#🏞｜general-with-images
static tusk
lol...I just watch a video from Olivio on the BEST lora training. At the end his result was garbage and he had to merge it will realisticvision to get what he wanted.
static tusk
sure things change. thats the nature of SD, but he said the BEST...
hearty karma
aparently chrome iron man is nsfw
static tusk
yet you seem to be adverse to me pointing it out. I can still criticise him for it
clickbait or not
hearty karma
wym zovya sucks at nsfw?
yeah okay but thats specifically trained on tons of nsfw
something strange about what model? zovya?
timid heron
Try dreamboothing SD1.5 and merge that into Zovya
😛
The normal V1 is a little more suited to NSFW in my testing
it does. it's why some people do that when they've overfitted. it's a real balancing act to make it work. but looks like that might be the easier route
timid heron
here it goes
they're right. it's a lot of junk, I've tried several ways to delete the junk and leave the versatility but I gave up. So if you have the machine to run it, it's a little icing on the cake. but the normal V1 model will suit 99.9% of users.
concerning your tests though....
hearty karma
im pretty sure someone tool kitted the 16gb and it went all the way down to 2 gb
timid heron
I spend literal weeks and 1000s of generations of all subject types and scenarios, including upscaling, controlnet usage, outpainting, etc. when mixing the models. The ultra just had better flexibility all around. from product photography, macro, architectural, cars, industrial, background coherancy. So I decided to offer it to those wanting to tread new ground.
I've been creating and innovating in this space since SD initially released. it's been a fun ride so far. I think of the ultra model as just another experiment to see where it goes. It's always fun to pick and pull at the latent space and see what comes out of it. Like magic almost.
There's been some very vocal detractors concerning its large size and junkiness, but I'm not going to stop experimenting and seeing what I can make.
I am offering a full refund to anyone that's dissatisfied with the Ultra model however. Spread the word. 😛
tranquil grail
Characteristics of a Leo male
tranquil grail
/ Characteristics and movements of the Leo male
timid heron
I appreciate that level of scrutiny. I don't think I noticed it much. For me, that's an easy fix in an image editor. I think I was more focused on textures and the coherancy of objects, stuff that's harder to correct.
hearty karma
timid heron
my theory is, it's throwing out everything that's not part of the standard SD array
which might make it untrainable. But I've always been the type to train, then merge.
hearty karma
yeah i have ended up now only using zovya
Sen I seem to be having a problem with upscaling
im upscaling to 2048
yesterday it gave me 49 tiles for a 1024x1024 - 2048 image
now its giving me 16 and im confused
timid heron
lol. thanks guys I appreciate that. let's hope I can keep a good model going. it moves so fast, there'll be a new hotness in a few days
static tusk
now its giving me 16 and im confused
the size of the tiles eg 512, 768 determines the number required for the final resolution
timid heron
yeah, I can't wait to start playing with XL
hearty karma
i think yesterday after setting the tile width setting you gave me
it didnt actually set it to that
and stayed at default 🤔
oh nvm
valid marsh
Controlnet is a blessin'
Also found out a weird way to "outpaint"
Generate 512x512 image. Put in controlnet. Now wanna make horizontial expansion? Gen in 768x512. And 512x768 for vertical. Rinse and repeat. Upscale when needed.
dry crow
Also found out a weird way to "outpaint"
Generate 512x512 image. Put in control...
In img2img or in txt2img?
valid marsh
In img2img or in txt2img?
