AI Image Prompt Glossary

Plain-English definitions of the terms you'll meet when writing prompts for Stable Diffusion, ComfyUI, Automatic1111, Midjourney and Flux.

Prompt: The text description you give an image model. In tools like Stable Diffusion it is usually a comma-separated list of tags or short phrases.
Negative prompt: A second prompt listing what you do NOT want (e.g. “blurry, extra fingers”). The model steers away from these concepts.
Attention weight: A multiplier on part of a prompt — written (token:1.2) — that makes the model emphasise (>1.0) or de-emphasise (<1.0) it.
CFG scale: Classifier-Free Guidance scale: how strictly the model follows your prompt. Low is loose and creative; high is literal but can look over-cooked. 5–8 is typical.
LoRA: Low-Rank Adaptation — a small add-on model that injects a style, character or concept. Loaded inline as <lora:name:weight> or via a node in ComfyUI.
Token: The unit a model reads. CLIP processes prompts in chunks of 75 tokens; a token is roughly a short word or word-piece.
BREAK: A keyword that ends the current 75-token chunk and starts a new one, keeping groups of concepts from blending together.
Embedding (Textual Inversion): A trained keyword that stands in for a concept the base model did not know, used by typing its name in the prompt.
Sampler: The algorithm that turns noise into an image step by step (e.g. Euler a, DPM++ 2M). It affects speed and final look.
Steps: How many denoising iterations the sampler runs. More steps can add detail up to a point, then plateau.
Seed: The number that initialises the random noise. The same seed plus prompt and settings reproduces the same image.
Wildcard: A placeholder like {red|blue|green} or __colors__ that resolves to one random option, for quickly varying a prompt.
Danbooru tags: A controlled tag vocabulary (from the Danbooru image board) that anime / illustration models like Pony and Illustrious are trained on.
Checkpoint / Model: The main trained weights that generate images (e.g. SDXL, Pony, Flux). Different checkpoints expect different prompt styles.

Keep learning

Build it without the syntax wrangling