#🏞|general-with-images
1 messages · Page 60 of 1
personally I am retiring my HDD
Intel use to support the faster memory but idk if that's true anymore
Bye cat person
With two evo 970 sata ssd I will get about 1GB/s so should be fine for this. If not I get another for about 1500MB/s
Speed is key 🔐
Jk I have no clue what I am talking about
P40 or 3060 someone make the decision for me so I can order it
Raid 0 is where the speed is, and also where none of the backup security also is ;P
yep
I boot from dropbox
personally I wont buy anything other than nvme now
I wish I could run another NVME M.2 gen 3 but B450
and riser cards for that are stupid expensive
Still lose my second slot
what other cards do you have installed?
none but a 4090 will zap so many slots.
controlnet tile is pretty useful for training images i've noticed. throw them through realistic vision to get a consistent image with refined details
Can't use it any way as I have no x4 slots on the two long ones and 4 x1
is 28s/it good
fantastic if 2.1 768x768
joking
no, that is pure rubbish
almost half a minute per iteration
@sterile temple Do you have recommendations about how to create a 3D style picture from 2D picture?
Woah
@static tuskI just moved a safetensor to my sata SSD and almost instantly it loaded it. Same file on my HDD takes forever.
Pop it in and test it
That's a 3080, what were you running before?
I ran a benchmark with it at his house, and it performed perfectly
But nice, 10G
I cant imagine the slowness of HDD loading checkpoints
3060 TI, I just picked this up for $320 nearby. I'm going to be selling my 3060 TI
I have several AI commissions right now, so this will speed things up nicely
It is horrid.
I tried it once, gave up and moved everything back.
My friend is actually using stable diffusion with a hard drive, and it's not nearly as bad as you would think, because it's continuous read and write, so most hard drives can still get well over 200 megabytes per second when loading a model
I even notice sata ssd to nvme. big difference there
with dreambooth training, how do you fully take advantage of prompt naming, most guides just say to use "photo of ohwx woman", is that enough or am i not taking full advantage of its capacity
My whole PC is now NVMe m.2, and I could never go back to having hard drive storage, as all the things that I do are extremely read and write random intense
I only got a 3070, but it's been mostly enough so far to do what I need, I can train most things on it and generate quite large images.
It takes me up to 3 minutes on my Barracuda 2tb but on my sata SSD <10s
well LLaMA is 221GB just for 65B
I can load models in like 5 seconds
have fun putting that on nvme 😄
I can't wait to see the it/s on a 3080 vs your 3060ti
i'm back now. i can probably run it for you
cache models into ram and its faster by a little bit again
Ah sweet, if you can run it and figure out how to send the necessary files that would be great. I already set up the huggingface page with the safetensors and yaml (I didn't make a model card yet though)
Yeah, I cache all the time so I had to find a model I don't normally use for the simple test. Originally the models were all on the sata ssd but I ran out of room and moved them to the HDD ages ago.
