Clothing LoRA of Grea's "Light" version from Granblue Fantasy.
Hi-Res fix HIGHLY recommended.
This LoRA can reproduce Grea without my Grea LoRA (there isn't that much difference - my LoRA helps a bit with hair, horn shape and tail) by using dragon girl, dragon horns, dragon wings, dragon tail,
(Use them in negative should these things appear unwanted)
Main trigger: gl_dress
See info sidebar for main tags per version (v0.2b has more than v0.1a)
Other useful tags:
P: tail ornament
for character with non-fluff tail (unstable)
P/N: hair ornament
for the wing-like hair/horn ornaments if they don't show (might need a few tries to get right. Alternatively put in negative. The tiara should work without these)
N: clothes pull, pulling own clothes
if wrist cloth is pulled instead of attached to bridal gauntlets
N: thigh strap
can appear from using lace
Issues based on my testing:
-
The cloth hanging from wrists is not stable (bit more stable with
hand up
; more stable withspread arms
) -
The cross badge thing on the belly is kinda unstable (mostly outside of upper body/cowboy shots) (worse in v0.1a)
-
Pelvic curtain part is biased towards keeping legs together/crossed in upper body shots if that matters (worse in v0.1a)
There is little playroom to create a thigh gap using some of these tags and maybe with higher weight, v0.2b doesn't seem to have that issue on certain concepts:
-
P:
thigh gap, legs apart
-
N:
legs together, knees together feet apart, crossed legs
-
Any of these tags could help although they shouldn't be needed: toeless footwear, high heels, jewelry, sideless outfit, backless outfit
描述:
First an foremost: it's not overtrained anymore.
Its a bit more stable and better in some details, i think.
ALSO: me goof has used the wrong model for showcase images (epoch 9 instead of the uploaded 8th epoch). There's little to no difference I'd say so i guess its ok.
The cross belly badge thing is now more stable and some outfit details should be too.
The issue with the wrist attached cloth is still there, i don't know how to fix it. I don't know if it's because of the training images, the tags or the training method.
If you are reading this and know how to train a LoRA (or LyCORIS or anything) with a small inconsistent or even smaller but consistent dataset please DM me on the site <3
otherwise i've kinda given up on this LoRA because i don't know what to change anymore.
训练词语: 1girl, gl_dress, white dress, tiara, hair ornament, sideboob, bridal gauntlets, lace, pelvic curtain, white high heels,potential helpful negatives below:,clothes pull, pulling own clothes, wide hips, thick thighs, full body
名称: GreaGBF_dress_outfit_v0.2b_e8r.safetensors
大小 (KB): 44720
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success