Glossary

Concise definitions of common diffusion terms.

How To Use This Glossary

Use this page for quick lookups, then follow links to deeper pages when a term affects workflow decisions or troubleshooting.

Terms

Adapter

A small add-on that modifies or steers a model without replacing the whole checkpoint. LoRAs are a common adapter type.

Attention

A mechanism that lets parts of the model relate different pieces of information, such as image features and text-conditioning features. Attention behavior is one reason prompts can affect both global composition and local details.

CFG

Classifier-free guidance. A setting that controls how strongly some workflows follow prompt conditioning. See Classifier-Free Guidance.

Checkpoint

A saved set of model weights, often used to refer to the main model file in image generation tools. See Checkpoints, VAEs, And Adapters.

Conditioning

Information used to steer generation, such as text, images, masks, pose maps, LoRAs, or embeddings. See Conditioning.

Denoising

The repeated process of moving from noise toward a cleaner sample. See Denoising.

Embedding

A learned token or set of tokens used through text conditioning, often tied to a specific model family or text encoder.

Guidance

The process of pushing denoising toward conditioning. Classifier-free guidance is one common form. See Classifier-Free Guidance.

Image-To-Image

A workflow that starts from an existing image instead of pure noise. See Image-To-Image.

Inpainting

A workflow that edits a masked region of an image. See Inpainting.

Latent Space

A compressed representation used by many diffusion workflows. See Latent Space.

LoRA

Low-rank adaptation. A small adapter that modifies a base model's behavior. See What Are LoRAs?.

Negative Prompt

Conditioning that tells some workflows what to avoid. See Negative Prompts.

Text Encoder

A model component that converts prompt text into conditioning representations. See Text Encoders.

Sampler

The algorithm that decides how to move through denoising steps. See Samplers And Schedulers.

Scheduler

The noise schedule used across denoising steps. See Samplers And Schedulers.

Seed

A number used to initialize random noise for generation. See Seeds And Reproducibility.

VAE

Variational autoencoder. In latent diffusion, it converts between pixels and latents. See Checkpoints, VAEs, And Adapters.