I restored my PATREON (or rather created a new one with the permission of patreon), if someone wants to support I posted there all my new models and LORAs, which were paid on BOOSTY in one archive. I will be glad if you even subscribe to patreon for free, I will know that you care about my work and you want to see updates.
NEW VERSION CrystalNoobaiXL_v2.006 and v2.010 on BOOSTY and PATREON
Comparison with the original v-pred model:
I decided to create my own version of NAI, but only with fixed problems with hands, proportions, eyes, improved perception of anime screenshots, photo backgrounds, improved color scheme, a greater variety of clothes, improved perspective in art and much more, since I experimented with the previous version of NAI, I did not have any serious problems BUT THE MAIN THING IS READ HOW TO USE IT, THIS IS IMPORTANT!!!
I also tried to improve the consistency when outputting a large number of characters (real characters from anime, games, etc.) to reduce the "mess" that the usual NAI produces, and it partially worked
Let's start with the fact that you will need forge or reforge. If the option is through Comfiui, but I do not use it, so I can not tell you. It is better to use reforge, because only it has a very important extension "RescaleCFG for reForge" built in, it removes this terrible contrast that appeared in the v-pred versions of NAI models (or use my Low Contrast LORA on Boosty or Patreon then you don't need to worry about where to use the model)
what the workspace looks like in general (greatly reduced so that everything is visible, including RescaleCFG):
Again, why reforge, and not forge - more samplers, but we only need Euler A Comfy, there is access to RescaleCFG, which greatly cuts the contrast.
Why Euler A Comfy and not just Euler A - it's a little better, tested many times.
Schedule type Align Your Steps GITS - IMPORTANT, USE ONLY IT (you can use other ays, but in general Align Your Steps GITS is a little better), also IMPORTANT Sampling steps 32 (if you use more, the image will become even more contrasty, if less, the detail will decrease, 32 forever)
install the booru tag autocompletion extension to use Booru tags, and restart webui, it makes the process of creating art much easier
Example of using ALL Schedule type, download rar archive, image too big:
negative prompts - I collected almost all bad arts with visual errors, with bad perspective or anatomy in one negative, no need to write words that are not in the dataset, this will only lead to an increase in contrast, in fact this is an ideal negative prompt, you do not need to add anything else, only if you want to remove some specific things (individually):
worst quality, low quality, (censored, bar censor, mosaic censoring), blurry, artistic error, bad anatomy, bad feet, wrong foot, bad hands, bad proportions, bad perspective, bad leg, bad arm, bad neck, bad vulva, bad reflection, bad ass, bad face, english text, chinese text, watermark, multiple views
all these words are present in the dataset and denote bad images to one degree or another
a normal prompt should start with:
masterpiece, best quality, newest, official art, absurdres, highres
you can remove the official art prompt if you get text on images too often, but overall it improves the image quite a bit
you can add very awa prompt - this prompt actually adds some styling that selects the best images according to waifu-scorer (not recommended, too much styling)
Resolution - the model is configured so that it can produce more than the standard 832x1216, and you can use 1024x1536 in simple scenes with simple anatomy, most likely the art will be perfect, BUT, in complex scenes there will be serious problems with anatomy and this approach is suitable for those who do not like to use Hires. fix, but want a higher image quality by default, the higher the initial resolution, the better the details, eyes, but instability increases, the choice is yours.
Example:
in some cases even at this resolution you can use Hires. fix with Latent upscaler (nearest-exact), but Denoising strength should be very high, and at high Denoising strength at this resolution the image instability increases (anatomy deteriorates)
I don't really like using R-ESRGAN 4x+ Anime6B as an upscaler because it basically does nothing but increase contrast and only slightly improves the image by simple upscaling, Latent (nearest-exact) is better, it adds more details and makes the image better.
RescaleCFG and CFG Scale are interrelated, the more CFG Scale you have, the more RescaleCFG should be to even out the contrast.
I use (as an example):
CFG Scale - 3.5
RescaleCFG - 0.55-0.6
How much to upscale the image? - 1.5-1.7, no more, I use 0.73 Denoising strength, at a resolution of 832x1216 it is quite stable. (You can see the working environment image above).
Here you can choose what you like best yourself, I just don't like contrast, someone else likes it better, therefore you can reduce RescaleCFG.
Reforge is installed in the same way as forge or auto1111, everything is simple, it is written on github, but I will copy the information here:
Installing Forge/reForge(Suggested) Clean install. Run cmd (command line) where you want the reforge to be installed and enter the commands one by one
git clone https://github.com/Panchovix/stable-diffusion-webui-reForge.git
cd stable-diffusion-webui-reForge
git checkout main
Then run webui-user.bat (Windows) or webui-user.sh (Linux).
When you want to update:
cd stable-diffusion-webui-reForge
git pull
Errors I know during installation:
1 you entered the commands not one by one, but all at once, then there may be problems during installation, copy them one by one.
2 You have too new or too old a version of Python, you need version 3.10.6 (I have 3.10.12, it works), if you have 3.12, 3.11, it may not install and will not start, delete the new version and install 3.10.6, first delete the venv folder to remove all dependencies and reinstall by running webui-user.bat
I was able to install RescaleCFG in a regular forge, but for some reason it didn't work, if you can get RescaleCFG to work, then installing reforge is not necessary.
Please post your work with or without comments, it will help me improve. Thanks! If you like my work, click on the heart above, I will be pleased :3