#🏞|general-with-images
1 messages · Page 95 of 1
"white square on white background" horror show
it's like it needs to average out at mediocrity
can't have pure white or pure black, no no
An yes, I love when I switch to a different model and suddenly "wild Jaguar" means "sexualized and perverted naked girls in a forest"
oh i love this, an asian detective getting caught manipulating evidence in the evidence room
(MJ)
tiger from earlier ckpt that knew what brightness and contrast means
he looks like he does know what brightness and contrast mean
perfectly positioned face to be flat
Oh I remember
Firefox is stupid and annoying in a very different way
Right, when you copy a seed, it copies
" 4120631616, "
you betrayed us
Space before. The comma, and the space after
naw just scrolling through the mj dataset
this is SD 2.1 tho
this is how my model does that
If chrome let you disable their stupid pull to refresh, and stopped reloading webpages you leave for 5 seconds, I would have no reason not to use it
imagine not using operaGX
SDXL looks like this at the moment, kinda
I'm not gonna use opera GX on my phone lmao
he only asked you to imagine it
oh, mobile? then you're kinda fucked
also; what's your people's favourite sampler
DPM++ 2m Karras
overall best with detailing and never disappoints, also one of the fastest
thanksies
from what I tried so far, ancestrals worked best for me so far, but I'll have a look
I'm okay with more steps and slower gen
if it means more reliable quality
ancestrals mean anything but reliable
Messing around again with animal realism
lil toy tiger compared to mine 😄
DDIM is good for img2img, but pretty bad for txt2img, i usually use txt2img, so for me DPM++ 2M Karras is the best
whatever works for you. i find the Karras stuff impairs a model that is otherwise properly functioning on more simple schedulers
DDIM is what I always use, for all models, for everything. I have run hundreds to comparisons, and DDIM wins in almost every category
i made a comparison once, the best ones are Euler_a and DPM++ 2M Karras. i usually switch between these two
It's the fastest, it needs the least amount of steps to look good, and it also does really good with resisting over fitting LoRA's
okay, well, i'm constantly making comparisons and DDIM receives improvements before other schedulers due to how simple its implementation is. highly recommended to just stick with that. Karras stuff is poorly maintained.
most of the samplers should just be removed from sd-webui as they do not receive fixes or were a bad idea to begin with, eg. heun
Yup
you use 2.1, i was talking about 1.5 models, I rarely use 2.1 models, next time i'l use a new architecture model will probably be SDXL
DDIM will continue to be my go to, only needs 15 samples to look great
What he said applies for 1.5 as well
Sytan is a 1.5 stan and they also support the use of DDIM. for what it's worth, DDIM didn't work with 2.1 very well until a month or so ago.
um, i don't know what to say, i really like DPM++ 2M Karras the best -_-
I mean, you can like whatever you want, feel free to
it's about your prompts too, if you're doing very good and long prompts that will look good on pretty much any sampler, maybe you do just fine with Karras. but simpler prompts i found work reliably on DDIM or DDPMScheduler, but DDIM is the only one that's also in sd-webui.
But lots of tests across lots of people and models have shown for pseudo and I that DDIM is faster, needs less steps, and adjusts to new information more reliably
yep and i hated DDIM until it received SNR fixes (and is currently the only scheduler that has received them)
huh, i always write a paragraph of prompts, so i guess that checks out.
👍🏽 not trying to make you doubt your decisions
i used karras for a while because it was the best option up until the DDIM fixes came along
but i was using them on the DPM Multistep Solver 2M which isn't in sd-webui either
know what? il make an xyz right now, talking the talk is not like walking the walk
that's the spirit
All of my images are 15 steps
@cyan snow challenge you to ada lovelace, 1982, neon club dance scene at night, high contrast rave molly mdma ecstasy tongue out
hmmmmmm
before training on widescreen images it's basically impossible for me to just get a single subject
and one of the subjects looks half-assed and deranged
oooh that might be the best one so far
ah maybe her sister 😛
thoughts?
euler a sucks lolol
ddim has one image that entirely encapsulates the galaxy in the jar
the karras almost gets there but not quite
try other seeds
il change it to just DPM++ 2m Karras vs DDIM
good idea
personally i'd use DDIM still because it's simpler and works with newer training techniques. Karras will not work with zero scale SNR models
i think DPM++ 2M Karras has better detailing tho
maybe in the future the DDIM will be better, but for now i think i stand by what i said earlier, but yeah, DDIM is great as well
the version of DDIM you have there is likely the old one
i'm on the dev branch with auto update on, unlikely
automatic1111 doesn't fix that shit
I would also use DDIM, as it's considerably faster, and usually needs less samples to look as good
it's the same speed as the other one in my comparison
Weird, what are you generating on? Cause karras is usually way slower
it's not slower here either
you mean the DPM (not karras) ones
Oh wait, I am used to SDE Karras
i think that's ancient
they are not as good DPM++ 2M Karras on every aspect
They look more crunchy and over processed
That's what I come to expect in general from karras
DPM++ 2M karras
@smoky oak those billionaires that got stuck in the atlantic ocean in the submersible are reportedly being subjected to Celine Dion's "My Heart Will Go On" as the computer onboard resets itself every few minutes on a loop
That's what I used
try generating something that exists in the real world lmao
give me a prompt, and i'l generate
try jack black riding a bear in the swiss alps and karras mangles his sweet little face
Awhat?
like a lora?
@shut sinew are you actually a Markov chain bot like TwoDukes told me once?
Its an addon that pretty much auto detects the face or any part you want and rerenders it
I am whatever you want me to be
that's how CodeFormer and GPFGAN work too
true, it can correct only minor errors eg. eyes
Feel free to try it and let me know
you get an extra prompt for the area you select
Its great
it's going to take me an eternity because i reimplement everything on my own, in a discord bot
so if that's a library i can make use of, that's a lot quicker
if it's an A1111 extension, ugh
so, it's img2img?
