#🏞|general-with-images
1 messages · Page 67 of 1

but I just did thatttt
and my PSU can't take it lmao
It may be able to handle this 3080, but its not gonna take more lmao
it'll take more if you undervolt and power limit it
trust me, Sytan, if you get a 4090 your power limit can go as low as 150W which is lower than most other cards
I was able to massive undervolt my 3080 without losing perf
psu explodes
for img2img what should i use
what
for the masked content bit
I am surprised that "tiger in the woods" did not trigger Tiger Woods lmao
Sytan just be sure to film it when you turn the system on with a 4090 on the 650W PSU
no reason why
oh, right. memories

it wouldn't break lol
But it would be a bad time haha
jesus fucking christ
it would catch fire
what thheeee
trippy eyes
man it's called self immolation please use the correct terminology.
what
😭
this guy loves screaming down old peoples ears
I'm about to defenestrate you
imma sacrisfy you
dont make me discombobulate you
defenestrate has to be one of my favorite english words lol
why is there a word for that...
lmaooo
Who doesn't need a word made specifically to mean the act of throwing somebody out of a window?
do a white one
so my suspicions were right, this model can do non anthro animals
l
LOL

what console is that for
Yeah, but the rest.
Nice
.-.
i just cant fix the eyes ;-;
yes

its the name for something below the belt
Welcome! Start by heading over to #1072220168534642768 to get yourself situated and help find the channels you are looking for! Please make sure you review our #✍🏼|rules-and-tos and feel free to assign yourself some #👥|roles as well! Answer any questions your may have at our #1072229020520947753. There are many ways of accessing Stable Diffusion, take a look at #1080946152318443610 to start your journey!
im sorry i cant read

what do those signs say?
@smoky oak Well, shit ASUS is doing another scummy thing it seems. After the blowing CPUs up because of their shit BIOs now they are reusing 3090TI heatsinks for their 4090. ASUS Tuf is in my cart too.
we seem to now be at 1.1.169 so 1.1.159 to 1.1.169 in 24h.
Jesus lmao
Yeaahhh... Asus is in hot water right now
Now I do not know what to do for the vid card. 😦
Not sure if you saw, but gamers Nexus and JayzTwoCents both cut them as sponsors, and LTT has put all of their deals with them on pause until they fix their crap
I knew ASUS with its history but figure 15 years later they should be good, and people praised them then all this at once.
I sure did see that.
There is no other company for the 4090 I can go to since Nvidia stopped making them.
They did Steve at GN bad which he explained today. They really are scum.
It sucks because from what I've seen, and what lots of YouTubers have been saying, Asus is still doing fantastic laptops and graphics card, it's just their motherboards that have been held
*hell
which is what zapped me. Now their insanity is screwing with 4090 cards.
They may be pin for pin compatible but I consider that a low blow. Let's stop using X to go back to Y to make more money. If Y was so great for the card then X would never have been made. Stinks
Maybe it is best to just buy a 7900XTX and be done with all this crap?
has to be better than my 1060, lol
Computex is in a couple of weeks at the end of May so will know more then. I just wish ROCm was ready right now for 7k cards.
oh, ffs, lol
Well, in a bit I will test this CN to see if anything has been fixed
10 version updates in 24h
Damn
yeah on my bot i made it so all of its images are automatically CTU'd with a badass prompt and then x4 upscaled

would you say that's better than MJ?
I'm trying to use multi-diffusion upscale with tile controlnet in Vlad, but I'm not sure which one it is. In "extentions" its already enabled
I don't like MJ so I am biased
Still messed up
A shiny spot on my source's forehead became like a greenscreen
This is new
i guess that's just the diabetes
its a little weird cause like, I can tell its upscaled, but it didn't really add much to the image
like its a lot of res and not much detail
that's what i was after tbh
just high res blur?
i don't think they're terribly blurry unless you really get into it
like, viewing at 100% zoom, they're great
I am just saying I do not think CTU is adding anything, just more pixels
like this is definitely not upscaled, just higher res, you know?
it would look better if it was downsampled
well i'm doing CTU before the x4 upscaler, friend
hmmm...
what is the point of the x4 if its just adding more pixels and no detail?
