#🏞|general-with-images
1 messages · Page 36 of 1
forgot my suit, just chillin' on mars
lol love it 😄
😄
I love how the helmet got cat hears
LOL YES
freaking amazing
I use blender and can generate line-art and depth maps from complex 3d scenes. But I don't know how to make the color consistent from different angles
I use ControlNet too
I'm not sure on that one, it seems like a complicated thing to get yeah
with some training maybe ?
nah
I want to make comics so training for each scene is way too much.
yeah it's really not practical to train each scene for the background itself
for the characters it can be though, quality wise
but to keep the colors on the background... maybe by img2img, populating by hand the good colors beforehand, but that again seems like quite the chore
For example this is generated in 3d, no AI used
I did a very small 4 panels comic, and the bg coudn't stay consistent, I needed so many seeds
very nice base for controlnet, yep
Scribble doesn't quite seem to understand the depth of scribble, but the depth doesn't get the intricate details of scribble. If I could get both that would be great
Yeah 3d+AI is really powerful. I just need a workflow that works
Drawn
you tried mlsd or hed with that ?
really good for architecture
I haven't. Generating lines is the easy part though
haha yep I see those
Hi there. I am new to SD (1st time). How do i start generating images?
hello there hardy
welcome around
we don't have a bot for it, but you got other ways
here is the faq
Currently, there is no bot on the server that generates images. 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!
How would that change when taken from other angles, distances, and perspectives?
Thanks🥰
well, right now, I used a very generic prompt, "a building in an american street".
To keep it more stable, I would need to over specify materials for a start
Then, I could use a second extension like "couple latent", that let me use multiple prompts on multiple parts of the picture, to be able to push for specific things in some places
Lastly, personally, I would also go into img2img, and "seed the image" with the right colors. Meaning I would use that image you shared as controlnet, and use it also as base picture in img2img after having painted it with colors I want, just "putting the right color palette" out there, so that the AI can morph those colors easily while following your controlnets
I see, thanks. I'm going to have to do a lot of experimenting.
Thanks but i still don't know how to start.
you may want to try some other tools, like comfyUI, if you are making a very specific but also very similar pipeline from a picture to the next.
It's a lot more complicated UI, but it lets you over specify how you want the picture made.
Like, you can pre sample the background, the characters separately, sum up the latents in a single bigger latent picture, finish the computing, while pushing different controlnets on everything.
Sure, you'll take 10 times more time setting your pipeline up, but then you'll just be spiting out content
ok so 3 possibilities
1/ you can use the website the bot linked
there you can do some free pictures for free credits
2/ you can install the thing on your computer
check #1080946152318443610 to find the right tool for your foot then
3/ you can, if your computer is too old to run it, use a cloud service like google colab to run it
I am getting so hyped for SD 3.0 man
if its even half as good as that SDXL demo, and it can be trained, SD is gonna be for sure on top of AI gen again
/prompt
I can generate line-art for characters and also their facial expressions (rigged characters). And color them (but without shading because with shadow they look bad). I think SD can't make facial expressions with fine control yet (I think there was one that came out just today though). So would character Lora be better or maybe just trying to use AI to color the line art generation? I'm just trying to think of ways to combine the different techniques
hopefully it won't take too long
Check out these videos for a great overview of the website and how to use it! https://youtu.be/014J2Yo1aGI | https://youtu.be/M3jCa6qTHpQ
about 3.0 exactly or sdxl?
@glossy herald I see the name change haha
So they confirmed that SDXL is not 3.0, but rather a part of it
yea I've seen that
one of the admin was in here chatting about it yesterday
so...what's exactly happening - SDXL release first and then, after some time - 3.0?
Or SDXL won't even be a thing on it's own?
so theoretically... We could be in for the biggest image gen improvement the industry has ever seen. Hell, even just that SDXL demo was a bigger improvement than all refined versions previously put together
for animation or comic book pages, I would recommend a lora per character and per change of appearance of that character (like change of clothes, or scars that would need to stick correctly, ...)
and to use open pose
yes, that new controlnet for faces that came out today seems very promising too, but I'm not sure how good it is yet.
