acceptodds
Under review as a conference paper at ICLR 2027

GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration

Abstract

Real-world image restoration (IR) is bottlenecked by the scarcity of high-quality paired training data. Synthetic datasets are abundant but often fail to model real-world degradations, while real-world paired datasets are expensive and difficult to capture. As a result, IR models trained on these datasets show limited generalization in real-world scenarios. In this work, we propose Generative Ground Truth (GGT) by using generative multimodal foundation models (MFMs) to produce high-quality (HQ) targets from real-world low-quality (LQ) images. We first conduct a systematic evaluation of nine state-of-the-art MFMs, including Nano-Banana-2 and GPT-Image-2, on images of various scenes and degradation types. The results demonstrate that Nano-Banana-2 with VLM-based adaptive prompting shows the highest capability to synthesize perceptually realistic and content-faithful HQ targets, which can serve as the GGT for the LQ input. We then employ Nano-Banana-2 to build a GGT synthesis pipeline, which involves multi-stage quality control to ensure data reliability, and construct GGT-100K, an LQ-HQ paired dataset comprising 103,707 training pairs and covering diverse scenes and complex real-world degradations. A test set of 500 image pairs is also established. Extensive experiments show that GGT-100K consistently improves the real-world generalization of a wide range of IR models, with particularly strong benefits for finetuning generative models for IR tasks. Our results suggest that MFMs can serve as practical tools for restoration-oriented data generation, and GGT-100K is a useful resource to expand the generalization boundaries of real-world IR models. Our dataset will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generative Ground Truths Dataset, Real-World Image Restoration, Multi-modal Foundation Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_4nt2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Generative Ground Truth (GGT) to expand paired training data for real-world image restoration. Generative Multimodal Base Models (MFMs) transform low-quality (LQ) inputs into high-quality (HQ) supervised objectives. After a systematic comparison of various MFMs, the authors selected the Nano-Banana-2 model and constructed the GGT-100K dataset using a Gemini-based adaptive cueing method.

The resulting dataset contains 103,707 training pairs and 500 test pairs, covering various scenes and complex image degradation scenarios. Comparing restoration models trained using this dataset with those not trained on it, the results show an overall improvement, especially for generative methods, where the performance is more significant.

Critical strengths
  1. Constructing supervised objectives directly from real degraded images provides a practical approach to extending training coverage in situations where collecting real low-quality-high-quality image pairs is difficult.

  2. Evaluation includes ten restoration models and external real-world benchmark datasets, providing evidence that the proposed test set is superior. Table 3 shows that FLUX-Controlnet and Qwen-Image-Edit improve on all six reported metrics when the results are averaged across the six external datasets.

Critical weaknesses
  1. The primary test set uses the generated target as a reference, which limits the effectiveness of reference-based improvement methods in terms of fidelity recovery. Predictions may be more similar to the generated target, but not more accurately reconstruct the original scene. Therefore, independent evidence is needed to determine whether training on GGT-100K preserves realistic image content and avoids reproducing changes introduced during target generation.

  2. OpenReal80K provides images for both dataset construction and external evaluation, but the distinction between these two uses is insufficiently explained. The protocol should differentiate between source image separation and image patch-level separation; otherwise, different cropped regions of the same image may appear during training and evaluation.

Questions for authors

How can we ensure that no source image is used for both training and testing simultaneously, especially for samples derived from OpenReal80K? Please describe how to check for crop origins and potential overlap.

Can the downstream model be evaluated using real reference images or a content-preserving human assessment? Such an assessment should check whether fixes introduce, remove, or alter scene details.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 4nt2 (1/2)

We thank the reviewer for the positive assessment and for raising these important points. We will respond to each point one by one.

1. Deduplication of GGT training and testing pairs, the usage of OpenReal80K (Weakness 2; Question 1)

  • Deduplication of GGT training and testing pairs. Our pipeline was as follows: we first collected LQ images without GT, removed exact duplicate images within degradation groups using pHash, cropped when necessary (if the size exceeds 1024×1024), and then applied the full GGT pipeline, including manual screening. After that, we selected 500 images as the GGT test set. For the cropped images, we explicitly removed same-source “siblings” in the training set: if a cropped image came from the same source image as any test image, it was marked as a sibling and excluded. Specifically, 184 of the 500 test images have siblings, and 1,011 sibling crops were excluded. Therefore, there is no identical image and no same-source crop shared between GGT training and GGT test.

