Image synthesis based on generative adversarial networks (GANs) is a hot spot in current computer vision research. The diverse condition image synthesis tasks are necessary to ensure that the generated images still have high diversity while meeting the input conditions. However, conditional GANs (cGANs) excessively rely on the input conditional information and ignore the noise vectors, which can easily lead to the mode collapse problem. Although many methods have been proposed to solve this problem, these methods still have some limitations. In this work, we propose an innovative model called DACGAN-TMS, which is used to generate diverse images without increasing too much training overhead. In DACGAN-TMS, an additional regularization term is added to the loss function of the generator model to encourage the generator to detect more minor modes during training, that is, to stimulate the generator to generate different pictures in the course of training. It effectively improves the problem that the original ACGAN is easy to collapse during training and difficult to generate diversified pictures. In addition, the standard convolution of the generator is completely replaced by dynamic convolution, and the convolution parameters are adaptively adjusted to further enhance the expressive capability of the model. Extensive experiments and statistical comparison results manifest that DACGAN-TMS is superior to the relevant GAN-based models in generating different images.