Diffusion Model
A type of generative AI that creates images (and now video) by progressively denoising random noise. Powers DALL-E, Stable Diffusion, Midjourney, Sora.
Diffusion models are a class of generative AI that produce images, videos, or audio by reversing a noising process. During training, the model learns to undo small amounts of noise added to real images. At generation time, it starts with pure noise and iteratively denoises it, guided by a text prompt, producing a coherent image.
Diffusion models replaced GANs (Generative Adversarial Networks) as the dominant image generation approach around 2022. They produce more diverse and higher-quality outputs than GANs and are easier to train.
Examples: DALL-E 3 (OpenAI), Stable Diffusion (open source), Midjourney, Flux. Video diffusion (Sora, Runway Gen-3) extends the same approach to temporal dimensions. Diffusion models are increasingly being explored for non-image tasks too, including protein design and molecule generation.