR difference map.

All maps are evaluated against the same human annotations. Table 5 compares the three methods using AUPRC for pixel-level localization and Prominence SRCC for agreement with human artifact-prominence ratings. Prominence SRCC measures the Spearman correlation between the difference in median responses inside and outside the artifact regions and the human ratings.

Table 5. Pixel-level localization (AUPRC) and agreement with human artifact-prominence ratings (Prominence SRCC) on SR-Prominence (Appendix G.5, Tables 24–25).

DatasetMeasureRGB-L1DINOv3-CosineFAIM-RCCB
OpenImagesAUPRC0.5230.4780.569
Urban100AUPRC0.3760.3900.411
OpenImagesProminence SRCC0.3360.3890.618
Urban100Prominence SRCC0.1600.1690.358

These results show that FAIM-RCCB provides more accurate localization and stronger agreement with human artifact prominence than the two simple discrepancy baselines under the same ILR–SR input setting. The maps provide spatial inconsistency cues without being trained to predict precise artifact boundaries.

W6: Figure Readability and Presentation

R: We have increased the font sizes of figure labels and legends and enlarged key visual examples to improve readability without excessive zooming.

References

[1] Y. Song et al. RefReward-SR: LR-Conditioned Reward Modeling for Preference-Aligned Super-Resolution. ECCV 2026.

[2] D. Chen et al. Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption. CVPR 2025.

[3] S. Su et al. Bridging the Perception Gap in Image Super-Resolution Evaluation. CVPR 2026.

[4] W. Ren et al. Hallucination Score: Towards Mitigating Hallucinations in Generative Image Super-Resolution. arXiv 2026.

[5] Ivan Molodetskikh et al. SR-Prominence: A Crowdsourced Protocol and Dataset Suite for Perceptually-Weighted Super-Resolution Artifact Evaluation. arXiv 2026.

Reviewer_q9vJOpenReview ↗

Rebuttal Answer

I thank the authors for their efforts that they put into the rebuttal and to update the script (e.g. the related work section and the implementation of the additional experiments). After having read the rebuttal (and the other reviews), I decided to keep my score. My main concerns still remain around the limited scope and the evaluation itself.

While, the new (small) external study on RealLQ250 shows potential for generalization across scales and methods, it is too small to be conclusive (I understand that it is not possible to perform a full scale study within the rebuttal period). Further, it shows new issues arise e.g. on DreamClear, a diffusion method unseen during training of FAIM. Additionally, the LR-conditioned methods (in the rebuttal one method, RefReward-SR, was added) seem to provide a stronger baseline, even surpassing FAIM RealDeg-test and RealSR in Table 1.

Reviewer_8K4ROpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FAIM, an LR-conditioned fidelity metric for real-world super-resolution, and FAIM-Bench, a human-annotated benchmark separating perceptual quality from compatibility with observable LR evidence. FAIM combines frozen multi-layer DINOv3 features with spatial and region-context discrepancy modeling. The main evaluation reports improved agreement with human fidelity scores.

Critical strengths

The authors provided access to the entire code repository and dataset, making their results easily reproducible. The task is relevant and clearly distinguished from perceptual-quality assessment. The held-out-method protocol provides evidence of improvement over the reported conventional baselines. Additionally, FAIM demonstrates zero-shot artifact localization, which is a promising result.

Critical weaknesses

The human-annotated dataset includes only 12 annotators and inter-annotator agreement for absolute fidelity ratings is limited (Krippendorff’s alpha = 0.307), so small differences between metrics should be interpreted cautiously. Aggregation of at least eight judgments per LR-SR pair reduces this uncertainty, but does not remove it. Furthermore, the dataset contains fewer than 500 original images and considers only diffusion-based super-resolution methods, limiting its diversity and generalizability.

While comparisons with standard metrics are informative, relying exclusively on full-reference with ILR and no-reference metrics may not provide a fair assessment of the proposed method. A more comprehensive evaluation should also include metrics specifically designed or validated for the task at hand, enabling a more appropriate baseline.