But my main model is only 2.4G, I don't need to increase speed too much
initial load of DD
Model loaded in 3.9s (create model: 0.1s, apply weights to model: 2.0s, apply channels_last: 0.3s, apply half(): 0.5s, move model to device: 0.8s).
then after its cached
Model loaded in 2.1s (create model: 0.1s, apply weights to model: 0.5s, apply channels_last: 0.3s, apply half(): 0.5s, move model to device: 0.7s)
Makes me so mad that newegg is back to playing games with my case. MIR offer ended May 1 then came back yesterday. I no longer do MIR since my bank is virtual.
thing is the mir brings it back to what it was pre mir
Oh, and my sata ssd on a load of the model barely made a blip but on the HDD it pegs it 100%
Great! Still don't know what exactly is working but i'm glad.
it created the proper config file and a safetensors file with idk, whatever it needed to add to it
adding: DigitalDiffusion/unet/ (stored 0%)
adding: DigitalDiffusion/unet/config.json (deflated 65%)
adding: DigitalDiffusion/unet/diffusion_pytorch_model.bin
to a zip file
But why do you need it in a special format, I mean.
You are running out of VRAM
even as low as 10 training steps per image with 10 images i get the error
12gb vram
@dense tapirAlright, installing drivers right now
to use it via huggingface_hub
The steps per image doesn't matter, it's more how big your images are and what optimizations you are using. I can train very good models on my 8GB 3070, using Lora and Xformers.
Oh, to run it using their GPUs?
it breaks the .safetensors file out into the vae, tokenizer, text_encoder, and unet
@hasty nova to run it on my discord bot network
trying to get people to use 2.1
it's called PseudoAI. i do not believe it is listed in any directories
ive tried getting xformers, but cant get it to appear no matter what, i can try lora though. Do you know what other settings I could adjust? I'm using 8bit adamw, fp16, and not caching latents
8bit adamw always makes me read it as
8 bit Ada Machine Wearning lmao
it's 2.83GB:
adding: DigitalDiffusion/vae/ (stored 0%)
adding: DigitalDiffusion/vae/diffusion_pytorch_model.bin (deflated 42%)
adding: DigitalDiffusion/vae/config.json (deflated 58%)
adding: DigitalDiffusion/tokenizer/ (stored 0%)
adding: DigitalDiffusion/tokenizer/special_tokens_map.json (deflated 72%)
adding: DigitalDiffusion/tokenizer/merges.txt (deflated 60%)
adding: DigitalDiffusion/tokenizer/vocab.json (deflated 71%)
adding: DigitalDiffusion/tokenizer/tokenizer_config.json (deflated 69%)
adding: DigitalDiffusion/text_encoder/ (stored 0%)
adding: DigitalDiffusion/text_encoder/pytorch_model.bin (deflated 41%)
adding: DigitalDiffusion/text_encoder/config.json (deflated 45%)
adding: DigitalDiffusion/scheduler/ (stored 0%)
adding: DigitalDiffusion/scheduler/scheduler_config.json (deflated 46%)
so you can see how it violated your model file
😄
compressing that sucked ass
Werid, xformers is a big help with this kind of thing. Lora should also help a bunch. I don't remember any other things, except if you had "Unfreeze model" checked, uncheck it, I rememebr it causing lots of vram issues when I had it accidentally on.
btw, how the heck did you manage to train 2.1?
I should upload all that to Huggingface?
@dense tapir I am a wee bit concerned
Dreambooth + Lora