Thats how it looks
RealESRGAN gives us a library
Yes and no
Ok, I see now why the colors are so messed up with DPM++ blah blah
It's cause it needs more steps
doesn't that just smear it
i use 50 step usually, that might explain it
dood -_-
ddim converges long before that
if you're testing both at 50 you're gimping it

I still like the DDIM look more 
i just fought with a ziplock bag to get it open for 15 minutes
i have STONK gpu, it doesn't matter for me, usually takes a few secs
yeah, Sytan is obsessive about getting over 69 it/sec 😛
both fucked up in here
the motorcycle 

I am not sure what happened there lmfao
`1234567890-=asdfghjkl;'zxcvbnm,./
No, I type without looking on my phone, and it appears as though I hit the menu to take me into character kmao
No secret codes 
an enigma
So I was looking away and typed that, but I guess I misclicked into that menu lol
to the code cracking-mobile
`1234567890-=qwertyuiop[]\asdfghjkl;'zxcvbnm,./
Yeah, so I'll stick with DDIM cause it's better at 15 steps
why does that matter to you, what GPU do you use
that was my takeaway too
well, for me i run a discord bot and image gens queue up pretty deep sometimes
3080, but I prefer to iterate way more images and find the one I like
I do like 16 images at a time, pick the one I like, then upscale from there
i can make higher res images that look nice in fewer steps while using the same time to make the image that it took to make a lower res image with DPM
Just wasted electricity for basically the same output
Thats my style but with a 4090 
i do that too, but with an A100 80G
as you do
For a single image, about 17it/s
But I usually do elongated gens, so about 9it/s
elon? 
for me it's 26 it/s
are her eyes open or closed?
EW bro what happened
idk
she doesn't know, that is why she is screaming
Good for you I guess, but it's still just a waste of electricity for the same image, and it still takes longer

RNGesus
that is about 1 second bro
ok i'll probably catch flak for this buuuut worrying about a few centiseconds of electricity use when flying across the country 😛
yeah, electricity isn't a problem, do you know how much U-238 there is on this planet?
ENOUGH
i use solar and i don't care at all if it takes 3 or 15 seconds to gen
Yeah, the power isn't really that big of a deal, but waste of any capacity is waste
¯_(ツ)_/¯
You could do 3x 15 steps images of the same quality in the time it takes to do just 1 50 step image... Of the same quality
well in that perspective your GPU is being underutilised if you aren't using the heaviest scheduler
the main problem with solar is that it's unreliable, the sun shines in a single place half of the time, that means you need batteries
bro i live in the pacific northwest, it's been rainy for 4-5 days
tell me more myths and lies please
wait, then how am i wrong?
If it's actively running your meter back, then you don't need batteries
i have a surplus of 400-500W all day today
overall i'm getting about 900W-1kW of power and i have a TINY roof, just 10 foot by 8 foot, which is as full of modules as it can be
all a solar panel does is convert photons to moving electrons, if there is no sun, the solar pannel does nothing
and yes, You Need Batteries. so what? you are going to probably complain about the mining that's done for them. oh well. uranium is mined up and purified too
I always prioritize being faster. Especially when there is no benefit to wasting more time
i have 10kWh of power storage and it weighs just 160lbs
Hell, you don't even need power storage if you are just in it for the money savings
the pipe dream of nuclear power is so that no one has to change their consumption habits
It will run your meter backwards at all times when power is generated, and then you just start running it forwards at night
but reducing consumption is a good thing regardless of how much power we have available.
yeah, that's why the most efficient way to generate power is to use the core of this planet
Yup
it's called a geothermic power plant, it exists
that's... not the core 
Yeah lmao
that's barely even the mantle
That's literally just heat lol
it's also really only something you can do with a water source too
like nuclear or solar concentration towers (heliostats)
it uses the heat of the planet, to make electricity
Yeah, that's not specifically the core
cellars use the cold nature of the planet to keep things cold.
This is a weird argument lmao
ANYWAYS I stick with 15 its on my images cause there are 0 benefits, and many drawbacks to going higher
¯_(ツ)_/¯
you can't tell me AI enthusiasts aren't passionate people tho
True lol
i bet you anyone here will argue about how to cook pasta
I want my hot boys, and I want them NOW
i hear that you should add olive oil to the noodles
also, chili never has beans in it according to the Texas Chili Commission
that's "Ohio Style"
i strongly disagree with the TCC on that one
but yeah one m^3 of U-238 can make enough power to keep humanity with enough power for probably a few months
they shut down the main street of my city downtown for 6 nights to shoot a 3-4 second clip in that movie
my uncle never shuts up about it
are we considered AI enthusiasts?

https://www.youtube.com/watch?v=MAu2e0VbvY4 this scene. a bunch of it is downtown victoria and they composited it into london
The Core movie clips: http://j.mp/15vJPSr
BUY THE MOVIE: http://j.mp/LlWk74
Don't miss the HOTTEST NEW TRAILERS: http://bit.ly/1u2y6pr
CLIP DESCRIPTION:
Problems with the Earth's core and electromagnetic fields cause birds to panic and attack London.
FILM DESCRIPTION:
An unlikely band of scientists and soldiers join forces to save the world fr...
Oh, I love Victoria :>
We were there for a day on our cruise, and I would love to go back
probably even YEARS, basically if we be smart about electricity, we basically have an infinite amount
you know how Uranium is formed? in heckin' supernovas, so yeah, we have electricity alright

anyways, here is master cheif
nuclear needs water
uranium fires are cool and all but i don't like the idea of them actually happening lmao
that shit's terrifying
that's true, but it's a closed cycle. it doesn't need replenishment
and i'm not anti nuclear, please don't argue for it at me
haha our plants here have a heat exchanger that uses lake water to cool it
the same in finland iirc where they've observed increased jellyfish populations near their reactors. might be somewhere else tho
that's stupid
bro it's just how it's done
they're not going to compromise on certain key points because it's more lucrative for their wallets
do nuclear reactors even make money?