Questions for authors
  • Can the authors compare FAIM with any applicable LR-conditioned metrics under the same protocol?
  • Can the authors test FAIM on unseen non-diffusion SR methods and other scales, reporting per-family correlations and confidence intervals?
4 comments
AuthorsOpenReview ↗

Response to the Review (1/3)

We thank the reviewer for the careful assessment and constructive suggestions. In response, we have expanded the analyses of annotation agreement and the reproducibility of the aggregated fidelity labels (Appendix B.4); added comparisons with the LR-conditioned RefReward-SR metric under both evaluation protocols, together with statistical and computational analyses (Section 5.2 and Appendices G.2, G.7, and G.8); and added an external evaluation using 20 LR images selected from RealLQ250, covering regression-based, GAN-based, and diffusion-based SR methods at ×4 and ×16 scales, including per-family correlations and confidence intervals (Appendix G.1). All revisions to the manuscript are highlighted in blue. Below, we provide point-by-point responses to the weaknesses and questions raised in your review.

W1: Reliability of human fidelity annotations

R: The reported Krippendorff’s α measures agreement on individual absolute ratings, not the reliability of the final aggregated fidelity labels. As described in Sec. 3.2 and Appendix B.3, FAIM-Bench adopts a two-stage annotation protocol. Stage I collects five-level absolute fidelity ratings, while Stage II collects relative fidelity rankings among SR outputs sharing the same LR input. Following Ye and Doermann [1], both sources are jointly aggregated into the final fidelity scores.

We evaluate annotation reliability at three levels.

  • For absolute ratings, we identified and corrected a calculation error that excluded the lowest valid rating category. The corrected interval Krippendorff's α\alpha is 0.359, compared with the previously reported 0.307.

  • For relative rankings, Stage II achieves a Kendall's τ-b of 0.312, measuring rank-order consistency between annotators, and a pairwise ordering agreement of 65.6%, indicating that annotators agree on which SR output is better in 65.6% of pairwise comparisons.

  • For the final aggregated fidelity labels, we conducted a split-half analysis across all 465 LR images. We independently reconstructed the hybrid fidelity scores from two non-overlapping groups of six annotators across 184 valid splits. The median macro-averaged SRCC between the two groups' scores is 0.571. Across all 13,020 within-LR output pairs, the median ordering agreement reaches 77.1%, exceeding the 63.0% obtained using absolute ratings alone under identical splits.

These results demonstrate that, despite variability in individual judgments, the final aggregated fidelity labels retain reproducible ranking information across annotator groups. We have corrected Table 4 and expanded Appendix B.4 to report the detailed agreement and reproducibility analyses.

W2: Benchmark size (fewer than 500 LR images), diversity, and generalization

R: (1) Benchmark size and annotation effort.

Each of the 465 LR images in FAIM-Bench was processed by eight SR methods, yielding 3,720 SR images. The two-stage annotation required approximately 186 annotator-hours, representing substantial human effort.

(2) SR-method diversity and generalization.

We initially focused on diffusion-based SR because its outputs may look clear and detailed but fail to preserve the input content, making fidelity assessment particularly important. Regression-based SR struggles to recover realistic textures, while GAN-based SR shares diffusion’s perceptual goal but produces less realistic details in our comparisons (Sec. 3.1; Fig. 8 and Appendix H.1).

To test generalization beyond diffusion-based SR, we conducted an additional human evaluation with seven annotators on 20 LR images selected from RealLQ250 that do not overlap with FAIM-Bench, covering regression-, GAN-, and diffusion-based methods at ×4 and cascaded ×16 scales. At ×4, FAIM achieves an SRCC/PLCC-log of 0.638/0.664. Its SRCC exceeds the strongest baseline, Hallucination Score (0.484), by 0.154, while its PLCC-log closely matches the best baseline result, RefReward-SR (0.668), with a difference of 0.004. At ×16, FAIM achieves 0.757/0.763, surpassing the strongest baselines by 0.099 in SRCC (Hallucination Score: 0.658) and 0.077 in PLCC-log (TOPIQ-NR: 0.686). Detailed results and per-family analyses are provided in our response to Q2 and Appendix G.1.

