Abstract: Aiming at the problem of low quality and insufficient diversity of the generated sample details of the stacked generation confrontation network, a text-to-image generation model combining maximum Pearson correlation coefficients is proposed.The method improves the discriminator so that it can be encoded and discriminated, then uses the discriminator to extract the feature from the generated sample, and calculates the Pearson correlation coefficient between the input vector and the feature vector and adds it as a reconstruction term to the loss.Maximize optimization.In addition, in order to enhance the coding consistency between samples generated at different scales, multi-scale joint loss is proposed.Experiments on the CUB dataset demonstrate that the proposed method can effectively improve the diversity and image quality of the generated samples.