NSFW AI Glossary: Every Term Explained for Beginners (2026)

New to NSFW AI and drowning in jargon? This glossary explains every term a beginner meets, in plain English with a one-line example each: checkpoint, model, LoRA, VAE, CFG, sampler, steps, seed, denoise, hires fix, upscaler, inpainting, img2img, ControlNet, negative prompt, token, booru tags, and NSFW versus SFW. Bookmark it and look words up as you go.

When you first open an AI image tool, it throws a wall of unfamiliar words at you. Checkpoints, samplers, CFG, denoise, LoRAs. It sounds like a foreign language, and most guides assume you already know it. This page does not. Every term below is explained as if you have never seen it before, with a quick example so you can picture what it actually does.

One rule applies everywhere in this niche, no matter which term you are using. Subjects you generate must be adult (18+), fictional, and AI-generated. Never real, identifiable people, and never minors. The example prompt near the end shows the safety tokens that belong in every negative prompt for exactly that reason.

You do not need to memorize any of this. Skim the quick-reference table, then read the definitions for the words you keep bumping into. Keep this open in a tab while you practice. When you are ready to put the words to work, you can open our free generator and try them live.

Here is a friendly way to think about the whole vocabulary before you dive in. The checkpoint is the artist. The LoRAs are extra skills you hand that artist. The sampler and steps are how carefully the artist works. CFG is how literally the artist takes your instructions. The seed is which blank canvas they start from. Once those five ideas click, every other term is just a refinement on top of them. So if any definition below feels heavy, come back to that picture and it gets simpler.

Quick-reference table

Term	One-line meaning
Checkpoint	The main AI model file that makes images
Model	General word for the AI that generates images
LoRA	Small add-on that teaches a style or character
VAE	Decoder that handles final color and sharpness
CFG	How strictly the AI obeys your prompt
Sampler	The method used to build the image step by step
Steps	How many passes the AI makes refining the image
Seed	The number that sets the random starting point
Denoise	How much an existing image is changed
Hires fix	A two-stage trick for sharper, larger images
Upscaler	Tool that enlarges an image and adds detail
Inpainting	Redrawing only a selected part of an image
img2img	Making a new image based on an input image
ControlNet	Guides pose or composition from a reference
Negative prompt	List of things you do not want in the image
Token	A chunk of your prompt the AI reads as one unit
Booru tags	Comma-style keyword tags anime models understand
NSFW vs SFW	Adult content versus safe-for-work content

Now the full plain-English definitions.

The core building blocks

Checkpoint

A checkpoint is the main AI model file that actually makes your images. It is the brain of the whole operation. Different checkpoints are trained on different art, so one might be great at photorealism while another excels at anime. Picking the right checkpoint is the biggest single decision for your style.

Example: you load a photoreal checkpoint to make lifelike portraits, then swap to an anime checkpoint for illustrated art. Our roundup of the best Stable Diffusion checkpoints for NSFW sorts the popular ones by style.

Model

Model is the general, everyday word for the AI that generates images. People often use model and checkpoint to mean the same thing in casual talk. Technically a checkpoint is one kind of model file, but if someone says “which model should I use,” they usually mean which checkpoint.

Example: a friend asks what model makes the best realistic results, and you point them to a photoreal checkpoint.

LoRA

A LoRA is a small add-on file that teaches the main model one specific thing, like an art style, a character, or a clothing type. It is far smaller than a checkpoint and stacks on top of one. You can mix several LoRAs to combine effects, each at an adjustable strength.

Example: you add a “film grain” LoRA at low strength to give photos a vintage feel without changing the base model.

VAE

A VAE is the part that decodes the model’s internal work into the final visible image, handling color and fine sharpness. When colors look washed out or grey, a missing or wrong VAE is often the cause. Many modern checkpoints include one built in, so beginners rarely touch it.

Example: your output looks dull and faded, you switch on the recommended VAE, and the colors snap back to life.

The dials you actually turn

CFG (CFG scale)

CFG scale controls how strictly the AI follows your prompt. Low CFG lets the model be creative and loose. High CFG forces it to obey, but too high makes images harsh and fried. For most models, a value of 4 to 7 is the comfortable beginner range.

Example: you set CFG to 6, and the result follows your prompt while still looking natural.

Sampler

The sampler is the method the AI uses to build your image step by step from noise. Different samplers give slightly different looks and speeds. You do not need to understand the math, just pick a reliable one and stick with it while you learn.

Example: you choose DPM++ 2M Karras, a popular all-rounder, or Euler a for a clean, fast result.

Steps

Steps are how many passes the AI makes refining the image. Too few looks unfinished and noisy. Too many wastes time with little gain. For most beginners, 20 to 30 steps is the sweet spot.

