#🏞|general-with-images
1 messages · Page 61 of 1
the most horrendous result was when i prompted the img2img with "blended and smooth lines"
Just Sayin'
Digital Diffusion really improves on SD 2.1, man
still has a heavy duplicates issue
although... i mean... that's like... i wouldn't say no to those three
that one is pretty close to the prompt, too.
dark and mischevious portrait of a maine coon cat sitting on the mantle of a fireplace during halloween, with halloween decorations visible around the log cabin with bear skin rug
like im impressed it threw all of that in there
now i need to get DeepFloyd properly running on the bot because i'm so curious. i haven't had a good interface to run great tests on it with
the one poopy thing is that Sd2.1 doesn't seem to work as a base for ControlNet
realismengine, artius, prmj, illuminati diffusion, are all good 2.1 models along side digital diffusion
if they're not on huggingface hub it's harder for me to try them, but i can also convert and upload them to HF. but i'll have to get to that tomorrow, thanks for the suggestions
illuminati you may have trouble with as it was removed by the author from many repositories, but the others should be
which is better?
1 is a little less noisy
the 2nd one has hyperrealistic and the far right's cat's eyes look better
it's not a noise strength change, but keyword change
pretty small difference though, mostly down to personal preference i think
yep
Tile size: 1024x1024
Tiles amount: 54
this upscale is gonna take a while 
@oak osprey
SD really struggles with only drawing 1 fireplace
More cats
Ty
🗿
so cute
yeah i thought SD2.1 had an issue with duplicates, reattempting that prompt on 1.5 is like 
Meow
Digital Diffusion makes SD 1.5 feel like a damn toy @hasty nova
What is it
Show me

I must eat it
Hello everyone, i hope you have a wonderfull day! I was wodnering if you can help me with prompt to create an image like this (ultrasound photo of fetus)
i have legit spend 4 hours and still cant get any close result
thanks in advance everyone!
what is your actual goal here
do it from scratch? only with prompting?
even with img2img it's going to be hard or impossible
only with prompting yes its going to be the finish of the circle of life
a close up of a person's ultrasound, by Scott M. Fischer, happening, symmetry! human fetus, official screenshot, hyperedetailed photo, group photo
yes got the result and tried to reporduce it but still ididnt got any good result
thank you tohught!
i mean, your best bet will be to make a LoRA of Ultrasounds
have you ever used Dreambooth before?
no, i actually got into 2 days ago i just grind it cause i find it super intresting
sad 😢
well, fine-tune it with dreambooth
are you fabricating a new person from birth to death
yes but in reverse. My video starts from death and finish to birth
oh a video nice, kaiber ai or something else?
deforum
cool havnt experimented w that yet but looks interesting
will be nice when the animation gets smoother and less choppy
this space moves fast so im sure thatll come around sooner than we think
i have a question, how do you guys add a little bit roughness of the image effect or imperfections to add a more realistic feeling?
ai generate images too clean
i am quite sure that with all those available settings to work around, you can create a very smooth experience. I am stiill new but i trully believe it
controlnet tile upscaler
best sport car
@fading whale this is after strength 0.3 Tile Upscaler
Damn the P100 and 3060 very close 
Do Shrek swamp for discord Shrek party
Glad you are getting good results!
you have been having issues with oversaturated images?
i got very detailed and very normal on 6
@bold cave do you keep changing your pfp
I am confused 
I like img2img best with low noise and cfg
Just make an image pop like professional adobe artist
does anyone use 4090 just for SD and keep their old GPU to run the display and everything else with? Wondering if I should keep my 3070 when I upgrade and try to connect both cards
I want to dreambooth with you tonight
I mean, if you spend 2k on a 4090 and your old gpu will only sell for like 200, it's not a huge difference money wise
That's true
I'll take your 3070 for 200
250 yen aka 15$
I'll do 200$ and you pay the 30$ shipping 

boc
4090 is freaking huge
Im annoyed that I need a bigger case for the 4090 in the first place, going to get the fractal torrent one, hoping it's big enough at least until like 5090 ti
it's big enough that I need a new case at least, and I didnt have a tiny one or anything
Do you need some support for it also
like 35% bigger than the previous GPUs I think
i mean yeah, all our gpus are big, the 4090 is just even bigger
this big apparently
these are done by a the ctu bot from shrek jar dudes server
450 watts, so yeh I also need a new PSU
What girls think 6" is
need minimum 850 PSU recommended I think
You only suppose to hit 70% of capacity of psu
yeh for 2 cards I'll get the 1200 PSU maybe if I go that route
what kind of psu should i have for a 3060 ti?
