Core Concepts
The authors explore the naturalness of AI-generated images by benchmarking and assessing visual naturalness. They propose JOINT, an objective evaluator that aligns with human ratings.
Abstract
The study delves into the challenges of assessing naturalness in AI-generated images, introducing the AGIN database and proposing JOINT for accurate evaluation. The research highlights the impact of technical and rationality distortions on image naturalness, providing valuable insights for future developments in this field.
The proliferation of Artificial Intelligence-Generated Images (AGIs) has expanded the Image Naturalness Assessment (INA) problem. Different from traditional definitions, INA on AI-generated images faces diverse contents affected by technical and rationality distortions.
The AGIN database collects human opinions on overall naturalness, technical, and rationality perspectives to understand how these factors influence visual naturalness. The proposed JOINT model significantly outperforms baselines in providing consistent results for naturalness assessment.
Overall, this study contributes to understanding human reasoning in visual naturalness evaluation for AI-generated images through a multi-perspective approach.
Stats
AGIN contains 6,049 images collected from five generative tasks.
907,350 human opinions were collected for technical and rationality perspectives.
MOS = 0.145MOST + 0.769MOSR correlation observed in overall naturalness score approximation.
Quotes
"We take the first step to explore the naturalness of AI-generated images." - Zijian Chen et al.
"JOINT significantly outperforms baselines for providing more subjectively consistent results on naturalness assessment." - Zijian Chen et al.