NovelAI Diffusion V4 for Noobs

Prompts

Prompt Box

A recommended prompt structure is as follows:

1. Tags for the number and type of subjects (e.x. 1boy, 2girls, 1other).
2. Series tags, if creating fanart (e.x. boku no hero academia, pokemon).
3. Artist/style–related tags. This is usually in the form of artist:[artistname], but some series tags can also work as style tags, especially if you're using V4 Curated (the SFW version of V4).
4. General tags related to the overall style or composition of the image (e.x. western comics (style), blurry background, wide shot, cowboy shot).
5. Specific tags related to the setting or action of the image (e.x. outdoors, night, anal, fellatio).
6. If you want to include a prose description (either to explain complex concepts or to reinforce what you tagged), do it in a new sentence (or sentences) at the end of the prompt. (E.x. 2boys, outdoors, night. A middle-aged man sitting on a red park bench at night, wearing a trenchcoat and reading a newspaper. A clown stands behind him, holding a large rubber mallet and preparing to hit the man.)

Undesired Content Box

Like with quality tags, NAI invisibly adds negative tags to prompts that you can turn off.

Honestly, undesired content is hit or miss. A lot of times it has a weak or no effect so you don't need to worry about it too much. It's generally better to describe what you do want to see in the prompt fields rather than what you don't want to see in the undesired content fields. E.x. no headwear and bottomless in the prompt is better at removing a character's hat or pants respectively than putting hat or pants in undesired content (although it doesn't hurt to do both).

Character Boxes

By far the best part of V4 and why it's so incredibly based. Character boxes are very simple to use and basically work the same as the regular prompt and undesired content boxes, except you only describe things relevant to that specific character.

My go-to character prompt for porn is:
boy, [name], [series name], mature male, muscular male, topless male, bottomless, pubic hair, girthy penis, foreskin, huge testicles, sagging testicles + pose tags (e.x. standing, looking down) + action tags (more on those in a minute).

If you have multiple characters, you can specify their rough positions by deselecting "AI's Choice" under the bottom character box and then clicking the position box in their character box.

Action tags are used to specify who's doing what and to who and have the following syntax:
source#[action], target#[action], mutual#[action].
For example, for a blowjob, you could write in Character 1's (the top) box:
target#fellatio, source#hands on another's head
…and in Character 2's (the bottom) box:
source#fellatio, target#hands on another's head
To indicate that Character 2 is giving Character 1 a blowjob while Character 1 holds Character 2's head.

And if you wanted two characters kissing each other, you could put mutual#kiss in both of their boxes.

You can add up to 6 character boxes (any more than 3 is likely going to be a challenge for the AI though), and you can toggle characters on and off with the check mark icon or delete them entirely with the trash can icon.

You can also reposition them with the arrow icons; NAI recommends ordering the characters roughly from left to right, top to bottom.

AI Settings

Steps

I use 28, but it doesn't really matter that much.

Prompt Guidance

I use 6, but again, it doesn't really matter that much.

Variety+

I leave this on most of the time (I mostly turn it off when inpainting for specific things) since it can make the compositions and poses more interesting, but it's a matter of personal preference.

Seed

This is a great way to test the effects of different settings or prompts and something you should make use of a lot if you want to learn how to make better generations. Whenever you generate an image, there will be a seed in the bottom right corner that you can click to copy. You can then make adjustments to your settings/prompts and regenerate. Testing how settings/prompts impact generations when using the same seed is the fastest way to improve your understanding of NAI.

Sampler

The default (Euler Ancestral) works best in my experience, although you can play with other samplers to make some interesting artstyles.

Advanced Settings

Prompt Guidance Rescale

I use 0.4, but—you guessed it—it doesn't really matter that much.

Noise Schedule

The default (karras) works fine. Some people like to switch to exponential or polyexponential for inpainting, but I haven't noticed a big difference either way.

Legacy Prompt Conditioning Mode

Generally works best off, although certain styles work better with it on, but overall it tends to have a pretty small impact either way.

Text

I'm just going to copy-paste NAI's official explanation of text in V4 because I don't really use text:

The best way to prompt for text is to include the text, english text tags in your prompt and end it on a period followed by “Text:” and whatever text you want to show up in your image.