and the 3070 itself is smaller so maybe it's doable
Just do modular psu and you set
I think you need over 600 PSU for a 3060 ti
im already running on a modular 750w
yeh even 3090 is fine on 750
I just plug my gpu straight into the wall idc
🤨
Seasonic psu
sd cant do silver foxes :-)
no it got the general fox right
it just made it a white haired fox instead of what i wanted
lmao CivitAI is down
HTTP request sent, awaiting response... 503 Service Unavailable
2023-05-10 15:41:22 ERROR 503: Service Unavailable.
Just put white fur? 🤔
thats... thats not white fur
Fox pure white fur
depends where they live, guys
Silver fox change color
foxes in snowy places go white
yeah im not making an artic fox
arctic foxes are orange in summer
highres fix or CTU which is better?
the first one is more eye catching and sharper but the 2nd one looks more realistic
currently no i cant
im trying to do ctu on my ui
instead of through a bot
what image are you upscaling?
tbh im really just hoping this works
if it does then ill be able to ctu on my own
what would you say is a good amount of denoise to get more details?
depends how good / what res the initial image was
well ill tell you if hires to ctu is going to be better
ive abandoned the ctu thing i was doing
and im going to upscale my hires fixed fox image from earlier
and i just forgot to bloody set the tile resample to the fox image 💀

cleavage
oh good, my friend is awake and generating images now. ironically i don't think i will ever sleep again
anyways
im going to disappear for a bit and let that fox image get upscaled to 2k
so set it to 128 then?
i will make a second version of the upscale
simply because i already started it with padding 32
thanks

man... 735 samples 💀
ddim is confusing me for this
i have the sampling set to 15 but its showing 14 for the actual upscaling
but the total number of samples is acturate to the 15
yeah
that
ultimate upscaler's code is confusing
but i have achieved perfection
i dont even care his finger is missing, i chose that for that reason
alrighty ill take your advice and just ignore that then
but yes you can, it'll be like a mandala to me, lost to the winds of time
i love how it ended up with water marks
i don't have watermark in my negative lmao
ok welp im over halfway done with this upscale
but i fed the clown back into CTU 3 times with different base models
the glasses on the right look a tiny bit screwed up
on the frame for the left eye but thats an amazing result 💀
so ur only plugging into ctu now?
oh so u ctu it and then ultimate upscale it?
waiting is so damn boring...
its definitely a shitty aspect ratio
that image was pretty bad
idk if you guys know what bad truly is
it does look like the clown is breaching the hull of the sphere he is encapsulated whithin'
well he's not gonna accept that result, i know that right now
ngl the viewport thing is kinda deceiving
who calls eyes viewports? you robot
im not calling the eye a viewport
oh
im calling the box showing off the upscaling the viewport
it works on one tile at a time bro
mmm sniffer
Don't drop the dun dun dun
its a really shitty VAE used on a temporary tensor aiui and the preview doesn't even work with SD 2.1 models (at least, not with ControlNet)
that's really not the model's fault
i fuckin love 2.1
the text encoder is amazing
no its not
far better prompt results than 1.5
currently at 550
and using the fine-tuned 2.1 models makes it like a world's difference from base 2.1, just like any other 1.5 model. you all claim so much love for 1.5 but none of ya using the base
36 minutes just to get to like 550 / 735
well my first tiling attempt on an a100 took 10 hours to look like complete garbage, the image was about 1GB in size and something like 56,000 x 56,000 pixels
but i made my own tiling algorithm and it sucks
I like this confidence
know your limits, and then bypass them, as i always say
lmaaaaoooooo where'd the 4th monkey come from
or like, their faces in the reflection

it must be trying to make a metaphor
like, we really hide from ourselves the most!