I am actually asking, I am not trying to be a jerk
it's mostly for experimenting
fair enough
FFS
keep in mind i'm doing this all in python and it's a fair amount of nudging a boulder required to move the functionality that much and suddenly be able to upscale all my images 
I decided to down sample your image, and it does indeed look sharper when downsampled
oh yeah for sure, I was just asking out of curosity
not downsampled
1024x1024 without any touching
downsampled
what is that an hour for another update?
get up to go to the bathroom and when you get back, it'll be ready
there's the comparison between no CTU and what i was applying
Left side better
it's just subjective, there's no better between the two imo
to me i prefer the softened look
agree. left looks better
No, not really as you can see when you switch. Left side has more texture to it and right side smoothed it. Nothing subjective about that.
whut? the right side has more detail
clothing texture look more natural, face is clearer
Which is precisely what I homed in on
Why the colour shift?
Now that is subjective and for me I like the other better.
yeah it's a pretty high quality model that needs little help tbh
alrighty I am going to try and use SD for some of my school project
lol freeze, FBI

It's not an art project or anything
It's for english
first thing I want to generate is a two faced head (like a head with a normal human side and the other side is like a monster or smthn) to represent the theme of the book
I have to do like 4 objects and the cover of the book in one thing to make a "book bento"
like a bento box
I've already done all the writing
Or, most of it
WAIT i could use that multi diffusion thing right? Where you can prompt multiple things in one image
oops, let me do that again lul
original, "best quality" @ 1.0, "best++++ quality++++" @ 1.0, "<image prompt>" @ 1.0, "<image prompt>" @ 0.3
the one where i pass the image prompt through definitely has closer resemblance to the original image
i prefer the very last image, actually
it has the original prompt used for the first image, and 0.3 strength
the reason i like CTU is because it removes some of the blur from the original img :/ the ferns have fewer issues, and the gravel ground
got a decent result!
pretty much exactly what I am looking for
Though I did go to 1.5 for it
I wonder what I get with the same prompt on DD
have you ever considered doing transfer learning from 1.5 to 2.1
maybe if I added variation seed
wdym, like useing images generated with 1.5 to train 2.1?
yeahh
with a variation
Im sure a lot of the images in the DD dataset are ai generated
.3 variation strength
yeah I am just gonna go with the one generated with dreamshaper
wait so a good CTU prompt is what
I went with this
only thing I don't like is the eyes but oh well
without a background!
I don't think I am going to generate all of the images, but I will use inpainting to maybe combine them all
I have an extension called "Download All Images" on firefox that might also be on chrome, it would probably be very helpful to you
i run this stuff on a dedi so i don't have to heat my space up here
it's 93F / like 38C inside
😄
eh fuck it, that's a pretty good idea JR
i am using the KellyC downloader
some of them are terrible 
that lighthouse one isn't bad
yeah maybe it's still a bad idea tho
seems like a good way to burn all the progress made on SD2.1 lmao
I just realized, if I can come up with a different item to represent the theme, I don't have to spend all the extra time doing this with SD
Because I have all these other items
Yeah, I give up on that reference only
what a silly folder
Anyone have any ideas for items to represent this theme?
The difference between monster and man is near-indistinguishable, and often leads to utilizing one's morality rather than physical observations.
the painting of dorian gray
that's a story
not bad but I dont have those
I got the picture I needed, no SD!
So all that SD work was for nothing
I think it looks ok. Most of that stuff is my sister's, exluding the book and globe. That's even her phone lol
I forgot an item
I am not going through all that again
I give up imma just use what I got
its like they were there the whole time...
why can't I get SD to ever under what and (not AND) means?
because it's not hooked into a NLP - it just takes tokens and converts them to numbers without too much context
Man, I can't wait for it to finally be able to do that.
if I say oil and water it will do the oil and skip the water. Sometimes will skip the oil and do the water instead.
well oil and water don't mix 😛
What does mix is RCNZ Optimus Primate, and Disney Pixar Cartoon Type A - this is the end result 😄
and doesn't mean to mix but AND does in sd
https://twitter.com/TheRRRealist/status/1648789518244274178 any ideas what model/workflow was used here? nsfw warning
it's r18 so can't see it
idk if im allowed to send the pic here, its not full nude or anything but still
Man, 2.1 is missing the seg model too
I had to turn off my filter to see it then saw it and was like wtf?
