Inspired by Image on civitai.
Using Incantation Extentions.
This extension implements multiple novel algorithms that enhance image quality, prompt following.
and separate two different subject.
So, I'll try two charactrers .
Mabye 4 character is later.
First attempt.
promnpt :
wide angle shot, wide_shot, long shot, (2girls:1.2), full body, side by side, look at viewer,
__hair_color__, __hairstyle__, __color__ bodysuit, globes, boots,
muscular body, abs,
Decided pose, dynamic pose, BREAK
__hair_color__, __hairstyle__, white shirt, (upskirt, wind_lift, panties), (highheels),
curvy body, huge breast, narrow waist, huge pelvic,
battoujutsu stance, Fighting Pose,
simple background, strong wind,
(realistic, detailed, best quality:1.2), QuickHands,<lora:QuickHandV2:0.6>
negative :
(sandals:1.4), mutated, extra arms, (worst quality, low quality:1.4), (Watermark, Tex,signature),
without Incantation Extentions.
Oh?...! Not bad? haha;;
add prompt : outside, (strong wind:1.4), storm,
negative : (nude, breast out, nsfw),
Now on Incantation Extentions.
Incantation Extentions setting :
PAG scale : 5
CbS Correction Strength : 0.5
Alpha for Cross-Token Non-Maximum Suppression : 0.8
Multi-Concept T2I-Zero , Append Generated Caption ON
Seek for Incantations active
Append Generated Caption check
(others default)
?
?
???
Quality even worse and why 8pic ? I set 4 pic.
See the console
Um...I don't even use extensions that modify the CFGDenoiser.
Activate one by one
The first, only "Perturbed Attention Guidance" Active
Same error, but show 4 pic.
The second, only "Multi-Concept T2I-Zero" active
No console error but didn't separate two girls.
Changing prompt order and adjust weight.
[ (2girls:1.2), (full_body:1.2), side by side, look at viewer, wide angle shot, wide_shot, long shot ]
The third "only Seek for Incantations" active
"Seek for Incantations" causes make 8pic not 4pic.
And
The another issues.
Remove Delimiter ' BREAK '
BREAK location change
[ Decided pose, dynamic pose
BREAK
__hair_color__, __hairstyle__ ]
Rearrangement prompt order
(full_body:1.2), (2girls:1.2), (side by side), look at viewer, wide angle shot, wide_shot, long shot,
__hair_color__, __hairstyle__, __color__ bodysuit, globes, boots,
(muscular:1.2), abs,
Decided pose, dynamic pose
BREAK
__hair_color__, __hairstyle__, white shirt, (pleated skirt, wind_lift, panties), (highheels),
(curvy_body), huge breast, narrow waist, huge pelvic,
battoujutsu stance, Fighting Pose,
outside, simple background, (strong wind:1.4), storm,
(realistic, detailed, best quality:1.2), QuickHands,<lora:QuickHandV2:0.6>
Only "Seek for Incantations active" again.
Ah that images are before and after!
So I look again previous images,
Quality is slightly better or not, separate two girl is not perfect separation but just a little better or not.
"Seek for Incantations" active and "Append Generated Caption" checked
Mabye ignore consoler error messages, and adjust value setting .
"Append Generated Caption" disabling.
"Perturbed Attention Guidance", "Multi-Concept T2I-Zero", "Seek for Incantations" on
Redesign test plan
Make 3 image each time.
All the default setting and adjust just one setting value.
"Seek for Incantations" , "Perturbed Attention Guidance" active
PAG Scale from 0.5, each test 0.5 up.
PAG 0.5
I used Dynamic prompt, but for check cleary just see the defference of result, hair color and style fixed.
" blonde hair, twintails " , " black hair, wavy_hair "
clothes color fixed too.
" black_bodysuit " , " white shirt, pleated_skirt "
PAG 1
Two girls are in defferent space. Remove " side by side ' prompt.
Boots => sneakers
PAG 1.5
Start to show one girl.
PAG 2
Rearrange prompt.
" full_body:1.2), (2girls:1.2), look at viewer, wide_shot " is after girl description prompt.
Start to show 3girls.
Remove " two girls " prompt.
PAG 2.5
Only shows one girl.
PAG 3
PAG 3.5
Get weird.
Control PAG value can't seperate two girl.
PAG is 1 fixed. " Seek for Incantations " and adjust " Multi-Concept T2I-Zero " value.
" Alpha for Cross-Token Non-Maximum Suppression " from 0.25, each 0.25 up
" Alpha for Cross-Token Non-Maximum Suppression " 0.25
" Alpha for Cross-Token Non-Maximum Suppression " 0.5
" Alpha for Cross-Token Non-Maximum Suppression " 0.75
" Alpha for Cross-Token Non-Maximum Suppression " 0.99
Get weird.
" full body " prompt weight up. (1.4)
0.99 is not the value I need.
Multiple test arround 0.75
Still shows one girl.
Reduce 0.5
Keep cause CUDA error. Every just 3or 4 times test.
Not even close settings causes CUDA error. (hires 1.5x)
Multiple test arround 0.7
0.65
'curvy body' prompt weight 1.2
0.64
0.62
CUDA error is so much.
Test terminates.
Provisional conclusion : " Incantations " extentions is not for seperate similar two object. It takes more time to make a image and causes CUDA error in my settings. Quality up is unnoticeable.
PS.
one last images without Incantations.
??
Comments
Post a Comment