#🏞|general-with-images
1 messages · Page 94 of 1
@smoky oak this is me rn
As you do

Currently struggling to communicate with the Mexican side of my family here in mexico

I will end your bloodline
i will go marky-mark on yo ayuss
i'll like, flex in the high contrast street light illumination with some sweat drops catching the light just right and my hollywood-inspired physique will distract long enough that i can disappear into a puff of fart cloud
like Batman
You genuinely just gave me brain damage. Continue with that and I'll probably forget how to breathe and suffocate to death
that's my next level-up power 
inapinting chickens into my home
it's glorious
the true meaning of AI
unknown bird has now claimed my bathtub as it's own
the chicken army is taking over
monday morning.. might be time for coffee #2 already
someone needs to make a mr burns excellent with merged fingers as an AI meme now!
i think i scared everyone off
@smoky oak the 5800X3D i upgraded to in my ML rig really flies during the aspect bucketing and image processing stages
Just exported a ton of great images off of my computer
here’s one
And a different one
Made in Stable Diffusion, if you can believe it
which you probably can because the image is not great pixel art
but it’s pretty cool that we can make pixel art. I’m using, I think, the pixelized image addon
How to fix?
try #🤝|tech-support
I am up so upset right now lmao
I know this has nothing to do with you guys, but we are in Playa Del Carmen Mexico right now, and there is a small store where there are itty bitty baby big cats, and they are so fucking cute, but they are charging $60 PER PICTURE
And you can't touch them unless you pay.
They are so small you can lay them on your shoulder, and they are so cute and floppy and playful
They have a black panther, a normal tiger, a white tiger, and a golden tiger
And the guys are all assholes to you and tell you to move on if you aren't willing to pay
probably better not to support that anyway right? there's a lot of animal cruelty in exotic animal tourism
can always get SD to create cute pic for you 😉
ew that's some tiger king shit, they probably kill off the cubs when they get too old to be cute for $60 pics
that's what happened at their zoo anyway
crazy how there's more tigers in captivity in the US than exist in the wild at all anymore
My only hope is, that it is in a very public location, and there are a lot of resort circuses around, and if they belong to any of the adult big cats in their local shows
That would make me feel a lot better, to know they are their cubs and that they get to play with people while their parents perform
yes sure 
If they belong to any of the paradise hotels, then they have very good living conditions from what we have seen, huge enclosures, lots of shelter, better than most in zoos
That'sy only hope
i hope they just get returned to the wild and the place shut down

tony tiger approved
don't want to make you too depressed, but what you see is just the cost of doing business with rich people to make them feel okay about it. you don't see what really happens out of sight. what do the same people do with the animals in attractions for the local or poorer tourists? most people are ignorant to it unfortunately. unless you make it against the law and enforce it, people are going to exploit animals for money. we are lucky in US and Australia where there a lot of laws against animal cruelty and we can afford to buy free range animal products and what not. just not the same elsewhere 😦
Free range has been shown to be a lie too caught on video. If money is to be made there will be happy thoughts marketing to cover up what is really happening behind the scenes.
@smoky oak i'm sorry we can't enjoy your nice tiger story lmao i hope you're having a good evening
It wasn't a "nice story" by any means, but you guys sure did a great job of depressing me on the matter even more
Just trying to get my mind off of that before I spiral so yeah, I'd like to move on.
well i've been making lots of deformed images again
do you want to see some
its a painting
no its supposed to be a photo
hipster harry potter lmfeao
the meeps start kickin' in at the perty
this looks really pixelated on my remote connection, but hopefully not too bad
i don't see the pixelation here
full meep
i bet that looks great to you @sterile temple
Kinda looks like Steve Jobs
iWizard, hairy
cartoonized portrait of harry potter at a 1990s house party in bokeh, depth of field, professional photography, high resolution 1920x1080, highly detailed, intricate detail, portrait
tried to make him handsome lmao
how many fuckin jackets is he stealing from this party 
If you can get men's clothing to work, that would be like solving the hand problem
Buttons or button holes on both sides of the shirt, double collars, sleeves that stop and start again, watches that are just a mess, weird studded earrings, random bag straps, etc etc
sounds like a job for a negative embed, unfortunately
done
this model is so hard to use but dear lord the fucking wide gens can be funny
pure magic
Tripped on the rug
That feels very midjourney v3
@midnight sedge
tried that also
I even did a soft reinstall
One thing Im noticing is that in console menu it always shows it cannot find whatever VAE was selected in my drop down menu previously
hmmmm should i be concerned
it's a thing on sketchfab now 
cute
I think that’s a very selfish thing for them to do
Kinds weird generation but I like the style
damn, scrolling through some of your guys' images, yall have really advanced in your abilities. Crazy!
Ears not great but still like it
Now THAT is a true moral question lol
very good. none or all.
not quite right. either switch selection will end with all
a good approach would be to question the settings to begin with 🙂
Does anyone know an OpenCL-supported method of running Stable Diffusion?
do your chain hang low?
ask a more specific question
what are you trying to do
That was a very specific question. Lol
A method of running SD through OpenCL.
Hopefully a WebUI.
Who knows.
sorry, i am a developer and it doesnt make sense to me
are you trying to use intel arc? or radeon? cpu? something else?
wont work
I was wondering if there was a repo or a WebUI capable of properly running SD on Windows through OpenCL.
