Hi everyone, I trained another character, Lora. The character is G-lo, a jello girl from the show Villainous. It was a request, but I also wanted to do it myself.
Now, you can generate images with your favorite jello sidekick hero. No one knows which situation she will get herself into. Share the situations this poor sidekick gets herself in for everyone to see.
I recommend using an ADetailer and Upscaler to get better results .
Socials
I have a DeviantArt account and a Discord server (join here). If you have some transformation ideas and good images, please create a Lora for them. But I mostly do it for fun, so don't expect too much.
G-lo (Villainous): version 4.0
*Use strength 0.7 - 1.0 on the Lora for some good results.
Trigger tags
yellow skin, g_lo, yellow hair, ponytail (hair), yellow eyes, slime girl, goo girl: required
translucent head, translucent body, translucent face, see-through skin: transparency
ssuperhero costume, shirt, short cape, round goggles, domino mask, white turtleneck, blue shorts, shorts, blue gloves, blue domino mask, blue cape, g-lo, orange text "g-lo" on topwear, white waistline, blue topwear, white pattern on topwear, glo topwear, navel: hero costume
tiger costume, tiger ears, tiger tail, shirt, round goggles, (orange goggles:1.2), ponytail \(hair\), domino mask, white turtleneck, orange shorts, (orange domino mask:1.2), g-lo, translucent head, (translucent body), translucent face, see-through skin, blue text "g-lo" on topwear, g-lo, glo topwear, orange gloves, white pattern on topwear, white turtleneck, orange cape, (orange topwear:1.2), short cape, orange cape,: tiger halloween edition
nude, yellow pussy, yellow nipples: for nsfw
G-lo (Villainous): version 3.0
*Use strength 0.6 -1.0 on the Lora for some good results.
Trigger tags
yellow skin, g_lo, yellow hair, ponytail (hair), yellow eyes, slime girl: required
short shorts, navel, halftee, short cape, round googles, domino mask, crew neck, blue shorts, blue gloves, blue domino mask, blue cape, blue top, long sleeves: hero costume
tiger ears, tiger tail,short shorts, navel, halftee, short cape, round googles, orange domino mask, crew neck, orange shorts, orange gloves, orange domino mask, orange cape, orange top, long sleeves: tiger halloween edition
nude: for nsfw
G-lo (Villainous): version 1.0 to 2.0
*Use strength 0.6 -1.0 on the Lora for some good results.
Trigger tags
yellow skin, g_lo, yellow hair, ponytail (hair), yellow eyes, slime girl: required
sleeve halftee, short cape, round googles, domino mask, crew neck, blue shorts, blue gloves, blue domino mask, blue cape, blue topwear, medium sleeves: for her hero clothing
nude: for nsfw
Description
Changed to AdaFactor
Longer and bigger training
FAQ
Comments (12)
10.000 steps?! Jesus... this version is way too overcooked and overtrained especially for the Civitai generator, all my results have the same face with lots of image breaking on the most basic of prompts. The other version was much better if a bit undertrained.
Here's some training advice, you don't want the steps to be too high compared to the amount of images you have. For example, if you have 130 images for reference then you can raise the steps up to 1300, it's like Nr. Of images x10, anything higher beyond the number of images for the steps risks of overcooking the model.
This might be good for external generators but for Civitai? Not so much.
Thank you for your advice. Yeah I was trying some stuff out for difficult concepts it payed up. But for some simple lora's like that it is a overkill.
@MisterTwister69 yea simple characters with simple designs and simple images don't need too many steps.
@BecauseNoOneElseWill I will try other epchs and take the most flexible maybe do 2.5v. But yeah when you char becomes a litteral pudding or mannequin then it will be very difficult with 1000 steps. But I would recommend to pin the model I will be realising a more flexible version soon.
@MisterTwister69 I shared a few lora training tips here. Maybe it could help.
https://civitai.com/articles/7679/need-of-improvement-on-nita
@MisterTwister69 how many reference images do you have for G-lo?
@BecauseNoOneElseWill Thank you. I am always open to suggestions and tips. I will check it out later. I have around 110 images. The difficult thing is the clothing lines visible on her skin.
@MisterTwister69 I can help out with a few tips. Mainly if you want to translate something into a model you gotta have good captioning. For instance if you wanna capture a translucent body just write Translucent Body, or Translucent Skin in the captions, or Slime Girl, stuff like that.
Additionally, I would strongly suggest you separate screenshot images from the episodes and official art from the show/creators using Cartoon_screencap, Screencap and Parody captions, Screencap for obviously the Screencap images abd Parody for the images that are official art with her cartoon design art style.
@MisterTwister69 additionally, this is something you can do for any particular art style or artist, if you have enough pics of a specific art style from a specific artist, just write in a trigger tag in the captions which can be either the name of the artist or whatever you want for the ai to keep it in mind. Cartoon_screencap for instance is in itself a trigger tag to get a more accurate show looking style for a character. Just don't tag that to every pic like the ones who are not screencaps. Like for instance, the-butch-x artist drew lots of G-lo, you wanna copy his style for her? Just add the tag TheButchX to all images from him and boom. You got another style unlocked. You can have a variety of art styles all in one lora if you have a trigger tag for each of them
@MisterTwister69 and as for the steps, for a character like G-lo you should keep her at around between 600 and 1100 steps, but nothing higher than anything between these two values or you risk overcooking it. The number of Epochs is best kept at 10, increasing it adds extra steps but only a few and its pointless to temper with that one. The main ones you wanna temper with if you wanna add a medium or high number of steps are the Train Batch and Num Repeats respectively. Decreasing the Train Batch will add a medium amount of extra steps but you shouldn't have it at anything Below 3, and increasing Num Repeats adds a huge number of steps but again, you shouldn't go past the number of images you have x10. So if you have 110 images, your limit should be 1100 steps. For G-lo and her reference images that are available that I know of I'd say a good sweet spot would be around 800ish steps or so. As for every other setting Below those step increasing numbers, I'd recommend leaving Clip Skip at 1, turning on Shuffle and having the Kepp Tokens at 3
@BecauseNoOneElseWill Thanks for your advice. I just trained the Lora with fewer steps and had some fun generating stuff. I use clip 2 it works better for me. I hope you will have fun with this Lora and generate some cool images!
@MisterTwister69 oh yes, this new version is certainly much better trained. I'm getting much better results. A much better version overall and very malleable while keeping on model. Well done.
Details
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.