The Second Challenge on Real-World Face Restoration at NTIRE 2026: Methods and Results
Published in arXiv preprint arXiv:2604.10532v2 / CVPR 2026 Workshop, 2026
Abstract: This paper reviews the NTIRE 2026 real-world face restoration challenge, including the task setup, submitted solutions, evaluation protocol, and final outcomes. The challenge targets natural and realistic restoration while preserving identity consistency, without restricting computational resources or training data. Performance is measured with a weighted image quality assessment score and AdaFace-based identity checking. The competition attracted 96 registrants, 10 teams submitted valid models, and 9 teams obtained valid final scores, providing a snapshot of current trends in real-world face restoration.