how do I do that?
use Graph Mode if there's an option for it
@oak osprey Sytan only got a 3080, isn't the torch.compile thing better for 40-series cards?
it works for the 3xxx but the speedup is pronounced on the Ampere and Ada series
yea dont have unfreeze checked, i tried the --xformers command and in the loading terminal it would say its loading with xformers but in the ui the option was never available
I am just trying to get my new 3080 to perform as it should
I heard nvdia had an update that was better than xformers, torch or something along those lines, not sure about the legitimacy of that claim
i did torch.compile on a 3070 8gb
should be in the settings, in this drop down (sorry it's greyed out I am generating something)
oh wait what the heck, task manager isn't showing CUDA on this GPU
start from basics. wipe drivers and reinstall. wipe venv and git pull
👁️
@smoky oak well if you struggle with that you're not the only one but it's possible you have too many offload options enabled / not the right ones
oh come on, please don't tell me I have to do that bro
no worries, but yea thats where I check, its never there, only default option
i wish i knew more about the stupid way they've named options inside there so i could help you
Does it come in default with vlads ui?
nevermind, my dumbass was clicked on my iGPU lmao
Have you tried turning it off and back on again
its showing 100% GPU usage, and only 12it/s
honestly, I should lmao
new GPU and new drivers
brb

hmmm
Well :/
when changing gpu the minimum is wipe drivers with DDU and reinstall fresh
maybe
you want to be on torch 2.0 and use the latest cudnn
almost there... I am trying this new fangled controlnet tile upscale
something tells me 12-14 is what you will get
3080 should be way faster than that tho
making a big image, hopefully it works, it's been like 20 minutes or something
ok, I am redoing my venv, and it seems to still be using torch 1.13, perfect
I just do not want torch 2 man
I dont have an issue with torch 2
ew
Every time I have used torch 2, it has been slower
do you run that on win95
in Vlad, A1111, Comfy, and Ana
486 GX2
always been slower every time
@oak osprey I have just extracted the zip you sent me, and just dragged all the files in to the upload section on hugging face. It should be up soon, I guess I should also make a model card or something.
i'm stoked for you and your new GPU, Sytan
Thank you, I should be able to work on my comissions much faster
As long as this GPU starts running at its proper speed
Is it Satan or sytan
make sure you have this turned off too
Its my producer name, its based off Satan, but only in name
Saitan is a food thing
Yeah
You have the upside down star thing going on also
and Seitan is a Shakespeare character lol
lmao getting a real "Richard Ramirez" vibe all a sudden
Well gosh darnit, the controlnet thing smoothed the image a bit too much. My guess is that my tile size was way too small.
I am a satanist, but my name is just Sytan, its just similar to Satan
hes edgy
What tile size?
I am not edgy >:C
im sorry 
I am round and rotund :>
For the controlnet tile upscale thing
few edges here
Why is the new generation like this 
I know, what did you use?
only a sharp person could come up with a joke like that

Oh I think it was just 512
I have been twarted lol
@oak osprey
512 should work fine, what pixel upscaler did you use?
omgomgomg
ah damn it, middleware error @dense tapir
I changed my eye color with a prompt and now look like Oscar Isaac
trips and falls down the stairs running to test
Back to our pinned messages lol
bro, you were spawncamping lmao
@smoky oak do you use Twitter to promote your work or something
wdym
no
You are using it with ultimate upscale, right?
Yep
What upscaler did you use?
It's in the screenshot, though I think vlads has an issue where it defaults to one set in the settings, so I think it was SwinIR
Ahhh, shit. I was getting that and had to zap it all
oh right, that one you used has a propensity to smooth and blur things
I recommend ESRGAN 4X
its much sharper, but it adds artifacts
I am using a new one now
don't remember where I got it from
What about R-ESRGAN 4x+
btw, I went looking and a 4090 pytorch 1.x 8s for 512x512 eulera 50steps. Pytorch2 using vlads was <1s. Yep, made for the damn 4090 and higher cards.
omg its working 🥺
Ooo cool
That one tends to blur things
this model is so slow lmao
Fixed it
If you want some other testing with the bot I could try adding it to my server, it's very small and has like 3 active people, though technically 160 people are in there
now who spawn campin
Swampy slow
Does anyone know how to generate an image like this https://www.instagram.com/p/Cr_o8UFoU3k/ ?
so just this one
its not any faster brooooo
Im also gonna try a higher tile size, just to hopefully make it faster
i wonder if i need to make a fp16 version
@dense tapir Any guesses as to why its the same speed as my old GPU?
use a photorealism model, and either include "film grain" in the prompt or add it in post, which is what I presume this artist does
DD is already fp16
oh 😄
only 2.4 gigs
I am so frustrated >:C
i dont think your model is particularly slow

