Source-linked AI summary
GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration
Xiangtao Kong, Jixin Zhao, Lingchen Sun, Rongyuan Wu, Lei Zhang
TL;DR
Real-world image restoration lacks scalable, faithful paired data because synthetic degradations miss real complexity and physical acquisition is costly. This paper uses multimodal foundation models to build GGT-100K, which consistently improves restoration generalization, especially for finetuned generative models.
Problem
Real-world image restoration lacks scalable paired training data that faithfully represents diverse, complex degradations.
Method
The authors evaluate nine multimodal foundation models and prompting strategies, then generate and screen HQ targets to construct GGT-100K.
Results
GGT-100K consistently improves fidelity and perceptual quality across restoration models, with particularly large perceptual gains for generative models.
Takeaways & Limitations
Multimodal foundation models can provide practical restoration-oriented training data, and GGT-100K supports broader real-world restoration generalization.
Takeaways & Limitations
GGT-100K is an approximation rather than a perfect substitute for captured references and may contain artifacts or hallucinated details.
Abstract
from arXiv · showhide
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.
1 Introduction
Real-world image restoration needs paired data that reflect complex, unknown degradations, but synthetic data leave a domain gap and real-world acquisition is costly. This work evaluates generative multimodal foundation models and uses Nano-Banana-2 to construct GGT-100K, a large paired dataset for real-world IR.
- Motivation: Real-world IR targets complex, mixed, and often unknown degradations, unlike classical restoration methods designed for predefined degradations.The paper contrasts real-world IR with denoising, deblurring, super-resolution, and dehazing settings.
- Motivation: Paired training data are a key bottleneck because synthetic degradations fail to capture real-world image formation, while physically collected pairs are expensive and difficult to obtain.The two established construction routes are synthetic generation and real-world acquisition.
- Motivation: MFMs could generate restoration-oriented HQ targets from real-world LQ images, but structural distortion, hallucinated details, and prompt-dependent inconsistency make the task nontrivial.The paper frames whether MFMs can generate HQ targets reliably as a central research question.
- MFM evaluation: Nine modern MFMs are systematically evaluated across scenes and degradation types using fixed and VLM-based adaptive prompting, with assessments of fidelity, perceptual quality, VLM evaluation, and human preference.The evaluated models include Nano-Banana-2 and GPT-Image-2.
- Dataset construction: GGT-100K contains 103,707 training pairs and a test set of 500 image pairs, covering diverse scenes and degradations including mixed degradations, rain, haze, snow, low-light conditions, and old photos.The dataset collects LQ images from existing datasets, Internet sources, and the authors’ own captures.
- Contributions: The work proposes GGT as a scalable MFM-based paradigm for constructing real-world paired IR data and systematically evaluates nine MFMs with multiple prompting strategies.The stated contributions emphasize practical insights for restoration-oriented GGT generation under diverse real-world degradations.
2 Related Work
Real-world image restoration has progressed across CNN, transformer, all-in-one, and generative methods, but training remains constrained by synthetic data’s domain gap and the limited availability of real-world paired datasets. Recent multimodal foundation models offer a promising route to generate high-quality targets, though real-world restoration additionally requires faithful content preservation.
- Real-world Image Restoration: Real-world IR methods have evolved from CNNs to transformers, all-in-one frameworks, and generative approaches.Lightweight CNNs remain attractive for efficiency, while heavier transformers often deliver stronger restoration performance.
- Real-world Image Restoration: Synthetic training data scale easily but fail to capture complex real-world degradations, creating a substantial domain gap.Existing real-world paired datasets use acquisition settings such as multiple acquisition and controlled imaging to provide more realistic supervision.
- Multimodal Foundation Models: Generative multimodal foundation models can produce content-aware HQ outputs from image and text inputs, making them promising for generating targets from LQ images.Unlike general image editing, real-world IR requires both perceptual quality improvement and faithful content preservation.
3 GGT-100K: Dataset Construction
GGT-100K is constructed by collecting diverse real-world LQ images, selecting a balanced MFM–prompting configuration, generating HQ targets, and applying multi-stage quality control. The resulting dataset contains 103K training pairs and 500 carefully reviewed test pairs at 1024×1024 resolution.
- Dataset pipeline: The pipeline collects real-world LQ images from existing datasets, Internet sources, and the authors’ own captures to expand coverage beyond existing paired data.The targets are generated after jointly evaluating candidate MFMs and prompting strategies for perceptual quality and fidelity preservation.
