A Survey on Quality Metrics for Text-to-Image Generation

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Image Processing

Background:

• AI text-to-image models offer fine-grained control, rivaling traditional rendering techniques.
• Assessing AI-generated images requires specialized metrics beyond traditional ones like SSIM or PSNR.

Purpose of the Study:

• To provide a comprehensive overview of quality metrics for AI text-to-image generation.
• To propose a taxonomy categorizing these metrics based on compositional and general quality.
• To cover benchmark datasets and identify challenges in text-to-image evaluation.

Main Methods:

• Literature survey of existing text-to-image quality assessment metrics.
• Development of a taxonomy for categorizing metrics.
• Review of benchmark datasets used for metric evaluation.

Main Results:

• A comprehensive overview of text-to-image quality metrics is presented.
• A novel taxonomy categorizes metrics into compositional and general quality.
• Key benchmark datasets and their usage are discussed.

Conclusions:

• Dedicated metrics are essential for evaluating AI text-to-image models.
• The proposed taxonomy provides a structured approach to metric selection and understanding.
• Identifying limitations and challenges guides future research in text-to-image evaluation.