  • Usage of OpenReal80K. OpenReal80K images are 512×512 and do not involve cropping. GGT (including train and test sets) uses only about 2.2K images from OpenReal80K (only from rain training split). The public test sets use the OpenReal80K test split, with 60 images each from haze, rain, snow, and night (align the image number of RealDeg). Thus, haze/snow/night are unrelated to GGT, and rain images come from separate train/test splits with no image overlap.

In summary, the separation is rigorous at the image and crop levels, no source image or patch is used for both training and testing simultaneously. We will add these details to the paper.

2. Independent evaluation of downstream content fidelity (Weakness 1; Question 2)

We provide an independent evaluation on ReLoBlur, a real paired local-motion-blur dataset that is outside GGT and outside the existing training data. Table ReLoBlur (see the next comment) reports results on 100 images of the ReLoBlur test set. Adding GGT improves PSNR for 8 of 10 models, with an average gain of +1.79 dB; most models reach PSNR above 30 dB, and most no-reference metrics also improve, indicating that GGT training does not damage content fidelity on real degradations.

To directly test content preservation, we asked three annotators and a VLM (Gemini-3.7-flash) to check the results under a fixed protocol: whether any object is added, removed, or structurally displaced relative to the LQ input. They found no systematic object insertion/removal or structural displacement; only minor texture-level differences were observed, which are expected from restoration and can affect LPIPS/DISTS but do not constitute content alteration. We attribute this to (1) after quality control, including manual inspection, GGT contains only a small number of images with substantial content changes, and (2) models with GGT are also trained jointly on synthetic and real paired data, we hypothesize that occasional inconsistencies are likely to be ignored by the network rather than learned as fixed patterns. Overall, models trained with GGT could preserve the original input content.

3. Clarification: predictions may be more similar to the generated target (Weakness 1)

We acknowledge that FR metrics on the GGT test set reflect similarity to the generator’s output. However, after quality control including manual inspection, the surviving GGT targets are statistically constrained to satisfy restoration criteria, so learning them and learning restoration are largely aligned. This is supported by VLM evaluations, visual results of the main paper, and the ReLoBlur evaluation with real GT.

This concern applies only to metrics computed against generated references. We will suggest readers place more emphasis on NR metrics and VLM-R, which use only the real LQ input and the restored output, and will clarify that FR metrics against GGT references indicate the degree of degradation removal rather than distance to a perfect target. Since restoration is ill-posed, the ideal HQ counterpart for a real LQ image is a distribution rather than a point; the GGT reference is therefore a good sample, not the unique or optimal solution. We will emphasize this more strongly in the revised paper.

We thank the reviewer again for these constructive comments and remain available to address any further questions or concerns the reviewer may have.

AuthorsOpenReview ↗

Response to Reviewer 4nt2 (2/2)

Results on ReLoBlur for the Response 2, the improved metrics are in bold.