- MFM and prompt selection: Nine MFMs and fixed or adaptive prompts are evaluated jointly because model choice and prompting strategy substantially affect restoration behavior.Performance gaps can be large across model–prompt combinations, and even the same MFM can respond noticeably differently to different prompts.
- MFM and prompt selection: Nano-Banana-2 with Gemini-based adaptive prompting achieves the best Avg. score of 0.84 and the highest human preference of 32.5%.It improves from an Avg. score of 0.76 under fixed prompting to 0.84 under Gemini-based adaptive prompting and provides a strong balance across fidelity, perceptual quality, VLM assessment, and human preference.
- Quality control: The quality-control pipeline combines metric-based filtering, VLM-assisted refinement, and manual verification to improve the reliability of generated HQ targets.The VLM evaluates restoration quality, object consistency, geometry alignment, content reasonableness, and color consistency, then guides issue-specific regeneration when needed.
- Dataset scale and verification: 103K training pairs and 500 test pairs are produced at a unified resolution of 1024×1024.Multiple researchers manually select the test pairs to ensure high fidelity, strong restoration quality, and no obvious hallucinated content.
4 Expanding Real-World IR Boundaries using GGT-100K
GGT-100K consistently improves real-world restoration across diverse IR model families and test sets, with especially strong benefits for fidelity-oriented and generative models. Multi-stage quality control further makes the generated supervision more faithful and reliable.
- Results on GGT-100K Test set: GGT-100K improves all fidelity metrics, MUSIQ, AFINE-NR, and VLM-R for every evaluated IR model on the GGT-100K test set.AFINE-NR is designed for generative-model-era restoration and is better aligned with human preference.
- Results on GGT-100K Test set: X-Restormer and PromptIR gain more than 3.5 dB in PSNR, with clear improvements in SSIM, LPIPS, and DISTS.These gains indicate that GGT-100K provides effective supervision for faithful reconstruction.
- Results on Public RealLQ Test sets: On public RealLQ test sets, GGT-100K consistently improves AFINE-NR and VLM-R for all models, MUSIQ for most models, and especially benefits generative IR.FLUX-Controlnet and Qwen-Image-Edit show more visible improvements than other baselines.
- Visual Results: Qualitative results show that adding GGT-100K improves haze removal across NAFNet, FoundIR, FLUX-Controlnet, and Qwen-Image-Edit.Models trained only on existing datasets leave noticeable residual haze under real haze degradations.
- Visual Results: Across desnowing, low-light enhancement, and old-photo restoration, all eight evaluated methods achieve noticeably stronger restoration after adding GGT-100K.The methods include MPRNet, NAFNet, SwinIR, X-Restormer, PromptIR, MoCE-IR, DA-CLIP, and FLUX-Controlnet.
- Quality Control: Multi-stage quality control further improves fidelity-oriented metrics for most models, especially PSNR, SSIM, LPIPS, and DISTS.Without quality control, generated data can hurt PSNR and SSIM relative to baseline, particularly for FLUX-Controlnet.
5 Limitation and Discussion
GGT-100K is a scalable, high-quality approximation rather than a perfect substitute for physically captured ground truth, and it cannot cover the full space of open-ended real-world degradations. The reported gains are conservative because experiments use widely adopted architectures and general training strategies rather than dataset-specific optimization.
- Limitations: GGT-100K may contain subtle imperfections, including minor artifacts or hallucinated details introduced by generative multimodal foundation models, despite multi-stage quality control.These issues are difficult to completely avoid when generative models synthesize restoration targets.
- Limitations: GGT-100K cannot cover the full space of highly diverse, mixed, and open-ended real-world degradations, so models may fail on previously unseen degradation types.The pipeline is considered scalable through more diverse source-image collection and more advanced methods.
- Discussion: The reported improvements should be viewed as a conservative estimate because the experiments use widely used model architectures and general training strategies rather than optimizing models specifically for GGT-100K.Specialized network designs, training objectives, and finetuning strategies may yield further benefits for different model families.
6 Conclusion · Appendix
The paper introduces GGT as a practical generative-MFM paradigm for real-world paired image-restoration data and releases GGT-100K, which improves restoration-model generalization. The appendix documents dataset collection, prompting, evaluation, quality control, user studies, training composition, and degradation-specific results.
- 6 Conclusion: GGT uses generative MFMs to construct practical real-world paired training data for image restoration.
- 6 Conclusion: 103K training pairs and a curated test set of 500 pairs comprise the large-scale GGT-100K real-world paired dataset.
- 6 Conclusion: GGT-100K consistently improves content fidelity and perceptual quality generalization across various restoration models, especially finetuned generative models.
