Here is the negative prompt:
text, watermark, low-quality, signature, moiré pattern, downsampling, aliasing, distorted, blurry, glossy, blur, jpeg artifacts, compression artifacts, poorly drawn, low-resolution, bad, distortion, twisted, excessive, exaggerated pose, exaggerated limbs, grainy, symmetrical, duplicate, error, pattern, beginner, pixelated, fake, hyper, glitch, overexposed, high-contrast, bad-contrast
I call this prompt "good", because I believe that it generally only improves quality (except for the eyes - working on it!) and each word in the negative prompt was tested for if it adds something or not, rather than just going with one of the big word salad negative prompts most people just copy from other people with no rhyme or reason.
I call it "universal", because I find that it can be used for any style - not just photos - as it does not appear to change the style too drastically - as long as you use it with only the G text encoder!
The best way to use it is by using it with only the G text encoder, as that allows to benefit from the negative prompt without changing the style or composition of the positive prompt too much!
How to do that:
double click into the ComfyUI background, search for "SDXL", add two "CLIPTextEncodeSDXL" nodes
right click each node, in the right click context menu of the first node select "convert text_g to input", and in the second node select "convert text_g to input" as well as "convert text_l to input"
double click into ComfyUI background again, search for "primitive", add two such nodes
add a connection from each primitive node to each of the CLIPTextEncodeSDXL nodes
write the positive prompt into one primitive node and the negative prompt into the other
connect the positive prompt primitive node to the first CLIPTextEncodeSDXL node that has both the G and L text encoders enabled
connect the negative prompt primitive node to the CLIPTextEncodeSDXL node that has only the g text encoder enabled
The final result should look something like this:
You can also just use the example workflow I attached to this article.
Known issues:
For some reason makes eyes worse - will figure out what part of the prompt causes that and fix it
I realise now that my eye problem was due to something else. The samples I posted and some further experimentation I did seems to result in no worse eyes than what you would already get anyway.
Comparison:
"full-length full-body photo of a woman"
base
with both the G and L text encoders enabled for both the positive and negative prompts
with both the G and L text encoders enabled for the positive prompt, but only the G text encoder enabled for the negative prompt
"closeup facial photo of a woman"
base
with both the G and L text encoders enabled for both the positive and negative prompts
with both the G and L text encoders enabled for the positive prompt, but only the G text encoder enabled for the negative prompt
My method:
I've spent quite a bit of time testing various potential negative prompt words as positive prompts, note their effect, and then choose which ones to include in a new universal negative prompt based on whether the output I got was generally worse than the average output or in other ways would probably negatively affect the image.
Ultimately I tested the following words (orange = not chosen, green = chosen):
filigree
morbid
zombie
mutilated
unfinished
amateur
multipack
moiré pattern
downsampling
aliasing
baroque
distorted
wrong
missing
child's drawing
TV Static
blurry
glossy
jpeg
jpg
evil
ugly
putrid
old
blur
artifacts
jpeg artifacts
jpg artifacts
compression artifacts
low resolution
facebook
twitter
pinterest
deviantart
artstation
book cover
album cover
picasso
avant-garde
punk
steampunk
cyberpunk
poorly drawn
mutant
monster
lowres
low-resolution
low res
low-res
impressionist
impressionism
expressionist
expressionism
impressionistic
expressionistic
extra
extra fingers
extra limbs
long limbs
cover
strange
unknown
found footage
bad
meme
many
many limbs
many fingers
abomination
distortion
deformed
deformity
multiple
twisted
twister
excessive
unhealthy
exaggerated
exaggerated pose
exaggerated limbs
boring
soulless
inexperienced
mugshot
lifeless
emotionless
bored
cold
meagre
sparse
simple
gaze
grainy
fanart
cosplay
symmetrical
clone
duplicate
mistake
error
pattern
beginner
low-budget
low budget
fixed
photoshop
pixelated
fake
cheap
copy
fax
perfect
saturated
oversaturated
gamma
overtuned
desaturated
hyper
unrealistic
illogical
glitch
floating
bad quality
bad-quality
low quality
low-quality
png
unnatural
outdated
pixiv
cover
Frida Kahlo
overexposed
selfie
snapshot
text
watermark
signature
abstract
burned
instagram
action figure
low-contrast
low contrast
high contrast
high-contrast
bad contrast
bad-contrast
puppet
cliché
stock photo
magazine cover