Example: you set steps to 25 and get a clean image without waiting forever.

Seed

The seed is the number that sets the random starting point of an image. A random seed gives a brand new image every time. Lock a seed and you can reproduce or refine the exact same composition.

Example: you love an image, so you copy its seed and reuse it while tweaking the prompt to improve that specific shot.

Denoise (denoising strength)

Denoise controls how much an existing image is changed when you feed it back in. Low denoise keeps it close to the original. High denoise changes it a lot. It mostly matters for img2img and hires fix.

Example: in hires fix you set denoise to 0.5, which adds detail without redrawing the whole picture.

Making images bigger and cleaner

Hires fix

Hires fix is a two-stage trick that first generates a small image, then enlarges and refines it in one pass for sharper, larger results. It is the easy way to get a detailed big image without the doubled-body problems of generating large from scratch.

Example: you enable hires fix with a denoise around 0.4 to 0.6 and your 1024 image becomes a crisp, detailed larger one.

Upscaler

An upscaler is a tool that enlarges an image and adds detail rather than just stretching it. You can upscale during generation or afterward as a separate step. It is how soft images become poster-sharp.

Example: you run a finished image through an upscaler to double its size for a high-resolution print. If your results look soft, our blurry image fix guide walks through upscaling in detail.

Editing and guiding the image

Inpainting

Inpainting means redrawing only a selected part of an image while leaving the rest untouched. You mask an area, then the AI regenerates just that spot. It is perfect for fixing one bad hand or changing a small detail.

Example: a face came out odd, so you mask just the face and inpaint it to fix it without redoing the whole image.

img2img

img2img (image to image) makes a new image based on an input image instead of starting from pure noise. The denoise setting decides how much it changes. It is great for refining a rough draft or restyling a picture.

Example: you sketch a rough pose, feed it to img2img at medium denoise, and get a polished version of that same pose.

ControlNet

ControlNet guides the pose, outline, or composition of a new image using a reference. It lets you control exactly how a subject is positioned instead of hoping the prompt gets it right. It is a more advanced tool but hugely powerful.

Example: you supply a pose reference through ControlNet so your generated character stands in that exact stance.

Prompt-related terms

Negative prompt

The negative prompt is a list of things you do not want in the image. It removes artifacts like extra fingers and blur, and it is where your safety tokens belong. Always use one.

Example: your negative prompt includes bad anatomy, extra fingers, blurry, plus child, minor, underage, loli, shota. Our negative prompts master list gives you a ready block to copy.

Token

A token is a chunk of your prompt that the AI reads as one unit, usually a word or part of a word. Prompts have a token limit, so very long prompts can get cut off. Keep prompts focused so the important words count.

Example: “soft window light” is a few tokens, and the model weighs each one when composing.

Booru tags

Booru tags are comma-separated keyword tags that anime models were trained on, like “1girl, long hair, smiling.” Anime checkpoints respond far better to this tag style than to full sentences. Realistic models prefer more natural phrasing.

Example: for an anime model you write “1girl, lingerie, bedroom, soft lighting” instead of a flowing sentence. See our Illustrious XL guide for tag tips.

NSFW vs SFW

NSFW means not safe for work, which is adult or explicit content. SFW means safe for work, which is everyday content you could show anyone. Many tools and models are labeled one way or the other so you know what they allow.

Example: a model described as uncensored handles NSFW, while a filtered tool only allows SFW output. Our best uncensored AI image generators list focuses on tools that allow adult content.

A safe example using these terms

Here is how several of these terms come together in a real, tasteful generation. Subjects must be adult (18+), fictional, and AI-generated, never real people and never minors.

Checkpoint: photoreal SDXL model
Positive: (adult woman, 26 years old:1.2), elegant lingerie, bedroom, soft window light, photorealistic, detailed skin, sharp focus
Negative: child, minor, underage, loli, shota, lowres, bad anatomy, extra fingers, blurry, watermark, text
Sampler: DPM++ 2M Karras | Steps: 28 | CFG: 6 | Size: 1024x1024 | Seed: 12345 | Hires fix denoise: 0.5

Read that block and you can now decode every part of it. The checkpoint sets the style, the sampler and steps build the image, CFG controls obedience, the seed makes it repeatable, and hires fix sharpens the result. That is the whole vocabulary working together.

How to use this glossary

Do not try to learn every term today. Pick the three or four words you see most often in your tool and read just those. The rest will sink in naturally as you generate. Within a week or two of practice, this entire page will feel obvious, and you will be the one explaining “what is a seed” to the next beginner.

