Tammy

It's everyone's favorite babysitter, Tammy!

Shoutout to Tekuho, who this Lora is inspired by! Please, check out his awesome work, and shower him with support!
https://www.patreon.com/user?u=6714576

Just add the lora @0.8-1.0 strength to the prompt using

Tekutammy

and optionally,

braces 

That's right, I got braces to gen.... kinda. I recommend genning at a higher base res if you're looking to have them look okay.

If you want Tammy proper, I recommend pairing her with my Tekuho Ultimate lora for a "correct" gen.

I could have just trained her with the style embedded so that you'd have no control over the style and you were forced to have Teku's style for all your gens, but where's the fun in that?

Despite having a 100% homogenous dataset all in the same style (an OC by a single person draws them a same way, who knew?), I was able to successfully isolate Tammy the character from the style she's portrayed in (more info near the end if you're interested). What this means is, you can have Tammy in every flavor you can think of! Throw any kind of style lora at her, and she'll gen with that style 90% of the time! With the added bonus that she still looks like her!

And of course, Tammy isn't complete without braces, luckily this lora comes included with them at no extra cost! An over $9000 dollar value, yours free!

Braces will gen if you prompt for them, but whether they're nightmare fuel, ugly, or acceptable is up to the gacha gods. I've really done everything I could to get them to gen perfectly, but I think this is the current limit of SD 1.5 models and (possibly) AI. Bad at hands, bad at braces. Still, they're okay enough that I can call it a success.

I've found that genning at higher resolutions (before hi-res fix) greatly improves the generation of them.

Speaking of hands... She really really really likes to have one of them up. If it bothers you, I recommend you put "hand up" in the negative prompt. Unfortunately, I don't think there's a lot I can do about this one. She's got her hand/arm up in roughly 70% of the dataset, and while I've tagged it, it will come up unprompted. I'm not entirely sure what's the cause (other than the dataset). It's not really a disasterous thing, but just something to keep in mind. Just think of it like a kid who can't keep from touching their face.

The magic of style separation...

If you have a dataset comprised of images largely from the same artist, generally that also means that the artist's style is gonna come uninvited with your character lora like an annoying person you can't get rid of. Luckily, there's a way to train a character lora from a homogenous dataset. Thanks to a tip from a kind anon on /hdg/ I've learned this arcane art.

First grab a set of candles, grab some salt and then you draw a pentagr....

But seriously.

The trick is to first train a style lora of the style you want to remove, on the model you're going to train the character lora on. After that, you merge that style lora into the checkpoint you just trained on. This will basically embed the lora into the checkpoint (kinda like calling up the lora normally, but the lora is on all the time). If the style is good, merge at 1.0 strength, otherwise merge at the most stable weight. Finally, train your character lora, against the merged lora+checkpoint model. When the lora is all done baking, the style will be subtracted from the model when you gen with the normal un-merged model. Magic! If you're genning sample images when training, they'll be in the style you're trying to remove, so just a heads up if you think something is going wrong.

That's it!

Let me know how you like this one, any problems you have, and any other comments!
Karen model coming soon????