AuthorsOpenReview ↗

Response to the Review (2/3)

W3&Q1: Comparison with LR-conditioned metrics

R: We include RefReward-SR [2] as an LR-conditioned baseline. RefReward-SR aims to assess the overall quality of SR images by considering both their consistency with the LR input and their perceptual quality. In contrast, FAIM focuses exclusively on LR–SR fidelity, assessing whether restored content contradicts observable evidence in the LR input, regardless of its perceptual quality. This distinction is also reflected in their training objectives: RefReward-SR learns from human preferences for SR quality, whereas FAIM is trained specifically on human LR–SR fidelity judgments.

To investigate this distinction, we compare the two methods under two FAIM evaluation protocols: (1) leave-one-SR-method-out, where one SR method is excluded from FAIM training and validation in each fold, and (2) known-method LR-held-out, where FAIM is trained on the evaluated SR methods but tested on unseen LR images. We further compare their computational efficiency.

We report SRCC and PLCC-log. Within each dataset, the first protocol averages per-method correlations, while the second computes correlations over pooled SR outputs, which explains why the RefReward-SR results differ between Tables 1 and 2.

  • (1) Leave-one-SR-method-out evaluation.

Table 1. SRCC/PLCC-log under leave-one-SR-method-out evaluation (main-text Table 1).

MethodRealDeg-testRealSRRealPhoto60Average
SRCC/PLCC-logSRCC/PLCC-logSRCC/PLCC-logSRCC/PLCC-log
RefReward-SR0.441/0.5430.455/0.5310.340/0.4370.412/0.503
FAIM (Ours)0.401/0.4740.412/0.4750.562/0.6380.458/0.529

FAIM achieves higher average correlations than RefReward-SR, although RefReward-SR performs better on RealDeg-test and RealSR. On RealPhoto60, FAIM significantly outperforms RefReward-SR (ΔSRCC = +0.223, ΔPLCC-log = +0.201; Appendix G.7, Table 35). These gains are particularly relevant because RealPhoto60 has the lowest mean human-rated LR quality in FAIM-Bench (2.473; Fig. 3(b)). Such degraded inputs provide less reliable visual evidence, making fidelity assessment particularly challenging (Appendix C.3–C.4). FAIM's advantage on this dataset suggests that fidelity-specific supervision is beneficial under such conditions.

  • (2) Known-method LR-held-out evaluation.

Table 2. SRCC/PLCC-log under known-method LR-held-out evaluation (Appendix G.2, Table 21).

MethodRealDeg-testRealSRRealPhoto60Average
SRCC/PLCC-logSRCC/PLCC-logSRCC/PLCC-logSRCC/PLCC-log
RefReward-SR0.416/0.4430.392/0.4170.144/0.2060.317/0.355
FAIM (Ours)0.798/0.8520.661/0.6720.698/0.7740.719/0.766

FAIM consistently outperforms RefReward-SR across all three datasets, improving average SRCC from 0.317 to 0.719 and PLCC-log from 0.355 to 0.766. These results further support the effectiveness of learning from fidelity-specific human judgments rather than relying on general SR preference supervision.

(3) Computational efficiency.

We compare computational efficiency on 2048 × 2048 images from RealPhoto60 using a single NVIDIA RTX A6000 GPU with a batch size of 1. “Avg. time” denotes the average inference time per sample.

Table 3. Computational efficiency comparison (Appendix G.8, Table 37).

MetricRefReward-SRFAIM (Ours)
BackboneQwen3-VL-8BDINOv3-Large (0.3B)
Avg. time (s) ↓4.11210.8771
Peak GPU memory (GiB) ↓18.9151.520

RefReward-SR requires approximately 4.69 times the average inference time and 12.44 times the peak GPU memory of FAIM in this benchmark. Moreover, FAIM introduces only 0.046M trainable parameters beyond its frozen DINOv3 backbone (Appendix G.6, Table 34).

