Diffusion Model

A diffusion model is trained by taking real images and progressively adding noise until nothing recognizable remains, while learning to reverse each step. Run that learned reversal from pure random noise and it produces a new image that never existed.

Text guides the process. The prompt steers each denoising step toward an image matching the description, which is why small wording changes can shift a result substantially — you are nudging a path through an enormous space of possible images rather than looking one up.

The same approach extends to video, with the added requirement of coherence across frames, which is considerably harder and is where the field is currently advancing fastest.

Diffusion explains the characteristic behavior of image tools: why generation takes seconds of visible refinement, why the same prompt yields different images each time, and why fine details like hands and text have been persistently difficult — they require global consistency that a local denoising process does not naturally enforce.

Because output is stochastic, the workflow is generate-several-and-select rather than get-it-right-first-time. Most tools expose a seed value that makes a result reproducible, which is what you want when iterating on a prompt rather than exploring.