AI image and video generation glossary: 60 terms from checkpoint to VAE, explained in one line each

Topic: Prompting

By Captain

Published 2026-04-15

Overview

Every community invents a vocabulary and this one invented a large one quickly. The glossary below is the reference the other guides on this site assume; each term gets one sentence, the family or tool it belongs to where that matters, and the longer guide to read next. Terms are grouped: models and architectures, files, settings, techniques, prompting, video, hardware.

Families on this site (

Flux

SDXL

Wan

and so on) are explained on their own hub pages, and every model page shows the terms in use: its type, its family, its trigger words, its recommended settings. If a word on a model page is not here, the community chat linked from the FAQ is the place to ask.

Reference

Name	Type	What it is
Checkpoint / base model	file	The full model weights in one file (or a few files for Flux-era models). Works alone. See checkpoint vs LoRA vs embedding.
Fine-tune	model	A checkpoint retrained on new data (DreamShaper, Illustrious , Pony). Same architecture as its base.
Merge	model	A checkpoint made by averaging others. Inherits the family.
LoRA / LyCORIS / DoRA	file	A small add-on that changes a checkpoint's behaviour; needs a trigger word and a weight. See the LoRA guide.
Embedding / textual inversion	file	A learned token for the text encoder, typed by name; mostly negative embeddings on SD 1.5 and SDXL.
VAE	component	Encoder/decoder between pixels and latents; wrong VAE means washed-out colours.
Text encoder (CLIP, T5, LLM)	component	Turns the prompt into numbers; decides the prompt dialect (tags vs sentences).
U-Net / DiT / MMDiT	architecture	The denoiser design: U-Net in SD 1.5 and SDXL, diffusion transformer in Flux, Qwen , Wan, Z-Image .
MoE (mixture of experts)	architecture	Several sub-models of which one runs per step ( Wan 2.2 high/low noise experts, HunyuanImage 3).
Diffusion / flow matching	training	Learning to remove noise step by step; flow matching is the straighter 2024+ formulation. See how models work.
Distilled (Turbo, Lightning, Schnell, Klein, Hyper)	model	Trained to need 1-8 steps and CFG 1; faster, less flexible.
Guidance-distilled	model	CFG baked in ( Flux dev ); a guidance value replaces CFG and the negative prompt does nothing.
fp16 / bf16 / fp8 / NF4 / GGUF Q4-Q8	precision	How weights are stored; lower precision means smaller files and VRAM , slightly lower quality. See the VRAM guide.
Safetensors / ckpt	format	Safe tensor-only format vs code-executing pickle. See safe downloads.
Steps	setting	Denoising iterations; 20-30 normal, 4-8 distilled.
CFG scale	setting	How hard the prompt steers; 5-7 SDXL, 1 Flux. See settings explained.
Sampler / scheduler	setting	Numerical method and noise curve (Euler, DPM++ 2M; Karras, simple, beta).
Seed	setting	The starting noise; fixed seed, repeatable image.
Denoise	setting	In img2img and inpainting, how much of the input may change (0-1).
Shift / sigma	setting	Where a flow model spends its steps (Flux, Wan, Z-Image).
Latent	concept	The compressed working image (1/8 or 1/16 size) the denoiser edits.
Resolution / native size	concept	512 (SD 1.5), 1024 (SDXL, Flux); multiples of 8 or 16.
Hires-fix	technique	Second diffusion pass at a larger size during generation. See upscaling .
Upscaler (ESRGAN, SwinIR)	technique	Pixel enlargement network run after generation.
img2img	technique	Generate from an existing image plus noise. See img2img and inpainting.
Inpainting / outpainting	technique	Regenerate a masked area / extend the canvas.
ControlNet / T2I-Adapter	technique	Structure control from pose, depth or edges. See ControlNet explained.
IP-Adapter / Redux / reference	technique	Image-based style or identity conditioning.
Face detailer / ADetailer	technique	Automatic face mask plus inpaint at higher resolution.
Tiled diffusion / Ultimate SD Upscale	technique	Big images in overlapping tiles with a Tile ControlNet.
Prompt / negative prompt	prompting	What to draw / what to push away from (SD and SDXL only). See prompt basics and negatives.
Trigger word	prompting	The token a LoRA was trained under; required for it to act.
Weight (word:1.3)	prompting	Attention emphasis on CLIP models.
Quality tags / score tags	prompting	masterpiece, best quality (Illustrious); score_9 (Pony). Anime SDXL only.
Token limit (77)	prompting	CLIP chunk size; the reason long prompts get cut on SD and SDXL.
Prompt bleeding	prompting	An attribute attaching to the wrong object.
T2I / T2V / I2V / TI2V / V2V	video	Text to image / text to video / image to video / both / video to video.
Frames and fps	video	81 frames at 16 fps is 5 seconds on Wan; memory scales with frames.
High-noise / low-noise expert	video	Wan 2.2's two 14B models for early and late steps; LoRAs target one.
First / last frame	video	Keyframe conditioning (FLF2V) for controlled motion.
VACE / Fun ControlNet	video	Video editing and control models in the Wan family.
Frame interpolation (RIFE, FILM)	video	Generating in-between frames to raise fps.
VRAM / offloading / block swap	hardware	GPU memory; moving parts of a model to system RAM when it does not fit. See the VRAM guide.
CUDA / MPS / ROCm	hardware	NVIDIA / Apple / AMD compute back-ends.
Workflow ( ComfyUI )	tool	A saved node graph; embedded in ComfyUI PNGs. See the ComfyUI beginner guide.
Custom node	tool	A ComfyUI extension; installed with the Manager.
Heat score	this site	How hot a model is right now, from downloads, runs and searches; explained on the FAQ page.

