Char - CivArchive (CivitAI Archive)

Char - v2.0

NSFW

### Char-v1.0 by AI_Characters

A 1024 res SD v1.5 trained (not merged) resolution model for various high-quality artstyles.

Huggingface link: https://huggingface.co/ai-characters/Char

#### Disclaimer

I have spent a ???? ton of money (https://imgur.com/a/zdISBEb) on trying to create a high-quality generalised model, free for all.

I have failed.

I don't know what it is. During the last three months I have tried every possible combination of hyperparameters, caption styles, training images, and what have you. I have trained literally hundreds of different models.

Yet each model has issues. No model is how I want it to be.

This model is no exception. It is not what I wanted to release. But I have run out of money to train for now, and the photography part is at least passable, so I am releasing the model now.

Maybe someone can create something better out of it via merging or training on top of it.

I can (mostly) guruantee that (most) photographs will look good. I cannot guruantee that every prompt will look good or even work. I cannot guruantee that any of the other trained tags down below will work well at all. Use those at your own risk.

#### Social Media

KoFi: https://badgen.net/badge/icon/kofi?icon=kofi&label)](https://ko-fi.com/aicharacters)

Instagram: https://badgen.net/badge/instagram/ai_characters/purple?icon=github)](https://www.instagram.com/ai_characters/)

Twitter: https://badgen.net/badge/twitter/ai_characters/blue?icon=github)](https://twitter.com/ai_characters/)

If you want to help me offset some of the big costs involved in creating this model, a KoFi donation would be much appreciated!

Feel free to tag me on Instagram or Twitter with your creations with this model!

## Recommended settings

These are my recommended settings for quick high quality results. These are also the settings on which I tested the model.

- Sampler: Euler_a

- Steps: 50

- CFG: 7

- Resolution: 1024x1024 (no high-res fix)

Some prompts may require some prompt engineering (I much recommend using positive prompts instead of negative prompts) to get rid of biases towards certain ethnicities (e.g. the model loves to create Indian women).

I also recommend inpainting faces, especially for full-length images. It works great.

## Trained tags

Caption style: "(beautiful lighting), POV, (from behind), style, image description"

Examples:

full-length, photograph, woman

beautiful lighting, medium-shot, photograph, woman wearing tshirt, standing in city street

- photograph

- full-length

- medium-shot

- closeup

- longshot

- headshot

- beautiful lighting

- from behind

- silhouette of

- street art of

Use at your own risk (doesn't work very well):

- lk artstyle (Legend of Korra artstyle)

- hg artstyle (realistic 3d renders)

- cw artstyle (StarWars: The Clone Wars artstyle)

- kd artstyle (Ghibli movies artstyle)

- wy artstyle (Makoto Shinkai artstyle)

- tn artstyle (Disney 3d artstyle)

- hq artstyle (high-quality digital art)

- aa artstyle (digital line art)

- dd artstyle (Darkest Dungeon artstyle)

- toc artstyle (Turtle of Canada artstyle (with their permission))

- (skimpy) sg outfit with cape (Supergirl outfit)

- gs outfit with cape (generic superhero outfit)

- Videl wearing sy outfit with cape and helmet (Great Saiyawoman outfit)

- cor (Korra) (doesnt translate well into other artstyles)

- with sd hairstyle (season 1-3)

- with sh hairstyle (season 4)

- wearing df outfit (season 1)

- wearing sc outfit (season 1 winter)

- wearing pb outfit (season 1 probending)

- wearing fs outfit (season 4)

- wearing ek outfit (season 4 earth kingdom)

- riding on Naga

- oyla with ay hairstyle wearing hz outfit (Aloy wearing standard outfit) (doesnt translate well into other artstyles)

- akos wearing sw outfit (Ahsoka wearing season 7 outfit) (doesnt translate well into other artstyles)

- humanoid animal (e.g. humanoid fox)

- woman with pointed ears (elfs)

- half-x half-y hybrid person/animal

- anthropomorphic animal person (e.g. anthropomorphic fox)

- female/male x furry (e.g. female fox furry)

- dragon

- person with feathered/dragon wings

- girl with x ears and y tail (kemonomimi)

- abandoned, flooded, overgrown landscape

- fantasy landscape

- cyberpunk outfit

- futuristic outfit

- futuristic armor

- fantasy armor

- fantasy outfit

- tribal fantasy outfit

- cl outfit (post-apocalyptic outfit)

- giant woman (works best with "hg artstyle") (is biased towards ????? women)

- tiny woman (works best with "hg artstyle") (is biased towards ????? women)

There are technically more trained tags, but they work even less well so I won't write them down here.