or something

thats ctu
this is the original
i dont see a difference
nvm
when i compare them directly i do
i see the improvements without even zooming in
yeah well im just an idiot so yeah :)
now hold on while i get the one ur bot made
high res | pseudos bot | ui ctu
im attempting to get better results on ctu now
a bot pseudo has
:-)
no...
pseudo as in shrek jar guy
basically stable diffusion stuff as well as being able to ctu images
anyway sen i started up a ctu with those settings you recommended
yeah i will
but i just want to get the fox in higher detail
yes
i mean yeah ik
but i want
some of the interfaces aren't as easy/intuitive as i'd like but it lets you run CTU with an arbitrary image and base model, step count, positive/negative prompts
i just put some code in this morning that redirects any incoming requests for CTU from a 2.1 model to use a default 1.5 model for that instead, makes it easier since you don't need to remember to switch
fawk man why would you switch from canny to anything else
does it have to be one-shot?
why can't you go canny to inpainting. two-shot
one-shot is everything in one shot, eg. perfect results on attempt 1
n-shot means you're using n steps to produce the output
I see the day coming fast, as I stated with the 3k series then the 4k is even bigger, that the video card will be in its own case attached to the computer via a riser like cable system.
I blame @oak osprey
i got an Antec case that has amazing cooling and holds like 10 HDDs
Pretty soon the card will hold the computer not the other way around.
10x 16tb RAID-10 for ZFS. it is a NAS.
But it's inside your tower
i use a laptop
What
that is a server tower
The fuck
No
So you use laptop to run on server? And server has all the gpus?
You just no longer have to do the math
yeah the server has the 4090, all the big storage, etc
How do you run it, just cli? Or do you use vm or something
uses more power than i want for a daily driver, so, it turns on when i need it and sits in S3 sleep mode otherwise
it's got currently, Ubuntu 22.10 installed i think. it's just used via SSH
running a graphics server on a machine learning rig just wastes VRAM
But how do you setup the automatic1111
i don't, it runs some code i wrote for a discord worker network. multiple slaves run on eg. an a100, a6000, 4090, 3070, and all link up to the master, which distributes jobs to them by capability
So you have to run scripts hmm.. 🤔
eg. the system with 128g ram can register to run Llama.cpp jobs and the A100 has so much memory it's registered for Vicuna LLM tasks etc
no, i don't run scripts
it is a Discord bot
Yeah but the bot build the processes to run on server
the bot manages queues and sends commands to servers that can do the generation
sniffer 2.0
I would rather tunnel in and build it with automatic1111 and turn off display when disconnect
i have friends that i generate images with and we do that in discord threads and it just manages automatically which worker gets used, so, you can both generate one concurrently if there's two workers online
Yeah but missing out on all the settings that cant be passed through the UI
My friend does that too. Nice but hard to finetune through discord imo
i mean tbf his bot is pretty good
oh pretty much every meaningful setting can be tweaked through the bot
some of those are easier to use, because they're more important/common things to swap
yeah except your bot cant do the one thing that ctu is mostly used with 💀
the ultimate upscaler?
yis
i can't do that, it'd be like giving heroin to irresponsible teens
what you trying to say?
you trying to say that uu is like a drug? and that its addictive?
Makes sense but you already running Ubuntu so just setup a tunnel for display to run big batches and tweak them
Here we go againhttps://www.youtube.com/watch?v=_rKvAmVqZpY
And set when disconnect turn on headless
aitrepeneur dropped off ngl
Lol
tiling can go pound sand
Not sure what happened to him, but he went from bad to worse to the worst
yeah they have shitty tiling too lmao
hes been replaced 👀
How do you load new models into your server?
some alien replaced him and cant figure out why the channel was good
Hey im new here
they're on hugging face hub and it simply downloads them
How do i prompt
ear
AutoModel.from_pretrained(model_id)
Help please
By typing a sentence.
Man that still sounds like a pain in the ass to run cli for each process
We do not do art bots
Masol, you have to spin in a circle three times while saying, "I AM INCREDIBLE, I AM COURAGEOUS, I AM CAPABLE" and then, confidently state your prompt while looking up at the heavens. your image will be in the mail in 6-8 weeks
Bye
Man, don't tell him life's cheat code
Bye
bye
shit, he's phase-shifting
I am unsure why we keep getting people in this discord looking for free hand jobs, errr image creations. IS that a thing on other discords?