Txt2Img. Have a very strict prompt tho.
hazy fog
Hi everyone, where could I find this AI model? I've seen a lot of photos like this on Twitter lately
fading whale
wow is the tile upscaler built in or i need to add that extension?
static tusk
wow is the tile upscaler built in or i need to add that extension?
its part of controlnet 1.1
you need to get that extension
fading whale
its part of controlnet 1.1
dang my laptop crash whenever i use controlnet, it always says gpu is full
static tusk
yeah probably not enough vram for the models
dry crow
dang my laptop crash whenever i use controlnet, it always says gpu is full
Whats your gpu?
fading whale
Whats your gpu?
rx560
dry crow
Oh okay yea that wont work
sterile temple
Does controlnet have a memory leak? Seems like if I run it a few times I start getting out of memory errors even on regular txt2img generations without it
oak osprey
lol. thanks guys I appreciate that. let's hope I can keep a good model going. it...
omg do a 2.1 model next plz
oak osprey
Does controlnet have a memory leak? Seems like if I run it a few times I start g...
I find that happens on 8gb systems where many offload tricks are used but not on larger systems where none of those are used
sterile temple
I find that happens on 8gb systems where many offload tricks are used but not on...
oh ok, thanks
oak osprey
haha blue Willow pretends they invented this stuff but they're just using controlnet on artius
sterile temple
haha blue Willow pretends they invented this stuff but they're just using contro...
Never even heard of 'em.
smoky oak
Hi everyone, where could I find this AI model? I've seen a lot of photos like th...
Looks like chill out mix, or one of it's several equally unethical and problematic clones
smoky oak
haha blue Willow pretends they invented this stuff but they're just using contro...
Blue willow would die if their entire service had to go up against one of us Stabel Diffusion Disciples lol
Blue willow always has been and probably always will be mid
sterile temple
wow, put 'female focus' in the negative prompt and suddenly the model starts listening to your other prompts instead of generating pretty woman portraits
smoky oak
Yeah, it's pretty incredible how much crap gets overwritten by the sheer amount of female (typically underage female) focus on these models :/
smoky oak
wow, put 'female focus' in the negative prompt and suddenly the model starts lis...
I desperately want a model focused specifically on diversity.
I really really crave being able to type "woman" or "man" and receive unique and individual people of different ages and ethnicities without having to force feed NOT blonde hair blue eye bimbos into it
Or hell, it does the same with men too, blonde hair blue eyes
sterile temple
It's frustrating when you want a picture with a male character, and you have the put (man:1.7) male 1boy etc in the prompt and then about 5 variations of female in the negative prompt just to get some models to stop generating female portraits
smoky oak
It also depends hugely on the model you use, I find that most good realism models can flip between male and female 100% reliably
sterile temple
And then if you have preview on, you can see it generate a female in the lower steps before switch to a man. Then you end up with a man that has a weird open shirt showing off his nipples 😅
smoky oak
Not sure what model you're using that's being THAT difficult lol
But regardless, I am really over the objectification of women in these models, especially when most of them are aimed at clearly underage women likeness
Like, if you wanna generate a big busty blow up doll looking woman, by all means, go ahead, but I hate that that's the default for so many models
sterile temple
Well it depends on the overall prompt I guess. If there are other keywords it's picking up on and thinking you want a female character
calm canopy
Loli, child, teenager in negatives usually fixes any underage issues that I've seen, I usually have those applied whenever I make a portrait pic
Some people really like their lolis 😒
smoky oak
It's disgusting that we have to do that to offset the perversions of models
Chill out mix is the worst offender I have seen
Taking prompts and LoRA's of women in their 40/50's with keywords of "mature" and "woman", and making them look like 14 year old school girl pinups. It's rancid and should legit be removed from Civit IMO
calm canopy
Tbh if anime models are trained from being scraped off of boorus then it makes perfect sense why lolis are so prevalent lmao
sterile temple
Ah, I think it might have been the text embedding I was using in the negative prompt that was increasing the bias towards women too: bad-picture-chill-75v
with bad-picture-chill-75v
without
smoky oak
Yeah, that makes more sense
glad wing
smoky oak
Oh, I quite like this new vintage style I have been experimenting with
glad wing
Woman with platinum hair, in a dress of ice flowers, a beautiful crown on her head, detailed face, detailed skin, front, background frozen forest, cover, unzoom, choker, hyperdetailed painting, luminism, Bar lighting, complex, 4k resolution concept art portrait by Greg Rutkowski, Artgerm, WLOP, Alphonse Mucha, little fusion pojatti realistic goth, fractal isometrics details bioluminescens : a stunning realistic photograph 30 years
crystal deltaBOT
sterile temple
<@504972304732651521>
what keywords do you use for that?
