AI art prompting has a vocabulary of its own, and a lot of it is thrown around without ever being explained. This glossary defines the terms you will meet most often in our guides and in the wider community. Where a term means slightly different things in different models, we say so. If you come across something we haven’t covered, tell us via our contact page and we’ll add it.

Prompt structure and language

  • Prompt. The text you give a model describing the image you want. It typically combines a subject, a style, and modifiers for lighting, composition and technical settings.
  • Negative prompt. A separate list of things you do not want in the image (for example “blurry, extra fingers, text”). Supported by Stable Diffusion and, in a different form, by Midjourney’s --no parameter. Used well, it removes recurring flaws; overused, it can strip out detail you wanted.
  • Modifier / keyword. A word or phrase that nudges the output — a style (“watercolour”), a mood (“melancholic”), or a quality tag (“highly detailed”). Modifiers are the main way you steer a result.
  • Token weighting. Techniques for telling the model that some words matter more than others, often with syntax like (word:1.3) in Stable Diffusion interfaces. Raising a weight emphasises a concept; lowering it de-emphasises it.
  • Token. The unit a model breaks your prompt into — roughly a word or word-piece. Models have a limit on how many tokens they read, so very long prompts can have their tail ignored.

Core generation settings

  • Seed. The number that initialises the random noise a model starts from. The same prompt with the same seed and settings produces (nearly) the same image, which is why fixing a seed is essential when you want to test the effect of a single change.
  • CFG scale (guidance scale). Short for “classifier-free guidance.” It controls how strictly the model follows your prompt. Low values give the model more creative freedom (and can wander off-prompt); high values force closer adherence (and can look over-cooked or harsh). Mid-range is usually the sweet spot.
  • Sampler. The algorithm that turns random noise into an image over successive steps (examples include Euler, DPM++ and others in Stable Diffusion). Different samplers can produce subtly different looks and converge at different speeds.
  • Steps. How many iterations the sampler runs. Too few and the image is noisy or unresolved; beyond a point, more steps add cost without meaningful improvement.
  • Denoising strength. In image-to-image workflows, how much the model is allowed to change the input image. Low values stay close to the original; high values reinterpret it more freely.

Composition and output

  • Aspect ratio. The width-to-height proportion of the image, such as 1:1 (square), 16:9 (wide) or 2:3 (portrait). Set in Midjourney with --ar. Choosing the right ratio up front matters because it changes how the model composes the scene.
  • Stylize. A Midjourney parameter (--stylize or --s) that controls how strongly Midjourney applies its own artistic sensibility. Low values track your prompt more literally; high values produce more stylised, “opinionated” images.
  • Upscaling. Increasing an image’s resolution after generation, often adding detail in the process. Many tools offer dedicated upscalers separate from the initial generation.
  • Inpainting. Regenerating only a selected region of an image while leaving the rest intact — useful for fixing a hand or swapping one element.
  • Outpainting. Extending an image beyond its original borders, letting the model imagine what lies outside the frame.

Models, styles and add-ons

  • LoRA. Short for “Low-Rank Adaptation.” A small add-on file that fine-tunes a base model toward a specific style, character or concept without retraining the whole model. Widely used with Stable Diffusion to lock in a consistent look.
  • Checkpoint / base model. The main trained model file that does the generating (for example a particular Stable Diffusion checkpoint). Different checkpoints have different strengths and default aesthetics.
  • ControlNet. An extension for Stable Diffusion that guides generation using an input such as a pose skeleton, edge map or depth map, giving you precise control over composition.
  • Style reference / image prompt. Supplying an image (rather than only text) to influence the style or content of the output, as with Midjourney’s image prompts and style-reference features.
  • Prompt weighting with ::. Midjourney syntax that splits a prompt into weighted parts (for example forest::2 city::1), telling the model which concept to favour.