It seems as though SDXL will not be its own thing
my guess at least
its probably just a base for 3.0
hmmmm
so really, we could have an even bigger improvement than the already monumental improvements with SDXL
And I don't say that lightly
yea both options sounds interesting anyway
SDXL is unprecedented levels of improvement
I'd go as far as to say with my testing, its text encoder is equal to or greater in power than midjourneys
it listens so well to what I ask from it. It as a base got less things wrong than all of the others got right combined
SDXL misses less than all the others hit
and I say that with ~75 multi version tests
I didn't test it that far, but yea, some aspects looks like a decent improvement
decent is an understatement. The improvement between 2.1 and SDXL is several times bigger than 1.4 to 2.1
Also, SDXL can do weapons
yea, tested it a bit too, but holding objects is still pretty bad
realllyyy good weapons
remember, its a base
What are the ? on 3.0 so far? Like the unknowns that could be really bad
1.5 is unusably bad as a base, and SDXL as a base kicks 1.5 finetuned in some ways
so imagine how far we can finetune SDXL
SDXL listens insanely well, from my testing
my imagination doesn't go too far
what I am saying is 1.5 base is so horrible that it gave this for "polygon vector art of a tree frog"
and even with it being that trash, we can get images like this out of it finetuned
yea I remember your images, I see where you're going with it
Listening to prompt and understanding it is one of most important things yea
that same trash base was able to be finetuned to give results like that
I think I had decent results with 1.5 or 2 base models when I started, but it was really hard to prompt it lol
so a base that was so unreliable its basically unusable can achieve near photorealistic quality and 8k+ images pretty easily when finetuned. So one can only imagine how much better we could make SDXl/3.0 with the same training
we'll see 
Cause to say SDXL blows the other bases out of the water would be a huge understatement 😅
This comparison between 1.5 base and SDXL 2.2.2 beta shows what I mean
It was supposed to be a night time photograph of a husky puppy sitting on a beach. All 1.5 got was husky puppy, while SDXL 2.2.2 got everything I asked for
That isn't nighttime that is dusk
its a properly exposed night time image in SDXL
then it sucks
it just doesn't have noise offset, chill out lol
still the only SD model that can do anything past sunset
When I ask for nighttime I want nighttime not that
(without noise offset)
Thats preference I guess
I still really, honestly, do not like noise offset
you're the first person to say that the SDXL image didn't look like a properly exposed night time image. And as a photographer of several years, its exactly that 😅
now if you wanted an under exposed night time image, then thats a different thing
Not in the mood to argue so you continue oin with the overhype train 24x7 and we shall see. I do hope it is as fantastic as you say.
Its not overhype if I am showing actual examples
I am not saying anything that isn't easily observable lol
it handles very specific prompting properly, something none of the others can even do on a basic level, so IDK
I have a feeling you'll find some reason to hate it anyways. Seems like everything with SD has a "I do not like" or "is pretty bad" attached to it, regardless of its comparable use or quality
like noise offset. It does what is needed, and it does it reliably, but you don't like it
- to that part
we shall see. I do hope it is as fantastic as you say.
I found a way to basically gain a view into what images a model was trained on
generate with no positive or negative and the biases literally show themselves
protogen infinity
faetastic
much more flat, interesting
Anybody know any good alternatives to ESRGAN_4x?
Something that can give extreme detail like ESRGAN_4x, but without the absurdly oversharpened look
I prefer swinIR_4x and more recently LDSR
Oh, I was meaning ones that were external. None of the built in ones have nearly the detail of ESRGAN_4x unfortunately
LDSR is very vram intensive, so I need tiled vae for that
I tried esrgan_4x for a while and found shadows to be too harsh
yeah, it makes thinsg crunchy, but damn is the detail out of it so much cleaner and higher density than the others
ESRGAN 4x looks higher detail than all the others at 2x the res
ideally you want to upscale at 4x not 2x with an upscaler that has_4x at the end
a 4x upscale at the same time using ultimate SD is a good way to get huge seams
I incrementally upscale
yeah you def wont get the best quality that way as the upscalers are built for 4x as mentioned
or they are made for 2x etc
Thank you very much 😊
Four images designed after snapshots taken within Second Life.
Barrack O'Bummer
cute studen
weird, I am training now with the same settings and the images are all jacked up
I'll see when it hits epoch 1
@dense tapirBTW, are you training on collab again?
