Based on the prettiest anime to ever exist.
This was trained purely on screencaps of Bishoujo Senshi Sailor Moon!
Big thanks to https://sailorsoapbox.com for being an amazing resource!
Will do the Manga's style at some point too, because it's pretty in a completely different way. I love Naoko Takeuchi's style.
If you want to support me buy me a coffee!
Weight: 0.8-1
Triggerwords: None.
Notes:
I noticed some color bleeding/streaking in a lot of the images.
I'll retrain this anyhow once I have some more screencaps, maybe it fixes itself.
The second version ended up worse than the first, so I am not going to upload it! So for the near future I hope you can tolerate that it sometimes goofs up.
Feel free to share some of the pictures you create with me and post them here!
Description
FAQ
Comments (23)
If I may ask, is there a reason you didn't train in 128/128 ranks?
Could you elaborate what you mean?
I'm still rather new to this (started a month or so ago), so I usually just use the same settings that generally tend to work for me and don't change too much unless stuff doesn't work, there's probably better settings for sure x)
So to answer your question, probably because I didn't know any better!
@dustystu 128dim is not necessary and a lot of people prefer smaller dimensions for storage reasons. Additionally, larger LoRAs are easier to over bake and occasionally have problems if you use multiple LoRAs at one time with one of them being that many dimensions. @Beelze It's just a setting for when you are training LoRAs. There is no inherent correct choice, so if your outcome looks good then there's no problem.
@Beelze as @thesauceer said, if it looks good, that's the most important thing.
As a rule of thumb, styles are complicated and often need 128 and low learn rate to capture all the details.
Characters are much simpler and only need 16/8 or 32/16 rank.
Concepts can fall anywhere depending on how complex it is, from16/8, up to 64/32.
I haven't tested your Lora yet, but the "overcooked" test is if it can
1. combine with any other Loras
2. compatable with more than 1 model
3. Allows you freedom of prompting. EG, every girl doesn't have blonde hair, blue eyes, and a bmi of 18.
4. If you use weight=.8 or higher, is there artifacts or other abominations. If so, it was probably over-trained.
I'll test it more later and give more feedback if you're interested. Also Come join the civitai discord and say hi. https://discord.gg/mMY5JhPt
Question I never though about until now, why are people so concerned about balancing around 1? Is it just a familiarity thing? I guess having a max of 1 allows for more intervals between fully on and fully off. Just figured I'd ask since you seem pretty familiar with all of this.
Oh alright! I didn't know u were talking about network dim, didn't know there was another term for it x) so I was a bit confused there.
But I generally use 32 because as I said I didn't have many issues with it yet and it was recommended to me as a good startingpoint and the filesize is nice and compact, but thanks for explaining @thesauceer @dustystu!
I tested it with a bunch of characterloras and a few concept/pose loras and it seems to behave quite well, at least in my tests.
I joined the discord, but I doubt I'll be very active on it x) (because there's so many people on there and I'm shy)
@thesauceer
Good question. I'd ask indolentcat, RockerBoo, or hollowstrawberry on the discord.
My theory is like this: having the lora preform well at close to 1.0 gives you the most flexibility in using it. Also I think it's somehow better to "slowly train" over many epochs, rather than do 100 repeats and 4 epochs and have epoch 3 be completely different than e4. You could theoretically do 1 repeat and 100 epochs, but that would be space wasting. So 10-30 repeats and 5-10 epochs gives you a good balance of all factors when training.
Also you can always train more epochs if the Lora is too weak.
@dustystu Fair enough. Thanks for the explanation!
I'm surprised this is my first time seeing a Sailor Moon style LoRA.. it looks good!
I know right? x) I thought this was gonna be one of first things someone posted on here, because it's so popular and aesthetic!
Thanks! The wife is going to love having this one available on our SD server.
What is the trigger word? I'm sorry I'm pretty new to loras
There is no triggerword! When you train a LoRA you can decide if you want to train it with or without a triggerword.
For characters I usually add a triggerword, as what it does is basically ''absorb'' the tags you tell it to, to create a specific thing.
For example if you have Sailor Moon (the character) and train with her you would want the triggerword to absorb her basic traits like haircolor, haircut, eyecolor, etc, you can go even further and make specific triggers for outfits that way too!
This allows for more flexibility with characters. So you can make a triggerword for her schooluniform and her sailor moon outfit in the same LoRA.
I usually train my style LoRAs without a triggerword, because I want it to always be ''active'' once it's in the prompt.
So once you add the LoRA (in this case it would be <lora:smv2-10:1>) to the prompt and it's in your LoRA folder (models/Lora) it's active!
@Beelze thank a lot!
that's a great lora,but why is my picture overly pink
Since you haven't provided a picture, generation info or anything else I can go off, I can't really answer you x)!
If you want I can try and help you but I am no expert either, but I would need something to go off!
That said the show in general loves to use light pink in general. It uses it instead of whites a lot, so maybe that's why. (For example on buildings in the background etc.)
@Beelze Thanks for your answer,i will send pictures,The characters are exquisite, very cute, the only shortcoming is too pink :)
This is so good! I love this art style so much. You're a real one!
That's one of the most awesome things I have ever seen trough my AI amateur path. It works just perfectly.
After months, still the best trained loras I've ever seen.
Thanks, really appreciate reading that even when I don't actively upload stuff anymore. It's nice to know people still enjoy my stuff!
Still one of the most impressive loras I've used, even though I've moved on to SDXL for a while now. Really wish someone would make a XL version of the original style that could rival this
sailor moon romance
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.