Text-to-Image　(T2I)　synthesis　refers　to　the　computational　method　of　translating　normal　human　text　description　into　images　which　have　alike　semantic　meaning　as　text　with　the　use　of　keywords　or　sentences.　T2I　has　achieved　great　success　in　a　few　areas,　especially　in　generating　vivid,　realistic　visual　and　photographic　images.　Recent　years,　the　development　of　deep　learning　has　brought　some　new　methods　for　unsupervised　deep　learning　area,　which　provide　some　models　to　generate　visually　natural　images　with　suitably　trained　neural　network　models.　However,　T2I　now　still　faces　some　challenges.　For　example,　current　T2I　models　are　unable　to　generate　high　resolution　images　with　multiple　objects.　And　it　can　be　difficult　to　reproduce　the　results　of　many　approaches.　Since　present　methods　are　mostly　based　on　GAN　(Generative　Adversarial　Network)　models,　this　paper　will　focus　on　the　methods　depending　on　it　and　systematically　clear　up　the　developments　of　T2I　based　on　GAN.　Also,　datasets　always　play　a　supporting　role　in　task　development,　several　current　T2I-related　datasets　will　be　discussed　in　the　paper　to　explain　more　about　T2I＇s　development.　A　brief　discussion　of　the　future　work　and　challenges　will　be　showed　at　the　end　of　the　paper.　©　2022　IEEE.