This was interesting: I asked for the text "Great Work" in a image, and as normal AI did fail (SD 2.1), but I locked the seed and ran it through "Hires. fix" and forgot I had LATENT on so it became unsharp, I then selected 4x-Ultrasharp and the text was fixed. So is I first run a upscale with LATEN and then Switch to Ultrasharp I do get better result.
yeah, on my old account.
ahh
I am using my port of the kohya_ss
since colab is headless anything designed for a display will not work
having ripping good losses now in 2.1 for 2.x, lol
I mean 0.1x kinda losses
instead of 0.4x
that really is interesting. Hmmm
not sure how it managed to fix it post fact though
anyone know how to make kohya change sample steps for its images?
Man, it is suffering for some odd reason
02325 = Original Image.
02327 = Image use LATENT trick to upscale.
02326 = Image upscaled with 4x-Ultra sharp
4x-Ultrasharp have more details but the image mage with the LATEN trick is closer to the original.
yay
@dense tapirPardon my botherance, but could you please remind me how to get that Ai layer that makes image gen several times faster?
I don't remember where we talked about it, otherwise I would search for the link to it
oh, I think I found it
VoltaML
Yep, volta
4070 is due out in a couple of weeks and it is almost priced right for those who want 12GB cards.
50 less and it would be the best bang for the buck (price to performance) of all the 4k lineup
I wish more tech review channels on YT were focused on SD and reviewed all the cards as they were released. sadly they only do gaming
any requirements for it on GPU series?
There's also this thingy
https://github.com/ddPn08/Lsmith
I just ran across that a few days ago but didn't really read it. Sounds a bit of more of the same at a cursory glance.
While I imagine the 4070 would be considerably faster than a 2060, is there be something the 2060 can't do that the 4070 can?
whats the cheapest gpu for 24gb vram either 3090 used or a 7900xtx?
just to squirt out pics then stay 2060
But the 2060 has a 12 GB version
got a 2080ti and it cant do shit
What kind of shit are we talking 
unless you know Linux then go for it
@dense tapir had a 69xt didnt try sd on it
rocm only works in linux
reee
Because if the difference between 4070 and 2060 (12 GB) is just speed, then I wouldn't need to spend as much... I'm used with slowness on SD already
comfyui is so good but theres no face restoration... auto111 is so slow for me and crashes my xorg a lot
if I had any other card I could buy that would gen and train and had at least 16gb that is a new card I would seriously look at it. Frankly Nvidia has all of our parts in a vise and they know it.
I can get two of those three but not all three
Is Intel Arc usable with SD at all?
You don't need face restoration if you're using updated vae and not generating persons too far away.
But well, even if you actually need to fix face - you can just throw image you want in SD or here
https://huggingface.co/spaces/sczhou/CodeFormer
Buuuut again, it doesn't work for all styles and images, often it's just better to get right resolution for your gen.
If you know you're going for full body - go for higher resolution. It'll fix face
kk
generated some images using stable horde it was horrible
no idea what that is
distributed computing
sounds weird 
I'm divided on whether I should delete some pictures. I like so many of them. But there are thousands of them and even the file manager is struggling
(ignore the ingrish)
wth? My lora keeps chopping off her legs :/
It's basically GPUs together over a network. So people basically pooling their resources
she has no legs below the knees
For $50 you can get an external HDD
Were there pictures showing the feet?
It's going to be the sixth one, but I see no way
The pic above her was less steps and look at it
This one? There is no feet too, but it makes sense on this one
trying a different prompt
upper body \ face works
but in this case you shouldn't specify clothes or details below upper body or below face
and probably shouldn't go too low on cfg
Anyone know if comfyUI can save presets for different types of workflows?
Saves them in the pic
Different prompt
Can't you load them based on images?
I'm just assuming things, someone said before you can throw an image to load nodes
maybe I'm saying dumb things
I haven't tried. But I thought it might be possible to have a preset based on if someone wanted certain ControlNet nodes, models, etc already ready. Instead of having to make new nodes each time.
