GGT-100K: Real-World Restoration Dataset

• GGT-100K is a large-scale real-world image restoration dataset that uses generative ground truth from multimodal foundation models to create high-quality supervisory signals.
• It comprises 103,707 training pairs and a 500-pair test set at 1024×1024 resolution, covering diverse, entangled degradations such as rain, haze, low-light, snow, and old photos.
• The dataset employs a rigorous multi-stage quality control pipeline—including metric-based filtering, VLM-assisted prompt refinement, and manual verification—to ensure reliable and realistic restoration targets.

GGT-100K is a large-scale real-world image restoration dataset in which the high-quality supervision signal is not captured through a paired acquisition setup and not synthesized by a hand-designed degradation model, but generated from real low-quality images by a multimodal foundation model. Introduced together with the concept of Generative Ground Truth (GGT), the dataset contains 103,707 training pairs and a 500-pair test set, all at 1024 × 1024, and is designed to address the central bottleneck in real-world image restoration: the scarcity of realistic, diverse, and scalable paired training data for generalization beyond synthetic benchmarks (Kong et al., 29 May 2026).

1. Conceptual basis and problem setting

Generative Ground Truth (GGT) denotes a high-quality restoration target generated by a multimodal foundation model from a real degraded image. In this formulation, a real low-quality (LQ) image is taken from the wild, an MFM is prompted to produce an idealized high-quality (HQ) counterpart, and the generated HQ image is then used as supervision for restoration training (Kong et al., 29 May 2026).

The motivation is a two-sided data limitation. Synthetic paired datasets are scalable because they start from a clean HQ image and degrade it with a model, but they suffer from unrealistic degradations and a domain gap. Manually captured paired datasets provide realistic supervision, but they are expensive, difficult to align, hard to scale, and often infeasible for degradations such as rain, haze, motion blur, low light, or mixed distortions. GGT is positioned as a practical middle ground: more realistic than synthetic degradation pipelines and far more scalable than physically captured pairing.

This framing is especially relevant for real-world IR because the target domain is intrinsically open-ended. The six categories used in GGT-100K—General Mixed, Low-Light, Haze, Rain, Snow, and Old Photo—are explicitly not restricted to isolated single-degradation settings. A rain image, for example, may also contain blur, compression artifacts, and noise. This suggests that GGT-100K is structured around restoration under entangled real-world degradations rather than narrow task-factorized corruption models.

2. Data collection, degradation scope, and dataset composition

The source images for GGT-100K are collected from three channels: existing datasets, internet sources, and self-captured images (Kong et al., 29 May 2026). The internet component is gathered via web crawling from CC0-licensed sources including Unsplash, Pexels, Pixabay, and Flickr. The self-captured portion uses different cameras and phones under blur, noise, low-light, and related real-world conditions. All images are normalized to 1024 × 1024.

The final training set contains 103,707 paired samples, partitioned as follows:

Category	Training pairs
General Mixed	66,058
Low-Light	9,786
Haze	7,822
Rain	7,177
Snow	6,759
Old Photo	6,105

Within General Mixed, the appendix reports 37,463 from internet images, 19,438 from public datasets, and 9,157 from self-captures. Additional category-specific provenance is also specified: Rain uses NTURain, OpenReal80K, Flickr, Unsplash, and Pexels; Haze uses ACDC and RTTS; Snow uses Unsplash and Snow100K-Real; Low-light uses DarkFace, ExDark, and self-captures; Old Photo uses Flickr.

The raw collection process is substantially larger than the final released dataset. The appendix states that the authors screened more than 1,100K image patches and retained about 120K low-quality images before the final MFM generation and quality-control stages. This indicates that dataset construction is not simply a one-pass generation procedure; it is a filtered pipeline in which raw diversity is intentionally narrowed into a reliability-controlled paired corpus.

3. Multimodal foundation model evaluation and model selection

Before constructing the dataset, the work performs a systematic evaluation of nine multimodal foundation models under multiple prompting strategies (Kong et al., 29 May 2026). The nine MFMs are:

• FireRed-1.1
• Qwen-Image-Edit-2511
• FLUX.2-dev
• Kling-Image-O1
• Seedream-5.0
• GPT-Image-1.5
• GPT-Image-2
• Nano-Banana-Pro
• Nano-Banana-2

These include 3 open-source and 6 closed-source models. The evaluation does not treat model choice and prompting as separable. Instead, the paper emphasizes prompt sensitivity and compares three prompt regimes: fixed prompts, fixed prompt with no-change constraint (Fix-NC), and adaptive prompts generated by a VLM such as GPT-5.4-Pro or Gemini-3.1-Pro.

