Glossary
Diffusion
Short answer
How image models work: building an image by progressively removing random noise.
Diffusion models learn during training by adding noise to images step by step and learning to reverse that process. At generation time they start from pure noise and progressively remove it, guided by the prompt, until the image emerges.
This method is why the same prompt gives a different result each time: the starting noise is random. The value that fixes that randomness is the seed.
The step count determines how many stages the denoising takes. Too few steps gives a blurry image; too many gives slow generation with no added benefit.
See also
- SeedThe number setting a generation's starting randomness; the same seed with the same prompt reproduces the same image.
- StepsThe setting for how many stages the denoising takes; it directly affects quality and time.
- Latent SpaceThe internal space where a model represents images in compressed mathematical form.
You will meet this in
- Stable Diffusion — Unlimited Generation on Your Own MachineOpen source; installed locally it is free, uncensored and completely controllable.
- Flux — The Model That Leads on Photorealism and TextThe strongest open-weight family for photorealistic people and for text rendered inside images.
- Midjourney — The Aesthetic Quality LeaderThe subscription generator that delivers the most consistent artistic and aesthetic results.