st-AI-le by AI_Characters

CivitAI link: https://civitai.com/models/97393/st-ai-le

If you like what I do feel free to support me on KoFi so that I can offset some of my training costs! https://ko-fi.com/aicharacters

A comprehensive account of my VastAI expenditures can be found here: https://imgur.com/a/nFbyUwh

st-AI-le is an all-in-one model incorporating multiple popular characters and distinct styles within a single ckpt. It was trained using a 512 resolution on the base StableDiffusion 1.5 model, without incorporating merges of any kind. It is thus wholly trained "from scratch" so to speak.

Formerly known as "Char", st-AI-le is basically a whole new model. I spent 6 months and around 4400€ generating hundreds of test models with different hyperparameters and datasets to arrive at this moment in time.

st-AI-le is designed to be able to generate multiple very distinct styles and characters as true to their likeness as possible, while still being diverse and flexible in its output, and offering the ability to transfer those characters into other styles. st-AI-le is by no means perfect. It has its issues. It does not always generate the most diverse output, it has its own biases, it can be incoherent at times, and sometimes it fails to understand a prompt correctly. However, by my own opinion it is currently one of the models with the highest amount of different styles and characters available within a single ckpt, without merging multiple models together, while also still offering good likeness, diversity, flexibility, and lack of bias.

This model also offers concepts not seen in any other model yet, namely a person transforming into an animal ("morphing") and giant/miniature people. Both these concepts however are highly experimental right now and likely will only produce bad output, or at the very least be very dependent on the prompt/style used.

I will continue to improve st-AI-le moving forward.

Recommended settings for interference

These are my recommended settings for quick high quality results. These are also the settings on which I tested the model. I cannot guarantee that settings other than these - or textual inversions/loras/negative prompts for that matter - will work well or at all with my model, as I have not tested them.

Sampler: DPM++ 2M Karras
Steps: 35
CFG: 7
Resolution: 512x512 + using high-res fix to 1024x1024 with a denoising strength of 0.4
Textual Inversion/lora: none
Negative prompt: instagram, pastel, dada, zombie, ugly, surreal, text, watermark, abstract, old, fat, jpeg, black and white, vintage, amateur, film grain, evil, damaged, concept, unfinished, model, cover, clay, figure, toy, pixelated, bad, inexperienced, illogical, random, oversaturated, overexposed, rough, fake, unrealistic, sloppy, artificial, low budget, unprofessional, cropped, out of frame, low-quality, poorly drawn, deformed, bad proportions, malformed, imperfect, unnatural, extra, rushed, weird

For general directions on how to best prompt the model, see the samples at the end of this page.

I also highly recommend using at least ControlNet-OpenPose with this model to remove any potential anatomy incoherences!

In case you generate some beautiful artwork using my model, feel free to tag me with it on Twitter and/or Instagram! https://www.instagram.com/ai_characters/ https://twitter.com/ai_characters/

Trained tags

These are the styles, characters, and concepts, with which this model was trained. Thus they exhibit a different behaviour within my model than they do in the base vanilla 1.5 StableDiffusion model. To prompt a certain style or character or concept, use these tags.

Artstyles

photo
digital art
lok artstyle (Legend of Korra artstyle)
realistic and detailed render (realistic 3d renders)
clonewars artstyle* (StarWars: The Clone Wars artstyle)
ghibli artstyle (Ghibli movies artstyle)
anime artstyle (generic anime artstyle influenced by Makoto Shinkai and SAO)
disney artstyle (modern 3D Disney movie artstyle)
realistic and detailed digital art (high-quality detailed realistic digital art)
darkestdungeon artstyle (Darkest Dungeon artstyle)
turtleofcanada artstyle (Turtle of Canada artstyle (with their permission))

POV

full-length
medium-shot
closeup
longshot
headshot

Misc

beautiful lighting
from behind
silhouette of
street art of
humanoid animal (e.g. humanoid fox)
half-x half-y hybrid person/animal* (e.g. half-turtle half-duck hybrid animal)
anthro animal person* (e.g. anthropomorphic fox)
female/male x furry* (e.g. female fox furry)
dragon
person with feathered/dragon/insect/butterfly/fairy wings
giant woman*
miniature woman*
person morphing into X* (e.g. woman morphing into dragon)

