The author notes the sudden rise of AI art surpassing human artistic abilities with models like Stable Diffusion. The application of denoising diffusion models demonstrates a unique approach to generating samples from an unknown distribution by blending noise over many steps. This method simplifies the complexity of previous models by focusing on a one-directional mapping. The use of denoising diffusion models has shown promising results in learning reverse distributions effectively through the application of statistical physics principles. The detailed breakdown of the training objectives and loss functions provides a clear understanding of the process involved in utilizing denoising diffusion models for image generation.
https://andrewkchan.dev/posts/diffusion.html