Now in DoRA quality!

DoRA (Weight-Decomposed Low-Rank Adaptation) is the latest and greatest, and seems to be all-round better at virtually everything a LoRA can do, especially fine details... and there are a lot of those in the lolita fashion!

This is a new edition of my original Sweet Lolita XL LoRA, feel free to check that one out too! This release was meant to be version 5.0, but I decided that switching to DoRA deserved a new page. For this reason I will continue with the versioning of the original, starting with Version 5.

---

Version 5!

• Converted to DoRA for greatly enhanced details and quality!

• Increased dataset, including for swe3tjsk and sweetop triggers!

• bl0use trigger remains the same for now, will add to this dataset for the next incremental update

• New training settings and procedure, slowing dialing in the best settings for my PC

I changed the ideal strength to 0.7, and will stick with this moving forward, for better multi-LoRA/DoRA compatibility. If it seems a bit weak, you can just bump up the strength closer to 1.0, but 0.7 starting in version 5 should be the same as 1.0 in previous versions.

Also note that while this model does work for anime checkpoints (Pony untested), it seems to require higher than normal strengths, above 1.0, due to being trained on actual photos. YMMV, let me know if you find an anime/cartoon checkpoint that works great with this DoRA!

---

Please leave some feedback, and I look forward to seeing your images! <3

Notes:

1. A weight of 0.7-1.0 seems to work most of the time. Depending on checkpoint, you might reduce or even increase a bit.

2. Trained on my own RTX 3070 8gb! Limitations apply, but I'm very proud of the results!

3. Trigger words: the dataset is totaled and split simultaneously into 4 categories: sweetegl is the entire dataset (minus some blouse-specific images). swe3tjsk are the images from the dataset that include jumperskirts, and also include some blouse images. sweetop are OP (one piece) dresses from the dataset. bl0use is still fledgling, but includes images of blouses. This last trigger has very limited utility still, but will be expanded in a future update.

Note on use of the word "lolita" on CivitAI:

Civitai pretty much bans the "lolita" tag on their website. When uploading images with this DoRA 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 DoRA, feel free to drop some coin my way:

Ko-fi.com/mysticdaedra

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 with OneTrainer: https://github.com/Nerogar/OneTrainer

描述:

Version 6

• Major Update: All triggers except sweetegl have been retired as of now. Their individual datasets are preserved in case I decide to split the training or spin them up as separate DoRAs.

• sweetegl is the only trigger starting in Version 6.

• Added several images to the dataset and spent an inordinate amount of time testing and re-training.

• Postures, poses, and general cohesiveness have been improved due to the inclusion of more worn dresses and less mannequins.

• 0.7-1.0 work great, results vary checkpoint-to-checkpoint but 1.0 seems to be the sweet spot unless other LoRAs/DoRAs are also being loaded.

训练词语: sweetegl