Faputa - Made in Abyss

The embodiment of value.

How to use:

Add the embed trigger to the prompt if you just wanna boogie and don't care too much about accuracy

FPTA

Add the full prompt if you wanna get closer to the proper character.

(FPTA:1.1), (ears back, floppy ears:1.1), (dark-skinned female:1.2), very dark skin, (fluffy, fuzzy:1.2), (pink inner ear:1.2), (extra arms, 4 arms:1.2) (small nose:1.2), (monster girl:1.2), (thigh fur, leg fur, waist fur, hips fur:1.2), (white legs:1.4), (huge hips, huge thighs:1.1), multiple tails, outie navel

The extra shit is mainly just making up for various shortcomings, and honestly a lot can probably be removed (weights tweaked based on your CFG). A slightly higher weight seems to get the weird layering of ears and stuff on the head a bit more correct with the added caveat that the metal bra shows up if clothes are added to the mix (I'll probably train a version without the bra as well).

If you care about the nitty gritty

The embed gets close (closer than you'd think for a tiny ass embed), but there are some issues present. That being said, most of these issues seem to crop up in other Loras of the character, so I'm calling it a success. Not to mention it's all of 300kb, so...

The fur on the legs is hit or miss on whether or not it goes up to the waist.

It's pretty gacha, and not really something I'd be able to fix. The full prompt seems to mitigate it a little bit though. Mostly it's just placebo...

Extra arms proc about 80% of the time.

I don't think I'll be able to do much about this either, since people have spent so much time optimizing image gen models to NOT generate extra arms and limbs. That being said, when they do show up, it's a real treat to see them acting independently instead of the same position doubled up. Having actually 4 arms is gacha. 3 or 6 is common as well as 4.

The ears stand up rarely.

Every example I've fed the trainer has the ears down, but something during gens make them stand up occasionally. The "floppy" bit of the full prompt is my attempt at nipping it in the bud, and weighting the embed higher really fixes it. Then again, the character's ears do move in the anime, so not all is lost.

Pawpads and other abnormalities can sometimes appear
I think that the models think it's a furry due to the character's proximity to furry attributes, so in some cases it'll give her pawpads on her hand...claw...things. Rarely as well, they'll be discolored.

Again though, we've got a 300kb filesize compared to 300mb. It's essentially free.

Now let's address the elephant in the room...

"Why an embedding??? Aren't Loras better?"

My question to you is; are they? Are they really??? What do you actually know about loras? Likewise, what do you know about embeds except that they're an older training method and old=bad?

Loras are useful for introducing NEW concepts to an existing model. In about 75% of the cases though, most loras aren't introducing new concepts, these already exist in the model in some capacity. Which is what embeds bring out from a model.

But I hear you saying... "But my waifu is unique!!!! She's speshul!". Unfortunately though, she's probably not. If you can get around 40-50% of the likeness of the character from prompting alone ("yeah, that kinda looks like X!"), an embedding will most likely take care of the rest. Pulling yours out of the hat via an embed isn't magic, it's what embeds are good at. This is because even though your waifu might not be present in the model if you called her out by name, the model already has a pretty good grasp of what waifus of all shapes and sizes looks. All the embed does is direct the model on how to use what it knows to poof your waifu into a gen.

"But what about concepts! I can't prompt her getting fucked how I want! You can't this position with an embed." Well... you probably can though. Just like there's a shitload of waifus in the model, there's an equal amount of them getting fucked in every which-way imaginable. Same rule applies as before. There's enough information in that several gigabyte model to infer what you want it to do if you train an embed on it.

Same goes for clothes, expressions, camera angles, compositions, hairstyles, most things actually.

Not to mention, embeds take up several hundred KILOBYTES, compared to several hundred MEGABYTES. Sure, a 200mb lora might not be a lot on its own, but they really fucking add up. They can also be a doozy on VRAM as well if you're not packing. SDXL models are already hefty enough, on top of adding several unoptimized and badly trained loras on top that tank your performance. That's not even getting into badly trained loras that are inflexible and have some super fucked style baked in that you can't change. Embeds even work more flexibly across different models of the same generation, whereas loras are more or less constrained to the model it's trained on.

"Well then if embeds are so good, what are loras even good for?"

Like I said, adding NEW concepts into a model. Styles are a great example. Styles are something that take advantage of how the lora bootstraps a model with new info, putting those dims to good use. If you've got a model that does great humans, but for some reason its idea of an elephant is some medieval interpretation (or it ignores the tag entirely), a lora will do well, as generally there's not a lot to reinforce with an embed to make that elephant appear. Pair it with an embed and it reinforces it even more.

I don't hate loras, far from it. I just hate that it's become a catch-all for training literally ANYTHING on the model that someone can't prompt. "I want my waifu to smile a certain way!" Better train up that 500mb lora! It's rediculous at this point that people are fine with 200mb+ models for something that a 3kb embed can do without any downsides like style bleed.

Basically, give embeds a chance, Jack.

Thanks for attending my schizo Ted Talk rant.

Hopefully you like the embed.