Characters

korra
nausicaä
san
kiki
videl
aloy
ahsoka
sadie sink
emma watson
zendaya
millie bobby brown
maya hawke

Outfits

sweatjacket
supergirl outfit with cape (Supergirl outfit)
vdo outfit with cape and helmet (Great Saiyaman 2 outfit (Videl's outfit))
kr outfit (Korra season 1/3 outfit)
watertribe outfit (Korra season 1 winter outfit)
probending outfit with helmet (Korra probending outfit)
finale outfit (Korra season 4 outfit)
earthkingdom outfit (Korra season 4 earth kingdom outfit)
zerodawn outfit (Aloy default Nora outfit)
mandalore outfit (Ahsoka The Clone Wars season 7 outfit)
nso outfit (Nausicaä)
mononoke outfit (San)
ko outfit (Kiki)

Hairstyles

sidecut hairstyle
ponytail hairstyle
lh hairstyle (Korra season 1-3)
sh hairstyle (Korra season 4)
nora hairstyle (Aloy)
lekku hairstyle (Ahsoka lekku)
nh hairstyle (Nausicaä)
kh hairstyle (Kiki)
snh hairstyle (San)

*experimental, likely will produce bad results

Samples

lok artstyle image of a woman standing in a city street

Negative prompt: lok artstyle, anime, cartoon, digital art, cgi, render, 3d, drawing, sketch, instagram, pastel, dada, zombie, ugly, surreal, text, watermark, abstract, old, fat, jpeg, black and white, vintage, amateur, film grain, evil, damaged, concept, unfinished, model, cover, clay, figure, toy, pixelated, bad, inexperienced, illogical, random, oversaturated, overexposed, rough, fake, unrealistic, sloppy, artificial, low budget, unprofessional, cropped, out of frame, low-quality, poorly drawn, deformed, bad proportions, malformed, imperfect, unnatural, extra, rushed, weird

Steps: 35, Sampler: Euler, CFG scale: 7, Seed: 3771075846, Size: 512x512, Model hash: fd7ce58be7, Model: st-AI-le_st-AI-le_v1.0, Denoising strength: 0.4, Hires resize: 1024x1024, Hires upscaler: R-ESRGAN 4x+

darkestdungeon artstyle, car

Steps: 35, Sampler: Euler, CFG scale: 7, Seed: 3771075844, Size: 512x512, Model hash: fd7ce58be7, Model: st-AI-le_st-AI-le_v1.0, Denoising strength: 0.4, Hires resize: 1024x1024, Hires upscaler: R-ESRGAN 4x+

medium-shot photo of ((korra with lh hairstyle wearing kr outfit)), shot in 4k high-quality with a Fujifilm X-T3 camera with natural lighting and f1.6 bokeh applied

Steps: 35, Sampler: Euler, CFG scale: 7, Seed: 3771075845, Size: 512x512, Model hash: fd7ce58be7, Model: st-AI-le_st-AI-le_v1.0, Denoising strength: 0.4, Hires resize: 1024x1024, Hires upscaler: R-ESRGAN 4x+

medium-shot anime artstyle of young woman with ponytail hairstyle wearing a kimono, beautiful lighting, evening, dusk, traditional japanese festival, lanterns

disney artstyle man

full-length highly detailed and photorealistic intricate digital art in 4K high quality trending on Artstation and CGSociety of young woman standing in a busy medieval market place

Steps: 35, Sampler: Euler, CFG scale: 7, Seed: 3771075847, Size: 512x512, Model hash: fd7ce58be7, Model: st-AI-le_st-AI-le_v1.0, Denoising strength: 0.4, Hires resize: 1024x1024, Hires upscaler: R-ESRGAN 4x+

turtleofcanada artstyle hamburger

clonewars artstyle woman