AtHom: Two Divergent Attentions Stimulated By Homomorphic Training in Text-to-Image Synthesis