A generative model that creates data by iteratively removing noise from random input using learned denoising steps.
Also known as: Denoising Diffusion Probabilistic Model (DDPM), Latent Diffusion Model (LDM)
In plain English
A diffusion model makes images or sounds by starting from random noise and gradually clearing it until a clear result appears, like wiping fog off a window. You type a prompt and the model refines the noisy picture into something recognisable. Many popular image generators use this approach, so you just see the final outputs.
In practice
Diffusion models power many image-generation tools and are chosen for high-quality, controllable outputs in design and content workflows. They matter in practice because sampling requires multiple inference steps, so latency, GPU cost and choice of latent vs pixel-space models affect deployment, scaling and cost.
Under the hood
Diffusion models learn to reverse a forward process that progressively adds Gaussian noise to data by predicting the noise or the score at each timestep; training uses score-matching or variational objectives (e.g. DDPM). Sampling performs iterative denoising with a neural network (often a U-Net) conditioned on timestep and optional embeddings; latent diffusion operates in a compressed latent space to reduce compute and memory.
Example
Can we use a diffusion model to generate high-resolution product mockups, and what GPU budget and latency should we plan for?