So this one I wanted to make, but before I knew anything about lora training DeskGrenade agreed to make this one. He did a good job with it, and while I liked it, it had issues which I aimed to correct. The other reason I wanted to make this is that she's one of my favorites, so here she is.
Just doing retrains lately and v3 I decided to share, threw prodigy in the trash a while ago and perfecting training with adamw, turns out it yields much better results if you know what you're doing. Keyword there is "if".
IL v2, v3 Trained on IL XL 1.1, v2 extended the dataset, v3 full rework and further extended the dataset
Outfit: hairband, earrings, necklace, arabian clothes, tube top, off shoulder, harem pants -no specific outfit trained, but somewhat pieced together her original outfit (or close to it) and style
v2 for some reason it defaults to see-through pants, not really sure why, but it's not too much and fits well imo.
Generated with: Euler a - automatic, 25 steps, CFG 4.
Schedule type: automatic I use No embeddings, NO negatives and NO quality tags at all!
Recommended Resolutions: (SDXL standards) - Pony/IL too
640 x 1536 / 1536 x 640
768 x 1344 / 1344 x 768
832 x 1216 / 1216 x 832
896 x 1152 / 1152 x 896
1024 x 1024
Description
FAQ
Comments (7)
Very well done, i tend to find some Dall-E based characters to look a little uncanny with painted on eyes but this is perfect. Doesn't really need upscaling or face detailing and I don't know whether it's the word 'arabian' in the triggers or what's baked in but i didn't even have to prompt for appropriate looking backgrounds either. Great work on this one!
this isnt even my final form! thank you nonetheless!
The whole dataset was created by me so she's also one of a kind!
@xJollyboyx She truly is, I had a lot of fun with her. Just noticed you have an Ariel model too, I'll definitely be giving her a go soon enough, unless there's a impending update i should wait for?
@NeilSylver i was thinking about it, not sure yet, that model is definetly not up to standards as jasmine is now
@xJollyboyx It does have that slight uncanny look i was talking about in comparison but i think i'll take a crack at her. A little weight lowering with some facial feature tags to tweak it might work for me, and if you ever update her i can show off the improvements anyway :)
Couldn't resist trying out a Jasmine LoRA. I found it pretty flexible and kept the strength at 0.80. Messing with the quality tags also helps dial in how much of a 3D-rendered look comes through. Nice LoRA all around.
exactly how i try to adjust my training settings, to get this effect on any character