What do you mean "Won't Work"
nope. pytorch does rocm and cuda
IPEX on WSL.
Or OpenVino.
Both of which work with SD.
But they aren't OpenCL-based.
Which was my question.
Lol
You say you're a developer but you for some reason think that Pytorch can only support two backends.
Lmao
i do not bother with windows
that is why i told you to ask a more specific question ahahahahahahahahahahaha
Entirely different distributions and compatabilities between operating systems.
good luck kiddo
all I know is that OpenCL has been falling out of a lot of programs in favor of Rocm, like blender
They removed all OpenCL from blender for just ROCm
Kinda interesting because OpenCL 2.2 bundles with ROCm.
In that case I am not sure, but I do know blender doesn't use just OpenCL anymore, and has no support for it like it used to
opencl is obsolete
apple got rid of it too
amd lost it
nvidia doesnt care about it
You used to be able to run hardware through OpenCL, CUDA, OptiX, or none
And now it's ROCm, CUDA, and OptiX now
Tho I am not sure why they support just raw CUDA anymore, as OptiX runs better even on non RT/Tensor cards
theres possibly a way to use vulkan. but no one cares about arc. i dont think any higher up pytorch devs even own one
optix is only maxwell or newer
isn't Maxwell 7xx?
No idea how the hell this server STILL doesn't have any spam protection, like at all
selfish Dan
Wonder how long it will take the mods this time
huh
Can't even read chat now, great
its probably Sen
the mod system here is weaksauce
He’s gone
That's a generous understatement lol
oh yeah im sure they wont come back
Time to once again set up SD in wsl
well thanks to whoever banned them ❤️
i got banned from a twitch stream because my user name was "sus" to say out loud. it is Nibbloid ahahaha
i didnt get it at all, everyone else was pissed
Hm?
Clyde is a pathetic bot
state of the art tech
Ah.
Maybe stank of the arch
dank of the starch
Cornstarch and water
filled an olympic pool with that once
we got to run across and see who makes it further
I can only imagine how that went
yeah was the indoor pool for the high school as a fundraiser
aquatics
we also made cardboard boats and raced them on the water
Careful, might ruin their water infrastructure, doubt it's any better than their nonexistent electrical infrastructure 
I love swimming. Running was great. But slip and slides have learned me that I....would not make it on an oobleck surface
How old is old, really?
Older than the other party
You don't have to be old to have that haha
My bones have ached since I was born
Kinda same
But I'm an outlier
Ah, genetics
A lottery of questions
If you're over 40, your bones start to ache
If you're over 50, then they really start to change with the air pressure and weather
If you're over 30, and and work manual labor, see 40's and 50s
I'm in my thirties and have an adjustable bed and a lazyboy recliner 😁
and a chair lift to take you upstairs?
I want a chair to take me upstairs lmao
No upstairs in my house
We just upgraded to a single story house
upgraded?
From a basement?
from a shed?
I have a........dollar store adjustable bed, aka, this giant pillow behind me with arms
From a 2 story house. 2 story houses are not good at all IMO lol
They have like 0 benefits
thats a hot take
There was no top floor--only a basement! (Congrats on your better situation!)
It's only hot if you have no AC and are, in fact, upstairs
Want an inconsistent temperature house, where the upstairs is too hot, and the bottom is too cold?
Get a 2 story house
Wanna be massively inconvenienced any time you need something from downstairs? Get a 2 story house
Wanna have 0 chance of automated home cleaning? You guessed it!
in a single storey house, the downstairs becomes.that hot
Get a dumbwaiter and paid staff
no one here has ac
Or train and animal to get said things for you, naturally
I will never happily choose to go back to a 2 story house, there is no reason to, it's just a waste of knees
Ah, the joys of getting older
he's a teen lmao
Owning a home isn't really a viable option for most anymore, not getting my hopes up on that lol
Look here, fine, we can all build tree houses together okay
they will log the trees
If it becomes Bladerunner, I guess they'll be metal trees, but we'll make do
and if you thought a 2 storey house was annoying to Sytan, wait til he has to haul a bucket of poop up and down the tree
i guess not up the tree, only down
the bucket will be empty going back up
No, no, the outhouse has to be downstairs
Like the outhouse down the road from me with the solar panel on it
you have to have some way to charge your cellie.
Before the obligatory roaming planet charges kick in 🤣
We already explained you that we have spam protection iirc, but some stuff just slips through
Unfortunately, this is true. I stepped away for two seconds, and something happened--such is life.
Esp on mobile
almost wore a groove into my phone screen
I'm just not gonna interact with this conversation before I get heated again lol
You know what's fun? When Discord gets stuck in a loop, and it thinks you're trying to open the panel to change channels, and it just opens, closes, open, closes.
Now that's a fun time.
Damn, that's a new one, never had that one
It does it to me all the time.
Never fully settling on what panel it's going to open.
Terribly exciting, really.
Very strange. Discord is likely to be the worst piece of mainstream software I have ever used, but never had that specific one. That's interesting
Tech is difficult.
I am having to use a very outdated discord version on my old and buggy phone, so I don't hold anything against discord at the moment
But back to our tree houses!
Tech is difficult, but discord is inept lmao
You've never used Teams?
I'm certainly having a garden.
