noun
A diffusion model is a machine learning technique used to generate high-quality images, text, or other data by reversing a gradual noising process. Unlike older methods like GANs, it adds and removes noise in small steps, making it more stable and controllable.
It works by training the AI to predict how to remove noise from an image or text, step by step. Over time, it learns to transform random noise into coherent outputs, like a photo or a sentence. This approach is widely used in AI art and text generation.
For creators, it changes workflows by making high-quality generation faster and more predictable. For example, Stable Diffusion can produce detailed images in seconds, but it requires powerful hardware and careful prompting to avoid biases or unintended results. Costs include GPU time and fine-tuning for specialized use cases.
An artist uses a diffusion model to generate a fantasy landscape in 10 seconds, adjusting settings to reduce blurriness.
In other languages
See also: Generative AI · Latent space