UPDATE: Version 5 has been uploaded!
I've upgraded the model to DoRA, and re-uploaded to its own page: https://civarchive.com/models/645251/sweet-lolita-xl-dora
Please take a look and let me know what you think! Your support, suggestions, and feedback make this all worth it!
This LoRA version will remain up, but will be receiving no more updates moving forward. DoRA is simply superior for this kind of subject. That being said, v4.23 is a strong version, so if your image generation software/service doesn't work with DoRAs yet, feel free to use this one instead!
Version 4:
Interim update, expect version 4.x over the course of the next week or so. I've been adding a ton of new images to the dataset, so the LoRA currently is sitting at around 220mb in size. Lolita fashion is complex and has many details, so a diverse dataset is absolutely mandatory for high quality results.
For specific details on the latest version, see the version details ------->
Please leave some feedback, and I look forward to seeing your images! <3
Notes:
A weight of 1.0 seems to work most of the time. Depending on checkpoint, you might reduce or even increase a bit.
Starting with Version 4.0 all versions will be trained on my 8gb RTX 3070.
New trigger words! Using "lolita" in the prompt tends (depending on model) to be the highest quality method, but I trained the lora with three different triggers that allow you to bypass the CivitAI filter (explained below). These triggers are swe3tjsk for jumper skirts, sweetop for one pieces, and sweetegl for both/either, if you don't care about getting a specific style. sweetegl will always be the largest single trigger dataset for the most diverse and detailed results.
Starting with Version 4.0 there is a bl0use (that is indeed a zero!) trigger. YMMV, but this will hopefully add more detail and diversity to blouses used with jumper skirts, or uncommonly one piece dresses.
Speaking of jumper skirts... my dataset for jumper skirts (that I have rights to...) is very small, so it's difficult to get the LoRA to return those, even when using sweetjsk. Any jumper skirt images y'all can provide that are legal for me to train with would be greatly appreciated! (by legal for me to train with, I mean that you didn't scrub them off the internet... I'm perfectly capable of doing that myself!)
Civitai pretty much bans the "lolita" tag on their website. When uploading images with this LoRA in them (or even just in general), try to leave out the "lolita" tag. If it auto-tags and doesn't let you remove it, I've found that upscaling with "highly detailed" or alternatively scrubbing the EXIF data works. Lolita fashion is NOT CP! It is a legitimate J-fashion, and most lolita clothing is far too expensive for children to afford anyways! An average high-quality brand-new Lolita dress can cost anywhere from $150-500 USD!
I'm doing this for fun and my own use mainly, but if you enjoy this LoRA, feel free to drop some coin my way:
Finally, huge thanks to the folks over on the Civitai, SD.Next, and OneTrainer discords, I couldn't have gotten into AI art and Stable Diffusion without them! You guys rock!
Trained starting with version 4.0 on OneTrainer: https://github.com/Nerogar/OneTrainer
Description
Version 3.5
Added a few images to the dataset
Trained on my PC courtesy of OneTrainer, 18 epochs 738 steps, 896x896 resolution.
Removed references to "lolita" and "dress" in the captions. This seems to have been a net positive, I think all the images are coming out much cleaner, although the skirts still are a bit fuzzy and lack definition; I think this is mainly to due to the small dataset.
Dataset is still pretty small for a fashion style this complex. I look forward to seeing what you guys can come up with though :)