I have, a markedly better and more reliable experience than discord IMO
Discord has had the same problems for over half a decade now, and we have all just kinda gotten used to working around them lol
just wait til you're older and you've experienced even more trash than Discord 😄
pretty sure others my age in here were first on IRC
don't be mean to blender, it is open source
You mean...Treeowners associations?
Blender gets no passes from me, free doesn't mean overlook huge issues IMO
you're able to fix them just like anyone else can, which means you don't get a free pass on blender 😛
Blender?
I do like parts of blender, but man they are all over the place
As in the 3D software?
crack open the source code, sign the CLA and get to work!
That is true, but it's not my job to do so, so I have stopped using it for the moment
that's why it sucks, it's no one's job to fix
I just don't agree with their way of doing things at the moment
It's the job of blenders paid programmers lmao
they're getting paid by users to work on stuff those users care about
They are getting paid by the blender foundation, and they work as salary paid developers
there's a lot of money backing blender last i checked
A shit ton lol
blender foundation is a big non profit
they're paid poverty wages lol
Yeah, they get donations out the ass
I will agree with that, yeah, but still, it literally IS their job to fix blenders huge issues lol
wow. don't be a google expert. That's litereally not even the first search result. It's google's algorithmic response
And of course I don't dislike blenders parts equally, cause some of their sectors have their shit together, but lots of others do not
iceycold you weren't even here for most of this convo 
don't show up late and start throwing jabs
oh i didn't realize we were in a private convo and not #1072236442463518882 okay sorry mb
@limpid lichen I like your PFP BTW
MHA
Blender Software - Salary - Get a free salary comparison based on job title, skills, experience and education. Accurate, reliable salary and compensation comparisons for Canada
Yeah, those are much higher pay lol
I saw less than 30k when I googled it, so that's already a huge increase over what I saw
man at your age you really need to start having a better understanding of pay scale
lol those sites are always skewed lower because they keep data going back over a decade
40k is nowhere near enough, especially in vancouver or california
Yeah, I am not saying I agree with their pay, but it is exactly their job to fix blenders bugs
200k is like required to live in california
No need for insults, I know it's bad pay, but it's still a job
noone getting paid 40k in vancouver. That's full time minimum wage
it's just how life goes, no one teaches young people this stuff and then bam, you're taking a job for $35k because it feels like enough money, and it really isn't. i ended up in that situation
you took something i said personal. I don't regret that at all because i don't know why you're irked. You need to figure that out. moving on
ok and what's the goal of that then
No comment lol
why say that
same goal as this perhaps. Worry about you.
no need to be angry because you're young and inexperienced. be grateful people are explaining to you in advance lmao
Agreed here
oh so now you're just being a dick for fun i guess, all on your own, to show how good you are at it 🙂
Your advice isn't asked for, and it's not even advice, I have worked jobs, I make money actively, and my parents guide me up u til my father could take anymore. I come from money, and we know how to spend, save, and protect our assets, so please, it's not needed
well i thought we were guessing each other's ages suddenly
I agree with iceycold here
age seemed to be a prominent discussion point
Pseudo is going back into his patronizing arch ._.
seems to be very passive aggressively condescending yeah
yes, it's a big deal when in discussions about life and money, knowing that someone is 19 and hasn't experienced jack shit in life helps a lot
And you don't know what I HAVE experienced
i know how long you've had to experience it, that's enough
maybe you're projecting your own 19 year old self onto them?
19 years is a lot of years
you literally can't have experienced all of the things i have. there's just no way
actually you know what, I do that need to prove myself for anybody, this is just a waste of my time
there's no proving anything anyway, just take the advice and move on lmao
complaining about blender when it's open source and you contribute nothing to it just really smacks of entitlement
Your advice is not wanted, or beneficial, but thank you :heart
Again, you know nothing of what I have and have not contributed to blender
the kind of youthful entitlement that goes away as you grow and realise you can "put up or shut up"
But sure, continue to assume, ANYWAYS
Chop chop chop chop chop chop
I have said I wanted this conversation to move on for a while now, IDK why you can't just drop things
So like I said multiple times now, moving on
Hey guys, let's be nice
Anyways, I am on a boat
don't know why age is relevant to blender. i imagine when i'm older i'll want software to behave particular ways, and not want to flex software engineering skills to get there
I do a float float
i'll probably just use a different software before i open a source code file
It's not, like at all, but some people forget that
So, let's pick a different topic, shall we?
i'm not much of a MHA fan, but this character has rockeet feet
I worked hand in hand with Drachen Seele for his Super Image Denoiser, which is massively successful, and I also have supported several cycles bug reports and suggestions for fixing them, and some have been implemented
I used to be soooo active in the biggest blender server, before their moderators went insane and banned most of the active members
And that kinda killed my love for blender, that and the fact that they aren't doing what people want, and are choosing to add stuff nobody wants instead of fixing problems people want fixed lol
I would love to mess with SD in blender tho
Rockets on feet sound fantastic.