- Appendix: The appendix organizes these materials as supplementary documentation for the paper’s dataset construction, MFM evaluation, quality control, studies, training, and degradation analyses.
- Appendix: The appendix provides more details on source-image collection for GGT-100K, referring to Section 3.1.
- Appendix: It details prompt designs for MFM evaluation and the use of VLMs as evaluators and quality controllers, referring to Sections 3.2 and 3.3.
- Appendix: It reports details of the user study and training-data composition, referring to Sections 3.2 and 4.1.
- Appendix: It includes detailed results for specific degradation categories, referring to Section 4.2.
A Details of Source Image Collection
GGT-100K sources real-world low-quality images from Internet sources, public datasets, and original captures, organizing them into six degradation and scene categories. After screening more than 1,100K patches and multi-stage quality control, the collection retains 103,707 low-quality/high-quality pairs.
- Source image collection: GGT-100K collects real-world low-quality images from Internet sources, public datasets, and self-captured images, organized into General Mixed, Low-Light, Haze, Rain, Snow, and Old Photo.Source-specific filtering, cropping, and quality-control procedures address variation in content, degradation type, and perceptual quality.
- Internet sources: Web-collected images use category-specific collection and filtering strategies and are restricted to images released under CC0 licenses.Rain and Snow searches target visible weather phenomena, followed by duplicate removal and full-resolution manual inspection.
- Public datasets and captures: Dataset-sourced images include over 10K images from real-world restoration datasets without ground-truth references and over 1,000K images from other low-quality or bad-weather datasets.Metric-based screening, VLM-based screening, and subjective filtering retain images with clear, suitable degradations.
- Screening and quality control: More than 1,100K image patches are screened, about 120K low-quality images are retained, and Nano-Banana-2 generates corresponding high-quality counterparts before further pair filtering.The final dataset size reflects multi-stage quality control applied after counterpart generation.
- Screening and quality control: 103,707 image pairs remain after multi-stage quality control, with General Mixed contributing 66,058 pairs.The passage attributes the retained pair count to Nano-Banana-2’s strong generation quality.
B Detailed Prompt Designs for MFMs Evaluation · SYSTEM PROMPT
The appendix specifies fixed and VLM-based prompting strategies for evaluating multimodal foundation models, with Gemini-based prompts selected for their stronger quantitative performance and avoidance of facial blurring or anonymization. The prompts prioritize fidelity, degradation-specific restoration, and preservation of image content, geometry, and color.
- B Detailed Prompt Designs for MFMs Evaluation: The evaluation compares nine candidate MFMs under four prompting strategies, with the appendix detailing each prompt design.The fixed and fixed-prompt-no-change strategies use the simplest instruction form.
- B Detailed Prompt Designs for MFMs Evaluation: A simple “do not change the image content” instruction consistently improves content fidelity for most MFMs.This instruction is used in both fixed-prompt variants.
- B Detailed Prompt Designs for MFMs Evaluation: A VLM receives the input image and a meta prompt, then generates a detailed image-specific restoration instruction describing content, degradations, and their removal.The VLM prompt is illustrated with a haze image and uses GPT-5.4-Pro or Gemini-3.1-Pro.
- SYSTEM PROMPT: Most models benefit from VLM-based prompts, with particularly strong gains for Nano-Banana-2; Gemini-based prompts are chosen over GPT-based prompts for better quantitative performance and fewer facial artifacts.GPT-generated prompts often blur or anonymize facial regions.
- SYSTEM PROMPT: The general restoration prompt analyzes scene structure, degradation types, and regional severity before requiring fidelity-first restoration without changing content, geometry, or color.It explicitly covers noise, compression artifacts, defocus blur, motion blur, and combinations of these degradations, while requiring balanced full-image clarity improvement.
- SYSTEM PROMPT: The degradation-specific prompts adapt the analysis and restoration constraints to rainy, snowy, low-light, and old-photo images.They address rain and snow removal, stronger enhancement of dark regions, and conservative restoration of historical visual details while preserving scene fidelity.
- SYSTEM PROMPT: The low-light and old-photo user prompts require stronger denoising and clarity recovery while preserving original content, geometry, color, and style.The old-photo prompt explicitly forbids colorizing monochrome photos, while the low-light prompt permits stronger brightness and saturation when needed.
- SYSTEM PROMPT: VLM quality control evaluates restoration quality, object consistency, geometry alignment, content reasonableness, and color consistency before accepting or regenerating an HQ target.Rejected samples receive strengthened instructions and up to three total generation attempts.