Thank you for your reply, how would you go about the background as in my case it always gets blurred
I should be getting 16it/s minimum right now
you could add "blurred background, depth of field" to the negative
whys installing xformers so complicated, why couldnt adding --xformers in bat file be enough lol
damn
thats how it worked for me lol
that's usually all it takes
Have you tried controlnet? Put the 2d image in the controlnet image source, select the canny models and then type a prompt that describes the picture and also use '3d render'
love that high quality SD 2.1
man I love stina tattling and Goslino
i have --xformers --reinstall-xformers in my cmd args but i get this error: Launching Web UI with arguments: --xformers --reinstall-xformers
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
PyTorch 2.0.0+cu118 with CUDA 1108 (you have 1.13.1+cu117)
Python 3.10.10 (you have 3.10.6)
Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
Memory-efficient attention, SwiGLU, sparse and more won't be available.
Set XFORMERS_MORE_DETAILS=1 for more details
Very smart! Thank you
who in here loves data and charts? i had SD generate a chart of its performance
dude i love how everything in the legend is the same colour
Looks good
try deleting your venv? It looks like you have an earlier version of torch
well i literally prompted it 'SD2.1 Testing' to create the thread
delete the whole folder?
it's honestly promising that it works so well with so little inputs
Yep, it will recreate it when you reboot auto
cool, ill try that ty
You've used DD before locally though right?
I can't remember
i do nothing locally
Oh, so you've just been waiting for this
well i hadn't heard of the model before today
Oh 😆
went to try it, couldn't 😄
Yeah it's not a very popular model
i was like who's this goofball not on HF
well it's possible it'll pick up if you tag it correctly on hf as being a 2.1 model
It's a lot harder to make a model card on hugging face
they have a Space thing that does it 😄
oh cool, I found it
your spirit animal is a mage?
dunno
dumbeldore
Great, now you all have me in a chaotic mood
dumblebadass
dumbledussy
👁️ 👄 👁️
it hav mor ram tho rite
well that's weird
I do not understand why
are you runnin.. oh you wouldnt be in wsl2 right?
@hasty nova thank you so much, finally got xformers
Great!
what was it?
No, I am running straight up windows, no
delete venv, delete repositories.
change your ARGs to --opt-sdp-attention --opt-channelslast
git pull the latest A1111, dont install xformers
wht are your a1111 args now?
pytorch2 is where the party at
--xformers --listen --enable-insecure-extension-access
and then I also ran it with
--opt-split-attention --opt-channelslast
at the end
wait what are those
None, because the speed of your old card was damn high according to every 3060ti chart I saw. Your old one was an anomaly.
Cross attention layer optimization significantly reducing memory use for almost no cost (some report improved performance with it). Black magic.
ewww
but I SWEAR I remember 3080s being like 16it/s
pls try without this
already did, no benefit
didn't you tweak the shit out of it?
same exact speed
I am running the same tweaks on this 3080
--disable-opt-split-attention?
but it isn't the same card
when using SDP you cant have xformers installed as it screws with it
pure fire is telling me that I need to redo my whole new A1111 install
yes
I would too. save the old install though
that is literally the last thing in the world I wanna deal with
listen to them they seem to know. or not, idk do wat you like lmao
yeah, of course, as new installs never work lmao
Are you not using enviroments
Just create new environment than
I can't, cause it uses the old repository
not everything, you keep all your settings