ModelSettingPSNR↑SSIM↑LPIPS↓DISTS↓NIQE↓MUSIQ↑MANIQA↑TOPIQ↑AFINE-NR↓
MPRNetw/o GGT31.67130.92840.25530.18526.826237.30070.47170.2621-0.5543
MPRNetw/ GGT33.79960.92210.29880.19747.132442.01220.47280.2752-0.5718
MPRNetImprovement2.1283-0.00630.04350.01220.30624.71150.00110.0131-0.0175
NAFNetw/o GGT31.13550.92410.26360.19256.941836.46120.45560.2523-0.5381
NAFNetw/ GGT34.16310.92570.28370.18876.766244.53910.49550.2873-0.5840
NAFNetImprovement3.02760.00160.0201-0.0038-0.17568.07790.03990.0350-0.0459
SwinIRw/o GGT28.83540.91320.24470.17395.992337.57030.43880.2648-0.5427
SwinIRw/ GGT33.27890.91980.29580.18186.971740.42830.47200.2692-0.5505
SwinIRImprovement4.44350.00660.05110.00790.97942.85800.03320.0044-0.0078
XRestormerw/o GGT31.29910.92700.23240.17096.234439.58920.48890.2725-0.5701
XRestormerw/ GGT33.25700.91960.28600.18706.674243.96870.49910.2865-0.5903
XRestormerImprovement1.9579-0.00740.05360.01610.43984.37950.01020.0140-0.0202
PromptIRw/o GGT31.18410.92460.22070.16666.151938.89040.48340.2750-0.5761
PromptIRw/ GGT34.18750.92820.27040.18296.667843.78570.49780.2880-0.5867
PromptIRImprovement3.00340.00360.04970.01630.51594.89530.01440.0130-0.0106
MoCE-IRw/o GGT34.00110.93590.22300.17506.523938.52740.46970.2645-0.5541
MoCE-IRw/ GGT33.72850.92220.29360.19466.616545.65860.49540.3002-0.6088
MoCE-IRImprovement-0.2726-0.01370.07060.01960.09267.13120.02570.0357-0.0547
DA-CLIPw/o GGT31.17640.91300.21340.16084.961742.75870.51710.3052-0.5883
DA-CLIPw/ GGT32.73510.89700.20330.15075.627744.64200.52740.3089-0.6089
DA-CLIPImprovement1.5587-0.0160-0.0101-0.01010.66601.88330.01030.0037-0.0206
FoundIRw/o GGT28.95910.92590.23550.17265.987739.87840.49010.2748-0.5799
FoundIRw/ GGT32.00320.92360.26410.19486.672339.12990.46950.2776-0.5803
FoundIRImprovement3.0441-0.00230.02860.02220.6846-0.7485-0.02060.0028-0.0004
Fluxw/o GGT25.41460.77660.32430.19604.571156.44960.57390.4625-0.8089
Fluxw/ GGT26.93730.79470.34620.21643.838163.40840.62830.5300-0.8497
FluxImprovement1.52270.01810.02190.0204-0.73306.95880.05440.0675-0.0408
QwenIRw/o GGT31.71180.90610.17340.13754.731244.91760.52430.3398-0.6432
QwenIRw/ GGT29.17270.81880.29390.20013.996062.07430.62520.5288-0.8761
QwenIRImprovement-2.5391-0.08730.12050.0626-0.735217.15670.10090.1890-0.2329
AuthorsOpenReview ↗

Thanks & Friendly Reminder

Thank you again for your time and valuable comments. Please feel free to let us know if you have any remaining concerns.

Reviewer_QJb6OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Generative Ground Truth (GGT), a novel paradigm to address the problem of paired training data scarcity in real-world image restoration. By systematically evaluating 9 SOTA MFMs, the authors identify NanoBanana 2 with Gemini-guided adaptive prompting as the optimal generator for high-fidelity ground truths. Utilizing this pipeline, they construct GGT-100K, a large-scale dataset containing 103,707 real-world image pairs across diverse mixed real-world degradations. Through a multi-stage quality control process and comprehensive experiments on different models, the authors demonstrate that GGT-100K significantly enhances the real-world generalization and restoration quality of existing restoration networks.

Critical strengths

1.This paper provides a Comprehensive MFM evaluation for image restoration. It compares 9 MFMs with different prompts using FR, NR IQA and VLM-based IQA. This is helpful for the community to understand the current restoration performance of MFMs. 2.The article has a clear and sufficient motivation. In the past, there was no effective method for processing LQ that had been captured but lacked GT. GGT has provided a feasible new path to expand datasets to cover the degradation that was not covered. 3.This paper provides a real-world training and testing set that has undergone multiple stages of screening. As a supplement to the existing training data,this dataset significantly enhances the generalization ability of IR models (from CNN to the fine-tuned Qwen model), which is impressive.

Critical weaknesses
  1. Constructing GGT-100K relies heavily on MFMs and VLMs. Including a brief breakdown of the total computational resource requirements or API costs would help the community better understand the scalability and reproducibility of the data generation pipeline.
  2. The GGT-100K can be regarded as a kind of data distillation for the NanoBanana 2. Therefore, in the main table, the processing results using NanoBanana 2 should be included. This will enable readers to observe the performance of the teacher model, as well as the gap between the student models (various IR models) and the teacher model.
  3. 500 samples in the test set are sufficient for a single task, but are they insufficient for multiple tasks? I observed that some sub-tasks only have 50 test cases (such as rain, haze and etc.).
Questions for authors
  1. Could the author provide the API cost for the GGT-100K?
  2. Could the author provide the results of NanoBanana 2 in Table 2 and Table 3 ?
  3. Has the author considered increasing the number of cases in the test set?
  4. Although the recommended usage method by GGT-100K is to use it in combination with existing training data, I still wonder how the model (only trained on GGT-100K) would perform. Can GGT-100K be used independently?
5 comments
AuthorsOpenReview ↗

Response to Reviewer QJb6 (1/2)

We thank the reviewer for the positive assessment and for raising these important points. We will respond to each point one by one.

