Diffusion Model

A class of generative models that learns to denoise images, commonly used for high-quality image generation.

How it works

A diffusion model is trained by taking real images and progressively adding Gaussian noise to them over many steps until the image becomes pure noise. A neural network (typically a U-Net or Transformer) is then trained to predict and reverse this noise one step at a time. At inference, generation starts from random noise and the network iteratively removes noise over many steps, guided by a text or image conditioning signal, until a clean, realistic image emerges.

Why it matters

Diffusion models are the engine behind Stable Diffusion, DALL-E, Midjourney, and Sora, representing the state of the art in AI image and video generation. They produce significantly higher-quality, more diverse outputs than the GANs they largely replaced. Their ability to condition generation on text prompts has democratised professional-quality image creation, transforming creative industries and raising major questions about copyright, authenticity, and the future of visual content creation.