Components
VAE
Variational AutoEncoder — converts model latents to pixels. Different VAEs = different color/contrast.
Related terms
LoRA
Low-Rank Adaptation — a small trained file that adds a specific style or character to a base model.
ControlNet
Adds structural control: pose, depth, edges, scribbles. Essential for consistent character work.
Sampler
Algorithm that iteratively denoises images. Common: DPM++, Euler a, DDIM.
Checkpoint
A saved model file (.safetensors / .ckpt). Different checkpoints produce different aesthetics.
Embedding
Textual Inversion files that teach the model new concepts via short trigger words.