Last week, I was playing around with various LoRa training services. I didn't really expect anything publishable to come from it - afterall, I was just picking some training images, letting CLIP handle the captioning, and using only default settings provided by the service just to see the basic quality and accuracy I could get in doing so. I figured that if it came out even a little bit recognizable, then it might be a service worth paying for once I exhausted any free trial credits.
This one was generated at https://www.lorai.art and, quite surprisingly came out pretty darned good. It's far from perfect, but is as good as some LoRas I see posted on here and since there are no other LoRas for Missa now, I figured I'd put it up here.
It seems to have the most trouble with realistic images - especially those where there's lots of flesh visible on the body - the coloration of the head just seems a smidge off from the rest of the body. It does it okay - but the face tends to be a little "redder" than the body - I suspect because of her complexion (freckles) and the fact that it doesn't seem to carry those into the body at all? I'm not sure - and in a few weeks once I lock in on the trainer I decide to use, I'll surely improve on it.
I also seem to have screwed up the trigger token naming. Now... the LoRa will generate Missa without using a token at all, it does "hold" better in complex scenes if you use the obnoxious token that the system autogenerated for me: 100_c1acabe164da40d1b2f3ef89dde39fd7loratoken <- WTF is that? lol (I'll clean that up when I get ready to rework this, too).
In some cases, too - Missa's hair color can change from it's normal red/ginger color to other colors. I've found that calling her "ginger haired" seems to best keep it on track - though that might differ on other models.
Don't expect a lot of miracles with this - it's my first try and my "free trial" tier didn't allow me to make different epochs and change up other options to refine it. But it DOES look pretty good in the end and I've found through experimenting that a lot of the issues can be overcome with some prompting help.
Like it? Buy me a beer.
Join the XenoEngine Discord to discuss tips and tricks, complimentary Lora's and Embeds, and to post (and see) images not allowed here on Civitai.
If you like this model, please take a moment to leave a review.
描述:
Initial Release,
训练词语: 100_c1acabe164da40d1b2f3ef89dde39fd7loratoken
名称: Lora-MissaX.safetensors
大小 (KB): 4718
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
