IllustriousXL Explorations 🔦

"I first found out about IllustriousXL a few weeks ago, and I've been having a blast ever since! 🤩 It's an absolute treasure trove, managing to learn almost every character and artist up to 2023. This means you can pull off thousands of characters and styles right out of the box, and with high prompt accuracy, it's a real step up from PonyXL as a base model.

That being said, I've been spending a lot of time exploring it lately and figuring out what it can do. I'll be noting what I’ve managed to find in this article ☕"

Named Races ? :

okay so apparently there are some races that are tagged by name on the dooru database, and depending on much data they got up till 2023 they could triggered by name when prompting with IllustriousXL models ( can also be some what replicated using the franchise name and some of its unique traits ) here are some examples

1. direct race name Example:

" masterpiece, very aesthetic, volumetric lighting, dark, backlighting, (neocoill:0.7), 1girl, hatsune miku , standing, full body, simple background, solo, smile, looking at viewer, contrapposto, hand on own hip,grey background, dark background ," & "harvin"

2. direct race name and traits Example:

3. series name and traits Example:

" 1girl, megumin, full body standing," & "na'vi, alien, colored skin, james_cameron's_avatar,"

and with that being said and done I think be revisiting my alternate races wildcard collection and try to fill it up with possible named races .

Quality Control :

(all test result files attached here in the article )

and no I don't mean it as a production cycle phase, more like the tags you use in your prompt to guide the final gen toward your desired outcome quality wise, I rly wanted to know or see what cumulative effect that quality tags has on the final gen in IllustriousXL based models, so for that I'll be using these two popular checkpoints:

Illustrious-XL SmoothFT as an example of conservative fin-tune of IllustriousXL
NTR MIX V4 as an example of a style enhanced fin-tune of IllustriousXL

1st Quality Positives Tests

for this test I'll be gradually adding positive tags on top of each other with the base prompt " 1girl, rebecca \(cyberpunk\), contrapposto, standing, looking at viewer, dutch angle, holding weapon, handgun, jacket, shorts," /in txt2img /24 steps /Euler a / CFG 7 /no hires fix :

2nd Quality Negatives Tests

this is where most of the magic happens, now with last positive prompt currently sitting at " masterpiece, best quality, 1girl, rebecca \(cyberpunk\), contrapposto, standing, looking at viewer, dutch angle, holding weapon, handgun, jacket, shorts," / now i'll be gradually adding some known quality negatives (restrictive tags excluded like "watermark", "2koma" ect) in the negative prompt /in txt2img /24 steps /Euler a / CFG 7 /no hires fix :

3rd Style Fidelity Tests

up till now now you may have noticed that IL NTR Mix have huge lead on IL SmoothFT quality wise and that because like the base IllustriousXL model IL SmoothFT does not come with an out-of-the-box base style so it may not shine in with basic out of the box prompting like IL NTR Mix or any other noobXL mixes , and it stands for the reason that it doesn't any intention of introducing any alterations (or impurities ) to the base model's pre-existing knowlage of artist styles.
here's a quick test using the last result from the negative test but this time with some iconic artists names included in the prompt with and without positive quality tags :

bonus experiment

like any greedy human being I want to get the best of both worlds, so though why not use an enhanced fine-tune to achieve an appealing composition and over all shape in txt2img then a model with high style fidelity later on in img2img and try to re-anchor it the desired style or vice versa, and here are the results :

so over all choosing what base model to prompt on what quality tags you want to include rly falls back to what you wanna achieve at the moment pixel art, cartoonish or sketch gens might not work well with some quality tags while other more elaborately styled ones might require them that why its important to know the stuff you're using if you have better control over the final outcome.
- make sure try all sorts of basic or wild idea for yourself and form you own take on them as well as to share what you find it might go a long way into helping someone out there of staking a forward and sharing their own findings in return.

Art Styles:

Crafting your own:

you can literately literally prompt down your own styles using by mix and mashing a couple of aesthetic modifiers that the model have managed to pick up from the booru dataset imbuing your gens with a whole different vibe
- examples of modifier styling
- left pic style : "pastel colors, yume kawaii, pale color, colorful,"
- right pic style : "muted color, grey theme, vignetting, static,"
so far the modifiers tags that I have picked up on are :
- abstract, flat color, minimalism, no lineart, green theme, limited palette, yume kawaii, neon palette , pink theme, psychedelic , traditional media, yami kawaii, dark, muted color, grey theme, vignetting, dithering, 8-bit, pixel art,
- you can also make use of this amazing embedding map made by @epiTune to get a visual idea on effects of each style modifier (the map contains explicit content)
and that not even the best part of it as you can also you them in conjunction with artist names to pull off some amazing results
- examples of mixing modifiers and artist's styles
- left pic style : "(by porforever:0.5), acid trip, colorful , psychedelic ,"
- right pic style : "by f7\(eiki\) , limited palette, abstract ,"

Using games and shows names:

Now with most of the booru tags being efficiently learned we can explore wider frontiers further then just using artist's names as as there's a hint of consistency in actual you can capture that in a style
- first lets take a look as the useful tag "style parody" its commonly used in posts replicating an iconic artstyle that's usually related to a franchise, with that in mind I could be used to extract reoccurring art style when paired with a franchise name example "style parody, one piece, oda eiichirou \(style\), megumin"
- another method i used to extract styles from games and Visual Novels is by pairing the tag "game cg," or "official art," with the name of the franchise,
- I compiled those style in a rentry for now

Using artist names:

first notable thing is being able call almost any artist by name to apply their unique art style with to the prompt by simply adding "by TheArtistName" to the prompt
- artists that has less then 70 artworks on danbooru by 2023 a less likely to work
- artists with multiple styles are less likely to work just by adding the name

Basic take on Prompting:

enhancing positive and negative prompts are back
- for my positive I settled on "masterpiece, best quality, highres, absurdres ,"
- as for the negative its "lowres, worst quality, bad quality, bad anatomy, jpeg artifacts, bad composition, 2koma, 4koma, comic, multiple views,"
- these usually does the job without hurting the most art styles .
notably its also trained on Timestamp keyword as mentioned in the paper .
- its rly useful so navigate around evolving artstyles of certain artists with a long posting history like "Ilya Kuvshinov " to try to isolate certain sub styles.
- its also useful to recreate characters in different phases of their life for example "haruno sakura".a screenshot from the model's paper about trained score and year keywords .
as always is the model is more biased toward female characters then male (its natural given the what popular on booru sites) one but honestly it whole lot better compared to other anime base models
this one one is rly interesting apparently it has solid understanding of characters that you can them to cosplay other ones by prompting "chara1, cosplaying chara2"
this is mostly possible on character that has a strong presence in the booru2023 dataset