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

Researchers introduced Generative Ground Truth, a new way to create training pairs for real-world image restoration.

They evaluated nine multimodal foundation models and selected Nano-Banana-2 with adaptive prompting to generate restoration targets.

Using this approach, they built GGT-100K, a 103,707-pair dataset plus a 500-image test set.

Experiments show GGT-100K improves performance and generalization across many restoration models, especially generative ones.