Not based on image, because I'd want to make images rather than reverse engineer based on another image
I like SD's way, I'm storing my prompts as images, easier to see what it represents without reading
likeness is spot on there
Yeah, it hit the mark for sure
After I generated tons of demon girls I can't look at diablo 4 logo anymore, it looks like it's made by AI now...and it doesn't even look like good gen 🥺
I think the halo that news sites love to use aren't so subtle any more. It's something they slap on all over
as reference lol
lilith is def a 3d model
yea, looks like low poly 3d, it's just my associations put demons in 1 line now 😄
John wick futuristic
using controlnet with jsut a white sphere works to make very discord profile picture worthy images
sorry. i mean white circle
hh
?
reminds me of the galloping goose trail i used to walk every day to work
the circle control gets awesome framing on lora portraits too
or hair, i guess
Robot red
quite nice results
What in the hell haha
@dense tapirI am looking to trying VoltaML for insane speed gens
aiming for sub 1 second 512x512 images
controlnet a solid circle as a test to frame discord profile pics
ahhh haha
I see now lol
that explains the ROUND
@dense tapirWould you possibly be able to help me source tensorRT 8.4.1.2 and PyTorch 1.12 cu11.x?
I am not sure how to ensure I have those exact versions
that is a bit older version of both
yeah, I am just not sure how to force those specific versions in my SD install is all
anywhere I can put intentionally unaesthetic images like this
I wonder how many times multiplication I can expect to my SD gens
makes me feel violent

i seriously want to crush it so bad
to see if you have it pip show nvidia-tensorrt
is that a command I run in the root?
Can anyone tell me how i can generate images ?
pip install tensorRT==8.4.1.2
which cuda do you have?
in the root nvcc -V
thought so
I do not even see that version
So I meet the cuda requirement
and I meet the driver requirement
just need the pytorch and tensor RT
i got 11.8 and 12 installed but i dont think i needed 12. it was an effort to boost my ada card
new pytorch came out this week i heard
I need specifically 1.12 pytorch
python -c "import torch; print(torch.version)"
so
https://www.i-programmer.info/news/216-python/16180-more-pythonic-pytorch-2-released.html this one. supposedly is a big deal
Programming book reviews, programming tutorials,programming news, C#, Ruby, Python,C, C++, PHP, Visual Basic, Computer book reviews, computer history, programming history, joomla, theory, spreadsheets and more.
no
it is, but its useless for now
compiles models to optimal tensors
version literal
useless to us. distinctions. lol.
lol
its around 2it/s slower for me
there are branches with it i think
so type it exactly like that, I am not replacing any variables?
yes
2.0.0+cu118
shit, lol
this was the SD install I did for the pytorch 2 one lol
oh its just a pull request and it is filled with people using it and misunderstanding the purpose lol
brb
slower not good. we want that compiling feature, plus also speed on base models is good too. you got a 4090 yeh?
one day the webui with all it's extensions and sloppy ui goodness will know compiled models
What i am doing right now should make my 3060ti faster than a standard 4090 without this tweak lol
ooo 3060ti kay
Hi
this change should hopefully see my 3060ti above 40it/s
crap, ok. Can I just delete the venv and force it to install older pytorch?
but can you still use embeddings or train with it?
yes
yes
yes to what I asked, or what they asked?
both
ok
yes
lol
I am trying to install VoltaML, which can speed up SD by several times
wish my card had tensor cores but not on pascal cards as they came with Turing cards next gen
some things it speeds up by 10x+ over pytorch
volta is awesome yeah. works on 4090 tho too lol. it recompiles a model and i haven't seen a lot of research on how well that performs
I know, I just said it can make my 3060ti faster than a 4090 without it
so I should be able to do sub second image gens
@dense tapircan you please remind me how to force install pytorch 1.12?
I deleted the venv
stability was talking about fast diffusion too. i think its the same idea, compiled models. but the stuff is really buggy to work out maybe. tbh i been really busy doing image batches and experiments and haven't dug into much news on that topic lately
just rerun auto
--reinstall-pytorch but also deleting it does the job too
then check it
or --reinstall-torch actually, iirc
probably not
I am running it, making venv
ah, thats a problem
yes, then activate that venv and check
hope its an easy fix
ah, alright
should just be going into venv and force downgrade maybe
it was just a problem to try to downgrade 2.x to 1.x, wasn't it?
yeah, that 1.12 is old
changing the version in requirement_versions.txt maybe
I am just letting it do what it do
general awareness has yet to let me down on these mutual experiments lol
Pytorch is never implicitly mentioned only pytorch_lightning
They tell me how to do it, and I let them test vicariously through me haha
yeah if you don't specifically need 1.12
yeh lotsa paths to get there
All this shit is very restricted to a single version of things
Know what that says? they are hacked versions
pytorch is called torch in python, for some weird reason
I still don't believe it can be that fast 😅
update breaks the hack like game cheating
it shows "Shell infrastructure host" using a lot of CPU, not command line
is that normal?
i was reading that you can just make a new virtual environment in the same folder, call it something else, and point a copy of the launch at the new virtual env. but that's still out my level with python. i'm grinding to get there
I know nothing about code. Its something that has never clicked for me
luckily, I can rely on other people to help me understand the little I tend to need
what version exactly do you need of pytorch?
