I've been requesting that as well but Civitai admins won't listen

Seems like the lora version doesn't have as much of a default artstyle baked in, meaning it shouldn't interfere as much with features characteristic for whatever style you're trying to apply. In a way it's closer to the new NoobAI Epsilon-pred 1.1

@tkio Thank you for the answer, I understand now 👍

fr, it's hard to get rid of fisheye image

@kuraig it's not fisheye it's a main style impact with giant heads. look at every single image in the examples

@9868686 oh I see, I'm using style lora so it's not very noticeable to me

Why not just try it one by one? You can use https://github.com/DominikDoom/a1111-sd-webui-tagcomplete to get the hints

You might want to use this: https://civitai.com/models/653604/darker-than-dark

@anon5522 Doesn't work, but thank you anyway :(

I had the same problem and the change of negative prompt helped me:

lowres, (worst quality, bad quality:1.2), bad anatomy, sketch, jpeg artifacts, signature, watermark, old, oldest, censored, bar_censor, (pregnant), chibi, loli, simple background,

I agree. It got me hooked on AI generation again.

I assume this is due to merging in NoobAI EPS v1.1, the base style looks very much like it but more stable from what i've experienced

the versions were also rearranged to show v4 at the front again and v777 behind v2.1#

It's a bit harder to handle, requiring you get more familiar with NoobAI prompting, but if done correctly the ouput is indeed more stable and impressive than the previous version.

@vemn What's the differences between NoobAI and Illustrious prompting? I've been looking at some of the prompt from NoobAI epsilon and I don't see anything that stands out.

@stableskynet171 from my own experience there doesn't seem to be a major difference between illustrious and noobai prompting, its more pony vs noob/illu since with pony, simple prompts can give a good image but with a lot of noob/illu models simple prompts won't yield good results, but thats HIGHLY dependent on the model you use. base noob/illu are not very usable for a noob to get good images out of since they are not trained on aesthetics at all. there are models where it works tho.

Also NoobAI based models support simple natural language prompting, you can combine tags and sentences.

@stableskynet171 one of the biggest difference that I saw reading, according to a guide, illustrious can work better with more detailed natural prompting, and noobai apparently goes more to tags of booru tags and staff like that

@P_Universe other way around, iirc according to the paper onoma released, illustrious v0.1 is not trained on natural language at all. Only their unreleased 1.1 and 2.0 models are trained with natural language.

NoobAI is the model that is trained on natural language and native tags, as the model description CivitAI says

@DraconicDragon loool what, I just read this guide https://civitai.com/articles/8380/tips-for-illustrious-xl-prompting-updates

Claiming that, anyway,I guess I must try again noobai, although I don't remember natural language on its examples tho, anyway, do you know if it still 75 tokens or you can go further?

@P_Universe 

TLDR: illu v0.1 is not trained on NLP, but it can change images. NoobAI is trained on NLP and will actually understand the sentences better. using natural language on any of the two models should be a good thing

Ok i think i see the misunderstanding now, the article says it works with NLP, which all SDXL models do, including Pony actually, but both illustrious and pony don't understand natural language as well since they aren't trained on it. The article shows a section from the illustrious paper which shows the effect of NLP and long prompts but it explicitly mentions that illustrious v0.1 is NOT trained on NLP, the section shown is just showing the effect of what happens when you do use it.
(see "3.1.4 Limited Corpus" in the paper your article mentions > "Furthermore, the v0.1 model has difficulty processing natural language-based prompts, especially longer ones, as it was not well adapted to such formats."
And Table 5 under section "4 Training Setups" )

NoobAI on the other hand is trained on native danbooru and e621 tags as well as natural language captioning (at least they state so in the model description on civitai), it should actually understand simple sentences rather than simply making more sophisticated images (which isnt bad, its good but illustrious won't understand NLP as well as noob)

Regarding the 75 token limit, any sdxl model can go beyond 75 tokens, comfyui/forge/a1111 just split the prompt into 75 token chunks if its longer than 75 tokens so subsequent chunks just have less weight than the first one (they aren't considered as much)

There is also LongCLIP which allows for 248 tokens in a chunk i think but i don't think theres any sdxl model using it

@DraconicDragon first thank you so much to take the time to write all this, it's very helpful, and now I understand but I have a question, you see noobai V-Pred-0.65S-Version seems to be totally different to epsilon pred 1.1 as they seems to have different "guides" to use, Wich one you say is the one that will understand better with natural prompts with some tags?

@P_Universe Purely thinking of "understanding" of the prompts I'd say they are the same. The V-Pred enhances colors, composition and coherency, it might give you more desirable results than Eps pred.
Great article explaining vpred and the "issues" with epsilon pred: https://rentry.org/wtfvpred

I personally am holding off on using noob ai vpred since i want to see the finished version, the main issue i personally have is that its too hard to get good results out of either base model with simple prompting, be it illustrious v0.1 or any of the newer noobai (eps/vpred) versions. They just aren't trained for aesthetics (but both, mainly noobai, have a lot of knowledge making it for me more appealing than any Pony based or illustrious (base) models. I recommend using some style finetune/merge of noobai like this one - NTRmix - or any others that fit your preferences. (theres also hikari noob vpred which seems to be giving a more stable base style)

@DraconicDragon okay then, thank you so much, this will help me a lot

v777, or at least the lora version, seems unusable. v4 is great if only style loras could be applied in an effective way to change the default proportions and such.

It's closer to the new Epsilon 1.1 version of NoobAI.

agree, hard to do img2img, literally the best in composition

i like to train NoobAI 0.5 eps regardless of what happen down the line.
it ensure max compatibility over most of the merge.
for instance my own LORA works wonder with their respective details. i updated them all. i use Konpa Evo Mix tho...

same for me!

-> masterpiece, best quality, amazing quality, very aesthetic, absurdres, newest, scenery, 1girl, 1boy, hatsune miku, (ronald mcdonald), muscular male, , rabbit hole (vocaloid), action, battle, motion lines, comic, text, speech bubble,manga, blood on face, punching , dutch angle, dynamic pose, , masterpiece, best quality, amazing quality, very aesthetic, absurdres, newest, scenery, volumetric lighting

You may need to get an additional addon

@Fatbuns thanks. What addon?

@9868686 https://github.com/new-sankaku/stable-diffusion-webui-simple-manga-maker maybe

This usually isn't up to the model creators, unfortunately. Civitai makes those decisions.