Functionally speaking, this model will be worse at just about everything other than 2 specific functions: it makes thick lines about on par with other good models, and the only thing it excels at: making thick, angular eyelashes. I understand that this is confusing, so I will include a short series of questions I think I should answer - obviously feel free to ask more.
You could already get those almond shaped anime eyes with relatively thick eyelashes, and wasn't that good enough?
No. I want them pointy and huge.
Couldn't you just do this with inpainting?
You can inpaint with this to make them even thicker and more angular.
Why would anyone use this?
Well, if they want thick and angular eyelashes at any cost, obviously.
As much as this model is basically a one trick pony, I'm not forcing you (thank god) to use it - so if you want thick and angular eyelashes, I'm your guy, if not, feel free to use an, admittedly, better model made by someone vastly more competent. If anyone has any ideas on how to make the eyelashes even more thick and/or angular, please let me know.
I try to upload images as a normal generation resolution + remacri for sake of ease, but better quality can be achieved through img2img upscaling.
My personal suggestions for settings are as follows:
30 steps
Euler A/simple
CFG 6
1216x832/1152x896/1024x1024 (basic and boring, I know)
V2.0 currently has stronger eyelashes than V3.0Neo, but V3.0Neo is
a. More usable
b. an easier base to build off of
3.1XXL I promise you will have stronger eyelashes.
Happy eyelashing!
Description
Back to life like an EpiPen. This edition is a finetune and not a merge; I said in my last version that I would need to make my own dataset, so I spent a year (this entire last year) making a hybrid set paired with many of my own extremely painfully made synthetics. This model here started off by merging Raehoshi 0.7 x Danliuwei + NovaAnime at 0.3, then I did an unbelievable amount of trial and error in my own custom finetune suite (I cannot use Kohya, I'm so sorry) to find a balance, which ended up being all keys at 2e-6 using adafactor, but I have an alternate version I might work on which knocks out QKV keys. I will update my tool soon so you can play around with it if you want & have 16gb or more of vram. This is the base version, so there will be an XXL coming down the pipeline, I just have to figure out how I want to handle it (I'm going to have to do some LoRA merging, but I might use some custom DoRAs). This could also be a buggy mess, so I'd love to hear any feedback you have. If required, I can spend ANOTHER year fixing it. You may be able to guess I pulled for Aemeath.