Quality assessment of synthetic images via spatial distortion recognition

Tomoya Sawada, Marie Katsurai, Masashi Okubo

Abstract: The advancement of generative models necessitates reliable methods for assessing synthetic image quality, particularly when models are trained with limited data and may produce spatial distortions. Current AI-generated image quality assessment techniques often assume high-fidelity generation or target text-to-image models, and many are computationally intensive. This study introduces an efficient approach for the assessment of images synthesized by generative adversarial networks (GANs). We propose a convolutional neural network, trained to recognize spatial distortions. Our model learns from real-world images altered with random spatial transformations to create distorted examples, simulating common artifacts without relying on GAN-produced training data. The model’s confidence in identifying distortions quantifies image quality. We utilized bilateral filtering to emphasize shape information on natural images. Experiments show that our method distinguishes between synthetic and real images and assesses visual quality more effectively than existing non-reference methods. Furthermore, our method aligned with human perception for both natural and illustration-style images from GANs trained on limited data. Our approach offers a computationally efficient framework for evaluating GAN-generated images without relying on text prompts at assessment and any actual synthetic images at training time.

External IDs:dblp:journals/mva/SawadaKO25