Performance analysis of face generation using an integrated GAN and deep reinforcement learning framework

Generative Adversarial Networks (GANs) have demonstrated outstanding performance in tasks such as image generation, style transfer, and high-resolution face synthesis. StyleGAN has established a strong presence in the fields of controllable image generation and style transfer. However, the training process requires significant time and computational resources, posing a major challenge for consumer-grade hardware. This is due to the need for high-performance or multiple GPUs to reduce training time, as well as substantial VRAM to operate at a reasonable speed. In this paper, we aim to enhance training efficiency and resource accessibility by integrating the Q-learning framework from deep reinforcement learning into the StyleGAN2 training process. In the original StyleGAN2, eight V100 GPUs were used continuously for several days to reduce the FID score to 2.84. We explore the use of a Deep Q-Network (DQN) to dynamically guide training decisions, thereby improving the learning efficiency of the generator. The experimental results show that with a single A6000 GPU and after observing 3 million images, the FID score can be reduced to below 10—significantly lowering the cost compared to the original StyleGAN2. This paper demonstrates that integrating a Deep Q-Network can provide a more directed training strategy for StyleGAN2, moving beyond traditional adversarial training to improve efficiency and reduce computational costs.