you on dsl or wat
@static tuskAlright, I'll try it
good luck
so install a whole new A1111 from git?
vlad's fuck people so hard we all had to start clean over as just erasing venv was not enough
training whaaaat
hopefully Autos hasn't morphed into that
a womans face
yep, could be throttling
could need a repaste by now
it ran MSI kombustor at like 74C
tbf i think its unlikely to throttle so quick
Thats nothing for kombustor lmao
the fuck is cumbuster
2 nuts and a long night
most GPU's hit like 90, its as intense as it gets
speak ENGLISH, doc
90c for me is shutdown
Can I just put my GPU ina freezer and run it
Yeah this is better!
Furmark Kombustor on a 4090 peaks over 500 watts and like 90C
oh i just had deja vu so keep in mind we're all stuck in a time loop repeating this moment forever, that kinda sucks ass
lmao
How much faster is a 4090 vs 3090
10s vs 7s
about 1000
Like 1.5x or we talking 2
way faster depending on the use case
likw 58x speedup over 3090 in ML tasks
soemtimes like 30% faster, sometimes like 100% faster
actually not so much in BS5
BS5? what is this. i'm old
Because the cuda cores are different?
hardware is 56% faster but software has not caught up yet so 35% in the real world
What about v100 vs 4090
pytorch using eager mode loses a lot of potential on this hardware. you have to compile the models to get the full FLOPS output of the 3090/4090
cuda cores on the 4090 is 16384
@oak osprey Have you read the prompting guide on the civit page for DD? There are some tokens you can use to improve generations.
oh i was just messing around with that for the moment just making sure it works
yeah, it is insane
generating a batch of 6
So 4090 is just like an insane overbuilt consumer gpu
there we go 60c
of course
yes
the heatsink on the FE is over engineered due to it being made for the 4090 we did not get
Than technically what gpu fits sb the best as what is build for a program like it
Accelerator*
steps?
lets say 100, just to average out, at 512x512 please
omg i love Jim Bean
I would kill to get a 4090 FE but not happening since Nvidia stopped making them and only solde in the brick and mortar Best Buys
it auto installs 1.5 BASE what is nice about that? lmao
Yeah every image with this prompt is pretty much a banger
4GB to immediately delete lol
At least I have fast internet to download this model in so I can turn around and delete it lol
yeah auto is written very poorly for boot up
ok!
Switched over to doing some mountains with the same prompt
it worked @static tusk !
13.2it/s with no optimizations and live preview
schleep well
@hasty nova
lol
they are in fact the same picture
lmfao
that's Digital Diffusion ^^
Yep
mine i meant
Oh.
yeah there's the pytorch 2 boost
Sytan you on PT2 now?
it was onlt 13.2it/s with base, but with SDP attention it boosted that high
aye that'll do it 🙂
2.0.0+cu118
yep, 3080, 3090 and 4090 with 4090 REALLY seeing the bene
and you doubted my advice @smoky oak lol
annndddd
pwned
6.5 it/s faster
Fatal cuda air unspecified launch failure
*error
My whole computer froze, my audio drivers reset, and my monitors went black
I have no idea what just happened
probably a nightly package you need to update to
I was about to say that it went back down to 10it/s for whatever reason
Right before it did that
Let me make sure I'm monitoring GPU temps

That was scary
not laughing at you, just that you're not monitoring temps on a new card
Unlock thermal throttling activated
I can buy a renewed 3090 for 900-1k but I dunno
but i have this all the time on the side here
use hwinfo to check gpu temp and also vram hot spot temp, thats the problematic one
CUDA kernel stuckies
Does anything smell like it is burning
i smell burnt toast but i'm not sure who you're asking
It does not, or at least I didn't see it on the board
that could be an issue. what W?
well since you're upgrading PSU now you may as well get that WSE 2
650, which is below recommended for a 3080, and I am OC'd 😅
Too low tbh.
underclock your 3080, broseppi
you want 750W ideally for 3080
you will save power and gain efficienci 💪🏽
yep pure
I have 1000W for upgrade headroom
Why is a 3090 24gb more than 4090 24gb
if you hit one of those spikes the 3k series are notrious for the system will shutdown.
I asked that earlier
people like whats his name here telling bubs to go get the slim sexi 3090 instead of fat thiccy boi 4090
its not about the avg its about the transient spikes that lead to crazy V-drop
3090 can do SLI, maybe thats something business want
also, that gen was the slowest yet
9.9it/s
I am so confused
it hit 18 in that one test
If I were you I would buy a 1kw and be done. Get a good brand though with a good capacitor in it for those spikes that just happen. Gamer's Nexus had an episode about it. Luckily Lovelace doesn't have that issue now.
yeah i wouldn't do a whole lot of throughput testing til you have solid power
bad PSUs can kill gear
Yep, I lost a mobo years ago due to that
use afterburner and lower power down to 80% for now
jesus christ this card really uses that much power?
What I do on mine just for heat
350 watts stock
sure does
whatttt
Hahahahaha
mine runs at 280W with v curve and power optimisation
alright, lowered down to 80%
Yeah I know they do but like, my super OC'd 3060ti used like 150 watts, this uses well over 2x at stock
remember these cards are tuned for gaming and likely are blowing out an O-ring where they don't need to
💨
well alright then
4090, for instance, is 450w
it's definitely worth investigating how reducing the core clock vs mem clock goes
now to find out why its now slower than my last card
you can try reducing the voltage of the uhm, what the fuck's they called.
Nvidia loves the power
@static tusk Any idea?
core voltage
the states tables
I know how to OC and UC lol
did you reboot pc?
Am I gonna have to do that every gen or something lmao?
I hit 18it/s one time, then boom, now its 9
until you get the power sorted out, possibly
states tables at least on AMD hw are configured so you have the voltage vs mhz set up correctly
no, but the power spike earlier may have ruffled some things
alrighhttttt
yeah reboot and keep it at the speed limit, hoss