they could literally save SO MUCH water if they weren't so dumb
that stuff is hella expensive to do right
ah, no, it's not financially viable and often survives by government support
actually not, after the architecture is made, and you know what you're doing, it's basically infinite energy
lmao
couldn't read that text before i used the gif

look, i'm glad you nor i have any say in this because neither of us know what we're talking about
you think it's infinite energy and i have no idea how to fix core issues i know about, so, that's just how that is lol
people paid a lot more than you or i say "this shit is hard" and i believe them
I said it's almost infinite, and only when done right. like, certain states have WAY lower power cost because of this reason. they know what they're doing
a good example of idiots using nuclear energy is what happened in Chernobyl, they cooled it with local running water, and well, one day there was an outage and shit happened
Hi guys. I try to install SD but get following error:
File "C:\Users\rp89\stable-diffusion-webui\repositories\CodeFormer\basicsr__init.py", line 11, in <module>
from .version import gitsha, version__
ModuleNotFoundError: No module named 'basicsr.version'
Does anyone have experience with this and can help me with the problem? Unfortunately I am not experienced.
look, if they where smart about it nothing bad would've happened, but no, those fucks thought they can cool a nuclear plant using local running water
look, the solar panels i have will last me 30 years and they fulfill my needs. things become more efficient over time and my power becomes more flexible as a result. i don't discharge my batteries deeper than 15%, they'll also last me about 5-10 years, and i will simply end up discharging them deeper and deeper til they suddenly don't work (but probably one of them will do that first, and then the other will shortly after)
people could do this today, we don't need complicated nuclear reactors that no one understands, just simple photovoltaic systems. we don't need massive centralised power grids. you've seen the problems when there's outages. the one in 2003 in the northeast, the ones throughout California right now because of their massive solar farms with experimental equipment
tryed, no answer
alright, this will turn into #🤝|tech-support temporarily
when the Hz of the grid deviates too far it literally damages equipment physically and can make it very difficult to restart the grid. in fact, running a power grid is so difficult that to me it elevates into a form of art, because there is a delicate balance to maintain between the load placed on the grid, and the power being fed into it
GPU?
GTX 1060 6gigs
looks like their install is missing some file.
I hate to break it to you, but that won't generate high quality images. but it should be able to atleast run it.
what files do you see inside the basicsr folder?
no problem, its just to tet and then get petter GPU 😉
here is a workaround for you, @true badge
already tryed, but nothing of that helped
yeah, that should work, i also run into stupid errors like this when i started with python
clearly you didn't try it because that file isn't in there in your screenshot
one second please
take your time
you mean this
yes
@smoky oak MJ's best result for photo i can find in the data i have, but also remember i love that yucky grain
tis Ada Lovelace apparently
y'all like Halo?
kinda did her dirty and she looks like a clown 😦
It looks just fine
¯_(ツ)_/¯
Now I am facing following arror 🙂
seems like your install is broken though. i recommend redoing it
thank you very much will try
state of the art technology
the cornhole edition of the ram 6900 dually comes in 4 different flavours
the F350 comes with extra spare tires that are always ready to go because they're all installed on the rear driver side
now it works, thanks a lot 🙂
yay
now show us what you got
@true badge can you try making pictures of house parties?
try prompt, photo from a 1998 house party retro themed, ethnographic photography nikon d5200 polaroid style
use negative prompt:
(deep fried) watermark, cropped, out-of-frame, low quality, low res, poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation
i will try to set QR Codes in a picture, later
#qrcode #qrcodes #ai #stablediffusion
Generate AI QR Codes by enhancing regular QR codes with AI - Stable Diffusion and ControlNet. Convert or transform regularly created QR Codes into amazing AI QR Codes including Upscaling and Outpainting.
Note: I’m using an iPhone 13 Pro to scan the QR Codes and there are differences in terms of code recogn...
yo, what's up with people turning qr codes into other stuff
usually those trends are cool, but i don't get this
You need to fill out the template it displays after you type the dream command
☝️
qr codes are really useful but hard to draw people's attention to them. They're not enticing the same way nobody cares about a barcode. This makes them really neat.
Also, it takes a bit of finesse so if you can do it it's a flex
Bro really tried to slip it the center of there prompt
Well, I guess I'll be making AI barcodes tomorrow lol
This guy never heard of running the AI locally? You can literally make everything when running locally
It doesn't even need access to the internet, that's how dense diffusion models are
some people can't afford a decent computer to run SD on
Just out of curosity how would i run it locally?
do i need to download a program or something?
@smoky oak aspect bucketing is hard to implement in a fucking YOLO way ahahahaha
you make better choices in designing it early on and it's easier but damn
many sharp edges
there are a few different ways to run locally. Automatic1111 is the most popular due to the features and number of extensions. But there is also ComfyUI, Easy Diffusion, and others
Thanks ill check those out
Yeah, A1111 is by far the best interface, it's the most customizable, and has the most features by FAR
But the creator has a dirty past. I personally don't care at all, but I know many people do
There's issue with their readability tho...like my phone rarely picks them up
Challenge accepted
even ones in the article -
https://stable-diffusion-art.com/qr-code/
I had to sit there for like 5 mins to make my phone read 3 pictures (others didn't work at all)
can it read this?
my phone can't
maybe if I'll give it more time or something, but nah, noone's gonna do that in a real world, standing with phone staring at the picture hoping it would work
my phone can read it instantly
hm, idk
might depend on camera \ lighting maybe then
my android thing is pretty old + dim light in the room
shittiest dim room and looking at an lcd backlit screen
but it picks up normal QR codes instantly
trendmicro free qr scanner is the best but the google search one built in gets it too
could be the processor or software. i think the trendmicro one is meant to be a halo product. get their brand some trust
yea, but most people just use default scanning tool thing on their phones I believe
this one is a bit slower to read, takes a couple of seconds
oh this one actually worked instantly lol
hi 🙂
probably cause it has better contrast, hm
i'm certain that next os updates will adjust to pic more of these creative codes details
that one was funny, she just suddenly appeared in front of the barcode while doing an img2img
qr codes will explode into new territories
idk man, there are lots of people with really old phones like me
there's no reason for me to upgrade until it's dead
QR codes needs to be accessable for everyone, quick and easy to scan
my company has been warning against scanning qr codes in public as they could link to sus websites
yeah i warn people of that too. trendmicro actually has a good scanner for that
its also fast and no ads
@smoky oak i fixed the contrast and darkness 
i wonder if i have to alternate noise schedule sometimes to keep it from getting too dark
we shall see
who's coming to the AI Cryptoid party?
your invitation is in the mail
do you guys use prompt generators or any other hekping tools for prompt?
