Title: GGT-100K: Establishing Generative Ground Truth for Robust Real-World Image Restoration

The field of real-world image restoration (IR) is currently hindered by a critical shortage of high-quality paired training data. While synthetic datasets are plentiful, they frequently fall short in accurately simulating actual real-world degradations. Conversely, acquiring real-world paired datasets is both costly and technically challenging. Consequently, IR models developed using these existing resources often exhibit poor generalization when deployed in real-world contexts.

To address this challenge, this study introduces Generative Ground Truth (GGT), a method that leverages generative multimodal foundation models (MFMs) to generate high-quality (HQ) targets directly from real-world low-quality (LQ) images. The research begins with a comprehensive assessment of nine leading state-of-the-art MFMs, such as GPT-Image-2 and Nano-Banana-2, across various scenes and degradation types. The findings indicate that Nano-Banana-2, particularly when utilizing VLM-based adaptive prompting, possesses the superior ability to create HQ targets that are both perceptually realistic and faithful to the original content. These generated targets effectively serve as the GGT for their corresponding LQ inputs.

Building on this finding, the authors utilize Nano-Banana-2 to develop a GGT synthesis pipeline featuring multi-stage quality control mechanisms to guarantee data reliability. This process yielded GGT-100K, a substantial paired dataset containing 103,707 LQ-HQ training pairs. This dataset encompasses a wide variety of scenes and complex real-world degradation patterns. Additionally, a test set comprising 500 image pairs was created for evaluation.

Extensive experimental results demonstrate that GGT-100K consistently enhances the real-world generalization capabilities of numerous IR models. The benefits are especially pronounced when fine-tuning generative models for IR tasks. These outcomes suggest that MFMs are viable tools for generating restoration-oriented data, and that GGT-100K serves as a valuable resource for extending the generalization limits of real-world IR models.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC