I think it's a combination of many things. Very powerful compute, very large datasets, transformers and diffusion models and a ton of research by lots of researchers easily available on arxiv, etc.