Lol
see the little exhausts
See, you said rockeet, so I figured, mayhap, you meant rocket.
rockyeet
But approval nonetheless.
the anime sometimes call them rocket engines or car engines. he's got machines in his legs
Probably gonna use all data out on this island haha
Oh, I finally found a distinct and cool style I wanna train a LoRA on when I get home
I wish you luck
Well, speaking of work, I have things, dear folks, to get back to
Cheerio for the moment!
i've actually never really been interested in MHA much. It's a typical highschool orientated anime. But i love the idea of engines / rockets in legs
i need to watch fma i think. people keep telling me it has awesome machines
It really does, FMA brotherhood is dope
FMA is cool, but Brotherhood is amazing
Wow, I still have signal this far out i to the ocean haha
*into
Kinda crazy
Holding on to one bar
We are probably like 10 KM off shore
I have 2 bars now what
👁️ 👄 👁️
Ohhhhhh
The island we are going to is huge lmao
I couldn't see it cause it was at the front of the ship, they have skyscrapers and huge infrastructure, no wonder I never dropped signal lol
not working? like it gives an error?
That's weird, I'm not sure why
Any of you guys playing with MusicGen? Visuals made with SD
no, it's not open-sourced, atleast as far as i'm aware.
It is! Commercial use is permitted too if I'm not wrong. Training code is expected to release soon.
after a while of training on 1920x1080, i get just one rainbow now
before it has like 6
different directions and multiple waterfalls
1.5 fail
blinks repeatedly, trying to focus eyes
I like that there is a stable diffusion discord. I'm finding it nice to look at the #1103141787960344576 now and then (even though the pictures themselves are pretty boring)
I haven't worked on my project in the last week, but I was able to get outputs that look like babies on my last attempt (and I realized that the artist prompts I was using was unintentionally causing the character to trend older than I was prompting for)
I might try learning control nets next
compared to some of my first early diffusions in October, just 8 months ago
we were all working on welding baby photos a week back or so
nice, yeah I'm going for more of a fantasy theme
and I guess a lot of fantasy art is like, teenage characters
or adults
so certain fantasy artists unintentionally result in making babies very hard to generate
a lot of people purposely train babies/toddlers/infants out of their checkpoints for... reasons
I think that makes sense
Basically I had a notion that I wanted to create art of "families" of characters, including people who wouldn't necessarily be protagonists
but more like background characters/NPCs
you can try something like textual inversion which makes it possible to train a keyword on a particular set of images, to help the model find a vector that fits them
yeah, I was thinking of doing that, until i found this workaround (removing artist names). I'll probably come back to Text Inversion later as well.
I saw some great Loras that are trained on hair styles, and also control net poses, which I was also planning to play with soon
artist names are indeed pretty broken as you describe
and of course, still working on the holy grail "different color pants/shirt/hair"
very overfitted in many regards. and i would avoid them for many reasons, but that's just one of them.
Not my example, but I like this demo of the problem of attributing color to a specific thing
thinking of trying out regional prompts this week to see if I can get enough control
Sytan said (and demonstrated) that regional prompting makes their photoreal gens look like clownshoes
I'm a little worried, because technically you don't really know where the character is positioned. So at best you're guessing where the hair/body/legs begin and end
Do you have a link/image? might save me some time
if you use Compel you can do prompt segmentation
eg. ("a man with blonde hair", "a woman with black hair").and()
you'd have to look up how to enable that in Automatic1111
https://github.com/damian0815/compel
This thing?
transformers i don't think work with diffusion models
Tested and developed against Hugging Face's StableDiffusionPipeline but it should work with any diffusers-based system that uses an Tokenizer and a Text Encoder of some kind.
from the github.
But regardless, before I spend the effort to try to set it up, are there more examples? There is only 2 in the read me and it's not entirely clear that the benefit is worth the extra setup
especially if someone was demoing this recently, I'd love to see what they produced
tbh I think I tried this method the other week, and got mixed success.
zuckerberg, who is overfitted, can be dressed up as Belle from beauty and the beast, with prompt segmentation only
but I think I read a different tutorial about it, so I can check the steps here and see if there is something subtle I missed
you can't really get zuckerberg to do stuff he don't wanna do usually, without segmentation
you probably need a 2.1 based model for this to work properly. i did not have great results in 1.5 and i don't know why
howabout the t2i-adapter for color in controlnet?
you'd have to paint a color map first but it could control where differences should be
I haven't tried it, but if you recommend it I can look into it. I haven't gotten controlnet set up yet
or perhaps make a base image it and reprocess it using that net
it controls where colors on an image go pretty well
Here was one test I was doing, without any special techniques
damn even without zooming in i can see the super saturated hues lmao
switch your VAE since you're focused on Anime
i actually like the poppy vibrant colors. but yeah that's a vae thing not the prompt issue at hand
yeah in this case, the bright colors are fine (just testing whether I can control robes/scarf separately)
oh
(also, hair is green, so naturally it's 3 different colors in the prompt)
the base model dictates how well it listens to your prompt to be honest
last week I followed a similar tutorial, and this was the output I got (inconsistent shirt/skirt colors). usually got the shirt right, but the skirt kept changing
another example
out of curiosity, what do you do with these images? just wondering why the colours have to be so specific
right now, nothing, but I'm trying to work on a system to create characters with specific color schemes for a game I was thinking of making
And I find it interesting if I can make it reliable, so that I don't have to manually check and re-generate images many times
(because it's a game, bright vibrant colors are fine, and even encouraged, so you can tell characters apart)
@hasty nova does game dev with AI assets
yeah I made a board game this year (basically just for friends) using midjourney, and it was fun. So thinking of doing something for the internet that more people can enjoy
tingdadengdada
sounds like the old movies
So how do I stop them from looking so similar?
