Trained on the animefull-final model for 9 epochs at a network dim of 128, then resized to a network dim of 64. Contains most of her official alternate outfits, though I had to remove a couple that were causing issues.
As usual, since this was trained on the animeful-pruned model, it will work well with any derivative merge.
The following tags are used:
Character: rosalina, luma, crown
Default outfit: blue dress
Fire Flower: red dress
Halloween: witch dress, witch hat
Biker: bikesuit
Swimwear: swimwear, flower
Chef: chef, apron
Tennis: tenniswear, white pantyhose
Volendam: volendamse dress, white hat
You can freely change her dress color by just putting the color you want before 'dress'. I tried to primarily keep her main attire faithful to the original, which means some of her alternate costumes which had a much smaller dataset to work with aren't 100% accurate.
Description
FAQ
Comments (9)
Nice. I'm interested at how the training data looked to separate the concept and outfits. Do you mind sharing the structure?
Of course, I separated all the different outfits and concepts into their own folders like so: https://imgur.com/a/gmd3Fk5
It's most important to make sure that the repeats are well balanced, I personally keep them around 5 to 10, where the folders with more images have less repeats and those with less have more so they don't get completely drowned out. When training I went through a lot of revisions tweaking the repeats as the dataset for a number of the outfits was quite limited which caused them to overfit and create nasty artifacts when the repeat count was too high. If you're interested in the exact details I'd recommend downloading the "additional networks" extension as that allows you to see all the training parameters.
Apart from that, you need to give all the images a fitting activation tag(s), I'd personally recommend using an existing tag that's already somewhat similar to the outfit so that the model has an easier time learning it. Unique tags can work just fine as well but I'd personally recommend only using them if you have to, as existing tokens already have a lot of relationships with other tokens in the model, which might come in useful.
@dubby thanks for sharing and the detailed response. Surprised by the actual way the folders were named. You didn't even use one called Rosalina 1girl, and the folders were mostly Outfit Rosalina instead of the other way around. I do have additional networks installed so I will take a look at the tags from there.
@psnp The folder names don't matter too much for training unless you're not using any caption files (in which case they will use the keyword after the repeat number), it's primarily just for organization purposes. I believe you're not even required to use a class name either (which is the word after the space) but I just put it there to be safe since I happened to be using regularization images as well (which is the folder with 1 repeat).
@dubby I see did not know that. So that folder with 1 repeat was the regularization images? what kind of images are in that folder?
@psnp 250 images I generated using the animefull model, an earlier version of the LoRA with a low strength and wildcards. From what I understand regularization images help with preventing the model from focusing too much on certain unrelated aspects of your subject during training. This isn't normally much of a problem with larger datasets but in certain cases you might want to use regularization images to diversify your training data more by using extra pictures that look similar enough to your subject. I could be wrong about it though because the documentation is almost non existent but I just ended up liking the version with the regularization folder more so I ended up going with that.
@dubby I was looking into starting to use regularization images but wasn't sure about how to go about it. I'm training an OC with original outfits and I saw your Rosalina and Shiranui LoRAs doing what I was aiming to achieve. Though hearing that the folder names don't matter makes me question that my tagging is the problem holding back my progress.
@dubby definitely a good idea worth blogging in more details since a lot of character LoRAs are popping up
Requesting Peach + Daisy + Rosalina bike suits to be turned into an outfit LoRA
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.