I use dynamic prompts with wildcard files. And I used chatgpt to help with those wildcard files
I stopped doing that as it seemed to not be random enough so I just go to lexica
No matter how many I added the same ones kept popping up.
True it does seem to pick some more than others
E.g the file with about 100 celebrities in it, I got 1 person 4 times in 20 images
I have a file with over 600 entries and the same batch comes up far more than others so I moved them to a different spot in the file YET they still come up far more than any others.
I ticketed it but nothing ever came of it.
hmm this stylized 3d model is interesting
also man stable is so fast now on my new pc
night and day
What did you get? I'm thinking about getting a new desktop PC for running SD
I just got permission from the SO, but thought I'd wait to see how much memory SDXL needs to run before getting anything
Upscaling plus photoshop makes it even crisper :)
Oh I am running a 3060 12gb(the extra vram is nice, nvidia keeps insisting on releasing doa cards with only 8gb, i could get 8gb a decade ago smh) So i made sure to go for a affordable option with more vram
32gb ram ddr4 3200
amd 5600
One of those fancy new ssd's that go on your motherboard
Nice, my laptop has the laptop version of the 3060 with 6gb of vram
oh yeah the 3060 also has a normal 8gb version, lots of options lol
I upgraded from a 1070 8gb
but my ram, ssd and cpu in particular were big jumps
I was looking at the 4070 12gb, I think that's the new 3060, since the 4060 seems to have less thingymajigs than then 3060
oh yeah the 4060 is horrible
The 4070 is a bad joke too, I wouldn't advise buying it
tbh the 4000 series in general is disappointing
part of why I went 3060 it's a solid card for the price if you go 12gb and leaves me options for upgrading incase nvidia stops releasing meme cards
yeah the 4090 is good but so expensive
4080 is expensive too
the prices in general are 🗿 the 3000 series had some value and then nvidia was like ohhh ai, let's JACK THOSE PRICES BACK UP BAYBE
I've seen PCs with a 4070 12gb for the same price as the 4080
tbh i'm not sure how good the 4080 is, i know the 4090 is good but insanely high end, the 4060 is a literal joke, the 4070 meh at best
like the price to performance ratio just isn't here
Amd atleast gives out some damn vram lol, but then you miss out on nvidia stability for creative applications, sometimes amd has issues and other things
Back then I had to buy a 3050 cause money was tight and gpu prices were horrendous, did me alright for 1080p gaming but I wanna upgrade now. Got a 4090 at work and I really want one for home too
lol i bet
I'm liking the 3060 I can do 4k nicely now
Now I just have to sell my old pc for a good price to justify buying a new pc with goverment money I got
💀
I wanna get the 4090 also because I'll be set for a few years then I hope
The problem with 6gb of memory is I can't play around with training anything
shit yeah you'll be set for quite a while with a 4090
oh yeah
vram is a big one
That's the plan 
this is why I went 3060 12gb
8gb is literal doa on every front
I could get 8gb a decade ago
But big gpu and big screen, boom 3k bucks gone
games are going to increase vram costs too, never mind ai
yeahhh
I'd love to get a better screen one day
For my self built pc where I reused to 3050 I didn't even pay 2k 
but it's a miracle I was able to get a new pc to begin with lol
I paid like 950 for this one I think
if I can sell my old pc for atleast 500 euros it's basically the cost of a new console

I can afford to get a 4090, but it's like a lot of money ya know. Feels like irresponsible lol
my pc
I feel that, but you gotta look at what you're getting for your money too
i'm not saying get a 4090, just look at the value
100% agree, that's why I'm still hesitating
I held off for years before finally buying a OLED tv, and that's about the same price as a 4090 😂 my plasma tv was about 12 years old
Too much cash for me when I am saving we edge closer to a 5090 that is up to 2.7x faster for 1.5 times the price of a 4090.
I went 0 rgb and functionality to save costs and because I don't really look at it anyway. I'm amazed I built my first pc and it.. works
the real question is whether novideo will make the 5000 series just as bad for mid-low range
this my basic pc btw lol
When would the 5000 series come out? 2 years?
My pc sits eye height with me so I love some shiny stuff
I sure hope not but I ran into a training situation last night where 16gb was not enough I had to throw on the checkpointing. 24GB is a batch size of 2 or it was 3 I forget.
😓
I'm hoping I can train ok with 12 lol
I know it's about on the edge
I haven't tried any of the fancy stuff yet, it's been a while. Going to look into controlnet because it's great, and also training things
I am stuck as 24gb is my min now and it seems for consumers that is all we get AND very possibly less due to their new neural compression junk.
nvidia hoarding the vram
for the big boy hoppers, and A100 yes
GDDR6 is cheap now so 24gb is about 100-120 USD added on to a card.
in bulk I can only imagine
yeah it's cheap yet nvidia is refusing to add it
yep
Do we know when the 50 series launches?
they fear hurting their business cards
also, what if I go amd on the cpu and then some new fandangle ai thing comes out that needs intel mkl for whatever reason?