In summary, RefReward-SR aims to assess overall SR quality by considering both LR consistency and perceptual quality, whereas FAIM focuses exclusively on LR–SR fidelity. FAIM achieves higher average correlations with human fidelity judgments under both evaluation protocols, while offering substantially greater computational efficiency. These results demonstrate FAIM's strong effectiveness and practical value as a dedicated metric for LR–SR fidelity assessment.

AuthorsOpenReview ↗

Response to the Review (3/3)

Q2: Generalization across SR families and scales

R: To assess generalization beyond diffusion-based SR and ×4 upscaling, we evaluated FAIM on 20 RealLQ250 LR images disjoint from FAIM-Bench. This evaluation covered six methods from three SR families: regression-based (RCAN), GAN-based (MambaIR_real, BSRGAN, and FeMaSR), and diffusion-based (DreamClear and SeeSR). We tested ×4 and ×16, with seven annotators evaluating 120 LR–SR pairs per scale (840 fidelity ratings). Human fidelity scores combine absolute ratings (Stage I) and relative rankings (Stage II). Full details are provided in Appendix G.1.

We compare FAIM with 11 IQA metrics (four FR-IQA and seven NR-IQA) and four additional methods:

  • RefReward-SR (ECCV 2026) [2]: Fine-tunes Qwen3-VL-8B on human SR preferences to assess overall SR quality.
  • A-FINE (CVPR 2025) [3]: Adaptively combines fidelity and naturalness to assess image quality.
  • RQI (CVPR 2026) [4]: Learns relative quality differences to align SR evaluation with human perceptual preferences.
  • Hallucination Score (arXiv 2026) [5]: Uses multimodal large language models to assess hallucinations in SR outputs. We adopt its GPT-NR variant using gpt-6.1-sol.

(1) Overall performance across scales.

Table 4. SRCC/PLCC-log comparison of ILR–SR proxy-reference metrics (Appendix G.1, Tables 18–19).

ScalePSNRSSIMLPIPSDISTSRQIA-FINERefReward-SRHallucination ScoreFAIM (Ours)
×40.422/0.4490.407/0.4610.478/0.5310.443/0.479-0.295/0.3990.304/0.4410.468/0.6680.484/0.4970.638/0.664
×160.642/0.6360.607/0.6200.575/0.6010.608/0.658-0.578/0.6720.633/0.6600.290/0.3290.658/0.6660.757/0.763

Table 5. SRCC/PLCC-log comparison of NR-IQA metrics (Appendix G.1, Tables 18–19).

ScaleNIQEMANIQACLIP-IQAMUSIQQ-AlignTOPIQ-NRLIQE
×4-0.292/0.400-0.173/0.303-0.136/0.287-0.325/0.466-0.065/0.202-0.302/0.521-0.371/0.474
×16-0.561/0.644-0.270/0.327-0.484/0.590-0.446/0.479-0.189/0.262-0.551/0.686-0.577/0.641

As shown in Tables 4 and 5, at ×4, FAIM achieves the highest SRCC of 0.638, exceeding the strongest baseline, Hallucination Score (0.484), by 0.154. Its PLCC-log of 0.664 closely matches the best result of 0.668 from RefReward-SR. At ×16, FAIM ranks first in both SRCC (0.757) and PLCC-log (0.763), outperforming the strongest baselines by 0.099 and 0.077, respectively. These results demonstrate strong overall agreement with human fidelity judgments across both upscaling factors.

(2) Generalization across SR families.

We further examine FAIM's performance across different SR families and training exposure. Correlations are computed across 20 images per method and averaged within each group. Brackets show 95% CIs from 5,000 bootstrap resamples of LR images.

Table 6. FAIM correlations across SR families (Appendix G.1, Table 20).

