About this version
for zimage base
few tag I have that I think is imporatant for this lora: portrait, chinese clothes, bracelet, leggings, standing,
styles tag: spiked bracelet, bun cover, blue outfit with gold trim, blue and gold outfit, skin tight, hands on own hips (hands on hips,) , brown leggings, blue leggings,
Above are true tag that I used in lora training, but with z-image you can also use you make up word include: chunli, qipao, light blue qipao, word that you think should be in, will also do a huge weight on result, (but it will also result sometime base trigger word override lora trigger word)
works with sampler: euler, dpmpp_2m_sde and 3m, DDIM( with DDIM_uniform)
best working with Scheduler sgm_uniform, also works with simple.
CFG: 4-9
feels like lora is underfit but it has been fully pushed for more than 8 hours of training, sdxl normally takes 30min xD. ( if you have any idea why or how to fix it pls tell me)
ps: I don't really know is because this is a DIT model or my training problem it does show the uNet is not working, but it does work, and the trigger word does have effect on the result but it is very weak. So make sure you explain what you want well, or it will be override by the base.
-
-
-
-
-
-
-
-
-
-
-
Pony_version for GoodFitPony---------------------
Specially Tuned For GoodFitPony Checkpoint
Trained 15 different versions and I think this is the best overall one.
This is for the checkpoint I merged (should have better background + better result + better res) it will give a closer and cleaner result to what ingame looks like. And this Lora is specially tuned.
Able to make more different image ratios other than just 1024*1024. With this checkpoint, it should be able to support all different sdxl image ratio
images used Adetailer( highly recommend) + Hires. Fix It can give very close results to the sd1.5 level image quality.
A few prompts you can use with this (don't use it all together sort it yourself ):
"bracelet, bun cover, qipao, chinese cloth, double bun, earrings, hair bun,hair ribbon, jewelry, pantyhose, leggings, pelvic curtain, puffy short sleeves, puffy sleeves, ribbon, "
This version's face looks softer than the other pony_version(choose the one you like)
give it a try, hope you like it :)
share some posts :D
sd3 </3 :'(
Pony_version---------------------
a pony version of this Lora, nearly the same dataset as 1.5 but with few improved pictures, and better tagging for the Lora. should be an improve version overall.
So far best result on 2.5dPony V1.0 (you want close result you need to use this checkpoint)
images used Adetailer( highly recommend)
All train image are on 1:1 ratio, other ratio may give weird result
Same as all my lora, forces on the face, and style. You can change the outfit and hair to the ones you like.
-
A few prompts you can use with this (don't use it all together sort it yourself ):
"bracelet, bun cover, qipao, chinese cloth, double bun, earrings, hair bun,hair ribbon, jewelry, pantyhose, leggings, pelvic curtain, puffy short sleeves, puffy sleeves, ribbon, "
join the discord if you like
share some post :)
sd3 </3 :'(
1.5_version
I made a discord server, it's in my profile, if you want to chat, share some ideas, share some images, or if you want me to do some Lora, feel free to join.
This is the same as the Lilith Lora, 3D and realistic checkpoints are all working well, very large Lora range, but seems like if it is over 1 it will get those in-game light blurry,
the face is very close to the new Street Fighter 6, the high leg, and up leg, are all working.
give it a try, hope you like it :)
share some posts :D