In Stable Diffusion 2.1, it seems, every woman has the same face.
you mean stabilityai/stable-diffusion-2-1? or a fine-tune?
because i can say with certainty not all 2.1 models are the same in that way. not even close
The SD2.1 model in Playground AI.
if that's base 2.1 i can say you have incredibly good results, clean images like that are near unheard of
well i have no idea what they're doing to that model there, haha
There's a filter called "hyperrealism" or something.
I guess it always gives such results.
I'm not complaining, but the faces ARE very similar.
makes me wonder if it's a LoRA
At least it can also do a Korean one.
i've never heard of Playground AI before. it looks pretty. but probably overly restricts results by pumping through LoRAs and embeds so that you don't get "deformed" results
If you've never heard of it, try it.
i don't like giving all these fly-by-night companies my info 😛
same. my mindset: if it's not local, I ain't using it
yep and it just helps them generate new models using all that data they will never make freely available to others
parasites
like Midjourney is
well, that is how SDXL is currently being made, but this time, it goes open-sourced after finished cooking.
lmao sdxl crew isn't going to make their dataset available, they never do
it's too much of a liability
the model will be
that's the whole point
they'll tell us "we used subsets of LAION x, y, and z datasets"
which subsets, we'll not know
you can't qualify a model as open source if you don't get its dataset
that's why OpenAI's CLIP isn't open source...
they're just "available weights", not "source-available"
it's a problem for large scale fine-tuning if you can't validate that you're not adding duplicate images it has already seen many times over during the pretraining
yeah, i wonder how will finetuning SDXL will work
i'm sure most people just want this so they can find out if they can sue SAI
the same as every other model for the most part, but they're promising to actually create tuning tools that are released with it
huh, but isn't SDXL going to like, run VERY slowly?
it uses UniPC sampler, so, unlikely
SAI said the requirements for it will be 12gb vram, which is exponentially higher then last models
they're oscillating between 6gb-12gb
i honestly don't care if users on low mem GPUs can run it though 😛
so, it will run in the same speed for the most part? that's odd, since it has ~2.5 more parameters than previous models
it would be an improvement to the model for it to not run on low vmem cards imo
according to the SDXL devs, focusing on that is causing damage to its ability to generalise
still, doesn't that mean less it/s?
I can’t use it locally because I have literally the weakest laptop.
skill issue, i was talking about people like me for the most part.
huh, i just made this with the SDXL bot, and this looks like a VAE problem
one of the outputs has odd settings as they're testing stuff
so, if the other image is fine, it's fine.
both are like this
meanwhile i'm getting better and better results when training on 1920x1080
nice
looks like they cooked the hell out of the firearms while trying to fix the ability for it to create firearms better
whoops
or maybe, that was just my prompt -_-
Hey guys, I've been installing the OpenDiffusion project and it seems to be running, but I can't work out what the internal IP it's using is
anyone know how to find it
other than the pants messing with my brain to think it's some sort of flag for some reason. Is the bleached look intentional?
idk, look at pictures of mars lmao
the hyper contrasted look was not intentional, if that helps
i prefer mars looking real and washed out. mars in real life looks like total hell
so, this will be like how it is with llm's? with llms the only performance change caused with bigger models is that they take way more time to load, generation speed is the same
it is, or maybe im trippin'
here's another result of that washed-out checkpoint, which has plenty of colour and contrast left
Feels like eons ago. My first attempts at making body horror more palatable.
What are you lot up to?
Just got done with a long day on Isle Mujeres in Cancun Mexico
Only very mildly sunburnt, interestingly enough
The uv even here is like nothing compared to where we live lmao
Oh Lord, these look like a huge step backwards, what are you doing? Haha
we were talking about regional prompting earlier, have you tried it? what was your results
I just finished "a couple" portrait tests. And now it's time to go to bed before I fall a—💤
i'm trying to wake up and start my paid job
I have messed with it, and found some workflows that have yeiled good results
I have used it for anthro/digital art, and for realistic images
Realistic images are way harder as there is really no way to get them to look realistic out of the box
As you can see here, this is what realistic images look like right out of regional prompter, and they are notttt pretty
Not too sure why it nukes realism, but it sure does
After some correction and a small workflow, it got this tho
So yeah, it can be useful, but it can also be very limited
No, though I am interested in that extension. At first I thought it was regional as in dialect non ENglish so never touched it.
Ahhhh lol
Still not 100% sure what it is just a general idea that region means a region on the screen.
Yeah
So there are different ways to use it, attention, and latent
Haven't been able to get anything not trash out of latent lmao
I don't think it works for 2.1 though
Wouldn't be surprised
2.1 is kinda abandoned for all the cool things
Its kinda become the bastard version
Few people want it, few people use it
And this, few people work on it
*thus
hey guys.. i have a problem. I am really new to this. I am trying to do a composition which a tribal girl jumping on a big bear. I rendered images seperate then bashed it at photoshop and tried to render again to weld them all in.. but no matter what i do it looks awful..
this is the best result i got.. using image to image and softedges.. any ideas?
it just doesn't need those things like 1.5 does... 🙂
also, people pay to use 2.1 on BlueWillow v4
Makes me want it even more. 1.5 just never did it for me and still doesn't from what I have seen of the photo real images. I bet you SDXL is not going to go the way people think solely based on how much vram it will require then the tools for it too. I can only imagine control net with it. Of course the original author of it wants nothing to do with the new clip so might not even be made.
