This LoRa allows you to use the R-MK style.
YES!!! IT'S MUCH BETTER THAN V2!!!!
I've literally found the holy grail for training Style LoRas; here's the result.
It has a small flaw: it tends to create images where nudity is the main focus. Simply adjusting the clothing a little is enough to fix this.
V3 Version
V2 VERSION
V1 VERSION
Description
-New Settings
FAQ
Comments (8)
goated cook on the V3
The holy grail? What did you do to it?
Well, since Civitai's autotagger wasn't working, I decided to train the LoRa with just 5 tags: byr-mkstyle as the activation tag and R-MK, soft shading, glossy skin, and smooth shading as general tags.
No tags like 1girl or nude, you know, the typical tags.
And let's just say I experimented with settings until I found the right one using AI Toolkit (since Kohya wasn't working).
It replicates the style pretty well, although it has flaws like prioritizing nudity, but compared to all the attempts I've made to improve the R-MK style LoRa, this is the BEST one I've done.
And I've made other LoRas with these settings, but the general tags I mentioned before have a significant impact, and if one doesn't fit the style... well, it looks bad.
@Shockwave_ Maybe just cutting down on the tags helped out? I have seen the tags you use on other loras, and very few come up as danbooru tags in search. I remember hearing once that giving a lora too many unique tags on every image makes it treat them all as activation tags without contributing much.
Speaking of, why do you have two intended activation tags for R-MK?
@Diamond_Letter Yes, I think I also saw what you said about adding too many tags being a problem somewhere.
I usually use between 15-20 tags for each image, not counting the activation tag, although with the tests I've been doing, using fewer works better.
Perhaps it's the training settings that I haven't fully mastered yet.
Well, Valo, a fantastic Civitai user, I've often seen him use two tags: one with "by*insert artist name*" and the artist's name, for example, Joaoppereiraus's LoRa.
Although it depends. I've been retraining Burgerkiss's LoRa, and there's already a Burgerkiss tag in Danbooru, which is causing a conflict.
@Shockwave_ By "unique tags" I was referring to tags that were not in danbooru. They do not have training data outside of the lora, and seem to be treated as activation tags if applied everywhere.
Just having less tags sounds like an interesting idea though. I ought to try that.
As for the "by*[artist]" thing, I want to say that some image site tags things like that. I know that some people use that as a style tag, but I do not know where it came from. Either that or someone just made it up and it worked in generations by virtue of having the artist's name in there. I would not know.
I do not intuitively feel like having two activation tags (one with "by" and one without) in there would do anything, but it looks like Valo stopped doing that so he'd probably know. Maybe there is some niche use case for it like with what you were talking about with conflicts?
@Diamond_Letter Oh, okay, now I understand. The translator isn't reliable, sorry.
As for having fewer tags, I tried making a LoRa (another one from R-MK to test) using only the activation tag, and it came close, but it was missing those details in the skin and lighting.
That's why I kept trying until I got this LoRa.
The first time I saw the "by" tag was on Valo's LoRas. I saw that it worked for him, so I used it too, and since it always works, why change it?
And the problem is that Burgerkiss has a tag on Danbooru. There are four images, and they all have the SubscribeStar logo and name in them, so it's a 50/50 chance that the SubscribeStar logo and name will appear in the image.
@Shockwave_ Interesting.
Also, I recommend the Lama Cleaner extension for Forge's inpaint if you want to remove the logo. It works really well for anything you can not just edit out in Gimp.