## Current shortcomings of the model

- some prompts may require prompt engineering to reduce biases (towards certain ethnicities) and issues inherent in the model

- some prompts may just not work well at all

- there may still be some issues with the 1024 resolution

- negative prompts work very weirdly

- certain prompts may rarely produce an output that may have a watermark or overlay despite me having removed them from all images

- generations at a resolution lower than 1024x1024 will look bad as the model has been entirely trained on 1024, always generate at 1024x1024 (or equivalent resolution, e.g. 512x1536) or higher

- ????? people wont look good (no training data)

Description

All samples are raw txt2img generations with no ControlNet, img2img or other editing applied!

To be done still:

Finish this description
Update Huggingface page
Add more samples
Upload safetensors, etc

Char-v2.0 by AI_Characters

A 512 resolution model trained (not merged) on SD v1.5 for various high-quality artstyles and established characters.

If you would like to support my model creation (I have spent the past 5 months and over 1000€ on model creation to reach this level of quality today) you can donate to me on KoFi: https://ko-fi.com/aicharacters

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/

Changes from v1.0 to v2.0

Everything. It is a completely new and different model. Most of the issues from v1.0 are gone and the quality and likeness is on a whole different level. Forget that v1.0 ever existed. Delete it from your system. V2.0 is the model v1.0 was always meant to be.

V2.0 was trained on a 512 resolution instead of a 1024 resolution, so for optimal interference you have to generate at 512 and high-res fix/img2img to 1024 res. Still, the quality is vastly superior to v1.0 despite the lower trained res. This is because higher res training resulting in higher quality seems to have been a fluke, and the real issue were my training settings and data.

Issues with this model and how to fix them

Some prompts and artstyles are still biased towards certain ethnicities, character looks, or other things, use longer and more detailed positive prompts or negative prompts in this case to get what you want!
Negative prompts are still finicky and I only recommend to use them sparingly and only when needed, they can make a generation worse!
Just like with negative prompts, use prompt weights sparingly
Some prompts may still feature incoherences with the body proportions or other aspects of the image
Some SD v1.5 stuff like Emma Watson and such seem to have vanished from the model, despite the training shouldn't cause that, unfortunately I cannot reverse that at this moment without destroying my model's likeness, use img2img or ControlNet if you really want to do something with that in my model
Not all characters are 100% flexible in the outfits and hairstyles they can wear as well as the artstyles they can be portrayed in, however, you generally can still portray every character in at least some different outfits and artstyles
Not all prompts work well in all artstyles, e.g. giant people seem to work only in renders and digital art pieces, not for example photos

Generally I highly recommend using ControlNet with my model as it will provide the coherency and consistency, while my model can provide the quality and different artstyles.

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.

- Sampler: Euler_a

- Steps: 20

- CFG: 7

- Resolution: 512x512 + using high-res fix to 1024x1024

Recommend way to prompt this model

This model was NOT trained using danbooru tags, so danbooru tags are generally not recommend to be used with this model. This model also generally shouldn't need these huge and long positive and negative prompts to work correctly, and those in fact may make the model perform worse at times!

Generally, the model was captioned the following way:

"(beautiful lighting), POV, (from behind), style, image description"

Examples:

full-length, photo, woman
beautiful lighting, medium-shot, photo, woman wearing tshirt, standing in city street
gh artstyle screencap, fantasy castle upon a hill next to a river

Trained tags

Not all artstyle and outfit tags will have perfect likeness, but they will get close.

Artstyles

photo
digital art
lk artstyle screencap (Legend of Korra artstyle)
hg artstyle render (realistic 3d renders)
cw artstyle render (StarWars: The Clone Wars artstyle)
gh artstyle screencap (Ghibli movies artstyle)
ms artstyle screencap (generic anime artstyle influenced by Makoto Shinkai and SAO)
disney artstyle render (3D Disney artstyle)
hq artstyle digital art (high-quality detailed realistic digital art)
dd artstyle digital art (Darkest Dungeon artstyle)
tc artstyle digital art (Turtle of Canada artstyle (with their permission))
anime screencap (generic animated artstyle)