[HUGGINGFACE]score: 0.48
GGT-100K Uses Multimodal Foundation Models to Generate Paired Image Restoration Data
May 28, 2026
GGT evaluates nine MFMs including GPT-Image-2 and Nano-Banana-2 as sources of high-quality restoration targets from real-world low-quality images, finding Nano-Banana-2 with VLM-based adaptive prompting performs best. The approach addresses the paired data scarcity bottleneck in real-world image restoration without physical capture.
paper
HOW THIS AFFECTS YOU
●
builderYou can use this pipeline to generate large-scale paired training data for image restoration models without expensive real-world capture setups.
●
researcherSystematic comparison of nine MFMs as synthetic ground-truth generators provides a reusable evaluation framework for generative data augmentation in restoration tasks.