Training Text-to-Image Models Without a VAE