AI image generation is a technology that has been hotly discussed in the art and Deep Learning (DL) field. You must have heard of the AI Art Generator such as Dall-E 2 or NovelAI, a DL model that generates realistic-looking images from a given text sequence.
To explore this technology deeper, we need to introduce a new class in the generative model called ‘diffusion’, first proposed by Sohl-Dickstein et al. (2015), which aimed to generate images from noise using a backward denoising process.
So far, several generative models exist, including GAN, VAE and Flow-based models. Most of them could generate a high-quality image, such as StyleGAN-XL, the current State-of-the-Art image generation model. However, each has some limitations of its own.
GAN models are known for potentially unstable training and less diversity in generation due to their adversarial training nature. VAE relies on a surrogate loss. Flow models have to use specialized architectures to construct reversible transforms (Lilian Weng, 2021)
The diffusion model has provided a slow and iterative process when noise is converted into an image; this makes the diffusion model more scalable than the GAN model. Besides…