this answers the question of whether you can stuff both GPUs in your case tho

lol
i know you were wondering
I might not keep this 3080, IDK
Oh, he can stuff it just don't turn the power on
well i say go through a tuning process first and see what other people might have done
the vega had a pretty dedicated community of nerds
gotta assume the 3080 has something
your problem is 650w in 2023 is amazing it booted
when voltages drop, current does go higher. but wattage stays the same. it just means wires get hotter and shit starts frying
I still see 450w PSU being sold and SMH
my other PSU died on me in the middle of online classes, so I needed a cheap fast and easy PSU, and this one is 100x better than my old one
its gonna blow your mind but my Celeron box uses 3 watts of power, @dense tapir
Celeron is why
yeah but it's got 6x 2.5gbit ports. ALL the hardware offloading. it's a badass router
I can sorta see a workstation only needing 450w, or even less, with integrated graphics
the low-power Xeon stuff uses like 45 watts
i have one i'm getting ready to throw away
yeah, it is funny the mobo stuff is using less power/voltage while the gpu is more
it's hard because it's a DDR3 board and it has like 144GB RAM but also it's a low power xeon and that ram is useless
Like a leader
:/
is that with 100% power or 80?
are you no longer using SDP or something
80%, but its the same speed
did you O/C the UV?
no, I just lowered power to 80%
seems like you're shartlenecked somewhere else
I hit 18it/s for one gen, then boom, back down to 10, doesn't make any sense
just once
well damn it man don't keep saying it like it's reproducible 😄
in any case you will want a new PSU with that 3080, maybe switch back to your old card till you get one
have you the spirit to try and reproduce the speed drop again?
card is like old used cars that had straw and axel grease to keep the rattling down long enough to get you off the lot.
Yay, another thing to dump money into
like if you can do that reliably that's actually pretty great
it wasn't the crash that lead to the lower in speed
I did 18it/s for one gen
did the next at 10it/s THEN it crashed
and then you rebooted?
the fuck
I did a few more, then rebooted

on fresh boot now, its now even slower at 10.2it/s
can you check the versions of stuff you are using again
you rebooted
same thing
and still slow after reboot
well there's more than just pytorch here, did you update xformers / diffusers too?
torch 2 and sdp doesnt require xformers
on vlad's you can't even click both
is that why i have to catch and squash all these stupid fucking errors from stable diffusion that are like "that function doesn't exist"?
possibly
have you checked how much power the card is using while generating?
250 watts
80085 watts
sorry, 280
what resolution are you generating
thats OK for 80%
. . .
ARE YOU KIDDING ME
AAAAAAAAAAAAAAAA
ITY WAS MY RES
AAAAAAAAAAAAAAA