50 series was delayed to 1 quarter 2025

was supposed to be oct next year
Yeah I can't wait that long
amd is great because they actually give out vram but I just can't commit to amd, it's too risky to me
plus I do like dlss ngl
Nor I but if, rather when, I buy a 4090 it will be 12 years before I buy another. I used to be 3 gens (I am due now) but at that stupid price 6 gens
I hate AIO ngl

for a vid card for as long as I keep them the AIO will die
I had my r290x for almost a decade so yeah
But that won't happen again anyways probably
I go up to 9 years on the same card
same, though 3 gens is normally twice the speed min only I see 1060 then I see 4060 and laugh
I kept it that long cause it did alright and I was a poor student
that right face is pretty crisp
I made do with an intel graphics for years in my laptop 😄 could still play a lot of games (at low settings), but i don't usually play AAA FPS anymore
Macro photography so must have been homed in there
I need the 4090 for starfield 
lol 120 for base game in AU and 170 for the deluxe 😄
Tell you what most people into LLM stuff are not touching the 4k series
bethesda has always been reliable at launch /s
It will be a shit show but I'm all in for it
can't wait to see vids people capture of NPCs floating off into the sky or stuck in the ground
Yes
I know bethesda gets memed but I straight up have not run into a game breaking bug at launch so far
Jensen so afraid of hurting the hopper had all nvlink ports removed from the consumer grade cards so people are buying two 3090 for less than a single 4090 and using the nvlink for 48gb of goodness.
closest to that is an A6000 RTX for 5-6k
memory has always irked me in consumer devices. they always shaft you with the ram and then want rediculous money to upgrade it even though we all know it's cheap
Yep
I had 16gb until a few months ago now have 48gb
next machine will have 64 or 128gb
I just upgraded from a 1600 to a 5600 too. Now I see the 5600X is just 18 USD more than I paid for the 5600
I think I paid a couple hundred AU to get 64gb to upgrade my laptop after i bought it, if i bought it new with 64gb they would probably want another grand
all but gpus are dropping in price
Honestly, one thing is keeping me from buying a 4090, besides the price, is that I have to buy a new case to hold it as they are too long now (FE would exactly fit but Nvidia doesn't make FE now).
the new case I want I have to buy the fans for as well. Sheesh, over 200 USD
no, 3090 has it
reason price is remaining high as the ML people are buying them up
yep
btw, I just (shocked as I was) ran over a vid showing hash rates for a 4090 mining the coins.
2 weeks old
If mining comes back just kiss the baby and toss out the computer.
It probably won't come back I imagine 🤔
I only saw the first coin tested it was raven coin
it was a lot of coins he was mining
hmm
different coins I meant
140+ h/s
mega
I would love to buy a 7900XTX but so much trouble with it for this stuff.
biggest business opportunity for them is to get CUDA code working on their cards
well
forget everything else
I just learned the main issue with 7k cards is that Pytorch doesn't support the ROCm that supports them and ROCm really is Linux only\
latest, took a long ass time for them to advance to this, for pytorch is 5.4.0 of ROCm
Makes me wonder if Jensen is tickling them under the table just a little?
it'd be like toyota refusing to build a decent truck, just keep making camry's. good car, but no one wants them
That is Jensen, and Lisa to a tee.
People don't wish to pay your price so instead of lowering the price you simply stop making them to reduce inventory.
Considering how badly Nvidia is treating their "partners" I suspect, as others do now, another is going to leave them. 2020 35% profit for AIB, now 10% to next year, I believe, 5%
EVGA was my only brand for Nvidia so now I am rudderless. I so mistrust the other AIB which leaves Nvidia only but they no longer make their own cards.
If they did I could buy the FE and would not need a new case.
Experimenting with MBW.
Using CounterfeitV3 and amIRealV3.
TL;DR: A plain style transfer isn't always the best if you want to change the style.
Regular CounterfeitV3
The "quality improver" thing that AbyssOrangeMix2_sfw uses.
1,0.9,0.7,0.5,0.3,0.1,1,1,1,1,1,1,0,0,0,0,0,0,0,0.1,0.3,0.5,0.7,0.9,1
The "quality improver" thing that AbyssOrangeMix2_sfw uses, but inversed.
If it normally changes composition, maybe the inverse will change style.
0,0.1,0.3,0.5,0.7,0.9,0,0,0,0,0,0,0,1,1,1,1,1,1,0.9,0.7,0.5,0.3,0.1,0
Inversed "quality improver", but only where the values in the modified "style transfer" are 1, otherwise 0.
0,0.1,0.3,0.5,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,0.9,0.7,0.5,0,0,0
"Quality improver", but set values to 1 where the values in the modified "style transfer" are 1.
1,1,1,1,0.3,0.1,1,1,1,1,1,1,0,1,1,1,1,1,1,1,1,1,0.7,0.9,1
Inversed "quality improver", but set values to 1 where the values in the modified "style transfer" are 1.
1,1,1,1,0.7,0.9,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,0.3,0.1,0
"Style transfer", but modified to not change OUT09-OUT11 since they changed composition and color.
1,1,1,1,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,0,0,0
Another seed. QualityInverseWhere1 looks good.
I like playing around with MBW too. Been mixing a model for a while, add and subtract various other models. Had some good results mixing an anime model in while preserving the original model
so guys this i my character and i want her to have the same clothes in every generation so i can make a light novel or manga
how to achieve that?
did you try describe the outfit in the prompt?
No i haven't i thought maybe control net could keep it but no
for example i like the first one
how could i describe that in the prompt
I haven't had any luck and I read a shit ton on how it is done. Might be a 2.1 issue but I don't think so for this.
orange and white robes? it's going to be tough getting consistent multi colour clothes, probably need a lot of either generating lots of images and upscaling the good ones, or inpainting
why does kohya install so many files?
i downloaded torch during setup so why does this download again?
on colab it takes me up to 6 mins for all the downloading and building of .whl
sucks ass juice
when you only get 2h per day every single minute counts.
It is really hard to find a 4090 these days as so many are OOS
why would 'ASUS GeForce RTX 4090 TUF Gaming 24GB' be the cheapest model I can see? I thought ASUS is usually the most expensive
TUF is their low end but for me that card is OOS and on Amazon is 1999 USD ffs.
They really are trying very hard to push me away to AMD
You can't even find a MSI Suprim/X
So now MSI Trio X is 1599
This probably has to do with Nvidia diverting almost all chip making to Hopper
looks like the cheapest I can pre order is the 'Galax GeForce RTX 4090 SG (1-Click OC) 24GB' for 2599 AUD, or 1757 USD
Ouch, although Galax is a good brand I hear just not really in the states.