do you get bored and then just have to login to discord to complain about 2.1 or something
what is a good model for celebrities?
realisticvision and a nicely done Lora
what gui should i use im using makeayo rn is that okay?
play around with different schedulers and noise levels, it might change the image to something similar but a lot better looking. you can add the new image as a layer over the old and then erase the parts on the new image you don't like
will try.. ty
where do i get checkpoints?
huggingface and civitai
civitai, home to the worst shit on the internet
civi gave me a safetensors?
yes
is that fine?
the preferred file type
where do I put extras?
what's an extra?
idk something in makeayo
but ive switched to a1111
now
its like pictures that the ai follows to more accurately represent something
i think
The devs for SDXL already confirmed that you can generate 2048 X 2048 images on 8 gigs of VRAM
I will believe it when I see it as I believe nothing they have to say. Once burned twice shy.
whens its eta? like october someone said or was tha tjust a guess here
Thing is how hard will it be to train on it? As someone told me, and I concur, if it can't be trained on then they don't want it.
The preview I saw of SDXL was not all that good so my hope is it is really great when released AND we can train on it as easy as 1.5.
The latest gens coming from the bot look really good for a base model
where can u find good prompts?
I don't wander into those now and just wait until released
scientists: "AI is dangerous"
me: "yeah right"
AI:
fuckin text encoder is drunk
i don't understand waht happened to the woman on the right side
is that a visor hat?
whats codeformer?
a face detection and fixer
ahhh
im just messing with settings and was wondering why it stopped at 98% lol
My guess is august-september
but i turned the face thing on
if they plan on doing all the things they said they would, october is optimistic
Emad keeps counting down from 5, today is 4, on his Twitter. he's being cryptic. it's probably because the thing being released is DeepFloyd's stage 3, and he knows no one really has interest in DeepFloyd anymore, other than researchers 😛
if it's as good as the dog photo he posted today shows though, i don't care much about SDXL's release window, DeepFloyd-IF will be plenty of fun
For loras do I need to create a lora folder?
they will just release it unfinished is my guess
as its meant to be a stepping stone to 3.0
diffrence between batch count and size?
well, it's working pretty damn well now. all of the duplicates seem to be disappearing from my validation prompts
where is clip skip on A1111
it's so damn weird to not see multiples and multiples
settings > stable diffusion
feels like i'm about to get insulted by this android
ty
ok god damn that's so good
@oak osprey I really hope that Deepfloyd gets better, cause as is, it's roughhhh
why did no one tell me to train on 1080p before lmao
i was looking at it with some of the other sdxl testers and it really seems like once that stage 3 upscaler is out, the whole world changes
you can use deep floyd's stage 3 upscaler with 1.5
it takes 512x512 to 1024x1024 (or non-square images)
I am now afraid to buy a 4090 for any model.
Tho it will be interesting to see if stage 3 can fix the nightmare that is stage 2, cause stage 2 looks dreadfully bad
The connector is bad and even with cable mod burning up.
Jayz2Cents showed the issue and I can't believe they did it that way with that much current.
cablemod concluded is was a pebkac issue
Nope, they did not
I'm pretty sure they did. It's been a problem since 3000, you need to seat it properly
Jayz just said what the conclussion is and damn nice of Cablemod to do what they are doing.
did you watch j2c video a few hours ago?
an hour ago, yeah
even when snapped that damn wiggle
all because they used 1 clip instead of two (one on each side) or like cars use where the clips are wide
still requires some force pulling it away from the connector its not just wiggle
you can see jay's fingers pulling on it to get it to wiggle
which is emulating the force of the cable. Remember that cable is heavy going to it so exerts force. I am sure they will fix this with a new standard a V2
all it takes is to have a wider clip or two and done
Pretty sure I remember this same thing with the same connector on the 3000 series, and people realized you just have to not be inept lol
its a skill issue
It is the amount of juice causing it so you can't be as "inept".
I abused the hell out of mine and never had any issues
@smoky oak 😮 😮 😮
Nice
It is only the 4090 having this issue because of how much power it is pulling, and as Jay said a 4080 or less will never see a melting connector.
all good now
but the contrast was improving to a point and then got stuck
now it seems to have received a punch to the face, in a good way
this was coming out as a car wreck before the contrast fix
literally
two different cars all wrecked the fuck up
well, i've been training it on widescreen at the same time
so i'm not sure which fixed that part
I know some idiots I read have their 4090 and the cable comes out and they have to force shut their side glass panel. Yeah, buy a new case at that point or get ready.
Sytan how long til the widescreen training splits this family up? it already removed like 3 children
As cool as the 1080+ base gens are, I still don't see any reason to use them personally
this is 1280x720
Regardless, not sure I see a benefit
zero shot widescreen image gen without duplicates? lol it's awesome
I can already do that on my own, so IDK
Again, I see no benefits in generating over normal base res
Although, I could see a benefit for upscaling maybe
i mean, it lets you do hires fix better if the model is aware of high res gens 😛
hires 
oh weird, the "group of 4 handsome men" prompt now just shows four men
well that's a benefit 😄
apparently you can only get that if you train the text encoder and the unet on high res
equivalent image from training only the unet on frozen text encoder
jeez louise
just catching up a little, interesting emads doing a countdown i guess. hopefully something cool at the end of it and not just an announcement
If I already have the voice, how do I pass it to an avatar and have its lips follow the voice? is it made with AI? What tool?
deforum
That's just a single generated image with some animated eyes and and a mouth on top with extremely rudimentary warp framing
Yeap but what ai for the lip movement with voice
sweet extreme bokeh olympics fuck
fart noises
don't forget to apply your shrek butter before bed.