1. API costs of GGT (Weakness 1; Question 1)

The API cost of GGT-100K mainly consists of prompt acquisition using Gemini 3.1-pro (approximately 0.01 dollars per image), the inference of NB2 (0.067 dollars per image), and the use of VLM in the QC stage (approximately 0.01 dollars per image), so the API cost of one image is about 0.09 dollars. Based on the actual expenses, GGT-100K costs approximately 12,000 dollars in total (including rework and discarded images). This cost is already very time- and cost-efficient compared to actually collecting a real paired dataset with such a wide range of degradation. Moreover, GGT can obtain pairs that are often impossible to collect in real scenarios (such as paired images of rain and snow).

2. The results of NanoBanana 2 in Table 2 and Table 3 (Weakness 2; Question 2)

We provide Nano-Banana-2 (NB2)’s direct restoration performance on the GGT-100K test set and public LQ test set under two settings: (a) Fixed Prompt, which is also what an ordinary user would most commonly type; and (b) AP + QC — following our full pipeline with VLM-based adaptive prompting and the automatic quality-control loop (without artificial screen).

NB2 Results on GGT-100K test set

SettingPSNR↑SSIM↑LPIPS↓DISTS↓NIQE↓MUSIQ↑MANIQA↑TOPIQ↑AFINE-NR↓VLM-R↑
Fixed Prompt23.940.75890.21780.10654.4765.180.60950.5213−0.912438.6%
AP + QC28.260.82550.13360.06814.8560.450.58670.4689−0.874092.2%

NB2 Results on Public RealLQ Benchmark

SettingNIQE↓MUSIQ↑MANIQA↑TOPIQ↑AFINE-NR↓VLM-R↑
Fixed prompt3.797472.95960.71520.6573−1.267154.3%
Adaptive prompt + QC3.856570.68720.69700.6347−1.237088.0%

Two observations. First, under the simple prompt, NB2 obtains higher NR perceptual scores but much worse FR fidelity and VLM-R, consistent with the hallucination/color-inconsistency trend in Table 1 of the main paper. Second, AP + QC is substantially stronger: its fidelity, NR metrics, and VLM-R are above all student methods. The best L1-based students approach the teacher on fidelity, and the best generative students approach it on NR and VLM-R, but a gap remains. The same trend holds on Public RealLQ, which contains no GGT references. We will include this teacher baseline in the revision.

3. Increasing the number of cases in the test set (Weakness 3; Question 3)

We will consider expanding the test set to 1,000 images. One practical constraint is evaluation cost: at 1024×1024 resolution, evaluating 500 images already takes nearly 1,000 minutes (~16 hours) on a single GPU for models such as DA-CLIP and Qwen-Image-Edit. We therefore plan to provide a 500-image core set for minimal standard comparison plus an extended additional 500 images.

AuthorsOpenReview ↗

Response to Reviewer QJb6 (2/2)

4. The results of models which only trained on GGT (Question 4)

We provide models trained on GGT-100K alone, evaluated on both the GGT-100K test set and the public RealLQ test sets. Given the limited rebuttal time, we trained NAFNet, PromptIR and FLUX-Controlnet to represent different architecture families (CNN, all-in-one+transformer, and generative). “Baseline” denotes the existing 200K syn+real training data (the “w/o” setting of Tab. 2 and 3 of the main paper).

