A diffusion model is the technology behind most of today's AI image generators. It works in a surprisingly counter-intuitive way: during training it learns to gradually add noise to images until they become static, and then to reverse that process. To create a new picture, it starts from pure random noise and "denoises" it step by step, guided by your prompt, until a coherent image emerges.
This approach is why AI image tools can produce such varied, detailed results — each generation is a fresh path from noise to picture, so the same prompt gives different images every time. It is also why generating images takes noticeably longer than generating text: the model may run dozens of refinement steps per image. Diffusion techniques are now spreading beyond images into video and audio generation.
Related terms
Generative AI · Prompt · Computer Vision