Selection uses four complementary aspects: fidelity preservation, perceptual quality, VLM-based restoration success, and human preference. Fidelity is measured on 100 DIV2K validation images degraded with the Real-ESRGAN pipeline, using PSNR, SSIM, LPIPS, and DISTS; low-light enhancement and dehazing are excluded from this fidelity evaluation because acceptable outputs may differ in global illumination. Perceptual evaluation uses 200 real-world images—20 each for rain, haze, snow, low-light, old photo, and 100 mixed-degradation images—with NIQE, MUSIQ, MANIQA, TOPIQ, and AFINE-NR. The study also introduces VLM-R, a success rate defined as

$\mathrm{VLM\mbox{-}R}=\frac{1}{N}\sum_{i=1}^{N}\mathbb{I}(d_i=Accept)\times 100\%$

where $N$ is the number of evaluated samples and $d_i$ is the final VLM decision for sample $i$.

For score aggregation, the paper defines a normalized overall score over fidelity, perceptual quality, and VLM judgment. The intended formula is

$\tilde{m}_{i,j}=\frac{m_{i,j}-m_j^{\min}{m_j^{\max}-m_j^{\min},\qquad s_i^{a}=\frac{1}{|\mathcal{M}_a|}\sum_{j\in\mathcal{M}_a}\tilde{m}_{i,j},\qquad \mathrm{Avg.}_i=\frac{1}{3}\left(s_i^{\mathrm{fid}+s_i^{\mathrm{per}+s_i^{\mathrm{vlm}\right)$

with $a \in \{\mathrm{fid}, \mathrm{per}, \mathrm{vlm}\}$ and $\mathcal{M}_a$ the metric set for aspect $a$.

The selected configuration is Nano-Banana-2 + Gemini-based adaptive prompting, which achieves best overall Avg. score: 0.84 and highest human preference: 32.5%. The comparison with GPT-generated prompts is operationally important: Gemini-based prompting is chosen because it performs better quantitatively, while GPT-based prompts sometimes blur or anonymize facial regions. The broader implication is that GGT-100K depends not only on MFM capability but also on prompt generation as a control variable in data synthesis.

4. GGT synthesis pipeline and multi-stage quality control

The GGT-100K construction pipeline has three stages: collect real-world degraded images, evaluate MFMs and prompting strategies, and generate HQ targets and apply multi-stage quality control (Kong et al., 29 May 2026). The last stage is the core of the dataset methodology.

Stage 1: metric-based filtering compares no-reference perceptual scores between the original LQ image and the generated HQ image. If the generated output shows little improvement or worse quality, the sample is rejected. The paper describes this as a coarse automatic filter, especially useful for General Mixed images.

Stage 2: VLM-assisted refinement evaluates five criteria: restoration quality, object consistency, geometry alignment, content reasonableness, and color consistency. If a sample is judged unsatisfactory, it is not immediately discarded. Instead, the VLM feedback is used to modify the prompt, the HQ target is regenerated, and the result is re-evaluated. In this design, the VLM functions both as evaluator and feedback controller.

Stage 3: manual verification checks the final retained set for strong restoration quality, high fidelity, and absence of obvious hallucinations.

The appendix specifies an iteration rule of up to three generation attempts per sample: accept if the VLM approves; otherwise refine the prompt and regenerate; if still rejected, perform one more generation attempt with the same prompt. This operational detail is central to the dataset’s claim of reliability. The paper explicitly argues that raw MFM outputs are not safe to use directly; multi-stage filtering and VLM feedback are required to convert generative outputs into restoration supervision.

A plausible implication is that the contribution is as much about data curation policy as about generative synthesis. The dataset is not formed by treating MFM outputs as ground truth by default; it is formed by treating them as candidate ground truth subject to iterative acceptance control.

5. Benchmark protocol and empirical effects on restoration models

The work establishes a 500-pair test set curated by multiple researchers and manually reviewed for high fidelity, avoidance of obvious hallucinations, and suitability for assessing restoration quality and content consistency (Kong et al., 29 May 2026). Evaluation is reported both on the full GGT-100K test set and on category-specific subsets: General Mixed, Rain, Haze, Snow, Low-Light, and Old Photo. The paper also evaluates on public RealLQ datasets without ground truth using no-reference metrics and VLM-R.

