Reference Free

Reference-free methods aim to evaluate various aspects of data quality (images, videos, text, 3D models) without relying on a reference standard, addressing limitations of reference-based approaches in terms of cost, availability, and applicability. Current research focuses on developing and improving reference-free metrics using deep learning models, particularly transformer-based architectures and diffusion models, often incorporating uncertainty quantification or mimicking human perceptual mechanisms. These advancements are significant for applications where reference data is scarce or impractical to obtain, improving the efficiency and accessibility of quality assessment across diverse fields.