Trained on a dataset of barely over 30 images, yet it somehow works pretty well for both her "default" outfit and her swimsuit. It seems to have some degree of baked style, but mixing with other LoRAs alleviates that.
Main trigger word is gbfSig. "mature female" is recommended for anime models.
Strength: 0.8-1.0, depending on how much you mix it with other LoRAs.
Recommended tags for her default outfit:
black hair, blue eyes, long hair, big hair, sidetail, headband, braid, hair beads, feather, jewelry, necklace, brooch, feather earrings, necklace, large breasts, cleavage, bare shoulders, layered dress, frilled dress, red corset, long sleeves, arm belt, bracer, belt, leg armor, thighhighsRecommended tags for her swimsuit:
black hair, blue eyes, very long hair, ponytail, headband, braid, hair beads, jewelry, earrings, necklace, dark blue nails, huge breasts, revealing clothes, dark blue one-piece swimsuit, plunging neckline, cleavage, midriff, navel, single elbow glove, bracelet, white sarong, see-through, loose belt, thigh strapFor her swimsuit, I recommend putting "leaf, feather, arm belt" in negative prompts. You will likely need to play with weights to get the best results.
Samples were generated using Anime Confetti Comrade Mix v3 and MeichiDarkAnimV2_Lust. I recommend Expressiveh as a style LoRA, I find it dramatically improves pretty much every result.
Description
FAQ
Comments (12)
Thanks for the great work.
But I found that the “__” in LORA's name makes the plugin dynamic prompt hang with an error.
<lora:Sig“__”Granblue_Fantasy_PonyXL:1>
I sincerely hope that if you can change the file name and upload it again, I would appreciate it!
Well, I hope that worked, deleting the original file is irreversible. And thank you for the kind comment. It was my first time trying CivitAI's training system, and it turned out far better than I expected it to. She really doesn't have much art to work with, about half a dozen of the images in the dataset were expression variations of her in-game artworks, plus I split the swimsuit ones between full body (from the summer event) and the more detailed, closer zoom cutscene art.
Thank you for your kind and fast handling, changed the name of the file in OS, but the name is still <lora:Sig__Granblue_Fantasy_PonyXL:1> when referencing LORA in webUI, so the problem is not solved yet. I don't know if this is an irreversible thing.
I don't have any experience in making LORA, so I'm really curious how you made a LORA with such a small amount of material for a character, thank you very much for sharing your valuable experience!
@xyz0128jp Oh...that would probably require training it again, I'm sorry about that.
As for how I made it, I think it worked as well as it did because I tried to be specific with tagging the images. Other than that, it was just CivitAI doing its thing.
@mutagene I have realized that I am actually the one who should say sorry. The problem is actually on the dynamic prompt side.
Thank you so much for trying to help me! :D
Thanks again for making a LORA for such a charming character as sig, I hope sig's charm will be known to more people.
@xyz0128jp No worries, it isn't like you caused any real hassle. And yeah, she deserves more, she's one of the least popular female characters in the game.
@mutagene I'm guessing she has a lot of hidden fans, just like you and I. Haha!
EDIT: Per @mutagene, this only affects people who downloaded the model not long after this came out and the issue mentioned below has since been fixed, but I'm leaving it up here just in case someone who did do that but didn't use it (like I did for months) has this same issue where it "doesn't" work.
In case anyone downloads the V1 version of this model, be advised there is a minor issue that will make this LoRA fail if you use Dynamic Prompts, since the LoRA maker mistakenly used a double underscore in the output name, and Dynamic Prompts assumes this denotes a wildcard file.
Fortunately, there is an easy enough fix: Load the model into an online LoRA Metadata Viewer, click "Metadata Editor" at the bottom, look for all cases of Sig__Granblue_Fantasy_PonyXL and change it to Sig_Granblue_Fantasy_PonyXL. Click "Update & Download," and replace the original safetensor with the freshly downloaded one.
You should then be all set to use her in your generations!
That should have been fixed, I reuploaded the file as SigGBFPony,safetensors 2 months ago, after it was brought up very quickly.
Unfortunately there wasn't anything I could do about it at the start since CivitAI automatically assigns filenames based on the submitted title, and I wasn't aware of that issue before.
@mutagene I think I might've been one of the first folks to get it then, because god I've always thought Sig was underused and deserved better, so knowing me I raced to snag it as soon as I saw it. She's a great character.
If that's fixed, then that's good! I guess just in case the issue comes up, hopefully people will see this comment and have an idea of what to do.
Keep up the good work!
@Cult_of_the_Black_Buffalo No worries, man. I'm glad you cared enough to download it in the first place, going out of your way to try to help others like that is honestly really flattering and I'm having a hard time processing it. lol
As far as updating/retraining the LoRA goes--I'll 100% do it when she gets her SSR, and I'll also try to set up more user-friendly triggers for her outfits, like what I tried with my Silva LoRA.
@mutagene God I'd love it if she got an SSR. It looks like they're slowly starting to bring a bunch of the Rs/SRs out of purgatory. Robertina's got one now. Chloe got one recently. Balurga got one (though they paired her with Skull and she didn't really change looks).
There's quite a few chars stuck in R and SR purgatory, and Sig is one of those... but I think the summer stuff might've caught some eyes and brought up demand.
One can only hope!