HiRk Jark
just dropped an update to my model training guide ppl - "STYLE TRAINING [precision method]" - https://civitai.com/articles/397 take a look and feel free to ask the questions.
Too bad for 1.5 as that never gives me trouble to train in.
skill issue
tried over 100 images, still no text on it
pew pew
can i make this guys face as a marathon runner?
like this picture but with his face
lol getting there
That's not bokeh, that's blur 😅
Extreme blur Olympics
@echo mountain
😁
it's brokehn
i can do a single dove in the desert now though
the bokeh was because i asked for it a lot in the prompt
https://www.youtube.com/watch?v=5mhk252Zqm8
me, defending my obsession with model training to my colleagues
my masterpiece
Oh my God lmfakoo
These images all look a litttleee crunchy on the contrast, is that what you meant by cooked contrast?
honestly, looks like a VAE issue, but it's probably because of CFG scale
Other than that, they look phenomenal for 2.1 images
high CFG scale tends to look overcooked, i bet that's what's causing that
yeah it's not good
the darks aren't black anymore, same issue SDXL was having a week or so back
i adjusted my training data and my aspect bucketing code shit the bed and i realised ohh i had a pretty big issue with how i was doing the cache so.. i can't really get a new ckpt just yet
SDXL is pretty much a masterpiece, do y'all think it will release in 5 days?
well the contrast being over-driven is a result of the issue with the blacks not being black
the overall image darkness is averaged to gray
What happened to you using SNR?
there were code changes and it broke apparently
Ahhh
can I try sdxl in any way?
The bots, but it's not a very good experience
it's a similar concept to MJ, many people like that interface
but yeah, something might happen in 5 days, last time Emad tweeted numbers was that number amount of days before a release, so, who knows, maybe we could start finetuning SDXL very soon
needing a dev level opinion, any reason why the command line --medvram is slower than --lowvram?
embarrassing question, but how do I change the directory in Git Cmd? Like if I wanna put stable or auto my on my d drive
everything I try is messed up and even after googling it im funking up lol
like i have c:\users\chazzy right now and i wanna just make it on my d:
type D:
and hit enter
lol damn
A yeah, but the bot experience for SDXL is massively less good than MJ.
MJ is blowing me away. I splurged for a sub last week. The lighting it does on photoreal is pretty wild
leverage on google Bard, you'll be surprised
I also love how you can toss a bunch of images at MJ and get some cool results
I have still yet to see anything too impressive from it
The only thing we're I was impressed was the way it can do photos of people behind glass, and that's about it
i feel that way about stable tbh. it all looks kinda plastic-y same vibe on XL from what I played with. idk MJ photoreal just looks incredible imo
MJ photo real looks like decent prompting in 1.5 IMO
but I hate the cost and discord view. I couldnt disagree more but fair enough
I can pretty easily match their realism, and surpass in some cases
If you have some I ages to share, I would love to take a look
maybe when I get back home tonight
I am always open to being proven wrong, but it's happened very few times in regards to MJ quality, personally
yeah i noticed it happening but was wondering if it would resolve itself 
That's a fair assumption haha. SD stuff seems to not work until it works lmao
i adjusted my training data and i was caching aspect buckets by numeric index rather than a full string path, so, deleting one file just made the whole thing become useless. i also tripled my training dataset size and now it's taking 4 hours to cache the aspect buckets
sooo it'll be a while til i see the fruits of the labour but now it'll be more robust to me fucking around unnecessarily, as i do like to do that
Best of luck!
Chaz has me interested in these photo realistic MJ claims lol
I have yet to see anything but a hand full of images where MJ looked better than SD IMO
well i mean theres gonna be the fanboy angle here, mentioning mj always gets peoples ire up lol
I honestly could take give any shits about MJ anymore
Show me a good image, then it's a good inage
and have you used 5.1?
Nope, but I have had plenty of people try to convince me it's way better than all the other times 😅
Specifically when 5.0 came out, Sen was blowing smoke up our asses saying how amazing it was lmao
5.1 is magic but it's not real
Until we came together and handily outperformed it
i mean sound like you got your mind made up
show me some of the killer photoreals youve done that blow away mj
chaz i have like 38,000 MJ 5.1 photos in my training set and i love it but i know it sucks
so does sd, its why i dont even use it anymore
im not saying eithe ris perfect, but to my eyes MJ blows away sd
If I did, I wouldn't be asking for more chances to prove me wrong. I'm not gonna say that it's bad just cause it's MJ, but MJ has some fundamental issues in the way it does things that just irk me
true, MJ has surface level aesthetics out the wazoo, out of the box
like obviously the options are greater for auto or whatever but for a pure prompt to fool my eyes, id go with mj everytime
but zoom in and it's trash, like this flying burger photo i have in my training data where the airplane windows are like what the fuck, hardly matched up watercolour quality... and it often doesn't follow the prompt given
There have been some MJ images where I couldn't say anything reasonably bad, to which I was surprised and said so
I could totally see that for other aspects of V5, but it's realism still feels decently behind in a lot of ways
aite
this is the kind of stuff MJ gives people when they ask for photoreal
you must have it on the anime setting
Yeah, and it looks great, but it's clearly not realistic
I have even seen the specific photo real style, to which I still don't see it, personally
i don't have it on any setting, i have thousands and thousands of raw outputs with the full image settings
when i start with 35mm color photograh or 85mm ive never gotten anything like that myself so idk
random sample of MJ images and their prompts are their filenames
Like these images, they look carbon copied from like realistic vision V1.3 lol
that lower left green robot one is actually blurry
in fact, blurry describes most MJ images
alrighty
sorry i brought it up lol
i knew it would trigger people and i shouldnt have said anything
Chaz, we aren't triggered, this is just a conversation ._.