If the model struggles with spelling, writing your text in all uppercase can sometimes help. If you want the text in a specific location, you can use natural language to describe where the text should appear, unless your tags otherwise make it obvious (e.g. speech bubble). If you want your characters to say something, you can also try putting “Text:” in the character prompt boxes rather than the base prompt one. As a final bit of advice, it can help to repeat things for the model. Putting it all together, here is an example:

1girl, text, english text, speech bubble, white background. A pastel pink speech bubble with yellow handwritten text says “HAVE FUN!”. Text: HAVE FUN!

Tags

"That's all wonderful, anon, but what tags should I actually use?"

Danbooru tags.

Many people have had success using primarily prose/natural language with minimal tags in V4, and some have even stopped using tags entirely.
Personally, I find I still get the best results using Danbooru tags and supplementing them with prose at the end if I can't find tags for exactly what I want, but it's up to you what you want to do. In general, prose is likely going to be easier but slightly less effective at getting what you want, and both should work fine overall.

As for actually finding what Danbooru tags to use, there are five main methods:

1. From the Tags page

Open the Tags page and type in a concept you want to find tags about with an asterisk (*) on either or both sides, Order by Count, and click Search.

For example, if I'm trying to describe a concept related to balls, if I search *testicle*, the following tags will come up on page 1:
? testicles 95k
? large_testicles 2.5k
? caressing_testicles 2.3k
? huge_testicles 1.7k
? no_testicles 1.7k
? testicle_sucking 1.3k
? covered_testicles 1.2k
? testicle_grab 1.1k
? licking_testicle 772
? testicle_peek 561
? bouncing_testicles 461
? lipstick_mark_on_testicles 372
? sagging_testicles 281
? squeezing_testicles 242
? veiny_testicles 183
? penis_and_testicles_touching 180
? gigantic_testicles 168
? testicles_on_face 165
? small_testicles 157
? testicles_touching 127

Lots of tags for inspiration here, and since we Ordered by Count, we know the further down the list they are, the less likely they are to be recognized by the AI.

2. From the Related Tags page

Open the Related Tags page and type in a tag to see what other tags are commonly tagged alongside it.

The main use I've found for this is finding which characters the AI can recognize. To do that, type in the series' tag at the top and then set the Category to Character.

For example, if I search owari_no_seraph, these are the top characters:
? krul_tepes 342
? hiiragi_shinoa 318
? hyakuya_yuuichirou 193
? hyakuya_mikaela 157
? sanguu_mitsuba 58
? arukanu 47
? ichinose_guren 22

NAI has a decent chance of recognizing any of these characters. The rest either have a very low tag count (<20) and therefore are very unlikely to be recognized, or they have a very low Overlap percentage with the series tag and therefore likely aren't Owari no Seraph characters. (It can help to Order by Overlap and ignore any character with less than 90% or so.)

3. From the Posts page

While you're browsing the Posts page or viewing any post, there will be a search bar you can start typing in that will automatically populate with relevant tags (just be sure to use underscores instead of spaces—spaces function like commas when searching tags on Danbooru). This is basically just method 1 but a bit faster for looking up concepts or proper character tags.

4. From Tag Groups

There are a lot of these.

5. From NAI itself

Start typing into a prompt box in NAI and it will suggest related tags.

Tag Tips

I alluded to this above, but the general thresholds for NAI recognizing characters are:

• <15 taggings: Highly unlikely NAI recognizes them.
• 15-25 taggings: NAI may or may not recognize them. It's also somewhat likely V4 Curated will recognize them at least partially but V4 Full won't (or, rarely, vice versa).
• 25-35: NAI most likely recognizes them at least partially.
• 35+: Highly likely NAI recognizes them.

Note that this isn't guaranteed—I've seen it fail to recognize characters with nearly 100 taggings (e.x. lance (voltron)), while I've heard reports of characters with less than 10 taggings being recognized.

You almost always want to use Danbooru's official tag for a character. You can either search them up on Danbooru or type in the NAI prompt box and see what it recommends. If the AI isn't recognizing a character, make sure you're using the exact right tag. Danbooru tags can be unintuitive and inconsistent at times, so always double-check if you're having problems.

The majority of the time it's a good idea to include both the character's tag and their series tag. It can be especially important for characters with low tag counts, but even some characters with high tag counts confuse the AI if you don't include the series tag. Including the series tag in the character prompt box is usually enough, but in some rare cases including the series tag in the main prompt box can help the AI recognize a character. The main reason to leave out a series tag (other than laziness) is that it can influence style (for example, in the case of kurosaki ichigo vs. kurosaki ichigo, bleach).