C How to Use VLM as the Evaluator and Quality Controller
The pipeline uses Gemini-3.1-Pro as a VLM evaluator and quality controller, judging HQ outputs against LQ inputs for restoration suitability. It defines VLM-R from structured judgments and uses VLM feedback to screen, revise, and regenerate dataset pairs.
- VLM Evaluation: Gemini-3.1-Pro evaluates each restored or generated HQ image against its LQ input for restoration suitability.The VLM serves as both evaluator and quality controller in the pipeline.
- VLM-R Metric: VLM-R measures the proportion of evaluated samples judged Accept by the VLM.A higher VLM-R indicates more consistently perceptually plausible and content-faithful restoration targets.
- VLM-R Metric: The VLM judges HQ results across restoration quality, object consistency, geometry alignment, content reasonableness, and color consistency.For each image pair, it returns a structured JSON output used to determine acceptability.
- Dataset Quality Control: During GGT-100K screening, accepted LQ-HQ pairs are kept, while rejected candidates receive further inspection and corrective processing.VLM scores and textual feedback guide additional instructions to the original generation prompt before regeneration and re-evaluation.
D Details of User Study
The user study evaluates nine MFMs on 200 real-world test images with 20 participants, who choose results based on restoration quality and content faithfulness. Image-level majority voting produces human preference scores in 0.5% increments.
- Study setup: 20 participants evaluated nine anonymously indexed MFM restoration results on 200 real-world test images, alongside each image’s LQ input.Each candidate used its best-performing prompt setting, and results were randomly shuffled before presentation.
- Evaluation criterion: Participants selected one best result per image, considering clearness, naturalness, perceptual appeal, similarity, and faithfulness to the LQ input.The criterion penalized unreasonable changes that fail to preserve the original scene content.
- Voting aggregation: The model receiving the most votes for each image was designated that image’s winner, and win counts were converted into final human preference ratios.The aggregation treats the 20 participant responses as votes over the nine anonymous outputs.
- Score interpretation: 0.5% represents one image-level win because the study covers 200 images; therefore, 0% means no aggregated wins, not necessarily no individual votes.GPT-Image-1.5 could receive individual votes yet score 0% if it never became the image-level winner.
E Details of Training Settings
The training pool is constructed by explicitly balancing degradation categories and sample counts across existing datasets. Resolution-aware sampling preserves broad dataset coverage while preventing large datasets from dominating the distribution.
- Dataset composition: 120K pairs are assigned to the general mixed category, while the remaining 100K pairs are distributed comparably across low-light, haze, rain, snow, and noise.The general mixed category represents about half of the full training pool and is roughly consistent with its approximately 60% proportion in GGT-100K.
- Dataset composition: The resulting pool combines data from diverse sources, including DF2K with RealESRGAN degradations, RealSR, GoPro, RealBlur, LOL, RESIDE, Rain13K, Snow100K, and other datasets.Dataset-specific contributions range from 2K to 58K pairs among the listed sources.
- Sampling strategy: 512 × 512 crops are used for very high-resolution datasets such as RealSR and SIDD, while most other datasets retain their original resolution.The required number of images or patches is then randomly sampled from each dataset to form the final training pool.
- Sampling strategy: Controlled sampling preserves broad dataset and degradation coverage while avoiding severe imbalance from directly merging all available data.Without this control, a few large-scale datasets could dominate the training distribution and compromise fair comparison.
F Detailed Results of Specific Degradations
Category-wise evaluations across six real-world degradation groups show that GGT-100K generally improves restoration performance broadly rather than only addressing a missing degradation category. The results also reveal a Qwen-Image-Edit exception on rain and snow, where lower PSNR accompanies richer generated details, and emphasize that samples typically contain mixed degradations.
- Category-wise evaluation: GGT-100K is evaluated on general mixed degradations, rain, haze, snow, low-light, and old photos.The comparisons report training without and with GGT-100K for each category.
- Generalization across degradations: GGT-100K generally improves restoration performance across all six evaluated real-world degradation groups and different model families.The gains therefore do not merely compensate for missing categories such as old photos in the baseline training data.
- Model-specific exception: For Qwen-Image-Edit, finetuning with GGT-100K lowers PSNR on the rain and snow subsets, while remaining substantially above the official model.The authors attribute this behavior to richer, more realistic details that improve visual quality despite lower pixel-level metrics.
- Interpretation of categories: The degradation categories are not isolated settings because samples usually contain mixed degradations, with labels identifying only the most prominent visible factor.The overall multi-category trend supports GGT-100K as general-purpose real-world supervision for varied restoration scenarios.