Based on Illustrious-XL and NoobAi so please visit that page for a rundown of prompts, tags and other baked in features.
My Generation Settings for V5:
Sampler: Euler A AYS
Sampling Steps: 24
Upscaler settings:
1.5x Upscale
10 Hires steps
Denoise 0.1-0.6
Suggested Dimensions: 832x1216, 1216x832, or 1024x1024
CFG Scale: 4-6
ADetailer along with Anzhc's Face Seg Prototype
Inpaint Denoise: 0.2-0.5
Use Seperate width/height:
1024x1024
Inpaint mask blur: 24
Inpaint only masked padding, pixels: 64
Positive prompt base:
masterpiece, best quality,
Negative prompt base:
worst quality, low quality, bad anatomy, watermark,
Generation Settings for V4/4.1:
CFG Scale: 7-9
Rest same as above.
Generation Settings for V3 and older:
CFG Scale: 4-6
Rest same as above.
Thank you:
@mfcg for help with style training and that dope Pony Paintjob and now Mixed Grill for illustrious.
@PotatCat for helping fix some merging issues on a1111.
@FallenIncursio for getting my rear motivated to get this style made and for PaSanctuary ckpts.
@Anzhc for the AAA(AAAAAAAAAAAAAAAAAAAA) ckpt which made NoobAi v1.0 not look horribly contrasty.
@bluvoll for the assist with training style loras to use in V4.1 and V5 you the champ ;)
Important note: Baked in characters for illustrious are often style burnt so you may need to reduce weight so it doesn't overpower this ckpt's style.
This previously was a merge of a style lora I trained off data from my MorgyMix ckpt. Aim was to get a cleaner and more consistent base style for illustrious since I did not like the default look. V4.1+ is now a multi lora merge since got closer results to what I truly wanted.
The model may vary based on which character prompted so don't forget to play around with tag weights since illustrious is well nicer for that than Pony was.
Check out the samples above for inspiration and/or ideas.
Description
Release version - Merge on NoobAi's "Epsilon-pred 0.5-Version"
FAQ
Comments (13)
Very morgy, such wow (～￣▽￣)～
❤️
Hi, could you help me with something? I would really like to create my own checkpoint. But I honestly don't know how to do it. How do you create them?
I trained a style lora first on data I had then I merged that lora into an existing ckpt, I did not make this fresh. I used this guide below to learn how to merge ckpts together or like this model merge a style lora into one. The merging part is really not hard but finding the right mix takes time and play with numbers which the guide explains well.
https://civitai.com/articles/2370/model-merging-management-how-to-merge-stable-diffusion-models-to-fit-your-style
@Manityro Ok,thx a lot :)
Loving this model! Follows prompts quite nicely and has a great and consistent art style!
Only issue I have with it is that it often colors things weirdly, like a character has white skin but the abs are red, looks like the skin is missing, kind of like the Colossal Titan. Or they have blue nipples, or weird designs around certain body parts.
This only happens sometimes, and only when using 30 steps instead of 60, so it seems like 30 steps aren't quite enough to finish rendering.
For reference I'm using Fooocus, which is a very simplified software, so it may be something to do with my setup. I'm open to suggestions on better software, as long as it's not crazy confusing...
I just use Forge to gen images locally, I have vaguely noticed that midriff being misscolored effect but was on a custom skin tone like a purple Oni in my case so I plan to make a pass of the dataset and see if tags are messed up. Was planning to fix some stuff for a V2 anyway so works out.
Ty for the feedback and glad you enjoy.
@Manityro Damn, I just got Forge up and running and it looks a whole lot better now! Although again, it was probably mostly an issue with my settings on Fooocus. It's still not looking quite as good as your outputs, but it's a start!
I'm really not familiar with how the whole training and whatnot works, but are you planning on updating this once NoobAi-XL updates his model or is there no point on doing that?
@Edenko That will be an easy update since I simply trained a style lora on base illustrious and merged into into NoobAi. So unless anything crazy happens with noobai I should be able to do the same when it updates. Also feel free to yoink the generation settings from my images since they should contain the metadata unless civit makes a mess of it. Just drag an img into the "png info" tab on forge and can make it past all the settings use to make the img.
@Manityro Good to hear!
So I grabbed a few of yours and put them on the "png info" thing, I've gotta say I'm a little confused with the results. The quality seems similar enough so it's not a huge deal, but shouldn't the result be exactly the same, since I'm using the same seed too? Some of them are pretty close, but others are totally different, like the Cloud one. https://i.imgur.com/bx6L9sN.jpeg
Is there anything I'm missing with my settings here? Maybe a VAE/Text Encoder, since that doesn't seem to be selected automatically? https://i.imgur.com/KqH5HJ5.png
@Edenko It seems to be using a different sampling method than what I had, mine is Euler A with Align your steps schedule type. Tbh I got an older version of forge and that is prob why it didn't auto select the correct sampling method.
One last thing the odd body coloring I ain't sure how to fix but adding tanlines to the negative fixes it for me on color skin prompts.
@Manityro Alright all good then! I just wanted to be sure I had the right settings. Thank you very much for the help!