1.12 cu11.x
oops
that cu11 that isn't even listed on pytorch.org
I paused the command line lmao
thats why it was frozen lol
there we go, now python is moving
see, 11.6 was the last version of cuda for that ancient 1.12.1
so I am gonna have to downgrade cuda?
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116 try and see if 11.6 changed to 11.8 even works.
yes
this is why I hate fucking with this shit as it snowballs
alright so run that exact command when the venv is done?
if you have the 11.8 sdk installed, it should work fine with torch's 116
yeah
should and does is two different things
ahh, ffs
this is the third time this has happened with my SD installs lol
do we remember the fix?
auto is a big pile of shit but it is the only real shit we have yet
I don't
I keep having to redo venv a lot lately
it fixed it before
I really wanna get sub second gens for the hell of it haha
it fixed it :>
no more middleware error. Tis a good day
you can see what the ui uses by scrolling to bottom
yeah, do that
alrighty
huh?
the copy is saying its 230 GB
venv is 230?
lol
yeh the venv is what is operating really
inside root activate that venv
venv transplants are fun
once it is activated
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116
now those last two I do believe is in the requirements and he will screw you next git pull
I should be fine as long as I never update this install, right?
not via requirements it will not
lots of small files
its saying 0 bytes/s
and I have super fast drives

its a full speed gen 3 nvme
its saying 3 hours
dies dead
sighhh
pretty sure the copy just crashed
oh nevermind
it just froze one it got to the parts to replace
ok, this is at least faster
my drives are 4000/4000
ok sweet, the rest got up to 1.8GB/s lol
its done now
lots of small files. probably was determining how to cache that and you had just stopped the bigger ocpying op, might've been cleaning up after that in the bg still
@dense tapirok, i give up
even copied the venv
pytorch 2.0.0
its saying my main install is 2.0.0, which isn't possible
cause I never updated
bro
@dense tapir your command is pulling wrong info
it just told me my main install is 2.0.0, but when I ran it
did you activate the venv
its 1.13.1
as in run the bat?
no
oh-
"./venv/script/activate"
yes, but my SD install I copied over doesn't
its 1.13.1
correct
and your command even after activating it, is still saying 2.0.0
even tho if I run it and check it, its 1.13.1
that is why python uses virtual environments.
well, no idea but that is the command to check it
other one is to open a python teminal after you activate venv and type
import torch
print(torch.version)
import torch
print(torch.__version__)
or whatever. I despise formatting
this is proving to be a lot more tedious than I originally expected
there we go
alright, how do I open a python terminal?
ok, I'm in now
checking
I don't know man
still saying 2.0.0
when its not lmao
this is the install its saying is 2.0.0
assuming it was 1.13.1 like the program actually running says, what would I do from here?
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116
ok, i tried that and it says it needs an additional arg
well, wtf, that is directly from pytorch.org
now you run my previous commands it will say 1.13
weird how it just worked previously
also, the extra arg it needed was the URL lol
I didn't realize I needed them at the same time
it uninstalled torch 2.0.0 :>
yep
yep
ok, is there anything else?
beats me
yes
alrighty, lets see
should be the moment of truth I think
ok, its done
do I have the tensor RT now?
I think so
do they have anyway to see/test?
should be on their site some way to see
this is the site
yep, that is the site
oh my god
dude
They have a "fast installer" now

LOL
torch 2 doesnt need xformers
not on torch 2
ok, looks like volta is not working
same normal speeds
gonna try their fast installer, screw this lol
ticket it then because I suspect you must be be on 11.6 not 11.8 as a lot of this shit is not backwards compatible.
