Models
Diffusion Model
Also known as: denoising diffusion
The kind of generative model behind most modern AI image generators; it creates images by starting from noise and gradually removing it.
A diffusion model is the technology behind most AI image generators. The idea is surprisingly odd: it learns to turn random static (noise) into a picture, step by step. Imagine starting with TV snow and slowly cleaning it up until a clear image of exactly what you asked for appears. Because it starts from fresh random noise each time, it can create endless new images.
Diffusion models are generative models used mainly for images (and increasingly audio and video). They are trained in two directions: 1. Forward: take a real image and gradually add noise until it is pure static. 2. Reverse: learn to undo that, removing noise step by step.
Once trained, you start from random noise and run the reverse process to generate a brand-new image, guided by a text prompt.
Why they took over image generation
They produce high-quality, diverse images and train more stably than the earlier GAN approach. Text guidance (via embeddings that link words and pictures) is what lets you type a description and get a matching image.
The efficiency trick
Running this in a compressed "latent" space rather than on full-resolution pixels made high-quality generation fast enough for everyday tools.
Diffusion models define a fixed forward Markov process that progressively corrupts data with Gaussian noise, and learn a parameterised reverse process to denoise, trained via a variational bound that reduces to predicting the added noise at each step (Ho et al. 2020). Sampling integrates the learned reverse dynamics from noise to data; guidance (classifier-free guidance) steers generation toward a conditioning signal such as a text embedding.
Latent diffusion moved the process into the compressed latent space of an autoencoder, cutting compute by orders of magnitude and enabling high-resolution synthesis at scale (Rombach et al. 2022). Denoising networks were historically U-Nets; diffusion transformers (DiT) now scale competitively. Relative to autoregressive and GAN generators, diffusion offers strong sample quality/diversity and stable training, at the cost of multi-step iterative sampling, which distillation and improved solvers accelerate. The framework extends beyond images to audio, video and molecular generation.