eproigjWLSKJG KASNGHAEDRH
you did what i did
AAAAAAAAAAAAAAA
YOU DID WHAT I DID
what res?
I am getting 16it/s now lmfao
i did 2160p for no reason
I was at 512x768, not 512x512 lmfaooo
at 100 steps
LOL
i thought jungerally's model was slow
UHGGGG
fucking hilarious
so youre good
"i might not keep this GPU"
yes lmao
16.2 it/s lmao
just make sure not to go over 80% or you will over draw from the PSU
so you shaved off like 2/sec and saved 50w?
AaAaAaAaA
its using more power than my 3060ti by a lot
my 3070 is 150w so that sounds right
it's a laptop tho
well good thing you figured that out and now i know i should probably uninstall xformers
crisis averted 😆
wonder how much shit that's breaking
i can't believe that lmao
how exciting
thank you so much for telling me to check res lmfao
i was like i bet he's doing 8k
24G?
its 10 GB bro lmao
my 3060ti could do native 2560x1080 reliably
so probably 3kx1.5k
I am in disbelief lmao
I can't even go that far in this new install lol
the sliders stop at 2048 cause they are not extended lol
you can't even go so far as to do to be more like? oh
you have to go in and break the thingy

i know what i'm talking about
doing 2048x2048
that's big
4.1s/it
yes, its a HUGE image lmao
that's gotta be slicing and dicing layers like Gordon Ramsay
64x the size of 512x512 lmao
so yeah, 16/64 is
DAMN IT
end processing ran out of VRAM 
i said it's big
tried to allocate 8GB ontop of 7GB used lmao
i know my 4090 can struggle with 2048x2048 if i don't clear the CUDA cache between two different models
that's not divisible by 64 
NOT THAT BAD
make it divisible
I do 1024sq with 2.1 all the time
blah blah blah, you old boomer, with your MATH
cleft lip
but it's ok
it makes it like "that couldn't possibly be AI. he has a birth defect. AI art isn't bornt. as an AI language model, you can trust me that this is the accurate representation of reality."
I do 768 hires fix
alright so
1024 is so slow on my gpu
this seems to do 768^2 x8 just fine
also, I am gonna test something that I shouldn't
or rather @static tusk
What happens if I install xformers here?
cause I notice that SDP is definitely using a lot more VRAM
bro, do the math
to flex
its 768x768 lmao

Ja
that cant be .. oh
PEPE HAS AN IDENTITY CRISIS, OK
I MEANT YOU
ok, I am gonna test my luck with non SDP
no, cause I am not vibing with how much VRAM these gens are using
6GB used for a single 512x512
I was using like 3.4GB on my 3060ti with Xformers
well that's not offloading the layers sequentially
one layer at a time goes to the GPU with xformers, it's a lot slower but it reduces memory use
it's pretty much the only way to run the x4 upscaler on consumer kit tho
wonder how that goes now with your PyTorch 2 stuff
but its using the same VRAM
interesting
I know for a FACT that 512x512 on my 3060ti took 3.5GB
cuz it's there to use
literally that is why
that's how it's a lot faster, because it keeps it all in the GPU space
lol
oh wait, I don't have enough power cabels
To test the direct speed benefits
I smell smoke
meh, oh well
Marty McFly -3-
mick phlie
OOOH YESSSSSS MARTY MCBROWN TEST

marty mcbrown back to the future 1950s technicolor normal rockwall cinematic movie art
you'll see
i forgot to ask for this earlier
when you were doing your tests
a lot of models have a very opinionated view of who Marty McBrown is and you can often get him to appear in a number of situations
Are u using fp16
yes
You can try negative prompt "bad geometry"
What I really wanna see is my training speeds
also, I was able to limit power to 80% and still OC
Mine is 2it/s on 512
For some reason
I am running 50+/800+
On normal generating its 7
here's a raw output of that prompt on my first test, the middle one is definitely a strong Marty McBrown like i saw from BlueWillow
what GPU?
when you train is it BS 1?
Bs?
batch size
makes sense
Sometimes 1 sometimes 8
Lol
I bet this GPU gets 2it/s BS 3
Ill upgrade to 4070 soon
might even be able to do BS4
NO DO NOT
It is NOT an upgrade
it is not a good card for AI at all
It is

