Glossary

CFG — "classifier-free guidance": how strictly the model follows the prompt. Higher is more literal; too high burns the colours. FLUX.2, Z-Image, Qwen Image and Krea 2 are made for 1.

Checkpoint — a model file (or set of files) that draws pictures. Used interchangeably with model.

Denoise — with a starting photo, how much of it is replaced. 0 keeps the photo, 1 replaces it completely.

DiT — "diffusion transformer", the newer kind of image model: FLUX.2, Z-Image, Qwen Image and Krea 2. They read full sentences and draw any size from 512 to 2048.

Edit model — a model that takes a photo as a clean reference and changes it to follow an instruction: FLUX.2 Klein and Qwen Image 2.1.

Embedding — a tiny file holding one trained concept, used by writing its name in a prompt.

Family — a group of models that work the same way: SD 1.5, SDXL, Anima, FLUX.2, Z-Image, Qwen Image, Krea 2.

Flow — a set of nodes wired together for one job. The app ships ready-made ones; you can save your own.

HTP / NPU — the Snapdragon's AI accelerator (Hexagon Tensor Processor), where the models run. Its generation is shown as "HTP arch", e.g. v79 on a Snapdragon 8 Elite.

Inpaint — repainting only a masked area of a photo.

LoRA — a small adapter file that adds a style or character to a model without replacing it.

Mask — the painted area that inpainting redoes (red in the mask editor).

Node — one step in a flow: a prompt, a photo, a generator, an output.

Outpaint — extending a photo past its edges; the new area is generated.

Seed — the number that picks the starting noise. The same seed with the same settings makes the same picture; 0 means a new random one every Run.

Steps — how many refinement passes the model makes. More is slower; each model has a sweet spot.

Scheduler — the method used for those passes (Euler, DPM, LCM…). Mostly a matter of what the model was made for.

Upscaler — a small model that enlarges a picture and sharpens it.

VAE — the part of a model that turns its internal image into pixels and back.

VTCM — fast memory inside the NPU. Some older or smaller chips have less, which limits which model builds they can run.

Wire — a line joining one node's output to another's input.