The central empirical claim is that GGT-100K improves the real-world generalization of a wide range of IR models. The evaluated families include CNN-based methods (MPRNet, NAFNet), Transformer-based methods (SwinIR, X-Restormer), all-in-one / blind restoration methods (PromptIR, MoCE-IR, DA-CLIP, FoundIR), and generative restoration methods (FLUX-Controlnet, Qwen-Image-Edit).

On the full test set, adding GGT-100K improves PSNR, SSIM, LPIPS, DISTS, MUSIQ, AFINE-NR, and VLM-R. Representative examples include:

• MPRNet: 24.79 → 27.30 PSNR, 0.7637 → 0.8189 SSIM, 22.2% → 33.2% VLM-R
• NAFNet: 25.13 → 28.25 PSNR, 0.7708 → 0.8349 SSIM, 27.6% → 53.8% VLM-R
• SwinIR: 23.92 → 27.08 PSNR, 18.6% → 37.6% VLM-R
• X-Restormer: 24.69 → 28.23 PSNR, 30.4% → 54.6% VLM-R
• PromptIR: 24.57 → 28.18 PSNR, 24.8% → 49.6% VLM-R
• MoCE-IR: 24.86 → 28.21 PSNR, 25.4% → 55.2% VLM-R
• DA-CLIP: 25.88 → 26.69 PSNR, 31.6% → 51.0% VLM-R

The improvement pattern is differentiated by model class. The paper notes that fidelity-oriented models gain large PSNR improvements, whereas generative models benefit especially strongly in perceptual metrics and overall realism. For FLUX-Controlnet, the changes are 22.45 → 23.14 PSNR, 48.55 → 63.09 MUSIQ, 0.5157 → 0.5854 MANIQA, 0.3798 → 0.5013 TOPIQ, and 25.4% → 63.4% VLM-R. For Qwen-Image-Edit, the reported changes are 25.86 → 26.18 PSNR, 0.2813 → 0.2155 LPIPS, 0.1625 → 0.1183 DISTS, 51.42 → 62.55 MUSIQ, and 77.4% → 87.6% VLM-R.

On six public RealLQ test sets without ground truth, the addition of GGT-100K also generally improves no-reference quality and VLM-R. Reported VLM-R gains include MPRNet: 38.0% → 45.0%, SwinIR: 28.0% → 38.7%, PromptIR: 41.3% → 48.3%, DA-CLIP: 46.7% → 55.0%, FLUX-Controlnet: 47.7% → 59.0%, and Qwen-Image-Edit: 60.7% → 70.7%. These results support the paper’s claim that the dataset improves generalization beyond the nominal training distribution.

6. Ablations, limitations, and significance for real-world IR

A central ablation isolates the effect of quality control by comparing baseline, w/o-QC, and w/-QC training settings (Kong et al., 29 May 2026). The conclusion is that even unscreened GGT-like data can help, but quality control makes supervision more reliable and stable. The strongest example is FLUX-Controlnet: baseline PSNR: 22.45, w/o-QC PSNR: 19.42, w/-QC PSNR: 23.14. This establishes that unscreened generated supervision can actively harm fidelity, whereas the filtered dataset recovers and improves performance. For Qwen-Image-Edit, w/o-QC already improves perceptual metrics substantially, but w/-QC further improves PSNR, LPIPS, DISTS, MUSIQ, and VLM-R.

The appendix further shows that gains are not confined to a single degradation category; improvements appear across general mixed, rain, haze, snow, low-light, and old photo. An important nuance appears in category-specific tables for Qwen-Image-Edit: on some rain and snow subsets, GGT-100K may lower PSNR while still improving perceptual metrics. The paper interprets this as better generated detail and realism without exact pixel matching, reinforcing the view that restoration quality should not be judged by pixel accuracy alone.

The limitations are explicit. First, GGT is not perfect ground truth: even after QC, some samples may still contain subtle artifacts, hallucinated details, or minor inconsistencies. Second, degradation coverage is broad but incomplete because real-world degradations are open-ended. Third, the experiments are described as conservative, in that they mostly use standard architectures and training setups rather than algorithms specifically optimized for GGT supervision.

Within these constraints, the significance of GGT-100K lies in reframing restoration dataset construction. The paper’s main implications are that MFMs can be practical restoration-data generators, that restoration datasets can be scaled without relying only on real paired capture, that generative restoration models benefit especially strongly, that prompting strategy materially affects data quality, and that quality control is essential. In that sense, GGT-100K is not merely a larger benchmark; it is an explicit proposal for using multimodal generative models, adaptive prompting, and iterative verification as infrastructure for real-world IR data generation.