Results on GGT-100K test set

ModelSettingPSNR↑SSIM↑LPIPS↓DISTS↓NIQE↓MUSIQ↑MANIQA↑TOPIQ↑AFINE-NR↓VLM-R↑
NAFNetGGT only28.390.83980.30440.20546.8148.280.43860.3236−0.817457.6%
NAFNetbaseline25.130.77080.36530.22986.6742.010.46350.3118−0.701227.6%
PromptIRGGT only28.390.83930.30690.20796.8648.310.44010.3272−0.825654.0%
PromptIRbaseline24.570.75640.36300.23126.5943.230.46950.3202−0.717624.8%
FLUX-ControlnetGGT only19.050.69160.31940.17814.6965.750.60490.5593−0.972850.2%
FLUX-Controlnetbaseline22.450.69010.37730.21295.5648.550.51570.3798−0.691225.4%

Results on Public RealLQ Benchmark

ModelSettingNIQE↓MUSIQ↑MANIQA↑TOPIQ↑AFINE-NR↓VLM-R↑
NAFNetGGT only5.4759.390.57390.4297−0.900843.5%
NAFNetbaseline4.9157.970.59860.4576−0.855936.0%
PromptIRGGT only5.5259.620.57560.4410−0.915846.3%
PromptIRbaseline4.9357.180.60290.4556−0.850741.3%
FLUX-ControlnetGGT only3.9269.980.70190.6448−1.186755.0%
FLUX-Controlnetbaseline4.5060.940.64310.4795−0.877947.7%

In summary, GGT-100K is independently effective: trained alone, it's overall performance surpasses the 200K-pair syn+real baseline on the GGT-100K and Public RealLQ test set.

However, if handling synthetic degradation, models trained only with GGT are far inferior to those trained specifically on the corresponding synthetic degradation. This is expected: synthetic data can be scaled indefinitely and is the right tool when synthetic degradations are the target. GGT is designed as a complement for the much larger pool of real LQ images whose GT cannot be captured, and it extends the generalization boundary beyond synthetic and real-captured pairs.

We thank the reviewer again for these constructive comments and remain available to address any further questions or concerns the reviewer may have.

AuthorsOpenReview ↗

Thanks & Friendly Reminder

Thank you again for your time and valuable comments. Please feel free to let us know if you have any remaining concerns.

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

The paper investigates using generative multimodal foundation models to create restoration targets for existing real-world degraded images. It evaluates nine image-generation/editing models with four prompting strategies using synthetic-reference fidelity metrics, real-image perceptual metrics, a VLM-based acceptance measure, and human preferences. The selected Nano-Banana-2/Gemini prompting configuration is used in a pipeline combining metric filtering, VLM-guided regeneration, and manual verification. The resulting GGT-100K contains 103,707 training pairs and a 500-pair generated-reference test set spanning six degradation groups. Training or finetuning ten restoration models with and without these data demonstrates gains on the proposed test set and several metrics on public real-image test sets. Appendices provide quality-control comparisons, evaluator correlations, user-study details, and per-degradation results.

Critical strengths
  1. GGT-100K provides 103,707 training image pairs and 500 test image pairs, covering scenarios such as mixed degradations, low-light conditions, haze, rain, snow, and old photographs. It establishes training targets for real-world degraded images—for which obtaining clear, authentic reference images is often difficult—thereby serving as a valuable supplement to existing synthetic degradation datasets.
  2. This paper compares nine generative models and four types of prompting strategies, taking into account content fidelity, perceptual quality, VLM-based evaluation, and human preferences. These findings provide an empirical basis for selecting data generation schemes and offer guidance for future research on choosing generators and designing prompts.
  3. The experiments cover ten restoration models, including CNN-based, Transformer-based, general-purpose, and generative models, with comparisons conducted both with and without the proposed data. The authors also evaluate the models on real-world public images and report consistent improvements. For example, the MUSIQ score of FLUX-ControlNet increases from 60.9375 to 67.5600. These results provide evidence that the benefits of the proposed data are not limited to evaluations against self-generated reference samples.
Critical weaknesses
  1. Since both the training targets and the test references are produced by the same generator, some improvements in metrics may simply reflect the model’s output becoming more closely aligned with that generator’s output. While existing public benchmarks and human evaluations offer useful points of reference, a more compelling conclusion could be drawn by directly examining how well the downstream model preserves content before and after the integration of GGT.
  2. Some of the training images and the public test set are sourced from OpenReal80K, and the construction process involved image cropping. It is recommended to clarify the methods used for partitioning and deduplication at both the original-image and scene levels.
Questions for authors
  1. Data independence. Could the authors clarify whether the training pairs, generated-reference test set, model-selection images, and public test sets were separated at the original-image and scene levels before cropping?
  2. Downstream content fidelity. Is there any independent evaluation of whether models trained with GGT preserve the original input content, beyond comparisons with the generated references?
4 comments
AuthorsOpenReview ↗

