- 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
-
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.
-
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
-
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.
-
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
Response to Reviewer 4nt2 (2/2)
Results on ReLoBlur for the Response 2, the improved metrics are in bold.
| Model | Setting | PSNR↑ | SSIM↑ | LPIPS↓ | DISTS↓ | NIQE↓ | MUSIQ↑ | MANIQA↑ | TOPIQ↑ | AFINE-NR↓ |
|---|---|---|---|---|---|---|---|---|---|---|
| MPRNet | w/o GGT | 31.6713 | 0.9284 | 0.2553 | 0.1852 | 6.8262 | 37.3007 | 0.4717 | 0.2621 | -0.5543 |
| MPRNet | w/ GGT | 33.7996 | 0.9221 | 0.2988 | 0.1974 | 7.1324 | 42.0122 | 0.4728 | 0.2752 | -0.5718 |
| MPRNet | Improvement | 2.1283 | -0.0063 | 0.0435 | 0.0122 | 0.3062 | 4.7115 | 0.0011 | 0.0131 | -0.0175 |
| NAFNet | w/o GGT | 31.1355 | 0.9241 | 0.2636 | 0.1925 | 6.9418 | 36.4612 | 0.4556 | 0.2523 | -0.5381 |
| NAFNet | w/ GGT | 34.1631 | 0.9257 | 0.2837 | 0.1887 | 6.7662 | 44.5391 | 0.4955 | 0.2873 | -0.5840 |
| NAFNet | Improvement | 3.0276 | 0.0016 | 0.0201 | -0.0038 | -0.1756 | 8.0779 | 0.0399 | 0.0350 | -0.0459 |
| SwinIR | w/o GGT | 28.8354 | 0.9132 | 0.2447 | 0.1739 | 5.9923 | 37.5703 | 0.4388 | 0.2648 | -0.5427 |
| SwinIR | w/ GGT | 33.2789 | 0.9198 | 0.2958 | 0.1818 | 6.9717 | 40.4283 | 0.4720 | 0.2692 | -0.5505 |
| SwinIR | Improvement | 4.4435 | 0.0066 | 0.0511 | 0.0079 | 0.9794 | 2.8580 | 0.0332 | 0.0044 | -0.0078 |
| XRestormer | w/o GGT | 31.2991 | 0.9270 | 0.2324 | 0.1709 | 6.2344 | 39.5892 | 0.4889 | 0.2725 | -0.5701 |
| XRestormer | w/ GGT | 33.2570 | 0.9196 | 0.2860 | 0.1870 | 6.6742 | 43.9687 | 0.4991 | 0.2865 | -0.5903 |
| XRestormer | Improvement | 1.9579 | -0.0074 | 0.0536 | 0.0161 | 0.4398 | 4.3795 | 0.0102 | 0.0140 | -0.0202 |
| PromptIR | w/o GGT | 31.1841 | 0.9246 | 0.2207 | 0.1666 | 6.1519 | 38.8904 | 0.4834 | 0.2750 | -0.5761 |
| PromptIR | w/ GGT | 34.1875 | 0.9282 | 0.2704 | 0.1829 | 6.6678 | 43.7857 | 0.4978 | 0.2880 | -0.5867 |
| PromptIR | Improvement | 3.0034 | 0.0036 | 0.0497 | 0.0163 | 0.5159 | 4.8953 | 0.0144 | 0.0130 | -0.0106 |
| MoCE-IR | w/o GGT | 34.0011 | 0.9359 | 0.2230 | 0.1750 | 6.5239 | 38.5274 | 0.4697 | 0.2645 | -0.5541 |
| MoCE-IR | w/ GGT | 33.7285 | 0.9222 | 0.2936 | 0.1946 | 6.6165 | 45.6586 | 0.4954 | 0.3002 | -0.6088 |
| MoCE-IR | Improvement | -0.2726 | -0.0137 | 0.0706 | 0.0196 | 0.0926 | 7.1312 | 0.0257 | 0.0357 | -0.0547 |
| DA-CLIP | w/o GGT | 31.1764 | 0.9130 | 0.2134 | 0.1608 | 4.9617 | 42.7587 | 0.5171 | 0.3052 | -0.5883 |
| DA-CLIP | w/ GGT | 32.7351 | 0.8970 | 0.2033 | 0.1507 | 5.6277 | 44.6420 | 0.5274 | 0.3089 | -0.6089 |
| DA-CLIP | Improvement | 1.5587 | -0.0160 | -0.0101 | -0.0101 | 0.6660 | 1.8833 | 0.0103 | 0.0037 | -0.0206 |
| FoundIR | w/o GGT | 28.9591 | 0.9259 | 0.2355 | 0.1726 | 5.9877 | 39.8784 | 0.4901 | 0.2748 | -0.5799 |
| FoundIR | w/ GGT | 32.0032 | 0.9236 | 0.2641 | 0.1948 | 6.6723 | 39.1299 | 0.4695 | 0.2776 | -0.5803 |
| FoundIR | Improvement | 3.0441 | -0.0023 | 0.0286 | 0.0222 | 0.6846 | -0.7485 | -0.0206 | 0.0028 | -0.0004 |
| Flux | w/o GGT | 25.4146 | 0.7766 | 0.3243 | 0.1960 | 4.5711 | 56.4496 | 0.5739 | 0.4625 | -0.8089 |
| Flux | w/ GGT | 26.9373 | 0.7947 | 0.3462 | 0.2164 | 3.8381 | 63.4084 | 0.6283 | 0.5300 | -0.8497 |
| Flux | Improvement | 1.5227 | 0.0181 | 0.0219 | 0.0204 | -0.7330 | 6.9588 | 0.0544 | 0.0675 | -0.0408 |
| QwenIR | w/o GGT | 31.7118 | 0.9061 | 0.1734 | 0.1375 | 4.7312 | 44.9176 | 0.5243 | 0.3398 | -0.6432 |
| QwenIR | w/ GGT | 29.1727 | 0.8188 | 0.2939 | 0.2001 | 3.9960 | 62.0743 | 0.6252 | 0.5288 | -0.8761 |
| QwenIR | Improvement | -2.5391 | -0.0873 | 0.1205 | 0.0626 | -0.7352 | 17.1567 | 0.1009 | 0.1890 | -0.2329 |
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.
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.