take a screencap of the head and shoulders..?
tbh i dont care too much about the clothing, it can be regular clothes, but the skin texture is very realistic to me
The latest reference feature is just not that good
Well, I need a closer shot of the skin
looks like somone used the flashlightphoto LORA - model likely to be realistic vision or analog madness - would recommend trying ICBINP
😦 So close
I'm suggesting that clyde (the content filter) is more prudish...
same w this one?
just searched icbinp, no sample photos but good reviews, might try
It is. My BP was going up about it
i have a ton of realism models already, maybe workflow is more my issue
thanks for that.. post must not have uploaded.. shit!!
Why is your name in black?
no still no photos
amazing, i have 1305 images of vape clouds now, public domain
Even 2.1 has issues with smoking
Got the staff pick in the Challenge of the Weekend (cow) last weekend 🙂
@dense tapir yeah ti's one of my training goals, to introduce that concept to it
Oh, nice but the black.
Comes with the "Stable Society" role
Love to see it
Stable as if STABLE diffusion is stable, lol
The hypersmoke embedding for 2.1 is quite cool, but it isn't smoking as in nicotine sticks lol
where ya been?
Updating models 🙂 😛
yeah, cig in mouth having a Marlboro moment SD isn't doing.
lmfao
Now I have seen SD doing all sorts of stuff in mouths, but cig? Nope.
:3
I have to say after ASUS shenanigans I am probably going to AMD even if it is slower in SD
AMD can be faster. the problem is in software, even for nvidia you don't get the whole FLOPS capacity of the card
yeah
I am holding out to see if ROCm makes it stellar in 5.6.0 when they finally support the current gen of their cards.
Now that is whacked
My Sapphire Nitro+ is 50 over MSRP so they are coming down. Computex is in 2 weeks so things will be known then, I hope.
My fear is not the inference rather how well does it do with training. They are not the same so I can't say it does fine in inference and think that might be the case for training too.
1059 vs 1749 right now
I asked a dude who follows a lot of this on his livestream earlier and he said there will not be a 7950XTX. So many say there will be so I hope he is wrong.
prompt?
a breathtaking aerial view of a medieval castle on a cliff overlooking a vast ocean, surrounded by lush greenery and cascading waterfalls
notice SD skipped the waterfalls
a seed thing sometimes
The edge of this guy's face and sometimes ears keep coming out discolored, any idea why?
I love thisssssss
I still don't get this
Never use them
but they're UwU
Though I will eventually
I figure in a year I will be setup to do a whole lot more than I can now with all this PLUS the tech will have advanced so win-win.
🧓🏽 or you be left behind
Nah, I don't run with the herd.
the lone wolf?
pretty much. I do my own thing and never cared for that herd, "clique" mentality.
Now this is getting good
I think it isn't strong enough though and why I asked wth was that loop back shit
My honest reaction to that information
Did he even explain it?
I know with the new sliders I had to go to his community to even read about them and I left not finding out a damn thing
Stop using bad models lol
Yeah, this fidelity slider needs to go to 2 not 1
it's realism engine

That looks drawn
you know this learnt how to blur backgrounds from real photographs, right? 
dont gimme that shit
Depth of Field is a real thing.
man there's a debate that happened when I lived in Toronto. the city published a photo from the airport, showing the CN tower. everyone thought "THAT'S NOT POSSIBLE" but alas, with the proper lens, it is.
But this one looks ugly
Nmegaman
yeah trying some megamany styles
Boo
lol what the loli fuck
B
plz no
Start your engines. I am already finished. :/
Hmm if you don't think i tlooks like megaman Ill remove it. Ithought it was neat.
lemme try again
megaman?
The boobs look plastic
and the faces look 10
yeah, this sucks I can't force it to play right
Oh, and that loopback doesn't do anything
cries a coat of arms of a church with the following characteristics that has in the center an eagle, a bible and the sacred heart of Jesus, above the coat of arms a cross and crossing the horns at the back.