Response to Reviewer TZpU (1/2)

We thank the reviewer for the positive assessment and for raising these important points. We organize our response into three parts:

1. Data independence and deduplication (Weakness 2; Question 1)

We clarify the separation of the four data groups:

  • Model-selection images. We used 100 DIV2K validation images and 200 real LQ images. These are disjoint at the image level from all other data groups.
  • GGT training pairs and generated-reference test set. These were not split before cropping. Our pipeline was as follows: we first collected LQ images without GT, removed exact duplicate images within degradation groups using pHash, cropped when necessary (if the size exceeds 1024×1024), and then applied the full GGT pipeline, including manual screening. After that, we selected 500 images as the GGT test set. For the cropped images, we explicitly removed same-source “siblings” in the training set: if a cropped image came from the same source image as any test image, it was marked as a sibling and excluded. Specifically, 184 of the 500 test images have siblings, and 1,011 sibling crops were excluded. Therefore, there is no identical image and no same-source crop shared between GGT training and GGT test.
  • Public test sets. The public test sets consist of RealDeg and the OpenReal80K test split. They are disjoint at the image level from the GGT training pairs and the generated-reference test set.
  • Usage of OpenReal80K. OpenReal80K images are 512×512 and do not involve cropping. GGT (including train and test sets) uses only about 2.2K images from OpenReal80K (only from rain training split). The public test sets use the OpenReal80K test split, with 60 images each from haze, rain, snow, and night (align the image number of RealDeg). Thus, haze/snow/night are unrelated to GGT, and rain images come from separate train/test splits with no image overlap.
  • Scene-level deduplication. We removed exact duplicate images within degradation groups using pHash. We did not perform semantic/scene-level filtering or deduplication. Our rationale is that scene-level deduplication is uncommon in low-level vision, changes the data distribution, and has ambiguous granularity—e.g., if the test set contains a cat or a crossroad, it is unclear whether all cats or crossroads should be removed from training.

In summary, the separation is rigorous at the image and crop levels, including same-source crop removal between GGT training and test. We will add these details to the paper.

2. Independent evaluation of downstream content fidelity (Weakness 1; Question 2)

We provide an independent evaluation on ReLoBlur, a real paired local-motion-blur dataset that is outside GGT and outside the existing training data. Table ReLoBlur (see the next comment) reports results on 100 images of the ReLoBlur test set. Adding GGT improves PSNR for 8 of 10 models, with an average gain of +1.79 dB; most models reach PSNR above 30 dB, and most no-reference metrics also improve, indicating that GGT training does not damage content fidelity on real degradations.

To directly test content preservation, we asked three annotators and a VLM (Gemini-3.7-flash) to check the results under a fixed protocol: whether any object is added, removed, or structurally displaced relative to the LQ input. They found no systematic object insertion/removal or structural displacement; only minor texture-level differences were observed, which are expected from restoration and can affect LPIPS/DISTS but do not constitute content alteration. We attribute this to (1) after quality control, including manual inspection, GGT contains only a small number of images with substantial content changes, and (2) models with GGT are also trained jointly on synthetic and real paired data, we hypothesize that occasional inconsistencies are likely to be ignored by the network rather than learned as fixed patterns. Overall, models trained with GGT could preserve the original input content.

AuthorsOpenReview ↗

Response to Reviewer TZpU (2/2)

3. Clarification: improvements in metrics may simply reflect the model’s output becoming more closely aligned with that generator’s output (Weakness 1)

We acknowledge that FR metrics on the GGT test set reflect similarity to the generator’s output. However, after quality control including manual inspection, the surviving GGT targets are statistically constrained to satisfy restoration criteria, so learning them and learning restoration are largely aligned. This is supported by VLM evaluations, visual results of the main paper, and the ReLoBlur evaluation with real GT.