Learn by doing, not by reading

BitVector Prism labels its settings with the same words as this glossary and preloads the right values per model. Open it beside this page and try each term on a real image. No install.

Try BitVector

Step by step

Read a model page top to bottom with the glossary open: type, family, trigger words, settings line under each example.
Pick the three terms that confused you and read the linked guide for each; the guides are written so that this glossary is enough context.
When you meet a new term in a forum thread, check whether it is a file, a setting, a technique or a video term; most confusion comes from mixing those categories.
Come back when a new family appears; the hub pages add their own terms (Wan 2.2 experts, Flux guidance) and this list is updated with them.

Examples

A model page settings line, decoded

LoRA at weight 0.7 with its trigger word; native SDXL size; sampler + scheduler; 28 denoising steps; CFG 6; fixed seed

A Wan 2.2 workflow name, decoded

wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors -> Wan 2.2, image to video, the high-noise expert, 14B parameters, fp8 precision, safetensors format

Tips

Three words cause most beginner errors: family (a LoRA must match it), trigger word (a LoRA needs it) and denoise (img2img lives or dies by it).
When two guides disagree on CFG, they are talking about different families; check which.
"Model" means checkpoint in most sentences, but on this site "model" is any catalog entry including LoRAs; the type label tells you which.
Video terms are the same ideas with a time axis; if you know T2I and img2img you know T2V and I2V.
The cloud partners use the same vocabulary:
PirateDiffusion
's flags are /steps, /guidance, /seed, /size;
BitVector
's panel uses the same names.

Troubleshooting

Two sources use the same word differently

Why it happens

Terms drifted between SD 1.5, SDXL and Flux eras ("guidance" vs "CFG").

How to fix it

Anchor on the family: CFG for SD/SDXL, guidance for Flux.

"Model" on Civitai means something else than here

Why it happens

Civitai groups versions under one model; this site lists each version separately.

How to fix it

Match by hash, not by name.

A setting named on a model page is missing in my UI

Why it happens

Different UI names (denoising strength vs denoise; hires steps vs second pass).

How to fix it

The glossary row gives the aliases; the ComfyUI node guide maps node names.

Every term is a command

PirateDiffusion turns the vocabulary into Telegram flags: /steps, /guidance, /seed, /size, LoRA tags and [[negatives]]. Thousands of models, unlimited renders, fixed price.