Although　generative　adversarial　networks　are　commonly　used　in　textto-image　generation　tasks　and　have　made　great　progress,　there　are　still　some　problems.　The　convolution　operation　used　in　these　GANs-based　methods　works　on　local　regions,　but　not　disjoint　regions　of　the　image,　leading　to　structural　anomalies　in　the　generated　image.　Moreover,　the　semantic　consistency　of　generated　images　and　corresponding　text　descriptions　still　needs　to　be　improved.　In　this　paper,　we　propose　a　multi-attention　generative　adversarial　networks　(MAGAN)　for　text-toimage　generation.　We　use　self-attention　mechanism　to　improve　the　overall　quality　of　images,　so　that　the　target　image　with　a　certain　structure　can　also　be　generated　well.　We　use　multi-head　attentionmechanism　to　improve　the　semantic　consistency　of　generated　images　and　text　descriptions.　We　conducted　extensive　experiments　on　three　datasets:　Oxford-102　Flowers　dataset,　Caltech-UCSD　Birds　dataset　and　COCO　dataset.　Our　MAGAN　has　better　results　than　representative　methods　such　as　AttnGAN,　MirrorGAN　and　ControlGAN.