ScaleMetricRegressionGAN-basedDiffusion-basedDreamClear (unseen)SeeSR (seen)Five unseen methods
×4SRCC0.341[-0.179,0.738]0.592[0.303,0.780]0.104[-0.301,0.505]0.033[-0.464,0.522]0.174[-0.300,0.627]0.430[0.233,0.587]
PLCC-log0.488[0.260,0.805]0.719[0.589,0.890]0.547[0.355,0.789]0.501[0.201,0.813]0.592[0.344,0.868]0.629[0.544,0.793]
×16SRCC-0.132[-0.549,0.338]0.552[0.253,0.747]-0.208[-0.522,0.167]-0.368[-0.773,0.119]-0.047[-0.463,0.436]0.231[-0.025,0.452]
PLCC-log0.180[0.123,0.726]0.666[0.524,0.847]0.465[0.353,0.788]0.519[0.282,0.845]0.411[0.221,0.835]0.539[0.492,0.745]

FAIM achieves positive correlations on unseen GAN-based methods at both scales (95% CIs above zero). Across all five unseen methods, average correlations are positive in both metrics at ×4 and in PLCC-log at ×16. Together, these findings support FAIM’s applicability to unseen SR methods beyond diffusion-based SR.

In summary, FAIM leads in overall SRCC at both scales and PLCC-log at ×16, with positive correlations on unseen GAN-based methods supporting generalization beyond diffusion-based SR. Details are provided in Appendix G.1, Tables 18–20.

References

[1] Ye and Doermann, Active Sampling for Subjective Image Quality Assessment, CVPR 2014.

[2] Y. Song et al. RefReward-SR: LR-Conditioned Reward Modeling for Preference-Aligned Super-Resolution. ECCV 2026.

[3] D. Chen et al. Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption. CVPR 2025.

[4] S. Su et al. Bridging the Perception Gap in Image Super-Resolution Evaluation. CVPR 2026.

[5] W. Ren et al. Hallucination Score: Towards Mitigating Hallucinations in Generative Image Super-Resolution. arXiv 2026.

Thank you for the substantial additional experiments. But the scope of generalization is still a concern. In the new 20 image study, the strong pooled SRCC contrast with weak within-family SRCC for diffusion methods and for regression-based SR at 16x. Therefore I want to retain my original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Major Manuscript Revisions

We are grateful to all reviewers for their constructive feedback and valuable suggestions. The main revisions are summarized below and highlighted in blue in the manuscript:

  • Related work and motivation. Expanded the SR and IQA literature and added cross-paradigm visual comparisons to clarify the benchmark’s scope and motivation. (Sections 2.1–2.2; Appendix H.1)

  • Annotation reliability. Expanded agreement and split-half analyses to assess the reproducibility of aggregated human fidelity labels and contextualize model–human correlations. (Appendix B.4)

  • Baseline comparisons and statistical analysis. Added comparisons with A-FINE, RQI, and RefReward-SR, and evaluated Hallucination Score on the external RealLQ250 set. We also reported paired bootstrap confidence intervals for FAIM–RefReward-SR differences to quantify uncertainty. (Section 5.2; Appendices G.1–G.3 and G.7)

  • External generalization. Added human evaluation on 20 RealLQ250 images across regression-based, GAN-based, and diffusion-based SR at ×4 and cascaded ×16, including family-specific results and confidence intervals. (Appendix G.1)

  • Fine-grained sensitivity. Added controlled perturbation experiments to characterize FAIM’s sensitivity to localized inconsistencies and its limitations at small spatial scales. (Appendix G.4)

  • Artifact localization. Added RGB-L1 and DINOv3-Cosine baselines using the same ILR–SR inputs to assess FAIM’s localization and agreement with human artifact prominence. (Appendix G.5)

  • Computational efficiency. Expanded comparisons of average inference time per sample, peak GPU memory, and component costs to quantify practical performance–cost trade-offs. (Appendices G.6 and G.8)

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

Related papers

Discussion (0)

Sign in to comment.