tl;dr
Positive: 1girl, general, danbooru tags go here, highres, masterpiece, newest
Negative: lowres, worst quality, multiple views, comic, text, signature, watermark
Dataset
Illustrious v0.1 uses the danbooru2023 dataset. Trained at 1024x1024 with danbooru tags separated by commas. Besides danbooru tags it has some custom tags.
The danbooru dataset is just danbooru. It isn't other boorus. It isn't a certain furry booru used by PonyXL. It's just that one site called danbooru so if you plan to use it you better familiarize yourself with danbooru.
Danbooru has a lot of NSFW stuff in it and it's all in this dataset. Some of it is disgusting and the model can generate it sometimes because it just happens. If you're reading this guide because you want to teach minors how to generate drawings you should stop what you're doing right now and go buy them some pencils instead.
If you don't know what a tag means just check the danbooru wiki
Best illustrious model
The official illustrious fine tuned model is Illustrious XL. It sucks.
Don't use it.
Use other fine tunes like Hassaku XL because it's more consistent so if you change artist for example you get similar pose and composition.
So below is what Illustrious XL with 12 steps does
Prompt: close up, 1boy, kanzaki hideri, holding green apple, smug, frown, licking, half-closed eyes, traditional media, watercolor \(medium\), pastel \(medium\), watercolor \(medium\), from side, looking at viewer, forehead, green stile uniform, halftone background, highres, masterpiece
Negative: lowres, worst quality, multiple views, text, watermark, signature, username, wallpaper, dutch angle, vanishing point, comic, anime screenshot, game cg, sketch
And now Hassaku XL with 12 steps, same prompt same seed
This is illustrious XL with 60 steps
Some fine tunes add images/concepts that aren't in the v0.1 model so other models won't have it. Some loras have different results from model to model and some pony loras work in some illustrious models.
My V-pred model/lora doesn't work
If you're using a v-prediction model locally you need something that supports vpred. A1111 doesn't support vpred yet. Download the eps/epsilon version instead
Tag order
person count, character names, rating, general tags, artist, score range based rating, year modifier
person count = 1girl, 1boy, 1other, 2boys, 2girls
character name = not necessary
rating = not necessary, see below
artist = not necessary but probably good idea
general tags = everything else
score = probably a good idea
year modifier = probably not necessary
So for example you are supposed to type it like this
1boy, kinoshita hideyoshi, general, building, itomugi-kun, masterpiece, newest
So that is
person count = 1boy
character name = kinoshita hideyoshi
rating = general
everything else = building
artist = itomugi-kun
score = masterpiece
year modifier = newest
This doesn't matter a lot but changing the order of the tags will change what image is rendered because the thing was trained mainly to render prompts in this order.
btw you obviously can add as many tags as you want and you don't even character and artist tags if you don't want
1boy, building
Best prompt size
So when you type a prompt it's turned into tokens and SDXL can only do 75 tokens. If it's more than 75 it gets broken up into batches of 75 and that can really affect quality
If you want to render something very complex you'll need to inpaint it. civitai doesn't have this feature I think so you need to do it locally. and if you inpaint you want to soft inpaint the masked area instead of the whole image and tag only the things in that area like if you're inpainting the face you just tag the eye color and expression and maybe hair color
Are commas needed in prompts?
The model was trained with commas so yeah they are necessary because the commas are converted into tokens too
so instead of sleeveless striped collared shirt you need the tags sleeveless, striped shirt, collared shirt because striped collared shirt isn't a danbooru tag but striped shirt and collared shirt are.
Name order
Danbooru character names are in japanese order so it's kinoshita hideyoshi NOT hideyoshi kinoshita. If you don't know what that means go watch more anime
Upper case or lower case
Danbooru tags are in lower case so it's masterpiece not Masterpiece, no cap.
Illustrious tags NOT from Danbooru
These special tags added in illustrious v0.1 training don't exist in danbooru
Synonyms
Woman instead of 1girl. Don't use these. Just use 1girl like everyone else.
Quality tags
worst quality = 8% quality score.
bad quality = 20%.
average quality = 60%.
good quality = 92%.
masterpiece = 100%.
What is the score criteria? idk it doesn't say. Two possibilities. First is some AI rated them. Second is these are users ratings who clicked on the upvote in the site. Both sound terrible. But everyone puts worst quality in the negative prompt and masterpiece in the positive prompt so just do that i guess
Newest/Oldest tags
oldest = ~2017 and this physically hurts me
old = ~2019
modern = ~2020
recent = ~2022
newest = ~2023
Underscores in tags
On danbooru tags have underscores like open_mouth. It works without underscores like open mouth but sometimes you need them like +_+.