Ive been using torch 2 for a little bit now and instead of xformers I use --opt-sdp-attention
yes, run that
that Im not sure of
it says NVIDIA GPU with CUDA support
yeah, T4 listed as they told me it would be much faster around 60% on it
A100 is the 400%
great, fucking docker
grab it
xformers saves memory too I thought
I never have, no.
yeah the sdp attention command does the same thing
for me it allowed me to go from 768x1024 on my GPU to doing 2560x1080 gens no problem
oh, cool. I wonder if it is as unstable as xformers?
no issues here, but I dont train so not sure
even the devs of it on FB research said it jitters and can't be helped
like my gens arent visibly different
mine shows 1.13.1 still as well, even though I am on 1.12.1 now
2.0 sucked and made me so slow as I was already slow I ditched it
you have to rerun it all
what Im using now
python: 3.10.9 • torch: 2.0.0+cu118 • xformers: N/A • gradio: 3.23.0 • commit: 22bcc7be • checkpoint: 639d0db70f
why torch 2.0.0?
its slower
even auto himself said its slower
until he implements it
the difference is so small, so I left it installed
even the torch devs said that
once auto implements it correctly it will be a nice gain
for me it was like 20% slower
10.5 it/s to 8.5 it/s
really its almost unnoticeable when I tested
brb
I went from 20 to 32s and 40s to 56s (512x512 and 768x768)
maybe 5% if that
you must have a fast card
3080 10gb
1060 here sytan has a 3060
Grrrr
Windows updates
My bane
I have a *3060ti :p
The not slow one, but rather the VRAM cucked one
i installed xformers and now all is much slower (1070 Laptop) i delete the --xformers entry from webui, but its slower even weith out activated xformers 😦 someone has an idea what i can do? the basic 1.5 was faster for me.
I heard of that because I had that when auto first came out. I lost a smidge (like 2 secs) with 1.5
Ok, nevermind
This is just asking way too much
It needs me to install docker, but in order to install docker, I have to install Ubuntu, but in order to do that I need to install some virtualization in my PowerShell, which required me to restart. But now in order for me to utilize that virtualization, I need to turn off my computer and enable the virtual machine platform Windows feature in my BIOS. I don't see what part of this is an easy and fast installer lmao
Oh waitttt
I messed up
wtf? You have to be using wsl2 on windows to install this shit? ffs
I'm running the local install, not the docker install
Yeah, it was my bad, I went over to a different installation type that utilizes Ubuntu
I thought that was the link in order to install docker, but I clicked on the wrong hyperlink
Man, it's just an exe lmao
I was doing so much extra, and for nothing lol
see, I would love to install this on colab for training because the t4 has tensor cores due to it being a Turing based card
I am seeing if I can figure this all out
I am trash at code, but I am a damn fast learner, so I am following the guide
yeah, fuck that shit, and besides I can't use it locally
worse comes to worse, its just a program I have to delete
no clue
meh, apparently not
this is one of the errors I get related to OOM on the vramRuntimeError: Given groups=1, weight of size [320, 320, 1, 1], expected input[1, 2, 9216, 320] to have 320 channels, but got 2 channels instead
I X it out and restart and can go for awhile
oh weird
@dense tapirMan, this is a pain in my ass
Yep
i absolutely love having no info on all of this
It just says to set a username and password
and then says nothing as to how to do it lol
use all lowercase for username
yeah, i figured that out eventually, no thanks to literally ALL of the 6 pages of documentation I looked at
@static tuskPlease tell me you know how the hell I can paste something into linux
Cause I can't figure it out, and this is an annoying command
the limited usage ive had was just ctrl v
I wish it was so simple
shift insert I think
nope, that plays an error sound
shift insert, control insert, right click, control v, control shift v, they all don't work
ctrl shift + v. if neither then windows is fucking around
then windows is fucking around and why I would never run wsj
wsl
anything like this on the window?
thank god
thank you
@static tuskTHANK YOU
right lmao
99% of the crap people claim to be "how AI actually works" is just BS thats perpetuated to make more people angry
my ex boyfriend is a digital artist. He makes all money off coms, and he was sooo mad when I said I was using SD. But now that he has seen how it works, and I explained all the rest, he constantly asks me to make reference images for his art
once people know how it works, they typically chill out lol
@dense tapirI GOT IT
sweet
way too complicated. test it yet?