Import here though it is slowly creeping in
https://www.amazon.com/dp/B0BGV6LQYR?tag=pcpapi-20&linkCode=ogi&th=1&psc=1 says 5 in stock $1700
Amazon seems more expensive than other retailers for me
Good to see. Not as cool (temp wise) as MSI but still. My issue is my damn chasis as New Egg keeps playing MIR games.
is amazon price including tax?
I'd guess so
Now over 2 months it is the same pattern that repeats over and over when the MIR ends. Read the MIR it sounds scammy, and besides in this world if I did get the cheque I have no where to cash it
No, Amazon is not including tax
not in the USA it isn't
Ah, so the Galax excluding tax would be about 1597 US
I'd trade Galax for Gigabyte as Aussies are lucky having them.
In 6 more weeks if NewEgg continues this nonsense I will just pay more at Amazon. Sad that only Amazon, and NewEgg, carry this pc chassis at all. Must be an exclusive of Phanteks.
What is even worse is that it includes no fans so 40 USD more for the fans for it.
I like it only because it is so large it will be at least 14 years, as my current case, before it has to be replaced because it is too small.
The PC shop I'm looking at is using the ASUS TUF Gaming GT502 for their 4090 build
Did you see the new case where the gpu is up front where the radiator would go in an open loop wc build? It sits up and down in the vertical position as it is easier to build tall than deep.
you'd have to show me
you mean something like this with the radiator at the front?
no not at all
I mean the PCIe adaptor is from the board to the gpu and it sits like this in the front -> |
straight up and down.
case/chassis companies are having a bitch of a time with how damn big these cards are getting and one can only have only so deep of a case before it becomes unwieldly.
I am lucky as the new case is 530mm deep
my current case's door curves so is about that from front to back but much more I would have no place for it
I've seen a few with the bracket to change the angle like this
Yes, like that
not like that for the card
I mean that radiator that is what the card will be
oh i see
They are having to get creative now
I forget which company is bringing it out as it was on a GN episode at Computex
They didn't show the adapter though
OR how the card will get fastened.
geez there's a phantek case that costs 1399 😐
1400 for a case? wtf
oh, they can go to 5 grand
😐
Why? Because they are built like tanks made with rollers and all sorts of stuff but I would never buy it
More as a portable server really
made to take a beating
zombie apocalypse case can survive anything lol
YEP, lol
the expandable ITX case is pretty cool
179 AUD, about 121 USD
noice.
so nice waking up to just one training ckpt and a bunch of logs indicating it hit another bug in my aspect bucketing code lmao
69GB log file

look at the Thermaltake X9 Core
An E-ATX cube case offers an endless stackable and expandable cooling solution while allowing externally-modular upgrades for enthusiasts to create dual systems.
I think I posted a pic of that earlier because the way they mounted the GPU in the photo
you did?
buddy, just admit that you're not actually going to buy a gpu lmao you kept saying nvidia won't sell you one and here they are, available for purchase
I personally would never buy an MSI GPU directly, but I have had 2 from second hand over the years, and they have been pretty good
I don't support the company at all, personally, as they have a lot of problems with their products, and can be really scummy to reviewers, but I have had decent experiences with them

what i hear from person watching it closely, MSI is today one of best, times changing, and some good companies in past are using shi... materials.
i have not msi now, used to have 1070ti and cant complain, now i have gigabyte, and cant suggest it.
Sytan has spoken
@stoic pivot first try, no negative prompt. Probably good is Teddy-bear 🙂
Streamers buy those 1400 dollars. It's sold as a prestige product so that their RGB liquid cooling will shine. They'll give 1 or 2 to a popular streamer to kick start the fad energy.
In human size they are a bit creepy.
i bought a gigabyte 4080. i'm not concerned about problems. it's got a bracket. Any 4080 should have a bracket. holy it's yhuge.
first time i've bought a gigabyte in years though as their motherboards have had such shit components in the past
I am getting closer to what I wanted.
it will work, and thats the point 🙂 @limpid lichen
i wanted to buy an msi card but they were sold out. I like that MSI makes great free software tools like afterburner
gigabyte card works with afterburner too so whatever
Just new for me was it cant manage cooling bellow 40% of rpm.
Yes afterburner is great. I think all gpus are working with it
yeah i wouldn't expect a 300+ watt card to cool at low fan speeds.
but you know it could give you chance change 0rpm for say 10% rpm, but it cant it cant do it. So zero or 40% which is wasting energy imo.
With msi i could set 20% as lowest, not turning off and on, just fans keep going.
Lets see if Acer is going to get a partnership with Nvidia
And we see those Predator GPU cards etc
I am surprised my 3080 doesn't have a bracket, it's longer than a 4090FE lmao
But it doesn't say, so I suppose it's fine
*sag
It's not always about sag. The recent gigabyte problems are just caused because it cracks form strain at the pcie pins. Sag or not, all that weight still rests on that connection.
my ML server sits like a 1990s desktop horizontally so the concern i have is the motherboard taking the weight of the GPU lmao
didn't use the bracket at all. honestly wasn't sure how it goes on
My GPU is like 3 pounds I think
3.5
My GPU is huge lol
3080 longer than a 4090 moment lol
3 pounds it is weigth of a standard brick isnt it?
by whose standards? the IEC says a brick is 350lbs and OSHA defines it as 3.2lbs, while the US Carpentry and Trade commission defines a brick as the large cinderblock type, at 15.6lbs. the Canadian ministry of scientific measurements and standards defines a brick as 12kg which is about 26lbs, and i don't know what those look like
the DEA defines a brick as 50lbs but i think they're talking about bales of hay
Either way, my GPU is way bigger than it needs to be lmao
But luckily I have the case for it
TIL realistic vision is a 2.1 model

makes the same exact results as 2.1 does for the same prompt/seed
1.5 looks totally different
it doesnt matter, lets say OSHA then 🙂
hm actually RV 1.4 looks like 1.5 output
RV 2.0 looks like 2.1
they're all terrible in any case
they all got the same shape if you switch around
but not pseudo-journey-v2 
there's no people in togas but they were just deemed less important than otters by the model and i have to agree
i'm surprised that model's output is so good without negatives btw, i expected it to do poorly. the models i expected to do better, did very poorly.