If a term here connects to a setting you want to tune, our settings explained for beginners gives you safe starting values for each dial. And when you want to turn vocabulary into better images, how to get better NSFW AI results shows the quick wins. New to the whole thing? Start with our beginner generator guide.

A few bonus terms you will hear

These are not on the core list above, but they come up often enough that knowing them saves confusion.

Prompt. Simply the text you type to describe the image you want. The positive prompt says what to include. People often just say “prompt” to mean the positive one.

Artifact. Any unwanted glitch in the output, like an extra finger, a smeared face, or a stray watermark. Most artifacts are reduced by a good negative prompt and the right resolution.

Aspect ratio. The shape of the image, like square, tall portrait, or wide landscape. Match it to your subject. A standing pose suits a tall canvas, a reclining scene suits a wide one.

Weight. A way to tell the model how strongly to apply a word, written like (detailed skin:1.2). Numbers above 1 strengthen a term, below 1 weaken it. Use it gently, since heavy weights can distort the image.

Trigger word. A specific word a LoRA or model needs in the prompt to activate its trained style or character. The model page usually lists it. Forgetting the trigger word is why a LoRA sometimes seems to do nothing.

Batch. How many images you generate per click. Making several at once lets you compare and pick the best, which matters because AI image-making is a numbers game.

Checkpoint merge. A model created by blending two or more checkpoints to combine their strengths. As a beginner you just use them like any other checkpoint, no special handling needed.

None of these are essential on day one, but once you see them in a forum or a model page, you will know exactly what they mean instead of feeling lost.

Why the vocabulary matters

It is tempting to skip the words and just push buttons, and you can get images that way. But the moment something goes wrong, the jargon is how you find the fix. If your colors look dull, knowing the word VAE lets you search the exact solution. If faces melt, knowing ADetailer and inpainting points you straight at the answer. The vocabulary is not academic. It is the map you use to solve real problems quickly, which is why even a casual creator benefits from skimming this page once.

The fastest way to make these words stick is to use them. Open the generator, set a checkpoint, a sampler, your steps and CFG, paste a negative prompt with safety tokens, and generate. Every term you just read becomes muscle memory the moment you turn the dial yourself. Keep this page bookmarked, and look words up without shame. Everyone starts here.

Frequently asked questions

What is the difference between a checkpoint and a model?

In casual use they mean the same thing: the main AI file that generates your images. Technically a checkpoint is a specific kind of model file, but when someone asks which model to use, they almost always mean which checkpoint. Different checkpoints are trained for different styles, so picking one sets whether your output looks photorealistic or like anime.

Do I need to know what a VAE is to start?

Not really. A VAE decodes the model’s work into the final colors and sharpness, and many modern checkpoints include one built in. You only need to think about it if your images look dull, grey, or washed out, which usually means a missing or wrong VAE. For most beginners it just works in the background without any attention.

What is a LoRA and how is it different from a checkpoint?

A checkpoint is the full main model, the brain of the image. A LoRA is a small add-on that stacks on top to teach one specific thing, like a style, a character, or an outfit type. LoRAs are much smaller and you can mix several at adjustable strengths, while you only run one checkpoint at a time as the base.

What are booru tags and when do I use them?

Booru tags are comma-separated keyword tags like 1girl, long hair, smiling, that anime models were trained on. Anime checkpoints respond far better to this tag style than to full sentences. Realistic photo models, by contrast, prefer more natural descriptive phrasing. So match your prompt style to your model: tags for anime, natural language for photoreal results.

What does denoise actually control?

Denoise sets how much an existing image is changed when the AI works on it. It matters most for img2img and hires fix. Low denoise keeps the result close to the original. High denoise changes it a lot. For hires fix, a value around 0.4 to 0.6 adds detail and sharpness without redrawing your whole composition, which is exactly what you want.

What is the difference between inpainting and img2img?

img2img makes a whole new image based on an input image, changing the entire picture by an amount you set with denoise. Inpainting only redraws a selected, masked area while leaving everything else untouched. Use img2img to refine or restyle a full image, and use inpainting to fix one small problem, like a single bad hand or face.

How many steps and what CFG should a beginner set?

For most models, 20 to 30 steps and a CFG of 4 to 7 are reliable beginner values. Steps control how much the image is refined, and too many just wastes time. CFG controls how strictly the AI obeys your prompt, and too high makes images harsh and fried. Start at around 25 steps and CFG 6, then adjust from there.

What does NSFW versus SFW mean?

NSFW means not safe for work, which is adult or explicit content. SFW means safe for work, everyday content you could show anyone. Tools and models are often labeled one way so you know what they allow. An uncensored model handles NSFW, while a filtered tool only permits SFW. Whatever you make, subjects must be adult, fictional, and AI-generated.