A normalizing flow is a type of deep generative model that builds up a complex data distribution by applying a sequence of invertible transformations to a simple starting distribution, usually plain random noise. Because every step is invertible and its mathematical effect on probability can be tracked exactly, a flow can run in two directions: forward to turn noise into a realistic sample, and backward to turn a real example into noise while reporting precisely how likely that example is.
This exact-likelihood property is what distinguishes normalizing flows from the other two classic families of generative models. Generative adversarial networks produce sharp images but give no probability estimate, and variational autoencoders provide only an approximate likelihood. Flows give an exact one, which matters for tasks like anomaly detection, density estimation, and any setting where calibrated probabilities are needed. The cost is architectural: every layer must be invertible and structured so its likelihood contribution is cheap to compute, which constrains the design. The Glow model, introduced by OpenAI in 2018, is a well-known example that used learnable invertible convolutions to generate and edit realistic faces.
In raw image quality, normalizing flows were eventually overtaken by GANs and then diffusion models, but the underlying mathematics of invertible, continuous transformations proved durable. It resurfaced in continuous normalizing flows and in flow matching, a training method behind some of the newest image and video generators. For a business or general reader, normalizing flows are worth knowing as the third main approach to teaching machines to generate data, the one that keeps an honest, exact accounting of probability.