Hehehe
@dense tapir yeah Blue WIllow does it for more than like 20 million Discord servers
it's literally the only way to use Midjourney so yeah people hear that the way to do it is to write in a discord
I just saw a vid where MJ is no longer doing free
the fact that MJ is Discord-only is pretty lame, and i write a Discord bot for image gen.
what is that?
Why I left it to never use it. Seriously, only via discord? what is up with that?
it does add more detail
How tf did sytan get a 3080 for 350$
they hate people
I came back to SD more seriously last week, tried latest version (2.1) was like 'damn why is everything so bad' and just paid for Midjourney, tbf they are miles ahead
but a lot more customization available with SD
They drove to the dude and bought it last night.
@dense tapir well Blue Willow was cool to discover but their "Rookie" channels flood like MAD 24 hours a day when people are generating images.
eye!
I was looking at them and those are still selling used over 450$
i love how wild CTU can be when you just screw with the prompt too much
I honestly never tried them as I was told to stay away so I did. Forgot now why I think were were spammers or some such. I stayed with SD
it is
Hahahha
it was supposed to be Kid Rock in an apartment
does your ctu like automatically preprocess the image and then upscale?
CTU requires you to upscale the image using Pillow or something similar before you put it through the poopline pipeline
3090ti is same price USED as 4090 new within a hundred. At that 4090.
@dense tapir same here in Canada where i live
is CTU CN tile upscale?
yes
Yeah, right tell that to ebay, or Amazon refurb, or Newegg. Screw local as that is just for crackheads who stole a system and is looking to get a bag.
in Germany at least new 3090ti is 1400, new 4090 is 1700. Used 3090TI is maybe 1k
I'll take it
and if you mainly care for big ram gpu but dont mind taking a bit longer it's not too bad
you still have new 3090tis?
300$ for a 3090 is a steal
ngl, it is
we seem to yeah
I saw a vid where that seems to be the case for Europe while NA is now all used only.
You confirmed the vid
it's lower availability, and I've only seen palit though
Sen, just shut the nonsense up. You don't know jack shit about my local area.
2nd hand 3090 is maybe 800ish which is less than half of a 4090 so if you mainly care about being able to do big resolution and training but dont care for it taking longer it's definitely a sensible purchase
there we go
EU should just ban gaming gpus so I can get them cheaper
love how there's like random body parts in that mess
I can get a refurb on amazon of the 3090 for 849 and 1k for a ASUS tuf one. go to ti and double
Google should just start making new TPUs and selling them. We need a really big new player to make any dent in the way GPU availability and price is going
Meanwhile Intel fucks things up again
😦
If they haven't fixed it by Druid then I give up on them to never look at their gpus again

just buy a Cerebras WSE2 wafer. they're hella competitive
Battlemage is next then Celestial and finally Druid
Just buy the new Tesla whatever that has 80gb
Process 400,000x400,000 hentai pics for your wall
I rather fight driver issues, or coding, or w/e then to take it dry from Jensen like this. I hope in a couple of weeks we get to see what the 7950XTX is about.
i call this one, "Terrible Life"
Zotac bad

Scan the web you will understand why
msi gpu good?
I don't like msi gpus
Seriously ASUS and Nvidia are the best cards
Mine is a sapphire
I would buy a Nvidia FE but they stopped making them and only avail locally at brick and mortar Best Buys
Sapphire for AMD for me for life
EVGA for Nvidia then died
sapphire is amd, no?
correct
would the asus tuf 3060 ti be good then?
yes
there's nothing wrong with the MSI 4090 24G, that's the one i've got
and of course its too fucking large
it isn't a strix so takes what strix would pass but good cards
302mm max length
Snoop, welcome to my world all of these are so new case time for one I am not sure I can fit it
304mm for me
i can only fit 300mm cards
the new case will house up to 545mm long gpus
hands you a Dremel
I hate that I am going to lose 51/4 but I can always external it
nope, drives go there dremel no workie
so it would just about fit...
