Molecule generative models, particularly those leveraging deep neural networks, play a more and more important role in designing new molecular structures. Molecular generation involves training a model with known molecules and then using the learned model to automatically generate novel molecules. Current approaches can be categorized into four major types: generative adversarial networks (GANs), variational autoencoders (VAEs), energy-based models, and reinforcement learning (RL). GAN-based molecular generative models offer unique capabilities in generating diverse and realistic molecules. Other generative models, such as those based on VAEs and RL, provide alternatives, each with specific strengths and limitations. Comparative analyses highlight the advantages and drawbacks of GAN-based models against these alternatives, particularly in their capacity for capturing complex molecular distributions. Despite significant progress, challenges remain in molecular generation, including ensuring chemical validity, optimizing property-specific generation, and addressing computational efficiency. Future directions emphasize improving model interpretability, integrating domain knowledge, and expanding applications in drug discovery and material science. These advancements will be essential for enhancing the practicality and effectiveness of molecule generative models.