Shelly/Kelly. OC's by Aeolus06
V1: Requires heavy weight to look accurate, recommend: 1.4
V1.5: Tried to fix weight problems, reverted back to old way of training. Still recommend a weight of 1.1, but should give better results
Might try and find the reason later.
Main Tag:
skaeolAppearance Tags:
hair band, earrings
TO GET ONE OR THE OTHER JUST USE TAGS LIKE:
large/huge breasts
wide hips, large/huge assDescription
FAQ
Comments (6)
I've had similar issues with western styles; Illustrious tends to try to "help" images which can make them more anime or have too much detail...and that obfuscates the original style.
I'm not quite sure what helps out here. Sometimes increasing from 10 Epochs to 12 or 13 bakes the style in. Then the problem becomes not over-baking it. Another thing that has sometimes helped is keeping my dataset smaller.
If you find a good solution, I'd love to hear about it.
True, I had very anime'esque results. I changed training on small datasets to a batch size of 1 with network dim/alpha to 8/8. Which worked amazing on some of the models Ive been training recently, but for some reason not this one.
I reverted back to "classic" civitai training, but im still not really satisfied. Ive been messing around with various settings, but for some reason just cant seem to quite get it right, which is weird, because Ive never really ran into this problem before..
But you are absolutely right when it comes to epochs. Somehow it just works better. Ive tried steering away from 10 and increasing it to like 12. Which very often leads to epoch 12 being overbaked, but either epoch 8,9 or 10 being just right
Also, something ive been doing recently, which might work bad on western style characters is heavily reduce prompting. Before I would prompt things like hair/eye color, hair length etc. However if a character always has the same hairstyle/length/color or the same eyecolor, prompting it is redundant.
Sometimes, however. I feel like the AI/Illustrious has difficulty recognizing these things on (heavily)stylized western/cartoon characters
@Qneez Agree on all points; one thing that I have noticed is that sometimes including official style, style parody, parody in my prompts can sometimes trigger the model to focus more on preserving the original style of the images themselves. Mentioning it because sometimes it has helped.
I think that in the same vein of prompting, one thing I am always conflicted on is whether or not to tag hair styles, colors, etc. out of fear that the model will overtrain on the concept. Also fearful that if I don't tag it, the model will also learn it as an "implied" feature of the LoRA.
I'll see if any of those network DIM/alpha changes help out on my end, thanks for sharing.
@Oddities Yup, thats always the dillemma. Thing I always do now is that when its a complicated hairstyle, or a super unique one, I might tag it. If it just something like long curly hair or a hair bun I dont. Hair color I NEVER add to training anymore. Since you dont really "lose" anything by not doing it.
Even if a char has blonde hair and then sometimes gens with a different hair color, you can always just add "blonde hair" during genning, rather than during training.
@Qneez Might not be your issue, but in my opinion the new trainer that gets auto-enabled on-site made my LoRA noticeably worse. I have also found that including 3-4 reinforcement tags for styles can help a lot (i.e. minimalist, thick lines, cartoon, western).
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.