Quality Comparison Test Based on Denoising weight and facial expression prompt test

For fast prompt setting test.


Quality Comparison, two different Denoising weight


(Additionally, can stable Diffusion WebUI Recognized  'pouted lips, wink, smug, pursed lips, glare' prompt?)


Base setting is 


chech point : PL_mix3rdBakedInVae

DPM++ 2M Karras

Sampling steps 20

cfg 7

512*640 (4:5)

Hires.fix 1.5x


prompt : face of girl, (look at viewer), ({Caucasian|European|Spanish|korean|}), ({big_smile|pouted lips|wink|smug|pursed lips|glare}:1.4), (best quality, realistic:1.4),


negative prompt : glasses, eyewear, (worst quality, low quality:1.4), Asian-Less-Neg


(Asian-Less-Neg is textual inversion. in this test that i use is v1)


test denoising is 0.3, 0.4, 0.45


(Recommend view original size in pc)


Slight difference. 

Denoising 0.4 quality is better than 0.3.(more natural detail)

0.4 and 0.45, barely noticeable difference


'pouted lips' prompt didn't reconized.





0.3 and 0.4 barely noticeable difference.

0.45 is slight batter than others.

In this case, 'pursed lips' prompt didn't reconized.





changing prompt : face of girl => (close up:1.2) photo of a girl 



Barely noticeable difference 

'pursed lips' prompt didn't reconized. 

Add ' _ ' (pursed_lips)





Barely noticeable difference 

recognized 'glare'






Barely noticeable difference 


' pouted lips ' ......uncertain





more detail in denoising 0.45 than 0.3, but very slightly

'pouted_lips' prompt recognising is still uncertain

add ' _ ' (pouted_lips)





Ok quality test is done. 


There are some quality differences depend on denoising weight but very slightly.


' wink ' prompt is recognized.(not perfectly)









conclusion : prompt ' pouted_lips ' don't recognized.

(at least PL_mix3rdBakedInVae check point doesn't know 'pursed lips' perfectly)








Prompt ' smug' is uncertain but express something.








prompt ' pursed_lips ' don't recognized in PL_mix3rdBakedInVae






ps additional test

changing check point model MH V2

same setting

(close-up:1.2), face of girl, {20|30|40|50} yo, (look at viewer), ({Caucasian|European|Spanish|korean|}), __hair_color__, __hairstyle__, 
({orgasm|big smile|light_smile|Eyes narrowing wickedly, with a sinister grin|painful eyes, open mouth|shy, (blushing:0.8)|surprised|pouted lips|wink|smug|pursed_lips|glare}:1.4), 
(best quality, realistic:1.4),

glasses, eyewear, (worst quality, low quality:1.4),



smug







light smile







smug







glare







glare








light smile







surprised
???







big smile









painful eyes, open mouth








surprised
???






wink
???








surprised
????
(doesn't recognized cornrows) 






orgasm

tounge fused with lips, LOL







pouted lips








surprised







orgasm
????






eyes narrowing wickedly, with a sinister grin







surprised






surprised







pursed_lips






wink







shy, blushing






' wink, surprised, pursed_lips ' didn't recognized

Comments

Post a Comment

Popular posts from this blog

Solved: Facial Errors in Stable Diffusion Pony Models

Image
While it might seem minor, this issue has been bugging me for quite some time. Today, I've got a breakthrough. Let's take a look at an example first. They all have the same symptoms. All different check point models. Ponyxl About under 5% ratio show up like that face error. It's not low steps, After Detailer or promt issues. It's a sampler.  DPM++2M sampler => Euler or Euler a Problem solved~!!!!   2days after,  Problem not solved. It's not perfect. lower rate before but still exist...
Read more

Some of results (character full images)

Image
 Hires fix with my check point model PL3rdilxl semireal Sexy kalifa in fishnetstockings Hot hancock in stockings Morrigan aensland : Ready for battle!! Lilith aensland : Waiting for my sister. Naga the serpent in stockings Helen parr . Dating with bob Sohryu Asuka Langley Power girl!!!! Wonder woman in pantyhose Jessica rabbit Lady death Nico Robbin Kiryuin Satsuki  FAMT2A15thR_PLIOR v2 Ivy valentine Yamato Chacha Shadow heart Belldandy  Kasumi Lucy Yeah~~ pretty good results! 1200 * 1504 hires 2x  It takes about 10 minutes!!!!! hires 1.7 about over 5minute!!   All image has exif info. You can find me on  instagram  ,  deviantart  ,  pinterest  ,  patreon  ,  civitai
Read more

Todays simple discovery (uncannyValley_ilxl10Noob)

Image
check point model :   uncannyValley_ilxl10Noob Occasionally, it generates images with two heads. Why? maybe prompt "  two long pigtails " In my wild card has prompt " two ". Like hair style. ( two tone hair ) Add " _ " between prompt . Reduce double head~ (The problem isn't completely solved) Do you see any unnatural markings near the face? My wild card include Loras.  This model is like turbo model, so can't applied lora. Remove loras and problem solved~ Same issues. Foot AD(after detailer) is noticeable.  (Stable Diffusion still can't differentiate between hands and feet.) My setting useing pony hands lora. Remove hands lora .   All image has exif info. You can find me on  instagram  ,  deviantart   P.S
Read more