Diffusion

Diffusion models are a class of generative AI models used to produce images, audio, video, and other complex assets. They are trained by gradually adding noise to real data—a step known as diffusion—and learning to reverse this process. As a result, they can start from random noise and iteratively denoise it to produce high-quality outputs.

Applications of diffusion models include

• image generation,
• text-to-image and text-to-video generation,
• image editing and inpainting,
• super-resolution, and
• audio and music generation

Diffusion models are widely used in AI image generators such as Stable Diffusion (Stability AI), DALL-E 2 (OpenAI), Google Imagen, and Midjourney. They have become one of the leading approaches for creating realistic pictures, offering improved training stability and output quality compared to earlier methods such as variational autoencoders (VAEs) and generative adversarial networks (GANs).