Snoop, the card needs 6mm of clearance
aww man
oh wait itd be fine
Luckily no since the power adapter is on the side now not out the back.
although im def not getting a 3 fan 3060 ti
oh great asus has a two fan model as well
the fuck curerncy is that
pounds...
looks weird man
about 1.20 times the dollar in value
sounds like he's saving money by spending in pounds
buy when it is 1.12 and sell at 1.20
he can get two of them
no...
i live in england its my natural currency
Yeah, the writing is on the wall for everything from SD to gaming you need at least 16gb
Snoop how natural is a 5 pound note lmao do you find them in forests and rivers 
Pound Sterling
also all the games that will be based on AI for so-and-so graphics or physics or audio will definitely need more VRAM
the heck is SDXL?
Yeah, 256x256. Welcome to 2018
fine ill go look for another bloody gpu. ._.
eh video games usually need less, I dont think any games can use up the 3090 vram yet even, and you can go over that with ML easy
which is less than 3090
Snoop, it is your money but we are really trying to help you not steer you wrong.
A lot of games, even in 1440p is sucking up 16 gigs now. It is 100% insane and just damn lazy dev shit.
1080p is 11-12gigs for not even super duper everything? smh
For this sort of money yeah
yes but i am poor and cannot even find a 3090 anywhere
Back in the pre dawn days buying for now at 200usd was cool cause in 2 years you just buy another now, for me, 3 gens is now 6 so 12 years until I buy again.
The price, I don't care about headaches, is driving me to AMD.
The pull is immense
yeah I swear, PC costs are now more per year to keep up than buying and owning a car
It may be a little slower, or not, but I can get a newer card sooner since I didn't spend as much.
I am waiting to see the 7950XTX and ROCm 5.6.0
nah, the biggest crypto stuff cant be mined on a GPU again, it's really unlikely
and tbf it's not like prices dropped for the generation after mining dropped by 95%
Proof of Stake vs Work
it's more expensive than ever
Cause AMD, and Nvidia doesn't want to let that cash cow die.
Let it go it is 2023 not 2020/2021
Nvidia plays us all. 4090 MSRP 1599USD. How can you have a manufacturing suggest resale price if you don't manufacture any? OOOPs, lawyers said you gotta make just enough then you can't be sued. They did, they sold, now pay more for AIB partner cards.
but after deciding to get a 4090 I think I'll just start buying the newest one every time (and just skipping the TI versions) so I'll just buy 5090 at launch, too etc. and fuck it just precommit to spending and not worrying
If you have that kind of fuck it cash then yeah
i am now upscaling cat image :)
I mean if you buy immediately it doesn't work as bad because you can always recoup a decent amount for your last one
like when 4090 came out a 3090 was still a decent price 2nd hand
so it works out to maybe 1k per 2-2.5 years to keep on top of it?
Always in the hole like the people who buy a new car every two years by trading in. They never get out of that hole.
(though of coruse there's everything else other than the gpu too)
5090 at about 3k USD I would never ever even consider.
doubt
if it's 3k but like 48gb vram then maybe I would but yeh it'd be ridicilous and I'll probably pay it in installments or some bs
It could give me hand jobs, make my breakfast and kiss me good bye in the morning but for 3grand I am sorry, that relationship is never gonna happen.
if it's 3k but it's like 40% faster and only 6 gb vram more then yeah fuck it not worth
Per Jensen 24gb. Because it is already on the layout part now.
yikes if they dont raise the vram
Not on their consumer cards, nope
Now they have that Neural compression scheme so no more ram
eh, it's not like it doesnt all rise
With the advent of the neural compression thing it is signalling that Nvidia is NOT going to give more ram so shut up and give us your money.
AMD will do it though 🙂
ngl i wish nvidia went another way instead of going down this path 
to decompression it takes abit over 3 times longer than raw data so they figure around the 6090 it should be the same time or less even.
Well, I was watching Moore's Law is Dead channel yesterday and they said what I have been saying that gamer cards is not the future for nvidia they want to solely go head into AI/ML and will probably shutter gamer consumer cards within the next decade
...
