Frustration is gettin' bigger
Bang, bang, bang, pull my Devil Trigger
I really liked the atmosphere of Galbrena, so I decided to try to assemble a very rich LoRa model with a bunch of accompanying details, poses, and even kept the flying flame dah.
There are a lot of tags, so I don't recommend generating full-body images to avoid overflowing the SDXL models' CLIP encoder. Otherwise, this can lead to blurry images and artifacts. Some clarifying tags can be safely omitted to allow for greater scene richness; the features will still likely be preserved by the model.
Recommended weight: 0.7 - 1.0
Main tags: galbrena (wuwa), 1girl, lavender eyes, silver-blue hair, short bob, hair highlights, medium breasts, back tattoo, tacet musical symbol
Default outfit: snake earring, gold jewelry, gold choker, cropped jacket, long sleeves, black gloves, claw ring, feather trim, feather tail (right shoulder), shoulder accessory, right shoulder visible, criss-cross halter, blue crop top, snake buckle belt, black skirt, flame hem, black shorts, thigh strap, ring straps, ammo thigh belt (right), asymmetrical legwear, single-leg pantyhose (right), single thighhigh (left), black stiletto boots, blue boots soles
Half demon form: snake earring, gold jewelry, gold choker, cropped jacket, long sleeves, black gloves, claw ring, feather trim, criss-cross halter, blue crop top, snake buckle belt, black skirt, flame hem, black shorts, thigh strap, ring straps, ammo thigh belt (right), asymmetrical legwear, single-leg pantyhose (right), single thighhigh (left), black stiletto boots, blue boots soles, half demon form, single raven wing (right)
Hydra: spectral_hydra, floating familiar, golden faceplate mask, serpent sigil, ram horns, glowing eyes, fangs, ethereal energy mane, glow, no fur
Bad work:
snake earring - It often results in odd-looking earrings and only rarely works properly. I don't recommend using it, but it's even better to add negative earrings.
back tattoo, tacet musical symbol - It often leads to unpleasant artifacts on the back, it is better to refrain from using it.
When generating frames with hydra, you can omit the golden faceplate mask, serpent sigil, and glowing eyes tags if you're concerned these features might transfer to Galbrena. It's rare, but it does happen.
Optionals: It's difficult to pinpoint specific tags. They're primarily aimed at describing action scenes, poses, and certain interactions. Specific examples can be found in the showcase image metadata.
Description
FAQ
Comments (12)
dude, make smaller LoRAs. 360 MB is insane. Most people make 23-56 MB LoRAs for SDXL
If you set the ND to a lower value (precisely to reduce weight) at values like 16/8, precisely to get the look you want, then a lot of detail is lost. I don't like this.
@Sansenskiy
Thanks for replying, but not generaly true.
@novowels (https://civitai.com/search/models?modelType=LORA&sortBy=models_v9%3AcreatedAt%3Adesc&query=novowels) models are so very good with only 54 MB
https://civitai.com/models/2008564/korwa-granblue-fantasy-sdxl-lora-or-3-outfits-illustrious
https://civitai.com/models/1958186/kama-fategrand-order-sdxl-lora-or-13-outfits-illustrious
https://civitai.com/models/1392242/kamisato-ayaka-genshin-impact-sdxl-lora-or-5-outfits-illustrious
Or @LittleJelly https://civitai.com/user/LittleJelly/models?types=LORA
https://civitai.com/models/1741565/elaina-4-outfits-or-illustrious-or-majo-no-tabitabi
https://civitai.com/models/2012073/kazayama-kizuna-or-illustrious-or-tate-no-yuusha-no-nariagari
etc and many more. Both look very gorgeous and are small. I've seen similarly 24 MB LoRAs with similar level of quality
@TScitai Yes, the models you uploaded are probably not bad, I didn't go into detail or check them, just glanced at them. But the characters are simple, without asymmetries or many small details. Galbrena, for example, has several forms, a pet, an asymmetrical outfit with feathers, a wing, and so on. I don't see any problem with the lore weight, even if it weighs a 1 gb. I don't see a problem in waiting for the download to take 2-3 seconds instead of 1.
@Sansenskiy Well, if you have 2 LoRAS ever that's true. But a lot of people have a lot of LoRA. 1000 times 360 MB is 3.6 TB. 1000 times 23 or 56 MB is only 23 and 56 GB respectively.
While not everyone may be as extreme as my case, same applies:
I got 18038 LoRAs with a total size of 1.7 TB. Imagine, every one would be 360ish MB, that would be about 6.5 TB of disk space. Now 0.4 TB, with 23 Mb and 0.8 TB with 56 MB vs 6.5 TB disk space usage is a huge difference.
As for why, well. LoRAs come and down, people take it down or civitai for whatever reason. So to be sure you have them when you need them, you have to download them locally when available. Found a lot of models I liked/bookmarged either be gone (deleted by owner), just disappear (404) or set to archived/generation only, locking me out of it forever.
So the only save way is to download them localy if you like it (or rather the show) and then at some later point see which one works better/turns out your favorite
@TScitai Well, you have a whole NAS full of lore files, so to speak. I think this is a bit atypical usage :) I have about 100-300 lore files, not counting the ones I train, and they're stored separately from comfy UI.
I don't want to compromise on size, so as not to lose the detail I like. If people really want stripped-down versions, I can release them as an add-on. It's not hard to just train on the same dataset with reduced settings. But I don't want to give up the main ND32+ versions because of that.
I see there are still people who believe this crap that lowering Dim/Alpha reduces the end quality :kekw:
@Sansenskiy imagine there's a checkpoint that's 6-8GB in size and has 10k characters and style in it and it still produces the result with good enough quality...
you dont need more than 16/8 for a simple character
heck even 8/4 is enough to preserve a lot of the details and i mean A LOT...
The LoRA ends up only as good as your dataset. Trash dataset = trash generations... don't blame settings that have nothing to do with it lol
@LittleJelly My experience tells me otherwise. The neural network already has huge problems displaying asymmetrical body parts (say, Galbrena's arm is blue). At least those based on SDXL like NoobAI or Illustrious. Sometimes the neural network doesn't care at all, even though the dataset is perfectly assembled, with no flips or discrepancies anywhere. Even that doesn't help 100%; it'll still draw your arm on the wrong side, or even both. Why be greedy with memory in an already difficult situation?
@Sansenskiy you clearly don't understand that asymmetrical / mismatched stuff is fundamentally a huge issue for the Ai these days. It's not an issue of a dataset or your settings or freaking dim/alpha. It's just how it is and there's nothing you can do about that.
@LittleJelly I was just saying that I understand that this is a pain point for current neurons, and then I said that I am not happy with this and that there is at least some chance that the increased ND It helps, I'll use it. Maybe the translator didn't translate it well, and that's why my point wasn't conveyed correctly
@Sansenskiy but there isn't any chance for it to increase quality. That's not how it works.
This is the best Galbrena lora, so much details. Thank you for creating this.