Abstract: To address the issues that the design of the traditional Generative Adversarial Network (GAN) architecture and hyperparameters relies on expert experience and the high cost of tuning, a GAN automatic design and optimization method based on the Quantum-behaved Particle Swarm Optimization (QPSO) algorithm is proposed. Through the QPSO algorithm, the structure updates of the discriminator and the generator are guided collaboratively. A hybrid operation encoding scheme based on a modular search space is designed, supporting deep separable convolution, dimensionality reduction fully connected, and sparse fully connected layers, and an enabling mechanism is introduced to replace the traditional layer disabling strategy. In the initialization stage of the generator, residual connections and attention modules are integrated to enhance the ability of capturing multi-scale features. Furthermore, a multi-objective loss function based on weight allocation is constructed, jointly optimizing the adversarial loss, diversity loss, and perceptual loss. Experimental results on the CIFAR-10 and STL-10 datasets show that this method effectively improves the quality and diversity of generated samples while balancing the model performance and computational complexity.