You could tell that by listening to Jensen when he speaks.
nvidia really just fucking hates its consumers
When one H100 is 45k USD, yes
which im pretty sure like over 70% of its consumers are people who game
more
over 90?
around 80-85%
ah
Thing is they can make more money while doing less in the AI realm
one H100 is 27 4090.
in price
shrink Nvidia down to remove the excess fat and they will make a lot of money and their investors will be super happy
Coming from the business aspect it has a ton of merit. As a consumer for need of their cards it sucks
You know if Intel fails and Nvidia leaves AMD is not going to keep pushing the envelope, right? I hate this period in time
it doesnt matter massively since you arent allowed to use consumer GPUs in the cloud and for real business purposes
Yep, in their contracts
If they rip that consumer kidney out leaving the pro/business AI division kidney they will be super healthy as a lean mean money making machine. Gaming is sort of dead
whats goin on guys
Photobomb, lol
just hanging out with "The Quints". that's what we call The Quintuplets.
which one is best?
tbf i was paying more attention to the bear's expression than the humans ahaha
best boy
just wanna rub those fuzzy ears
had to do this one twice to fix the bear's eyes
oh damn it he's missing half his torso
yeah this i made at 1280x720 and then upscale to 1920x1080
i get better results that way
rather than making it at 1920x1080 natively and having like 80 bears in the shot
I also noticed that highres fix removes some of the blur as well. That's really 99% the reason I use it :P
Nice puppy
roflmeow
when your elderly bear biking friend doesn't understand selfies
tbh i think ctu on its own can be awesome
and yeah a bit of a dog vibe
Hahaha
I figured out if you find celebrities that look like you, and run that in prompt it helps
so thats what thats about...
ok should i turn that off then?
okie dokes
no wonder why its been taking 50 minutes for one upscale...
💀
loras are useful for people who dont have a lot of vram
like me because i am poor
no they arent
loras tend to be like 10 images worth of data
and proper trainings are like hundreds of thousands of images arent they?
ah
oh 🤔
do anyone knows how deforum have been built?
and how can i make extensions relying on sd?
things like temporalkit and discodiffusion
anyone have anyidea?
Damn, I just found out that the case I am going to get comes with zero fans. Fans are damn expensive these days for some odd reason.
Not the good ones.
I hate rgb
Reason I am going with this case is that it has a steel door not glass, or acrylic.
Oh, and because it allows up to 503mm long gfx cards
Which fans are for pressure and are cheap?
120 and 140mm?
I was asking for cheap pressure rated fans in those two sizes.
I never found any
oh, lord
Made for static pressure
nope
meh. High airflow on a radiator, or up against hard drives is crap
I even did an experiment and proved it to myself from numerous articles.
It really did make a big difference
I don't do closed loops
closed loop being AIO
No, as I said I tested it and it was a difference
I just run my machine mostly open out of laziness with whatever fans it came with
Mine comes with none, lol
For the back of the case (exhaust) any old fan will do 100% on a radiator or a heatsink I lowered my CPU temps by getting a static pressure one that had half the RPM.
It really did make a difference.
hey sen do you know if ctu can be batched?
my CPU at least will go up to the maximum temperature no matter what (in a stres test), the difference is what performance it will get
well, that makes no sense
it sounds like you are not using a stress test that pushes it as far as it can go
sure it does, it just does more until it reaches max, if it's cooled more it will push out more before reaching max temperature
I'll just grab some 140 Noctua fans made for static pressure and be done with it.
that sandwich was Digital Diffusion, this is Artius, same prompt/seed/etc
@hearty karma my bot now has Artius v2.1 😄
i uploaded it to huggingface hub
can be used in any of their Spaces now
very cool
wheres mine?
i just ate and looking at those did make me hungry too
lmao rainbow riot at the movie theatre prompt
its a good thing
idk bout you but i remember my grandpa by his smell and hugging his scratchy wool vest he wore often
it took away the fuzzy 😭
and added a watermark because of my activate windows that got into the corner of the image
LOL pwned
so much rainbow
@bold cave is it possible to batch ctu upscaling?

