From dust to darling
A diffusion model is an AI that learns to generate data, like images, by gradually adding noise to training data and then learning to reverse that process to create new, clean data from random noise.
Imagine you have a perfect sandcastle, and you slowly kick it until it becomes a pile of sand. A diffusion model learns how to do that kicking in reverse—starting from a random pile of sand and rebuilding the perfect castle step by step.
Diffusion models power many of today's best image generators, like DALL-E and Stable Diffusion, letting anyone create stunning visuals from text descriptions. They are crucial because they produce higher quality and more diverse outputs than older methods like GANs.
Many think diffusion models create images instantly, but they actually work through hundreds of small, noisy steps. Another common mistake is confusing them with autoencoders—diffusion models don't compress data; they destroy and then reconstruct it.
A diffusion model is a class of generative models that learn a data distribution by defining a forward Markov chain that gradually adds Gaussian noise to the data until it becomes pure noise. The model then learns a reverse Markov chain that denoises the data step by step, effectively sampling from the learned distribution to generate new samples. This process is typically trained by optimizing a variational lower bound on the log-likelihood.