1/5/25: Added WAI-NSFW-illustrious-SDXL version with new, improved data set, and a list of tags that were in data set.
12/17/24: Added Illustrious XL version.
A LoRA of Princess Jasmine from the 1992 film Aladdin.
Both versions produce a very nice likeness. For the 32dim I recommend a strength of 0.8 and for the 64dim I recommend 0.7 for best results. It likes to default to her iconic blue outfit, but the data set included each of the four outfits she wore in the film.
Like my work? Please consider leaving a tip to help fund future LoRAs or commission a custom LoRA for yourself. Higher quality versions of my LoRAs are also available for a monthly donation: https://ko-fi.com/artfulgenie.
Description
I used a new and improved data set for this one, as well as training it on WAI-NSFW-illustrious-SDXL v8.0 (https://civitai.com/models/827184/wai-nsfw-illustrious-sdxl). Because the WAI-NSFW model was trained at 1360x1360, I trained this Jasmine checkpoint at the same size. I was able to figure out a way to get SDXL to learn clean lines instead of focusing so much on the image artifacts. You should be able to generate large images with clean lines because of this.
You can call up her usual blue outfit or her red outfit using the provided trigger words, or you can dress her up in whatever you want.
FAQ
Comments (7)
For those of us too stupid to figure it out on our own, can you please explain the difference between the 32 and 64 versions, thanks.
Should get more quality from the 64dim. I trained the full SDXL checkpoint then used the kohya tools too rip off both locon loras (they are locon because it also adds in a little more knowledge than the basic Lora as it adds part of the unet into the model).
@artificial_genius I noticed that more people downloaded the 32 instead of the 64 which confused me because, I guessed, that 64 WAS better than the 32 but like everyone else, I wasn't too shy to ask. Thank you for answering and so quickly.
@AntUnderboot for sure, we usually drop the 32dim in front and lots of people just directly grab it. Believe it or not I'm a lot more used to people complaining about the bigger better one and then try to tell me how I can make it smaller lol. Thanks for the question and good luck. The model came out really good this time hehe.
Can you train a Noobai-XL V-pred 1.0 version?
what version of WAI-NSFW XL did you use? Wai keeps getting new versions, some of which can't even be used on civitai....
I've had a lot of success with v9. I'd try a few different ones though. As for training may want to try out runpod (20-25c) or masscompute (30c an hour for a 48gb card and a presetup kohya_ss env, so like you could train in full bf16 with a huge batch, with sdxl even with a size of 1360x1360 i can fit a batch of 8 on my 24gb card, so you could probably batch at like 30+) when i train i use v-pred and it absolutely rips, you start at a lose of under .2 in kohya_ss. I train the full model, aka the dreambooth section, and then use the tools section of the kohya gui to do a finetune-originalmodel=lora at my prefered size. It give me great results. On my 3090 it takes 2h for a basic one and if i had a batch of 30 instead of 8 it would be more like 45m max with vpred and your model would be absolutely cooked. Of course you have to guess there would be a learning curve in setup but I'd imagine it would be way better than civit. I'll be releasing training parameters soon.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.