sauce code
the coat is too fluffy 💀
well it is 💀
no, this is
are those fucking bots in the voice channels
lost in more ways than one 
bruh those are bots
man... looking back at old gens and comparing them to my new ones makes my old ones look like shit
look at their names
ikr I occasionally look back at what I used to think was good
yeah.. it all looks weird now
borat noises
now we have models that can do good faces straight off the rip
this is Artius v2.1
try same settings with DD, I am curious
that is meant to be christopher lloyd..
that is a good one
yeah only because i restored the face using gfpgan
if you put something like (perfect eyes, looking at camera)
makes it look a little better
that added eyes everywhere 
are you playing ovewatch and rendering at the same time?
interesting. I have been noticing my model is a lot darker than some, if you don't like the darkness I guess add "dark, moody" to the neg because those tokens were in a lot of training images
after feeding #3 through CN Tile via Sd1.5
forgot my prompt is still elderly fuzzy grandpa
it worked tho
how didnt you notice that?
it's a set and forget thing
so much power so little responsibility
Howdy heathens
hola amigo
mr clam chowder back for round 2
what is a clam chowder
best quality
It's a nasty
I love the color grading applied in that
i'm loving pretty much all the v2.1 finetuned models
His hands look like clam chowder
my new favourite baseline to go by
If you find a good one that doesn't result in a grainy / artifact the output, please feel free to let me know
most of the time it'll just produce crystal clear outputs and no refinement needed
yeah... on ur bot
grainy?
Gritty, more like
that sounds like you need to try the Karras sigmas
yeah enjoy your grainy shots then i guess 😄
I only use DDIM, it's the fastest, and it requires some of the least steps to get a good output
However, that may not be the case for stable diffusion 2.x
well i do stuff the MSE VAE on top of SD2.1
yis this is exactly why i use it
I usually use UniPC, it is fast and can be good in 10-15 steps.
amazing results with only 15 steps
Wait a minute, I just realized
My new graphics card has enough VRAM to run vicuna 13b
It's about to go down
@hasty nova the prompt is epic realistic, award winning photo of a skinny pretty mage woman posing in a dark forest, posing and smile, art, hdr, hyperdetailed, warm lights, dramatic light, intricate details, complex background
dark is in there
Vicuna 13b is about as good as it gets
Well, you actually might have enough VRAM to run vicuna 30
my bot has StableVicuna on it
it's really bad at programming, but what are you wanting to do with it
Stable vicuna on the other hand, I have not heard such good things about
oo ok, try
photograph of a skinny pretty mage fantasy woman in a dark forest, posing, smile, hdr, detailed, intricate, warm lighting, cinematic, dramatic, complex background
I feel like that might work better with the model?
How the hell do you download models off of hugging face?
Like I could have sworn there was like a button where you just clicked download, but it's not here anymore
go to the model card, click files, then click the file you want to download.
Oh God, I have to download all 15 of them individually?
This doesn't have some form of zip download function?
Aitrepreneur has a good one
15 of what?
15 of the files
Bleh, Aitrepreneur is a huge L IMO
One was for vicuna xD and in one he explains the oobabooga llm webui
who is that meant to be
the pagemaster
him
rip my brain
They look gross lmao
too much texture
GO BACK GO BACK
Looks like those sheer fabrics
thats a leaked photograph of me and you are all making me feel ashamed
ok im so confused...
this is what the upscaler currently looks like
it was not doing it like this before
oh 🤔
i think i might have thrown my cmd prompt box out of whack then
because every tile it kept adding another section
oh nvm its doing that again
that happens
what sampler you doing it with?
i usually do about 150 steps with PNDM
the default scheduler that comes with SD 
wth is pndm
no i dont think it is
EA is default
one thing im confused about.. is how my tiles have ended up so low
before i had like 49 tiles
now its 16
"scheduler": [
"diffusers",
"PNDMScheduler"
],
trust me i have this shit seared into my brain
as an AI Language Model, I am primed to provide ample feedback and funny memes and even boob jokes
make a boob joke 
Why does everyone hate boobs? ||they suck!!||
just off the top of my head right there, that one is free
oh is it going to cost money next time?
if i answer that, it'll eat into your allotted credits

btw snoop, try !generate-random-x on my bot and watch GPT3.5 prompt it better than you can

wondy's*
Deez nutz
@tiny violet You know I wonder why there are still cases being made, and sold, that can only use graphic cards to a max length of <10 inches?
Because sometimes people have big space restraints, and having a weaker and smaller graphics card in a computer that they can fit in their room is typically a better option than not being able to have a computer
Have you seen some of those "smaller" cards lately? Even they are getting stupidly long.
