KNK LumiNAI (LumiNoobai XL) Epsilon
None of the images in the previews were upscaled, detailed, or treated beyond adjusting the sampling steps.
This model born from the idea of get a NoobAI / Illustrious stable model, consistent in their base style but flexible enough to reproduce the native styles trained in Illustrious/Noobai. Including a semi-realistic controlled by the tag "realistic".
This model was trained originally on Illustrious0.1 using the dataset of
Around of 200 random cosplay images (SFW and NSFW)
Some anime images without artist tag for make an stable base style
But!... the resultant model was really bad and unstable haha. Additionally I didn't want waste the excellent work that the Laxhan Lab team did on Noobai.
So I swap the clips and I merge Noobai in certain percentage on this model. Additionally I extracted the Unet (and a small part of the clip) of my model KNK SoftShadding v2 and I did a lot of mixes between LumiNai (This model), Noobai 1.0 - 0.75, and KNK SoftShadding Unet, additionally, I had to train multiple loras for ajust the styles, specially the "realistic" style.
Was a loooong and iterative process but I got this model.
Main Features
A soft base style
Better limps and consistent backgrounds from the box
A semi-realistic style with the tag "realistic"
Non required quality tags on negatives or positive promp
Same features as Noobai1.0 (characters, tags and styles)
Good representation of native characters and styles (I refreshed several times the clip to ensure it)
Smells to lemon
How to use it
In general, it should be used like Noobai so follow their quick guide, remember this model is an Epsilon model, use as any other common xl model.
Additionally, you don't need any negatives, but It much less opinionated than HelioBlend, so I like add this negatives:
> worst quality, low quality, simple background
Reduce the character tag name for make easier change their original style.
Fun facts
The Noobai team included Cosplay sets on the model in ver 1.0, and the most of them had watermarks xd, the most was related with thighhighs. I fixed the mayor problem due I removed all the watermarks from the dataset.
I don't play Genshin Impact but after ChatGPT sugest a nearly name (something like lumiNust) I liked how it sound it and was changed to LumiNAI, and searching that name on google I realized that almost the same name of Lumine xd. In general I don't like use non original characters as main image in my models, but I liked the result of this one.
Today was the release of Noobai 1.1 and I'll need do the re-merge again ; u;
License
This model follows the Noobai Licence so check their model to see the updated one.
Description
FAQ
Comments (40)
This model looks great and a lot of fun! I just have one question, though. Does this model know the same characters and artists that NoobAi/Illustrious knows?
Yes it does. Maybe could be a rare character/style hard to get in Noobai could be a little harder to get in this model.
Illustrious in all its Glory. I forgot Pony even exists since I first tried KNK LumiNAI. Absolutely stunning results. Impressive to see how well Illustrious "understands" understand complex prompts. Insanely good inpaint abilities too.
Definitely a great fine tuned Illustrious model. Thanks for the share ;)
Thank you so much for your review, I really glad you like it!. I thought one of the most power-full feature of this models was something I didn't tried on my tests. It works rlly well with natural language (thought it is due the Pony Unet used) and mixing it with tags you can get interesting compositions and untagged concepts. I'm focusing in improve this feature using other SDXL models. The problem, all of them are vanilla SDXL trained models and mess EVERYTHING. Adds unwanted noise, realism, overwrite styles, etc. Is just a nightmare. I have worked on it since the first release, and I hope finish it on a few weeks more.
50+ steps? Why? :o
Bruhhh This Model is FANTASTIC. Easily the best Illustrious model by a mile. It has handled everything I've thrown at it with ease and anyone who knows me knows how much I love variety
IIs it possible to create the same soft style that is used by default in the KNK SoftShading model? If so, what tags in prompt should I use?
Anyway, thanks for the great models!
Hi, I glad you liked my work!. Actually when I started with this model, that was my first objective (make SoftShading with base Illustrious), but wasn't possible just extracting Unet. Since last year I 've been working on a unbiased model for use as base for those stylized models,: When it gets ready, I'll get back on a SoftShading Illus version.
I see. I found a workaround for myself, it's combining authors with fine-tuning like (authorName:1.0123) from these tips, but unfortunately it's not that deep. I'll be waiting for your future models and wish you luck
@kekus008 I found a similar one in a Chinese page. After some tests, I found the loras created with synthetic data from mi models actually works pretty well. I'll make it in the next week. I'll need make some gens for it and that take some time due Pony rlly had problems with hands and limbs in comparison with Ill.
Sup, I published a lora made with some gens from SoftShading. Works well in general but I think it needs more data for represent better the style and other features. If you want share some of your gens with the original model for improve it. I'll rlly appreciate it!
https://civitai.com/models/1175783/knk-original-styles-illustrious?modelVersionId=1322974
Hey, I was wondering if there are any plans to update this model? Has there been any additional work done? I really am in love with this checkpoint. It's very powerful, but I do feel like some tweaks could be made and udpated!
Hi!, I'm always searching for improve my models. These months I've worked on my model based on Animagine 4 but it looks like a loss of time (Complex poses, hands and other features are rlly messed, it works much better after several fine-tunnings and in general looks great. But illustrious based models are just much easier to train and gives better results).
