For version 2:
This is a totally new version for my Ina Lora, trained with from a reworked dataset. The previous version kept giving her donkey hears for some reason and it was driving me crazy. This version seems a bit less "on model" accurate, but those annoying donkey ears are gone so... yay I guess? I don't know why but Ina has been so much harder to train a Lora for than any other character I've tried at my end so far. It's like Pony just hate her or something. Anyway, this version is still not perfect but it's the best result I got so far.
Suggested parameters:
Prefect Pony XL
Strenght 1
CFG Scale between 7-8
Sampler: Euler a
Steps: 30+
(Adding "Big head" in the negative helps to give her a more accurate look. Otherwise it tends to make her look a bit to "cutesy" to my liking.)
Version 1:
A Lora for Ninomae Ina'nis from Hololive. Features only her first outfit, at least for now.
Seems most accurate with Prefect Pony XL.
Strenght 1.
CFG Scale between 7-8.
Sampler: "Euler a" is fine.
Steps: 30+ (You may go lower if you want but it looks much better with many steps IMO.)
Make sure to remove "tentacles" and put it in the negatives if you don't want them. Also for some reasons the thingies on her hear render as animal ears sometimes, putting "Animal ears" or "Donkey ears" in negative helps to prevent that.
Description
FAQ
Comments (6)
Yeah, I train ina for three times previously and all get bad result, especially because of the tentacles on back are so messy and pollute the general output. Some solution may be use photoshop to remove the tentacles as a part of trainset. You lora is awesome in general, but I think the face can be improved and make the lora less burned. If possible are you willing share your trainset, can also share mine, let me know is anything I can do because I want to train a better version.
I'd gladly share my trainset but unfortunately I didn't keep it. Yeah I just deleted a bunch of stuff on my computer to make some room without backing it up first, I'm just dumb that way... And I didn't upload it either, so it's just lost.
Anyway there was nothing in there that you can find by googling it. So you probably have most of what I used already. I also used a couple of screenshots from her 3D streams. The quality isn't amazing but it's pretty useful to get multiple angles. I do clean up the images a bit in Photoshop too. It seems to help to remove all the unnecessary information around the character when possible.
I try to avoid using fannart unless it's very accurate to the original artwork. My main goal for my Loras is to get the best on-model accuracy above anything else, even if it means overbaking them a bit. Because yes I know my Loras are probably overbaked to other people liking but whenever I try to go lower I greatly lose in details and accuracy, especially the faces and eyes. But going too high isn't better either. I've managed to make a more accurate Ina before but it was so overbaked that it barely responded to prompt, definitely not great...
I'm still trying to find the right balance between overbaking and keeping it on-model as possible. I've only started baking Loras less than a month ago so I still have a lot to learn to say the least. For now I guess I'll just keep experimenting until I find the perfect setting for what I'm going for.
@WeirdRandomGuy if you see my previous lora such as aymae or watame, yeah I also did overbaking at that time. But now I figure out some method for better quality, for example, train lycoris instead of lora, and decrease your learning rate (should be lower than 1e-4, I use 6e-5). Also, actually low-step model get high similairty, overback, in another way, affects the on-model accuracy. You can see my latest mumei lora and there is a xy plot that shows the different steps of lora. At least for my observation, for pony and training anime character lora, (batch size of 5 * 60 steps = ) 300 total steps will get best result. Hope these info helps you!
I am also welcome to discuss any further lora training technical details.
@SharkDreemur I'll try experimenting with lower specs then. But what are the advantages of making a lycoris instead of a lora?
@WeirdRandomGuy you can search yourself, try it and see what will happen.
@WeirdRandomGuy forget to mention, you should also try use came optimizer+rex scheduler, if you have more questions I can give you a link to a creator community discord
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.