Kawaii Makeup XL Version 2 is released!
Updated to Version 2, check side panel for details -->
Always looking for more dataset images, if you have some that you like that you think fit the style, please contact me :)
This version really likes to add the cheek marks for some reason. This is a valid j-fashion makeup style, but it is over-represented in the lora somehow, despite most of the dataset not having cheek marks. I suspect it's a caption issue; I think I've fixed it for the next version, but we'll see. In the meantime, lower strengths can reduce the effect, as well as inpainting.
As always, if you feel like my work is worthy of your praise, please leave a comment, share an image you generated with this resource, and/or stop by my KoFi page. I'm not in this for the money, but hardware and electricity to maintain this hobby are both expensive, and every little bit helps. https://ko-fi.com/mysticdaedra
Description
Version 2.0
Incremental update
Increased dataset, adjusted captions
50 epochs, 1350 steps
Kawaii makeup remains primary trigger, but "lolita", "eyeliner", "eyelashes", "lipstick", and other makeup terminology is used in captions as well.
As always, I'm constantly in the lookout for additional images for the dataset, if you have any or know of a good place to find high quality ones in this style, please let me know!
I can't wait to see your images! Please share :) Hope you enjoy!
FAQ
Comments (7)
I tried using this in easy diffusion, but got the following error - any ideas?
Could not find text_encoder_2.text.projection in the given object!
I did not train the text encoders, to make training faster and to prevent OOM on my 8gb GPU. TE for LoRAs usually isn't needed, and only marginally improves quality in most cases. Thus I decided I'd leave it out. As for your specific error, Easy Diffusion needs to allow models that don't have the TE trained to run on its software. AFAIK most other Stable Diffusion UIs like SD Next, ComfyUI, and A1111 should work just fine. Works on Forge too, although Forge hasn't been updated in like 2 months.
@MysticDaedra Thank you for the reply! Only easy-diffusion works on my AMD card, but I'm looking forward to trying your work once I upgrade cards!
@heaved_milk909 No problem. FWIW, SD Next does work well on AMD cards. It also has ZLUDA support, if that's something you're interested in. https://github.com/vladmandic/automatic
Any chance of this getting a Pony-specific version? It's hard to make it do much on Pony-family models at the moment, not least because you can't use the "lol" word in onsite prompts.
There's always a chance, but I don't personally use Pony so I'm not very motivated. FYI the LoRA was trained on the word "kawaii" which is not blocked, the word "lolita" doesn't appear anywhere in any of the dataset captions, so it's doubtful adding that would actually do much with the LoRA anyways.
I'm currently working on a Flux version of the model, if that interests you: stay tuned!
@MysticDaedra Thanks for the prompt response, much appreciated! Ok, completely understand, that makes sense. I'll carry on playing around, see what works. Might try it with some of the more anime-leaning XL models too. I use Pony a lot, but it's an open relationship 😊
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.