Uhhhh, no? It's a 1.5 model, I have mixed it with several 1.5 models, and even the page says it's 1.5
it's just weird that their output with all 4 models are identical composition etc 😛
the slant in the perspective is the same, the garbage noisy "people" it places
Yeah, not sure what to say, 2.0 always matches all of my 1.5 models lol
Yeah? Lmao
I have run probably over 100 comparisons across like 20 models
Realistic vision 2.0 is almost identical to 1.3
oh
i thought you meant SD 2.0
i was like, how have i not noticed this
could just be a fluke tbh, for this otter prompt
Probably lol
Cause 1.3 and 2.0 real vision are nearly identical
1.4 is better than them IMO, but it's still like 90% the same
It's very stupid, 1.3 and 2.0 are very similar, but 1.4 is considerably different comparatively
yeah
1.4 seems to have trained its text encoder a lot. it makes outputs, uh, i can't share here
XL will be fun to explore
horrible, horrible things have happened to those otters
grapes is a dangerous word maybe
can't feed it to anyone without negatives 
My Samsung watch is totalled, and I don't know what to replace it with ._.
get an iphone
No, I would never lmao
Just looking into other smart watch offerings at the moment
If there was a more recent version of my watch with wear OS, I would be sold
I HOPE this is satire
dude, those guy are they are definition of a rip-off.
siri is a spyware
and yet you look down and notice you have an iPhone in your hand. you look up and see your reflection. you're wearing Gucci and Nike apparel somehow.
Lol
it all begins to make sense. you look around and see the plebs with their android phones and you begin to smell Steve Ballmer on each of them
R/2sentencehorror lol
no spy could do that much in so little time
All listening devices are a form of spy wear
i just love how they named their assistant after a porn star
Bard is a sexy name i don't care who ya are
Google ASSistant
google's going to release a bigger language model variant called Big Bard
When I read bard, all I can hear is that damn 1700's drip meme lmao
all i can think of is George Carlin as the conductor, Sir Toppam Hat
or something
i have brain damage 
Indeed
It's so hard not to with so much evidence supporting it
when a pregnant woman asks "am i fat?" im curious, do you say yes 
I'm not around women, ESPECIALLY pregnant ones lol
can't relate more, i'm almost guaranteed to have dementia when getting older
I'm too busy cat calling their boyfriends 
oh she probably can't grab her sandal

that was rude and uncalled for i'm sorry
it's the brain damage
there's right and then there's being an ass
you can be as right as you want and people won't listen if you're an ass 😛
how do i know this? because i'm an asshole and no one ever listen to me
but i'm allllways right
Idk, I don't see anything wrong with that statement. Sure, I've seen you certifiably be an ass, but I wouldn't count what you said as being an ass lol
that's cuz you're one too 
it's like the blind trying to assess the vision of a fellow blind person
"looks fine to me!"
oh shit that's rude too, fuck

Yeah, you're really strengthening your brain damage case here lmao
took me a month to implement OAuth in a legacy application and get rid of static keys and deploy it while the thing is running with zero downtime today, i'm loopy as FUCK
as project lead i take on the "after effects", eg. everything that broke as a result of the changes
As much as I wanna be fake toxic and say "Skill issue", I can't bring myself to even type those words in any scenario in which they may be taken seriously lmao
there were zero issues in production
which i didn't expect
You just have yet to find them :p
developer machines however, like stumbling through the dark down some stairs
y'all like Halo?
it's true, there's always little subtle things that are either harder to test or just don't behave as you expect at full scale of operations but the main kit n kaboodle went smoothly
it's nice when you can slide some piece of technology in at a higher layer than everything usually happens at. and integrate it almost seamlessly
Every time I read "Kit n kaboodle"
It's "kitten, cab, doodle"
my friend says "kitty wumpus" to describe things that are clearly caddy corner
Oh, that's just a mix of kitty-corner and cattywumpus lol
When the term first made its way across the Atlantic, Americans conformed to the British way of speaking and said “catercorner.” But around 1883, the phrase had been corrupted to “catty-corner” in the South, and once catty-corner caught on, folks in the North and Midwest somehow managed to focus on the feline part of the word and turned it into “kitty-corner.”```
in canada, it's caddy with D's
sounds more like the people in Quebec say it
laughs en francais
oh no, again?
Yeah, we are going back
As much as I do like the vibe and feel of Mexico, I do like leaving it cause man it is not a safe place 😅
it's been more then 2 decades, why again?!
What?
Not sure why you turned me waiting for a plane into a 9/11 joke lmao
But okaayyyy
i won't bother explaining, i have a terrible sense of humor
Evidently lol
also, can you give me the Shrek sauce prompt? i want to see what my model makes
a jar of shrek next to a bowl of shrekuamole, that's all
oh
a delicious jar of shrek ...
delicious is in there
the mogwai shrek seems to be some rare AI cryptoid
i hate that he has nipples
aw a little shrek mogwai baby
why does it make that? that is not in the prompt
it's actually a result of deformations in the attention layers of the unet, as i understand it
training on high res reduces them a lot
same reason it makes 3 jars of shrek
the double exposure model gets super weird with this prompt
double exposure?
this is with shrek cereal
oh wait, this is shrek and hulk mixed together
capitalism, baby
it made pistachios again
oh, quaaludes is a drug LOL
it made another shrek baby
i like to imagine that the big one is happy to give birth to little food ingredient creatures that are happy to become food
shrek farming 
DUDE WTF, i swear i used the prompt generator thing once and it said something about ingrediants being happy to be food
yeah, that's so weird. it said something like: professional pizza food photography, the cheese is so happy to be food
im also in 1.5 and mine came out pretty good
mine was like, bulbous and happy creatures popping out of a tree trunk, lying down on the picnic blanket, excited to be eaten as food
but i connect GPT3.5 to it and tell it how prompting works with a few examples
async def random_image_prompt(self, theme: str = None):
prompt = f"Print your prompt."
