Advancing Text-to-Image Generation: A Comparative Study of StyleGAN-T and Stable Diffusion 3 under Neutrosophic Sets

Abstract

 Recent advances in generative models have revolutionized the technology employed 
for image synthesis quite significantly, and two paradigms—GANs and diffusion-based 
models—are leading the pack of innovation. This paper outlines an extensive comparison and 
analysis of some of the best models across both paradigms, namely StyleGAN-T, DF-GAN, 
AttnGAN, and BigGAN on the GAN side and Stable Diffusion 3 (SD3), DALL·E 3, Midjourney 
v6, and Imagen 2 on the diffusion side.                                                                                  
We systematically inspect the architectural design, training protocols, text-conditioning 
processes, and domain adaptability of each model, highlighting how they address text-to-image 
generation challenges differently. Through qualitative and quantitative measurements—such as 
FID, CLIP Score, human preference surveys, and compositional accuracy, the work reveals 
performance tradeoffs concerning speed, control, creativity, semantic alignment, and 
photorealism. We use the Neutrosophic Set model to select the best model based on these 
evaluation matrices. We have different scores for each model based on evaluation matrices. So, 
the neutrosophic set is used to overcome the uncertainty information. We use the COPRAS 
method to rank the models and select the best one based on the evaluation matrix weights. 

DOI: 10.5281/zenodo.15380944