That doesn't mean that previous generations aren't options. You can only get the best of what fits your needs
In the end it doesn't matter as most of those people with those cases are using internal graphics most times or on a 1050-1660
Just because it doesn't fit your needs doesn't mean it doesn't fit other people's needs
My new card is 328 mm long
So it's actually longer than a 4090 FE
And it fits in my case no problem, with plenty of room to spare
Yep, why I would love to get a FE but they don't make them now
I could probably fit a 400 mm GPU in my PC with no problem
And it's not an expensive case, and it's not a big case either
My limit, if I wiggle, is about 310mm
Most case I look at will not fit a 4080 AIB so I have to get a behemoth sized case (4080 and 90 are the same length most times).
The reason the cards are so long is due the heatsink not actually the cards.
Who is this chick
Ah yes, when the boobs look like they're made out of crocodile leather

?
whats the pinacle of 1.5
that is after adding more detail
ah right
im very happy now
the upscaling time is 18 minutes instead of almost an hour
lava ice cream
looks my old 2.1 embeddings aren't behaving like they used to. I wonder what it can be...probably everything :P
My friend just sent me this and I hate how accurate it is lmao
that last one gets me every time lol
chrome iron man
i hope the CUDA kernel hangs 
?
what does that mean
oh.
i wont have to worry about it
theyll be done by the time i wake up
Can't tell if robocop or ironman
you dont see it?
robocop is future
pretty much sums up the gpu market rn
Medieval Superman and iron man
do robocop
I think those are some of your best work.
ironman built a time machine and trying to blend in
Like that?
oh hell yeah
nice
thats whats up
Spider man
moar
Dark side Jesus
So, I have been debating making a Digital Diffusion v3, I just don't know what to improve. Any ideas?
the impossible caustics stuff like SDXL is capable of
portrait art of amorphic solids in the ferrofluid-infused lava pour, suspended animation, superposition structures within the geometric complexity of intricate positive and negative space, highly detailed, epic, cinematic, dark, twisted, impossible! mystical fog, rotational inertia, quantum feedback superposition
nice
I could try, I am thinking about expanding the dataset to like 500 images (would need 200 more) and that will take a while
So I could add some images with nice caustics
cool 🙂
I geuss I also need more images of certain subjects/concepts it struggles with
More animals, and specific like occupations I guess
And maybe more landscapes/environment photographs
oh it really struggled with monkeys and a lion in the same scene for some reason
Yeah it has no monkey images, so I can see that
but it does
It's not any good at making monkeys for me
Interesting
photograph portrait of a monkey staring at their own reflection in a pond in an african jungle
Well, I didn't put any monkey images so whatever you are getting is mostly using data from the base 2.1 model
Look at how many reflections there are....
yeah but your style training applies
the base 2.1 doesn't work that well with this prompt
I kind of want to train a lora on monsters from the Witcher, that could be cool in conjunction with DD. I have been playing that game and reading the books, I really enjoy it.
fourth monkey reflection
oh yeah all of the D&D monsters in general
ghost monkey
Pan's Labyrinth omg
LOL
I guess I could just do fantasy monsters in general, it's somewhat hard to get specific monsters with sd
"where the wild things are" prompt
my brain always wants to so badly understand "what IS that?" but it's not real
:/
一群达赖喇嘛在草地上玩飞盘
@rustic lily
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!
@hasty nova the model can already do some really cool stuff but it can be more intricate and fine-detailed
probably my fav example from SDXL: a stunning and impossible caustics experiment, suspended liquids, amorphous liquid forms, high intensity light rays, unreal engine 5, raytracing, 4k, laser dot fields, curving light energy beams, glowing energetic caustic liquids, thousands of prismatic bubbles, quantum entangled light rays from other dimensions, negative width height, recursive dimensional portals
funky
so, more details, got it. You can already get some nice details with certain prompts, but maybe I ought to specify it more in the captions
I feel like mine actually fits the prompt a bit better?
it possibly does
I don't see any liquid in the first one
SDXL is in very beta status
true
i guess i just don't know how to make impossibly detailed fractal machines with this 😄
man you are making me want to open up sd but it's way too late and I am working on homework