Emoji \:\) tags
On danbooru some tags are emoji like
:) = happy
:( = sad
:p = tongue
:d = :D because all tags are lower case on danbooru remember
If you use A1111 to prompt then prompts like (1boy:1.5) and [1boy:1girl:5] are special. So (, ), [, ], and : are special.
To use :p you need to write it like \:p
\:\) = :)
btw if you see ? in a prompt that's actually a tag on danbooru not a special thing. ! is a tag in too. v is a tag. w is a tag. seriously just try them sometime
Parentheses in tags
Same rule for parentheses in names of artists, characters, etc.
watercolor \(medium\)
astolfo \(fate\)
Rating tags
Posts on danbooru have a rating that tags it as nsfw or not except the tag isn't nsfw or sfw and it was changed in 2022
general = new SFW for all audiences
sensitive = everything that was rated "safe" before 2022 became this
questionable = sexually suggestive
explicit = adult audiences / NSFW
also on danbooru to search these you type like rating:questionable but I don't think you need the rating part in illustrious
Basically you could use general in your positive prompt and explicit in the negative to avoid rendering NSFW but because general was added in 2022 and illustrious uses the danbooru2023 dataset using general excludes like 99% of the whole thing.
absurdres, highres, lowres tags
Posts on danbooru are tagged according to their size
lowres = 500×500 or smaller
no resolution tag = larger than 500x500 and smaller than 1600x1200
highres = at least 1600×1200
absurdres = at least 3200×2400
incredibly absurdres = any dimension over 10000
Illustrious was trained for 1024x1024 remember so how did they use lowres and absurdres images in training if the size didn't match? idk but they probably stretched or cropped them and that's bad. so if that's bad why everyone uses absurdres?
so I did some tests and basically the tl;dr is don't use them at all because lowres will make your gen worse but absurdres won't make it better and any improvement you see is just the AI being random
also remember you only got 75 tokens so you don't want to waste 2 tokens on a tag that is unpredictable
Prompt writing tips
Avoid negative prompts
Negative prompts are kind of a black box because you don't know how the AI learned a concept so if you tell it to not generate the concept you never know what is actually going to get removed
like some tags can mess the background detail or change the saturation and if you have 30 negative tags you won't know what is doing what
There's a lot of tags in danbooru. Every single thing on danbooru affects your gens a bit. If you tried to exclude every thing you don't want in your gen the prompt would be immense.
Just write better positive prompts instead
Good negative tags
you'll probably want these tags if you are generating a bunch of pics
monochrome and/or greyscale so you don't get black and white images. it's spelled grey with an e on danbooru btw.
multiple views so you don't get the same character in multiple angles
text, watermark, signature, translation request so you don't get text
comic, 4koma, 2koma so you don't get panels
Prompt: 1boy, general, otoko no ko, monochrome, greyscale, multiple views, text, watermark, signature, translation request, comic, 4koma, 2koma, space, holding weapon, drill sidelocks, blue hair, aqua eyes, serious, full body
in case you're wondering yeah the letters look right but they're just placed randomly so it doesn't mean anything. if you put random japanese text in your gens and you can't read it it's probably going to say something you don't want
Use style tags
If you want to make some good gens and not just generic AI slop you need to use style tags
so on danbooru most posts are digital art so digital art is generic art
to change the style you need either an artist tag or a medium tag
if you don't know any artists go to this page type a tag you care about like landscape, artist category order by frequency. It's going to show a number next to the artist. That's how many images they got on the booru. If the number was more than 100 in 2023 you can probably use their style
Prompt: 1boy, full body, fog, cable, building, mountain, outdoors, solo, standing, tree, utility pole|asteroid ill
medium tags are tags like
traditional media
watercolor (medium)
graphite_(medium)
colored pencil (medium)
millipen (medium)
photo (medium)
watercolor (medium)
painting (medium)
marker (medium)
there's also 1980s (style), 1980s (style), 2000s (style), retro artstyle, pc-98 (style), game gc, anime screenshot
ofc the number of images varies on danbooru so the styles with more images are stronger
You don't need to make your gen 100% a style. Like you can use (traditional media:0.5) to make it only half traditional media.
Also you can combine multiple styles in one prompt like combining two artist styles together or two media
Prompt: 1boy, otoko no ko, artist's hand in frame, kirigami, photo \(medium\), pen \(medium\), looking at viewer, full body, ?