This concern applies only to metrics computed against generated references. We will suggest readers place more emphasis on NR metrics and VLM-R, which use only the real LQ input and the restored output, and will clarify that FR metrics against GGT references indicate the degree of degradation removal rather than distance to a perfect target. Since restoration is ill-posed, the ideal HQ counterpart for a real LQ image is a distribution rather than a point; the GGT reference is therefore a good sample, not the unique or optimal solution. We will emphasize this more strongly in the revised paper.

We thank the reviewer again for these constructive comments and remain available to address any further questions or concerns the reviewer may have.

Results on ReLoBlur for the Response 2, the improved metrics are in bold.

ModelSettingPSNR↑SSIM↑LPIPS↓DISTS↓NIQE↓MUSIQ↑MANIQA↑TOPIQ↑AFINE-NR↓
MPRNetw/o GGT31.67130.92840.25530.18526.826237.30070.47170.2621-0.5543
MPRNetw/ GGT33.79960.92210.29880.19747.132442.01220.47280.2752-0.5718
MPRNetImprovement2.1283-0.00630.04350.01220.30624.71150.00110.0131-0.0175
NAFNetw/o GGT31.13550.92410.26360.19256.941836.46120.45560.2523-0.5381
NAFNetw/ GGT34.16310.92570.28370.18876.766244.53910.49550.2873-0.5840
NAFNetImprovement3.02760.00160.0201-0.0038-0.17568.07790.03990.0350-0.0459
SwinIRw/o GGT28.83540.91320.24470.17395.992337.57030.43880.2648-0.5427
SwinIRw/ GGT33.27890.91980.29580.18186.971740.42830.47200.2692-0.5505
SwinIRImprovement4.44350.00660.05110.00790.97942.85800.03320.0044-0.0078
XRestormerw/o GGT31.29910.92700.23240.17096.234439.58920.48890.2725-0.5701
XRestormerw/ GGT33.25700.91960.28600.18706.674243.96870.49910.2865-0.5903
XRestormerImprovement1.9579-0.00740.05360.01610.43984.37950.01020.0140-0.0202
PromptIRw/o GGT31.18410.92460.22070.16666.151938.89040.48340.2750-0.5761
PromptIRw/ GGT34.18750.92820.27040.18296.667843.78570.49780.2880-0.5867
PromptIRImprovement3.00340.00360.04970.01630.51594.89530.01440.0130-0.0106
MoCE-IRw/o GGT34.00110.93590.22300.17506.523938.52740.46970.2645-0.5541
MoCE-IRw/ GGT33.72850.92220.29360.19466.616545.65860.49540.3002-0.6088
MoCE-IRImprovement-0.2726-0.01370.07060.01960.09267.13120.02570.0357-0.0547
DA-CLIPw/o GGT31.17640.91300.21340.16084.961742.75870.51710.3052-0.5883
DA-CLIPw/ GGT32.73510.89700.20330.15075.627744.64200.52740.3089-0.6089
DA-CLIPImprovement1.5587-0.0160-0.0101-0.01010.66601.88330.01030.0037-0.0206
FoundIRw/o GGT28.95910.92590.23550.17265.987739.87840.49010.2748-0.5799
FoundIRw/ GGT32.00320.92360.26410.19486.672339.12990.46950.2776-0.5803
FoundIRImprovement3.0441-0.00230.02860.02220.6846-0.7485-0.02060.0028-0.0004
Fluxw/o GGT25.41460.77660.32430.19604.571156.44960.57390.4625-0.8089
Fluxw/ GGT26.93730.79470.34620.21643.838163.40840.62830.5300-0.8497
FluxImprovement1.52270.01810.02190.0204-0.73306.95880.05440.0675-0.0408
QwenIRw/o GGT31.71180.90610.17340.13754.731244.91760.52430.3398-0.6432
QwenIRw/ GGT29.17270.81880.29390.20013.996062.07430.62520.5288-0.8761
QwenIRImprovement-2.5391-0.08730.12050.0626-0.735217.15670.10090.1890-0.2329
AuthorsOpenReview ↗

Thanks & Friendly Reminder

Thank you again for your time and valuable comments. Please feel free to let us know if you have any remaining concerns.

AuthorsOpenReview ↗

Thanks & Friendly Reminder

Thank you again for your thoughtful review and positive assessment. As the discussion period ends in two days, we would like to briefly check whether our response has addressed your concerns. If anything remains unclear, we would be glad to clarify.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.