Usage, regardless of underlaying model
(this is my recommended settings, you are free to test other stuff 😉)
Euler a + Beta
40 steps
CFG 4
Clip Skip 2
Beware of the denoising strength during Adetailer et Hires.Fix. Keep it low, like 0.15 max.
Negative prompt: IllusN3g, and optionally "3D" depending on the base model.
Positive prompt:
[whatever you want], Flatline, Flat vector illustration, Black outline, Vivid colors <lora:NJ_Vector_Style:1>
In the showcase, you'll find tests with my own checkpoints and several popular others to see the effect, including:
IXL_AnyLoRA <= for the more simple cute style ❤
AlterBan V3
Quirk V2
WaiNSFWillus
HassakuXL
SmoothMix
iLustMix
Animij <= gives very good results too 🤩
Basically, the simpler and more anime like the model, the "cuter" the effect. But with more "aesthetic" models, included NoobAI, you still get good 2D results :D
PS: it does not work well with a V pred model, sorry! You can still try, but then, be sure to switch to Euler + Automatic and the result will be... kinda psychedelic ^^;
The Story™
So, i really liked this LoRA by MrLoreTaker, but several points bothered me:
The overall hue is a bit too shifted
Strength 2.0 is needed and if you don't RTFM, you can miss it
The LoRA internal architecture was a mess! What The heck is an "sdxl_a1111_compat" LoRA?!? Why is a 16 dim LoRA so heavy? Why do i get some warning/errors when using it sometimes?
So, here is my manual rework on top of his great work:
Remade from scratch LoRA architecture, now a correct 16 dim LoRA and not a MB more.
Adapted weights for hue/saturation. SVD to the rescue. Manual Python and Pytorch stuff all the way, i really hacked into it for fun and learning a bit more this stuff.
Fixed needed strength. Works out of the box at strength 1.0
It took a while to make it cleaner/more to my own taste, Enjoy!
NB: when using this LoRA in a merge (if ever), please credit the original author MrLoreTaker instead of me (if you do things by the rules, otherwise, you aren't reading this anyway isn't it?)
Description
FAQ
Comments (8)
naaw, this style is so cute ♥ tyvm 4 crafting this ☺
Thanks!
in my defense I was a little drunk when I make the lora XDD, GOOD LORA BTW
Keep getting drunk and doing all the other amazing remix of Niji then :D
Protoneon is also one of my favorite of yours and you have so much great models 💕😍🥰
BTW: if you want me to take it down after all, feel free to tell me, i always comply with the sources creators requests :D
Thank you for the great Lora, I like it because the outline is clear, and it's one of my main picks.❤️
Thank you for your kind comment 🥰
Could you go deeper into detail on how you...
"Adapted weights for hue/saturation. SVD to the rescue. Manual Python and Pytorch stuff all the way, i really hacked into it for fun and learning a bit more this stuff."
Any articles or software you could point me toward. I'd love to learn more ;)
For Hue and Saturation, i changed values in the last out layer, following something similar to what supermerger do for checkpoints.
And as software... it was python interactive shell, "import torch", and some tests at intermediate steps when hacking into it (for example, i copied some layers from an other LoRA of mine to fit the target LoRA expected keys, deleted without consideration unexpected keys and renamed some of those, i merged it to one of my checkpoint and re-extracted it (That's the SVD part) and did a cosine similarity merge manually with weight layers).
I didn't really saved everything i did to get there, but if the occasion present itself, i'll take notes to do an article ^^
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.