Currently trying to figure out where ubuntu is located on my PC
@dense tapiroh goddddd
it can only use models it can directly download from hugging face
ugghhhhh
is there an compare about gpu spped from diffrent model avaiable?
speed
especially for sd 1.5
for a valid speed comparison you need to remove 512x512 from your mind and concentrate more on 768x768
3 times slower about
for me
768x768 isn't much slower for me
the system info extension has its own benchmark and results page but Im not sure how accurate it is https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html
SD WebUI Benchmark Data; Author: Vladimir Mandic
my point is still valid cause 1024 is coming where even your card will cry
@dense tapirwoah, VoltaML has samplers I have never seen
PNDM
I have seen that before
DEIS Multistep
dunno that one
LMSD
LSD?
UniPCMultistep
M is the new shit
here comes the moment we see if this was worth it
they were discarded by most
0.9s/img
👁️
not as fast as I was hoping, actually
WAIT
ITS
THE SAME SPEED AS A1111
BRO
better than nothin i guess. Big THX°
I'm getting 10.7it/s on DDIM
I just went to see current prices for a 4090 and FE 2300 USD. that is 1600 MSRP
like a real
Wonder what it tastes like?
heheh
alright, so looks like VoltaML was a complete and under failure heh...
its no faster for me
sure all the dependencies are right version?
I bet not
Its an autoinstaller in a VM version of Ubuntu
there was nothing for me to interfere with
slow cpu could be the cause maybe
that as well
I do have a slow CPU, let me see
like me and a 4090 I would bottleneck it all to hell
that is not how you know
you have 16gb?
32
doesn't matter if 1% usage what matter is how slow it is getting the shit back and forth from the stuff
it says very slight at the bottom you need to manually install xformers
WHAT? lol
yeah its single thread performance
well, it is taking me 2-3m per gen now
3m for a DDIM 30, 2m for a DDIM 20
I am really about to toss this shit
even dpm++ sde karras 8 is taking a min
I like that one and the colours.
after installing xformers a day ago i have same problem
do not load is no solution, its so damn slow now
by using "token merging" I was able to improve the performance twice
but for some reason the results get worse when the "--medvram" flag is enabled
I love this image lol
ah yes, download more vram
stable diffusion online cuz i cant
how the fuck did you make this with online tools
oh mah gawsh
looks right to me
I love the lighting in all of these
AI was a great idea to help humanity to be smarter, then it was released on internet and had to work with "Make a studio photo [[tiger|hamster]-[bird|human]] painting".
Master prompt No 346; "studio photo of a [[tiger|hamster]-[bird|human]] painting by Salvador Picasso".
Seekreet prompt recipe that no one will be able to guess 🙂
DownloadMoreRAM.com - CloudRAM 2.0
Can it be a renamed version of AnyPastel? https://civitai.com/models/15024/anypastel-anything-v45-pastel-mix
I do not own this, i am uploading this here to get it onto civtai for limited use. Owner's link can be found here https://huggingface.co/m4gnett/an...
The seekreet... censored.
at huggingfaces or civitai
you think i didn't search
read my message again
.
ah ok
You can Bake the vae in yourself with the checkpoint merge tab
sory misunderstood it
ah i understand
so he merged the vae to the checkpoint?
It seems so yes
No problem 🙂
English castle scene
Create a logo on the topic Logo for bullet resistant window and door company Logo Text
BallisticShield Aluminium Windows & Doors
^^there is no bot here, u have to install Stable Diffusion local or use the website
Currently, there is no bot on the server that generates images. 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!
Is the only way to do inpainting on Midjourney images,, is to download them into the img to img and do that?
img2img => inpaint , you can mask any areas you want
But you'll probably have to find your visual style model, since mj has it's own style thing going usually.
mona lisa is not overtrained in SDXL like she was in SD 1.x
theyre blurry cause i used 20 steps btw
Do we have more info on that btw?
like...how much more samplers need
no, I was just stating what I observed by doing a lot more higher resolution recently
I've been doing 1024x768 with 25 steps and it was just fine I think
I have had to go up to around 50-60 for something good to come up, when doing 960x540
on DPM++ SDE Karass
well, our millage seems to vary x)
What sampler was it?
weird, I thought it requires less steps then anything else
yea but euler a doesn't converge
it doesn't ? I tends to a given result though, it converges on something, it just doesn't start doing something else after step 100 for example
lemme find the thing again...
yea that's true too, I noticed it not doing anything after like 25 steps
on anime models even 20 is fine
that's what converging means usually, going to that point where any more step doesn't bring change