Use background/border styles
Danbooru has a lot of background/border tags you can use like
stripped background
argyle background
white background
ornate frame
lace border
So basically the reason your gens look so fake and generic is that 99% of art that gets posted has a white background but AI slop always has ultra detailed interiors/exteriors. it looks fake because a manga artist wants to draw cute anime boys they would never waste so much time drawing furniture
Prompt: 1boy, ciel phantomhive, watercolor \(medium\), averting eyes, ornate border, white background, eyepatch,
Use camera tags
if you want to control the camera
straight-on = front view
from side = side view
from above = top view
from below = tbh crotch view 9 out of 10 times
three-quarter view
isometric
symmetry
vanishing point
fisheye
full body
head out of frame
close-up
Prompt: 1boy, kinoshita hideyoshi, otoko no ko, standing, isometric, chibi, halftone background, grass
Use pose tags
you don't need control net to pose your characters just these tags like these and inpainting
sitting
standing
jumping
kneeling
on back
on side
knee up
leg up
standing on one leg
arms up
arms behind head
arms behind back
hand on another's head
hand on own hip
thumbs up
clenched hand
head tilt
Prompt: 1boy, otoko no ko, wariza, hand on own knee, v, smile, one eye closed, head tilt, from above, fisheye, graphite \(medium\), monochrome
Use effect tags
Some cool effects you can use
blurry foreground
blurry background
depth of field
glitch
bloom
spot color
chromatic aberration
Prompt: 1boy, otoko no ko, wariza, mechanical eye, glitch, chromatic aberration, blurry blurry foreground, blurry background, depth of field, anime screenshot,
Best boy tags
Always use these tags to make your gens better
otoko no ko = femboy
toned = muscular
collarbone = self explanatory
flat chest = male cleavage
shark teeth = ^wwww^
skin fang = self explanatory
groin = V lines
midriff
bare shoulders
fingerless gloves
very short hair
hair over one eye
Bad tags illustrious doesn't understand
Cargo cult tags
Basically if the tag isn't in danbooru and isn't a illustrious one illustrious won't understand it.
score_9, 8k, 4k, hdr aren't tags illustrious understands. Even high quality isn't a tag. Detailed isn't a tag. Many isn't a tag.
That doesn't mean they won't have an effect. Many does work sometimes but that's because multiple is a tag in danbooru I think. But basically assume that if it isn't a tag in danbooru it won't work even if it worked in other models like PonyXL
Low quantity tags
Since danbooru has like a billion images if a tag isn't used a lot illustrious doesn't know what it is but it does a lot better than other models which is why it's great.
Basically if a tag has less than 100 images on danbooru assume it isn't going to work. The more it has the better it will work.
Your favorite artist only has 3 images on danbooru? Too bad. You can't use them as a tag. Go make a lora.
You wish you could render something specific but those dweebs at danbooru don't appreciate it like you do so they never spent hours tagging hundreds of images with the tag so now illustrious doesn't know what you are talking about? Too bad. Create a lora. Better yet maybe you should go to danbooru create an account yourself and go on a tagging rampage so maybe the next model trained on danbooru will have your contributions in it and that benefits everybody.
But if a tag has more than 100 images on danbooru it's probably going to work. Unless those images were tagged after 2023 so it's not in the dataset.
So basically most artists don't work as tags because most artists don't even have 100 posts on their pixiv
But some artists will have even over 1k pics so you can use their style.
They don't post hundreds of screenshots of anime on danbooru so even if an anime exists on danbooru most of its characters probably won't work specially the less popular male characters.
Sometimes a tag works but it doesn't look like what you want because what's posted on danbooru is only what people will draw
Redesigns
Like bridget (guilty gear) has 3500 images but it got a redesign for strive and it's the same tag for both strive and xx. I scoured 2 boorus and the entire pixiv to make a lora. I found like 260 pics of the xx design. This is all pics of og bridget, ever. So like 90% of the pics are the new design. You can actually tell when strive was released just by browsing the booru. Start from the last page and go back a few pages then it's all the new bridget and you're like wait that's it?
Ambiguous tags
Some tags are used a lot but it's not obvious what it means or they are confused with more common tags.
Like contortion is used in all sorts of contortion poses so illustrious can't figure out what contortion images have in common
Danbooru has a tag for halo behind head but you get one over the head if you try to use it because just halo is more common
Conclusion
Now you know how to prompt
Go gen some cute anime boys