---zimage base - turbo
20+ step, around 6 cfg
Tested on dpmpp_2m_sde + sgm_uniform. (EULER, Simple is good too)
Main trigger: ruiel woman, (just use woman is good enough)
style trigger: braid, hair over shoulder, hair down,
clothes trigger: pajamas, high-neck dress, white dress,
ps: very happy with this dataset, the characteristic and the background are very flexable.
v5------------
have been training Lora for a few month, and recently upgraded some of my old Lora with better skills and settings. Hope this will have a huge improvement over v1.
Tuned many versions, working best with around 0.9, if passing over 1 will cause blury.
Share some posts :)
hope you like it :D
join the discord if you want I usually post images before updates, you can give me suggestions on there :3
v1---------------------
There is a Loha version, which can make a much closer face like what she looks in the movie, but since it is still a new thing some extensions might not work, and if you just want to easily use and have the style this should be just fine to use.
I trained her all day, but still couldn't find a checkpoint or setting that could remake her just like what she looks in the movie, I tried my best to make the face as close as possible but still not close enough.
A higher weight over 0.7 might get you better results, but on some checkpoints, it might be overfitting. The post I made is on Noospherev3 with cfg 8 and 0.8 weight. Noosphere holding this Lora pretty well,
hope this is good enough :3
give it a try and share some posts with me :D
Description
---zimage base - turbo
20+ step, around 6 cfg
Tested on dpmpp_2m_sde + sgm_uniform. (EULER, Simple is good too)
Main trigger: ruiel woman, (just use woman is good enough)
style trigger: braid, hair over shoulder, hair down,
clothes trigger: pajamas, high-neck dress, white dress,
FAQ
Comments (3)
Am I the only one concerned that one of the suggested trigger words is "skinsuit"?
you know what, it is weird, I thought its kind of like a bodysuit, but all works xD
I thought that because of the size it must be someone else who had made a mistake and that it was actually an SDXL, but no.
I find it curious that it works better on my realistic model than on my 3D models, XD
I just need to try it with my Frozen-themed model.
Note 1: When I write "my model," it's not my model, it's a downloaded model.
Note 2: SDXL models don't work for me (I don't have much disk space, and also because of the generation time, I use SD1.5). But despite the large size, it's worth every MB of the file.