Now, talking about LumiNai. I've tried to update few pre-models with Illustrious1.0 (I haven't tried with illust1.1), but I don't know what I'm trying to fix xd. I mean, in my tests, the original Luminai gives me better results than my tests.
So I would love to know which issues you have found and your toughs about what tweaks could improve it. Feel you free to send me a DM.
The model is fantastic. It always perfectly follows my instructions and makes results even unpredictably better. Like literally a n y t h i n g anime-ish I ask – it makes. It understands detailed instructions and makes even non-popular characters. Emotions, poses, clothes, backgrounds, settings, even mixing genres (e.g. Sci-Fi + succubus, Post-Apoc + Psychedelic, etc)
I use booru-style prompting (searching tags' examples on Danbooru tag list if needed), DPM++ 2M, Karras, CFG 7.
I wanted to find something similarly good, but other styles. Unfortunately, I couldn't yet.
Made tons of wallpapers for myself with it.
I glad you like it!. Actually I have another model named KNK NoobiComp, it's a derived version of this model that is more unbiased and makes easier represent the inherit or LoRa styles.
same. my first IL model and it awesome! ty author ))
V2 plz
Hi!, I already have a kind of "release candidate" but it feels like the same model. I'm not sure what kind of fixes this model could rlly needs. All my tries of include new illustrious models, just gives me worst results than the original model. So Idk what I'm trying to fix haha xd. Could you share with me your thoughs or opinion about it?
@Konoko o i dont know i just want more haha
Love the style, at least of the creations shown by people. Sadly I can't really get it to work... Are there recommendations on settings, dimensions and such for A1111?
It should work without any special configuration on A1111, dimentions you can start with 1024x1024, cfg 5 and 20 steps, the sampler you can use Euler or Euler A. If you can make work any xl model, this one should will work with the same configurations. PD: sorry for the late answer
Любимая модель. Спасибо за труд.
Я рад, что тебе понравилось!
Could you possibly make a 3d checkpoint with your methods for Illustious?
Hi!, I'm sorry but by the moment I don't have plans of make a 3D checkpoint. However, I recommend you use my model noobiComp. It's a derivated model that were created from the same base as this model. And just using a lora style, (compared with this one) you can get much easier any kind of style that you prefer.
@Konoko ok ill give it a try thanks
Can you upload this checkpoint to Tensor.Art too? Please?
I love how this looks. So soft and pleasant to the eye, the results are really pleasing and soothing.
I glad you like it c:
new updates bro?
Sup, actually yes. I'm working on a new (re-trained) version and I have plans for release the base version (no style loras) and the full one (with the loras that stabilize the model). I'm still doing few tests, due I've found some of those loras are burned. I'm not really sure if this new model is better than the first one but I'll try to improve it.
@Konoko nice bro! good job!
@Konoko i need to say, this is best model what i found ever, very good job
I have a LOT of checkpoints, many much newer than this one, I bounce around a lot... but I always seem to find myself back on this one, it's my #1 go-to model for well over a year now. It looks incredible, it accurately handles a huge amount of what I throw at it. I come back here it seems every few months to see if there is a newer version, and I'm excited to see you have plans for just that.
In another comment you asked for suggestions on how to improve. I have EXTENSIVELY tested this model, so I feel I have decent feedback to share, though my knowledge of what is possible is limited as I have not created a lora/checkpoint myself. In an improved version, the style should not change much, if at all. Instead, I would suggest focusing on improving versatility of subjects, building recognized characters, objects, scenes, actions, etc into the model. Accuracy of detail would also be nice, though this is already quite good; accuracy of fingers, eyes, limbs, poses, etc. - if these things can be improved upon, that would be nice. I guess, if it's possible, making it faster?
Again, the model is my favorite already, so please don't take these suggestions as criticisms. Instead, please accept this comment as the highest form of complement I can humbly offer.
KNK Luminai is so perfect as it is, it's gonna be hard for an update to make it better. It generates the cutest body shapes, by far. I tried many! IL trained checkpoints, or merged, a lot of so called excellent ones. And after a few images generated, I come back to KNKLuminai. Perfect bodies, insane (for some reason) ability at intricate multiple bodies scenes, great at accepting LoRAs, it understands ages very well, etc. I just love it. But hey, why not, I'm curious to see what an update could bring?
@javbukkake I'm actually still struggling whit it haha, In fact the multi-character scenes is a problem in this new one. but the loras seems to work a kinda better (NoobiComp actually does that pretty well and is an early version of Luminai), for that isn't enough to say "it's better". I'm sorry for make you wait guys, but I don't want release just for release something.
bro, are you still working on this model?
Yes I am. I just go a bit crazy with Anima and I was doing few LoRAs for it to see if the regular styles used as base of KNK Luminai makes a similar results on Anima. In my early tests, they did it. I promised the SDXL version so I will not abandon it. But maybe Anima could give better results without extra hardware requirements.
any updates soon?
how are you bro? everything is good?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.