GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration
Researchers have developed a new method called Generative Ground Truth (GGT) to create high-quality training data for image restoration tasks. This approach utilizes generative multimodal foundation models, specifically Nano-Banana-2, to synthesize realistic target images from low-quality inputs. The resulting dataset, GGT-100K, contains over 100,000 image pairs and has demonstrated significant improvements in the real-world generalization capabilities of various image restoration models. AI
IMPACT Enhances real-world generalization for image restoration models by providing a large, high-quality synthetic dataset.