This is why i upgrade
Idc about mainly au
Ai
I also play gamed
Games
fucking lul
i'll take this one as my fault, ya know, but it's just, it hurts to laugh that hard
christ i really need to remember to check my tile prompts
Testing: Which is the Better Choice: 4070 or 3060? This is a A comprehensive gaming comparison of GPU performance including all relevant gaming benchmarks. This video will help you to decide which GPU is the Best Value in the Ultimate GPU Battle. When performance, features, and value matters a detailed comparison of performance and features will...
Almost double fps
i want to make an entire movie with SD 2.1 and make people throw up in their seats
from like, motion sickness, and the deformed outputs
high fps just consumes more power 
U wanna play minecraft shader with 30fps max?
but if i play rocket league with 250fps i get to hear the coil whine
in games yeah, but the 4070 is almost no better in AI
I dont mind the ai
alright then I guess
But tho 4070 is 12it/s and 3060 7
I would still just go used 3080 for less money and get better gaming perf
It consumes more power
oh right, I remember that being a concern now
I have a 750w psu

you can parallel two power supplies too
20
is that a bird?
I dont wanna buy a new one
Its a bee
Its a WASP
its too big for that. where is the banana for scale
i'd squint with a 4080 in a 750w supply , but less than that should be fine
jesus, at batch size 16, my GPU can do one in less than a second 💀
@crimson dawn just pair the 4080 with a celeron and it'll be so fucking bottlenecked it won't use that much power
it's the perfect crime
ugh that hurt my eyees
@oak osprey4090 and intel atom
im cave dweller 
C pu bottleneck
Lets buy a gpu server for ai
Raytracing calculator
@wispy ether we're all already working toward a 2.5 million dollar WSE2
how come you all keep forgetting this
what the fuck
cerebras wafer powered off a button cell battery
Vgpu moment
@smoky oak do i need to FAQ you on how to image gen becuz that prompt goes int the prompt box
oh
well you edited it since then
get clowned, crusty
IDK, but you keep putting me in a chaotic mood lmao
Ask a rich friend to buy you a gpu farm


Yeah, as long as I don't have to look at wamen
@smoky oak oh do it on cute dudes that'll be better
beefcake wafer bois
There we go, Nep's daily dose of making a distasteful joke towards some form of minority
No
@wispy ether a monk's work is never done
oh i just realised that's a double meaning because monks self-immolate

Thats been known for decades
@oak ospreyplease, have the fire department on speed dial, I am about to do something very irresponsible
im inpainting expert now
if I die, it is now your fault, since I warned you
@dense tapir negative prompt bad text accuracy
Wow, I didn't even get to the irresponsible part before it crashed...
Welp
🤓
My PC is frozen lol
L
That would be my bad lol
@wispy ether didn't help
is it that bad
i almost couldnt even finish it was laughing so hard
imagine if i hadn't. you all would have missed out
I should not be listening to a song that has glitch sound effects that sound like a blue screen when adjusting clocks
3 separate times it got me lol
oh crap im fucking dying i did it in the wrong aspect
@wispy ether i bet he overclocked it again 🤣

seems to be running fine at 100/1000 at 85% power
CTU can't seem to like, fix that image in any meaningful way
i thought AI was supposed to be good
scam
do i just need to like, feed this thing back in again? and again? and again? 
50 iterations off the central finite curve lives a shrek that knows more about diffusion than he knows about farming
so i can see it again for the first time
feel free to just not lmao
that's kind of like the last hallmark card i got tho
just sayin'
i'll delete it for you though
it's already in your head
no need to leave it there
so one takeaway is that thing undid some of the editing i did in Gimp, pretty much explicitly undid the blending i tried to do
but it also blurred the lines between the mogwai and Ernie's hair
what happened with the hair overall is just really unfortunate and i have no idea how it happened