AI generated or no?
Yes
wrong
SNAP
hahahaha
Im a 2 grader

Wait
No
I thought u mean the number on tests
Bruh idk shit about english school
I finished school
all i care, is this more realistic
another nat geo photo, son
Wjat
Wolves live in groups called packs. A pack is a family of 7-8 wolves with a mom, dad, and offspring. The wolves' communication skills are very important to the pack's survival. Wolves work together to hunt, raise their young, and protect their territory. Wolves communicate with more than howls. They whimper and whine, growl and bark, yelp and sn...

Ew
wolves are animals and thus imperfect
coyote?
Why this one looks so blurry
im telling you, you're going blind
Not smooth ass
Green
Green.
so this is without any CTU even and, it looks synthetic to me when it's not slightly blurry
Tf is ctu doing
AI or Photo?
gonna laugh so hard if it's real
the weird hair fractal patterns
Photo
nope, AI 😄
Nope
yeup
LMAO
I mean no feet
there's snapping turtles that have pincers
Isnz that a squirtle
WTF>!> We do not even have a fucking style model for control net. SHIT
i forgot real name
any guesses?
Whqt
Just what I needed and we didn't get it
that's a weird dog
Photo
I see him drinking from a mug
pretty sure he's playing xbox
i dont know what i did but i seem to have broken it
This way yesterday with the same prompt, all of them look really weird today not sure why.. lol
Do you use a lora for twi'leks
glad you like 'em.. still trying to get a good one at 32:9
best so far but.. not great
Are you outpainting?
her headtail looks a little too jumbo on that one
do you generate at that native res?
also is that lora on civitai
I have a variety of workflows I'll use for ultrawide. For that it's generated at the proper ratio than hires'd
Yes, there are a few loras on civit I believe
ooo might have a good one
attempting hires
make sure you share whatever you create with me. gimme a tag or something on here so I can scope it out later 🙂
I'm thinking of combining it with one of my dreambooth models
we'll see if it works o:
I have a thing for twileks cuz back in the early days of midjourney only like 1/100 prompts looked half decent
things have sure come a long way
yeah for sure
paired up with certain models it is insane
or actually, hang on... lemme try a dual lora
eh i'd need more time to fart with this and ireally need to sleep 😛
was trying for a twi'lek tinkerbell hybrid...
lollololol "Matrix"
@smoky oak i'm using bicubic upscaling before running CTU and i think the output improves. LANCZOS is best for downsampling, it adds "ringing artifacts", oscillations around edge transitions, when upscaling
now vs earlier's
LOL, it finally gave me what I asked for.
peeng
I think they look about the same, that is to say, both look solid
DUDE
that looks hella dope
that actually is pretty good for a minimalist wallpaper.
Sen? Lmao
I figgered out why I have to use an extremely low CFG on my dreambooth. Someone told me that when u train the dreambooth model, the ENTIRE MODEL starts shifting towards your reg images
Yup, it is you lol
I made my reggies at CFG 20 to give them the most extreme detail possible... it backfired lmfaoo
so now even at cfg 2 the images look suntanned... I still like em, but if I were to need a more washed-out look, I would have to retrain the model ):<
sy I had a discussion with mods from the other SD discord regarding the whole lora vs. db debate
if you look at the distortion around her eyes, the issues with the old pre-CTU upscaling method are more apparent
cuz I gave u the benefit of the doubt when it came to lora since perhaps my info was out of date.
that's interesting though how much the noise from the upscaler that CTU recommends, contributes to its outputs.
Ah yeah, I see what you mean
i sat there flipping back and forth and it's unfortunate that the old noise from the upscaler led to more 'depth' in the patterns
because that's gone now
but i can fix the prompt etc
i probably can't use it lol
i wouldn't know what to do
hmm
ultrasharp makes them look too smooth
sytan, the other fix i did was to start doing what you do, and pass the original img prompt through to CTU
I like the more expressionistic results you have been getting
the tiny crinkles that one adds are so satisfying
both the positive and negative prompts
yes, the auto-CTU i want to keep because of how it relaxes facial muscles
so i'm pretty determined to make it look good
How is that working?
