Digital image steganography is a technique for hiding information in an image with the aim of hiding the presence of the information itself. A new steganography without embedding scheme is proposed in this paper to create stego images based only on secret information, in order to avoid image distortion caused by pixel modification-based embedding methods. In this scheme, according to the proposed interconversion algorithm between messages and pixels, the secret message is encoded into pixel format and placed in a blank image in a specified way, which is considered a broken image to be restored. The broken image is repaired by the training network to obtain the stego image. The receiver selects specific stego pixels from the received image and restores them to the delivered secret message according to the inverse of the transformation. In addition, the proposed communication framework supports the choice of different generative network structures, further improving the resistance to steganalysis. The basic generation model was trained and tested on the collected cat face images and bedroom images from the LSUN dataset, respectively, and the steganographic images were evaluated qualitatively and quantitatively. The final experimental results verify that the framework achieves satisfactory performance in steganography capacity and information recovery rate compared with other methods without embedding.