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

Released on 2026-05-28, this dataset contains 100,000 real-world low-quality and high-quality image pairs generated by Multi-modal Foundation Models (MFMs). It was created by researchers from The Hong Kong Polytechnic University and OPPO Research Institute to expand generalization boundaries in image restoration tasks.

Use Cases

• Training image restoration models based on the described low-quality to high-quality image pairs.
• Benchmarking model generalization on real-world data based on the dataset's stated purpose.
• Developing novel image enhancement algorithms based on generative ground truth data.
• Studying the properties of MFM-generated image data for computer vision tasks.

Strengths

• Contains 100,000 image pairs, providing a substantial scale for training.
• Focuses on real-world image restoration, addressing a practical application gap.
• Created by a collaboration between an academic institution and an industry research lab.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to the specific Multi-modal Foundation Models used for generation.

Provenance

Source

VCLab-PolyU (The Hong Kong Polytechnic University and OPPO Research Institute)

Collection Method

Generated using Multi-modal Foundation Models (MFMs).

Freshness

Last updated 2026-06-01 03:59:01; freshness should be verified.