you're seeing it
I have another tip for you I can share in DM's to get even better results at high res
the upscaler's detail it adds are generally more coherent, and keeps facial details eg. proportions, closer to source
or well, I am not sure if its as important now with CTU
i put READMEs on my two projects today so that it's possible to set this bot up, for yourself
and DM me if you want Sytan, i can't DM you, Discord sucks that way
oh thats weird, you should be able to
hmm
So pretty much the best way to add detail to an existing image is tile ControlNet upscale with loopback multidiffision? Or something?
with ultimate upscaler
@oak ospreyI DM'd you, not sure if you can respond or not
I personally like to keep the denoise low, but its personal preference
sent a friend request in case
I hear that incremental upscales in a loop helps keep the image consistent and adds detail better than big 2x upscale jumps
mah
nah
at the end of the day, whether u do straight to 4x or 2x then 2x, the end image still needs to be processed the same way
what is important is how much detail there is in the starting image
I might drop a video TuT explaining all the details because this tool is literally HEAVEN
And if the original image is large but not with much detail, then sending it to img2img with Multi ControlNet a couple of times with something like .5 denoise and keeping the image the same size, might be better?
nope I have a special trick for that o:
What I usually do is downscale it to the closest u can for the lowest rez to be 512
use photoshop bicubic sharper which works good for shrinking
then, do a 1:1 with the NKMD upscaler
and then process the image
see, unlike some selfish kids here, I'm an open book and share everything openly. Imagine being so conceited that you resort to DMing people your methods for fear that others might copy you
when you are using literally dozens of tools you had 0 involvement in
Yeah I pretty much share whatever I learn, it helps everyone overall
Basically I generate a render in Blender and a depth map in blender. But when I use that render in SD to make it look more stylized, it gets problems
depends on what the model is but for humans, that upscaler works really well for adding tiny skin details to make the img look more real
Like, the model tends to make the skin color lighter. And the character has discoloration on the face.
and you can crank up the denoise on the 1:1 + high cfg if you want even more details, I dont, I think after a certain point an image looks fake if it's overdetailed
Like for example, here's the render I made, no SD:
Then after a first img2img with Multi-ControlNet and .5 denoise with add detail Lora it is like this:
And after another pass also with face detect fixed it is like this:
I think for that type of image, the NKMD upscaler might not work since it is trained on human faces. there is another one that's called nickelbag or something that might be worth a shot
it's a combination of upscaler algorithm, denoise and cfg, and prompt obviously
I was thinking something like esrgan 4x anime because it's supposed to be made for something like that
yessss
I wouldnt know since I dont do anime, but I assume there are options out there
theres a TuT on CTU by stephen or some dude out there that's worth watching
https://www.youtube.com/watch?v=EmA0RwWv-os it's slow but this dude's tut is a good place to start
Render 8K with a cheap GPU! This is ControlNet 1.1 tiles for Stable diffusion, together with some clever use of upscaling extensions.
ControlNet 1.1 Tutorial and install guide: https://youtu.be/WZg3e6B2yPQ
ControlNet 1.1 Models: https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main
4x Ultrasharp model: https://mega.nz/folder/qZRBmaIY#nIG...
I tried but if the problems with the image aren't fixed, then the upscale will also have issues. For example, the second image has some discoloration on the side of his face, consistently. Like, every time.
Ah yes I've seen that tutorial
And the third image with the face fix did fix the problem but changed his skin tone too much and also his eyebrows shape changed.
Still using DD V2
You should take ICBINP v6 for a spin - see if you can get it close to the dd awesomeness!
What is that?
Updated - v5: Added Am I Real, Analog Madness, and TangBohu and the amazing add_detail lora at weight 1 to the mix. I think the details in the back...
it's a 1.5 base realism model
Oh, I use 2.1
I am downloading it now as I do swap in and out of 1.5 for face fixing. For film stars I go back to 1.4 sometimes
@dense tapir have you checked artius 21? it has issues with dishes this model.