Do note that while describing a character's basic features (like hair color and important articles of clothing) can sometimes help NAI recognize a character, in my experience, most of the time (not all) it either "gets" a character or it doesn't, and adding a bunch of tags related to them is probably just going to produce an OC that vaguely resembles them (which may be enough for your purposes).

Other

Enhance

This is a great way to polish an image and fix minor mistakes like bad eyes/fingers/toes. I typically do 1.5x Upscale Amount and 1 Magnitude. 1x Upscale Amount is unlikely to fix mistakes, and 2+ Magnitude often changes the style/faces noticeably, but is better at fixing larger mistakes.

Generate Variations

This is a great way to burn Anlas for no tangible benefit.

Upscale

This makes the image larger and crisper but it often gives it this subtle watery texture that I'm not a fan of (this is a matter of personal preference though) while also not being useful for fixing minor mistakes. I generally prefer to just use low-Magnitude Enhance, and only use Upscale if the image is blurry.

Use as Base Image

When used directly on a generation, this is a very good option to use if you're trying to get a consistent composition/pose found in one of your generations but not all. This can also be great for iterating on complex concepts: if the AI is struggling to understand your concept, you can perform an "Img2Img chain" where you keep selecting "Use as Base Image" each time a new generation gets one step closer to what you're trying to accomplish.

The default Strength of 0.7 is usually good, although if the image you have is very close to what you want, you may want to lower it to 0.6, 0.5, or even 0.4.

0.3 Strength or below is basically just going to spit back out the same image—this can be good for smoothing out seams introduced by inpainting, or getting rid of artifacts or other oddities.

Conversely, 0.8 Strength can spit back images that look nothing like the base image. This can be good if you want to maintain the general composition of an image but otherwise want it to be mixed up.

As for Noise, I leave it at 0. One of the NAI developers in the official Discord has stated a number of times since V4's release that Noise is a no-longer-useful relic and is probably just going to add random splotches of color to your image if anything.

If you're uploading an image to use as a Base Image rather than selecting one you generated, there are about a million uses for that which I'm too lazy to get into.

Edit Image

This lets you draw on a generated image. The point of this is to block out rough colors/shapes and then perform Img2Img. For example, if a character's shirtless and you want them to wear a blue shirt, you can paint their torso blue and then hit "Save & Close." (If you're wondering where your edits went—look at the Img2Img box under the Undesired Content box.)

Inpainting

If you want me to sperg about inpainting let me know and I'll update this later, but it's a complex topic and I'm getting tired and this guide is already 10x longer than it was supposed to be.

As for focused inpainting, I'm still learning how that works myself, so I'm going to direct you to NAI's official documentation for it. What I will say is that it can be very useful if you want to perfect small details (like eyes or hands in a non-close-up image) and Enhancing isn't getting it done.

Director Tools

Remove BG

Not only is this a waste of Anlas, it's not even very good at removing the BG. Just use another online service. (My favorite is pixelcut.ai. It's extremely fast and free and does a better job than NAI on top of that. (Do not use their mobile app though. It's very bad.))

Lineart

Mostly for funsies, not much to say.

Sketch

Same as Lineart.

Colorize

I haven't tried using this on black-and-white images much so I can't say how it works there. For already colored images though, you're probably going to want to crank the Defry up (4 or 5) and describe what you want recolored in the simplest terms possible.

Emotion

Honestly quite bad unless your image is optimized for it (i.e. a close-up of a character directly facing the viewer). The strength adjustment is also quite bad in a lot of scenarios. If you want to change a character's expression you're probably better off copying the seed and tweaking the prompt or inpainting (or focused inpainting) their face.

Declutter

Actually very good at removing text, watermarks, logos, borders, and other elements that pop up on top of an image.

Furries

To my knowledge, if you want to generate furries, put fur dataset at the very beginning of the main prompt and use e621 tags instead of Danbooru tags.

A Note on Styles

Unlike previous versions of NAI, V4 doesn't really have a "default style"—and in my opinion, V4 gens tend to be pretty aesthetically displeasing compared to V3 gens if you're not using good style tags. I highly recommend you experiment and find some styles you like if you want to use V4. These can be artist tags, series tags, medium tags, or a mix of any of those—whatever works for you, the possibilities are endless.

Here are two resources made by others where you can preview the effects of different artist tags to get you started:

NAI V4 Artist Analysis

Danbooru - Artist Tags

Or, you can just wait for Vibe Transfer which should be coming soon and should hopefully fix this problem.