idk dude you seem pretty passive aggressive to me
i was looking for the images from it that i loved
What have I said that could have been taken passive aggressive lmao
right
All I said is please, show me some images and I'll see if I agree or not, that's basically it
anyway, enjoy your stable images
Sytan just has resting asshole tone 😛 they mean nothing personal by it
as if id show you anythign now lol
And now you're going to choose to share nothing, and my opinions are gonna stay the same, oh well
guys
People make big claims about alternative platforms, and I would really like to see images matching their claims, that's all
that image is awesome
Images like this are some of the headline images for MJ 5.1's photo realistic capabilities, and I think they look pretty bad, personally
pixel art too but who uses that, honestly
Over processed, digital art esque, bad lighting
When I do pixel art, I gen normal images and then use the pixelizer plugin haha
It does a fantastic job
i think they look hardly any different from usual marketing materials, which is what it was trained on. overprocessed over-photoshopped images
yeah you won't get like, raw looking images out of MJ
that's not the goal of it
they try to avoid that, if anything
Yeah, they look like people put through face tune lol
One thing that I have noticed about portraits in mid journey is that they are always extremely shallow depth of field
Like, you only get detail on a very small plane of the face, and the rest of it is just a blurry mass
Which truth be told, is an aesthetic, and it is something that actually happens when you take extremely fast portraits, but it seems like a mid journey defaults to that always
It's like photo realism in midjourney is synonymous with having a small patch of fine detail, and everything else completely out of focus
Actually, you know what, I wish I had an MJ subscription so I could compare photo realistic animals, now that would be fun
this is what happens when it tries to make a black background right now in my model 
Ohhhh lord lol
that's what i'm referring to when i say, yeah, the contrast is just GONNA BE BAD in general
even when the image is otherwise good, its details are compromised due to the contrast issue as well
Alright, in that case I can overlook that as a temporary issue, to which your images once again look the best I have ever seen from a 2.1 model
LMAO you're too kind
i made you a thank you cart
card*
i cherry picked the worst one lmfao
going in deep without negatives enters you into the domain of weird shit
i want to try to get it to make a horse with an aftermarket wide body kit installed but i'm also afraid to see the results
portrait images are so chronically undertrained that it's just stupid looking sometimes 
Lmao
i kinda like that actually
it should be the logo for some shitty corporation that pretends to be moving forward while really looking backward
Lmao
I asked in prompt help, but could anyone help me out with the ai GIVING ME LEGS WHEN I SAID SPECIFICALLY NO LEGS
I'm gonna fool around with some photo realistic big cats, cause I have been focusing on hot big cats, rather than real big cats for a while : sweat
It looks nice and all, AI, BUT NOPE, STILL HAS LEGS
do an astronaut tiger
You just want like... Her legs chopped off?
I could with the furry model, but for realism I am going for realistic images lol
bro i can jam a tiger into an astronaut suit in real life. promise. all i need is a bucket of thorazine, some meat scraps, and a sewing kit. if we can get Olivia Newton John into those leather pants for Grease, we can do this for a tiger
don't make me make another one
it went a little heavy on the 1977 but i love the tiger ears on the outside of the helmet. that is a nice touch
this one really phoned it in lmfao
like
i do not have time for your shit today
it did some wish.com tiger woods 
Please, does anybody know how to disable the stupid fucken pull down to refresh on chrome mobile?
you can probably make it pull up to refresh
It keeps ruining all of my gens by reloading in the middle
i switched to vivaldi mobile recently as it does background youtube for free
has more customizability, the built in one is ♻️
I will probably need to switch back to firefox
Chrome mobile is pissing me the fuck off man
If I leave the app for more than 10 seconds, it reloads the webpage
put your phone in landscape mode
it won't fix any of the issues but overall i find it more pleasant to use
Then typing is a nightmare 😅
it's still making legs 
Idk what you want bro
Best of luck, not likely to do it
"Tony, I don't know how to tell you this, but you've got cataracts, you're severely overweight, and you have diabetes. It's gotta be the years of sugar frosted flakes."
🐯 "But they're greeeeeaaaaa--"
"Nope, don't say it, Tony. This is serious."
Alright, guess I have no choice but to go back to Firefox and play it's not as shit as last time
its like the more contrast the damn thing wants to make the worse off the image gon' be

must be confusing to the model to see reflections in visors like that during training
Firefox has the same stupid ass pull down to reload thing