I am not sure which version of that one I have
Artius 21 is nice 🙂
roasted duck and you get half living half baked duck
Just to see if that new model would work
but yes @tired basin i am satisfied
it is just oil painting
remids me babylon V.
try alien characters
to latest picture, alien giving weird results, alien character seems to me be better
If I type alien SD gives me Africans in Africa. Other models just give me Africans.
why still that hat, i think it was in that serie probably
Turning to the dark side....
on the edge 🙂
ewww boy
should i delete it?
you have special words for background?
this seems to be saint lous cemetery
I am so tired of the glossy skin. I need skin textures.
WTF? This is a ghost?
This is not going to do a ghost
yeah, worked on lexica but fuck it
on the first picture that ghost on the rim of photo is best
yes
it's this model it can't do it
trying base sd21
nope
must be in my neg
i am in that mentioned earlier
posting this because more things are happening there. But i imagine ghost differently
Ho-Lee-Fuk
That's senblend 1.8 coming soon to a theater near you ^ I would say is probably a tiny tad better than Zegovya but responds much better to dreamboothing
I've been playing too much Jedi Survivor...I'm in a Star Wars mood. lol
Indeed, hopefully they don't screw up the remake.
they were heading into that direction, but decided to scrap the company they were working with and start with a new one
A delayed game is eventually good, but a rushed game is forever bad
Bingo!
Sui
a ghost
A ghost
Hopkirk 😄
admire that rocket, i always got some weird, and broken
tickling tickling 😄
Currently, there is no public bot on the server that generates images. There is an experimental bot available for early server members & Stable Society. You can obtain the Stable Society role by winning our weekly events #⭐|pow-info #1087493421209485393! However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
Try haunting, instead of ghost
I am head over heels with this new model
what model?
2nd image is the new one. Fixed up the image a bit in Adobe Lightroom, made the guy look more like Anakin Skywalker too with the robe color change. Lightroom does wonders for SD images. lol
Took the awesome merge Disney Pixar Cartoon Type A , and merged it with RCNZ Optimus Primate using the Merge Block Weighting to get a look I'm real...
@tired basin thank you will try for sure
used to work and works on Lexica
f. 1.8 is clone? Probably long clone can make ghost like things
batman?
i thought it is at least bit transparent. 🙂
this is batman with same settings
🫣 oh my such a scary ghost!
thats the new model @tired basin mentioned @dense tapir
split personality
DDIM serving better hamburgers
yeah Euler As hamburgers look like those vegan ones 🤔
while DDIMs look nice and meaty
🙂
wow
the galaxy pic reminded me of the meme
That's the best way to taste sesame seeds
gotta lubricate those dry ass buns
shiny super man
Euler A burger
bon apetite
ew
open mouth 😄
yes they use string to make it eatable
@hearty karma i am their fan, and i remember they did it several times, shaggy did it and scooby eat it 🙂
https://civitai.com/models/48139/lowra A new tool to my arsenal o:
this thing works insanely well
what is that thing?
a lora to add more shade and darkness to ur gens
ah
hopefully this fixes my oversaturation issue
its up there with epiNoiseOffset
see now I gotta look that up
do you guys all use the 84000 vae?
I cant seem to find a varied source of vaes. im guessing there arent that many?
only 4 really, repackaged under different names
I think I'm actually better off not using a vae if I want a more washed-out look...
you are using a vae even if you're not using a vae tho
^
whats the default one called
it's just whatever's built into the model in that case
not necessarily like that, but all sd gens require a vae to decode/encode the image
so you can actually see it
kinda looks like a heart
kinda yeah
I see wat u mean
lmao
it's just a necessary step in the process, the default-default vae is whatever base sd models (1.3/4/5) use
SDE karras ++ it has support of bun in the middle
other than that there are 2 stability vaes, 1 nai, and 1 wd
you either have to take it apart or put it on a cob and munch it on the sides
looks like pixar movie
idk man it's creating some seriously good ambient on mine
the epi-watchamacall it idc if it's better, when you have to type a novel to remember the keyword I just bail.
dude has like 4 different encrypted keywords for his lora. sod off
thats a burger in a suspiciously dim-lit room
somebody likes it raw 🙂
gamer girl?
sorta. its from a really old model, just needed to test the lora
for being so old it still looks kinda kool
bet you can't generate that classic pizza slice cheese pull type photo