if theme is not None:
prompt = prompt + '. Your theme to mix in: ' + theme
system_role = "You are a Prompt Generator Bot, that strictly generates prompts, with no other output, to avoid distractions.\n"
system_role = f"{system_role}Your prompts look like these 3 examples:\n"
system_role = f"{system_role}a portrait of astonishing daisies, rolling hills, beautiful quality, luxurious, 1983, kodachrome\n"
system_role = f"{system_role}a camera photo of great look up a rolling wave, the ocean in full view, best quality, intricate, photography\n"
system_role = f"{system_role}digital artwork, feels like the first time, we went to the zoo, colourful and majestic, amazing clouds in the sky, epic\n"
system_role = f"{system_role}The subject must come first, with actions coming next, and then style attributes.\n"
system_role = f"{system_role}Any additional output other than the prompt will damage the results. Stick to just the prompts."
image_prompt_response = await self.turbo_completion(system_role, prompt, temperature=1.18)
i replace the prompt gen model with GPT-4 LLaMa
it usually makes great prompts
i like it to be fast 😛
you! you over there! yeah, you! i know ya see us shreks, rollin' around. we need help!
pseudo-journey-v2 without negatives is good times, man
dude that looks so bootleg
it seems that late stage elder magic syndrome has a venn diagram overlap with cryptoid mogwai shrek syndrome
and LOAB, of course
can't forget Ligma
you think SDXL will also have those syndromes?
SDXL has different forms of dementia but yeah totally
it has Sticky Pope Syndrome, and Wish.com Celebrity Syndrome
it's because there is no truly balanced dataset yet
there's too much of douchebag A, not enough of douchebag B
still, SDXL should be the best diffusion model made so far
would be neat to do dynamic concept bucketing as well as aspect bucketing
oh it is
it doesn't need negative prompts
the outputs here are all without any
I try and limit my own negative prompts, more is better, but it's not always needed to have any at all
i used to only ever test my in-progress models without negative prompts and it was very misleading, i'd greatly improved pseudo-journey-v2 and didn't realise how much until i added negative prompts and decided hell yeah it's good enough
but, when i run tests today without negatives, that model is clearly the best without the negatives. it gets left behind by one of the newer ones with a really good negative prompt though. it just doesn't know as much
i had one evolution i'd taken a single model, to get to that v2, and it was really good, and had a lot of training. like a lot. 15,600 steps
took 2 weeks to get those steps done
I mostly use negative prompts to detail the image. Such as removing a color.
the new lineage i've focused on photography skills
I also use negatives when the ai won't stop creating weird weapons :P
I don't see anything weird in this weapon, could easily see it in a game
hehe, well, thanks! But I'd maybe want it to be a little longer behind the character. But hey ho, one can always just try again :P
@smoky oak seems like i damaged the text encoder at some point 
left is base 2.1 text encoder, right is pseudo-real-beta that i based that unet training on
of course it would be able to make something of a silly prompt on the first try and not the serious ones I tried sooo much longer to get correct >:I
it's a prompt trying to get a dude playing guitar in a park at night. both are overly dark which i assume is just the model figuring out what "true dark" means
well, at night, illuminated only by party lanterns
Alright, when I get back home to America, I will have some commissions to do
Including a cute tiger guy :p
I'm actually very excited to get back home
@smoky oak the family broke up 😦
50% of AI training validation prompts end in divorce
is the contrast looking better?
huh, that's half the normal ones!
i really like the current lighting it can do, even though the darkest scenes aren't quite there yet
ah the moon colony prompt finally improved. that thing i feel like has been shitty for weeks
trying out multi-GPU training again
XL might drop tomorrow it sounds like. hopefully can train it
Precisely. The tables are flipping as they normally do.
I was never a fan of Acer as they were always Goodwill (thrift store) quality.
seems like it'll end up on the stable studio thing but i don't think it's going to be released
the images today are all showing up with watermarks
lol
i enjoyed using some of their products in the past
lets see what they will come up with now
some of other alternatives for me will be MSI and Corsair
not sure about ASUS anymore right now
ASUS has went full on rogue
I love my MSI motherboard and I love Corsair for power supplies and even ram.
I am not sure what is going on but in all my carts across various retailers AMD, and Nvidia, cards just went up 15-20%.
same with one of the online shops here
GPUs went straight +200€
but luckily only one of them
from what i saw
I have 7900XTX Sapphire Nitro+ and for 4090 various
4090 is my next stop, or 50 generation if it comes next year or the year after
although i might still stick to 4090
Well, 2-2.7x the speed of a 4090 at 1.5 times the MSRP for a 5090
128 GB DDR5 RAM is not a option now due to stability issues but i will get 2x 32 GB RAM of DDR5 and when they fix it i might upgrade to 128
and for processor, obviously i9 13900k
I am sticking with my B450 for as long as I can then when I upgrade all the wrinkles will have been straightened out with the 5th gen stuff
tomahawk?
for cpus I am all Ryzen
MSI Pro Carbon AC
Had it since Gigabyte burned up first gen Ryzen motherboards on people and refused to accept warranty returns. Saw Buildzoid on this board and he was right. Fantastic board with no complaints on it.
aah okay 😄 I was thinking there about either MSI MPG Z790 carbon or even Z790 godlike but the latter one really less likely
since very expensive
considering that i have to pay a LOT for art programs and plugins as well for example
and all the periphery for PC
monitor especially
for gamers thats okay, but creators almost always prefer Intel for creative sector tasks from what i witnessed
This is why I am anti Gigabyte and here we go again with them not accepting warranty returns on their 3k cards that snap and once snapped screw it you can't fix it
I will gladly accept a longer creative task for less price, wattage, and heat.






