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

Scaling Real-World Image Restoration via Generative Ground Truth

Creating training data for fixing blurry or dark photos is notoriously difficult. We rarely possess perfect "before and after" examples. Most engineers solve this by using synthetic degradations (mathematically simulated noise or blur). These often fail to capture the messy physics of a real camera lens or a rainy night. This paper proposes a way to bypass the scarcity of real-world paired data. It uses Multimodal Foundation Models (MFMs)—AI models that process both text and images—to generate high-quality "after" images from real-world "before" images.

The Question

Can we use the generative power of foundation models to create a scalable source of "ground truth"? Specifically, the authors investigate whether MFMs can synthesize high-quality (HQ) targets from real-world low-quality (LQ) images. These targets must be faithful enough to supervise the training of specialized restoration models. The core tension is fidelity versus hallucination. A model might produce a beautiful, clear image. However, if it changes a face or a building structure, it is useless as a training signal for a model meant to restore reality.

Why The Old Answer Was Incomplete

Historically, the field has relied on two imperfect paths. The first is synthetic data generation. Researchers take clean images and apply mathematical models to simulate noise, haze, or blur. While scalable, these simulations suffer from a significant domain gap (the difference between training and real-world data). They cannot replicate the complex, mixed degradations found in actual photography.

The second path is physical acquisition. Researchers capture real-world pairs using controlled setups. As shown in, even state-of-the-art models like FoundIR still struggle with artifacts when faced with genuine rain or haze. Physical acquisition is too expensive. It also lacks the diversity required to cover the infinite permutations of real-world corruption.

What They Did

The researchers did not just pick a model and start generating. They treated the selection of the "teacher" model as a formal optimization problem. They first performed a systematic evaluation of nine MFMs, including Nano-Banana-2 and GPT-Image-2. They tested them across various scenes and degradation types. Crucially, they discovered that fixed prompts are insufficient. Instead, they utilized VLM-based adaptive prompting. This uses a Vision-Language Model (VLM) to analyze the specific degradation of an input image. The VLM then writes a custom, highly detailed instruction for the generator.

To ensure the resulting data was not poisoned by hallucinations, they implemented a rigorous, multi-stage quality control pipeline, as detailed in .

This process involved: 1. Metric-based filtering: Automatically discarding samples where no perceptible improvement was made. 2. VLM-assisted refinement: Using a VLM to score candidates on five dimensions. These include restoration quality, object consistency, geometry alignment, content reasonableness, and color consistency. The pipeline then regenerates failed samples with corrective prompts. 3. Manual verification: A final human safeguard.

This pipeline resulted in GGT-100K. This is a dataset of 103,707 high-quality training pairs covering everything from low-light to snow and old photos.

What They Found

The authors report that GGT-100K consistently improves the generalization of a wide range of model families. The results are particularly striking for generative models. For instance, the paper finds that finetuning the Qwen-Image-Edit model on GGT-100K significantly improves both its visual appeal and its content fidelity.

The gains are not just aesthetic. In terms of raw metrics, the authors demonstrate significant improvements. Adding GGT-100K to the training pool of an X-Restormer model yields a PSNR (Peak Signal-to-Noise Ratio, a measure of reconstruction fidelity) improvement of +3.5397 dB. This represents a substantial boost in reconstruction accuracy. Furthermore, for the FoundIR model, the inclusion of GGT-100K leads to a 25.0% improvement in the VLM-based restoration success rate (VLM-R). This metric evaluates how well a model restores an image from a semantic perspective. As seen in, the qualitative difference is visible. Models trained with GGT-100K produce much sharper, more faithful textures under real-world haze and noise.

What This Changes

If this methodology scales, it signals a fundamental shift in how we approach specialized computer vision tasks. We are moving from a regime of "collecting data" to a regime of "curating generative teachers."

The implications are twofold. First, for practitioners, it suggests that the bottleneck for high-performance restoration is no longer architecture design. Instead, it is the quality of the supervisory signal. If you are building a product for consumer photo enhancement, investing in a VLM-driven data synthesis pipeline may yield higher ROI. This could be more efficient than searching for niche real-world datasets. Second, it validates the use of VLMs as sophisticated automated QA engineers in the data factory.

The obvious follow-up is to test the limits of this "synthetic ground truth." One should investigate whether the accumulation of subtle, systemic hallucinations from an MFM eventually creates a "ceiling." This ceiling might prevent specialized models from reaching true physical accuracy.