Diffusion Model

Diffusion models learn to reverse a process: start with random noise, gradually subtract noise, end with a coherent image. During training the model sees real images progressively corrupted with noise; during inference it runs the process backwards.

Diffusion is the core of every major image generator: Stable Diffusion, Midjourney, DALL-E, Flux, Ideogram. It's also become standard for video generation (Sora, Runway, Kling) and is making inroads into music and 3D.

Diffusion has different strengths from autoregressive (transformer) generation. It's parallel rather than sequential, gives you natural control over how 'noisy' or 'clean' an output is, and produces beautiful images. It's slower than direct generation, which is why much research focuses on speeding it up (consistency models, LCMs, distilled models).