Diffusion Model.
A generative AI model that creates data by learning to reverse a noise-adding process, iteratively denoising random inputs to produce high-quality outputs.
Diffusion models are a class of generative artificial intelligence models that learn to create new data by mastering the process of removing noise from corrupted inputs. Unlike traditional generative approaches, diffusion models work by first learning how to systematically add noise to training data until it becomes pure random noise, then learning to reverse this process step by step. This approach has become the dominant paradigm for AI image generation, powering systems like DALL-E, Midjourney, and Stable Diffusion, and is increasingly being applied to other domains including audio, video, and even text generation.
The core mechanism involves two phases: a forward diffusion process that gradually adds Gaussian noise to clean data over many timesteps until it becomes indistinguishable from random noise, and a reverse denoising process that the model learns to perform. During training, the model learns to predict and remove the noise that was added at each step, effectively learning the data distribution. At inference time, the model starts with pure noise and iteratively applies the learned denoising steps to generate new samples. This process typically requires dozens or hundreds of denoising steps, though recent advances have reduced this through techniques like distillation and improved sampling methods.
Diffusion models excel at generating high-quality, diverse outputs with fine-grained control over the generation process, but they come with significant computational costs due to their iterative nature. The multi-step generation process makes them slower than single-step generative models like GANs, though they often produce more stable and higher-quality results. A common misconception is that diffusion models are only useful for images; while they achieved breakthrough success in visual domains, they are increasingly being adapted for other modalities and even being integrated into language model architectures, representing a fundamental shift in how AI systems learn to generate complex, structured data.