Evolving a Better Scheduler for Diffusion Models

AI Generated Content (AIGC) is becoming phenomenally prominent and impactful. One of the key generative algorithms used in AIGC is the diffusion model which is widely used in generative images and audio. In comparison with other generative methods such as GAN (Generative Adversarial Network) and VAE (Variational Auto Encoder), diffusion models can generate samples of higher quality. To further improve diffusion models, especially in terms of sampling speed, we propose an evolutionary algorithm in this paper. That is to enhance the noise scheduler of the diffusion framework, thereby improving both performance and sampling speed. This is the first diffusion model that incorporates evolutionary algorithms. Our experiments show that evolved schedulers can bring concrete improvement in the generative process.