smoky oak
sterile temple
oak osprey
kind quartz
tiger diffusion?
sterile drum
Futuristic Lace dress, Fashion Concept, led lights. Sexy atractive pose. Hotpants. Beautiful, cinematic lighting, 8K, sharp details. Crystal clear feel, High quality, Fine-art print, HQ. Tron vibes. Futuristic background. colourful. Hypermaximalist, hyperdetailed, award-winning fashion photography, professional colour grading, clean sharp focus. Clean details.
crystal deltaBOT
oak osprey
tiger diffusion?
impossible caustics experiment by mc Escher featuring Bengal tiger
digital art portrait tiger made from 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 where the wild things are
that's the full prompt
Marty mcbrown 1950s technicolor
kind quartz
i like these images
oak osprey
i like these images
then your prompt is: 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 where the wild things are
kind quartz
thanks but no needed 🙂 i did for weird images photo of 3d diffusion if i remember corectly. Now now sure. But yes much simplier result
oak osprey
you have to use SD 2.1 because 1.5 based models give: https://images-ext-2.discordapp.net/external/Q1dTAxg3fEeVvI_SMnMm7L9DSIvOfseFLGr5gOMF8fw/https/tripleback.net/public/discord/1683811565.80516086058d4bab69679becca6c8df8abe7a09.png?width=752&height=1004
SDXL takes it further
kind quartz
sdxl seems very nice but for me impossible to use it
photography of diffusion and then soup like best quality and so. Going to try it with different model.
oak osprey
what are those from?
oak osprey
nice
sterile temple
good thing about abstract prompts, you can generate 1920x1080 images without worrying about repeating elements
oak osprey
tried to make a scene of researchers trying to cover up ancient alien tech
that's in 1.5 via URPM
it took 15 images to get this one which is pretty much what i was hoping it would make
first image out of 2.1
sterile temple
prompt?
oak osprey
cinematic epic photograph far-away perspective of researchers standing under a contorted ancient alien technology spaceship looming overhead in its grandiosity of radiant foggy other-worldly glow with sparking equipment and small brushfires visible on the sidelines
Artius v2.1
oak osprey
i find that to be dependent on resolution
it's also the looming overhead
when the giant booba ship comes to take you back to Boob World
oak osprey
Blue willow always has been and probably always will be mid
it's the result of tuning the 2.1 model to produce aesthetic outputs by default for even the newest/laziest of prompters
it really limits the variability of the model in favour of safe results
to me it seems like they've got a BERT model trained on sentiment analysis and have a series of datasets they've set up to point various prompt styles toward one or the other of their LoRAs
so you can "prompt" that BERT model with their prompt, and it'll return a single word, which is the LoRA that most likely applies to their prompt
BERT is hella fast and low-overhead, so this is pretty sensible to do. you can run that shit on CPU no problem
sterile temple
😂 really liked that oled keyword I added to see if it would boost contrast a bit
oak osprey
lmfao try HDR
oak osprey
the U.S.S. Cloaca Maxima
sterile temple
football players???
sterile temple
changed to (deep blacks:1.6)
alright i gotta stop playing around with this and go to bed
oak osprey
ah i just came up with the first good idea for adding a kid to an image because of that bedroom with the melting shit on the ceiling
oak osprey
Alyson Hannigan: Space Explorer
@smoky oak going through and thinking of CTU'ing some of my old gens lmao
or just deleting others
that's back to the future alright
pure rapids
dense tapir
SERIOUSLY? WTF? This is the old ASUS back from 2008/2009 that bought inferior Chinese caps that exploded on everyone. https://www.youtube.com/watch?v=cbGfc-JBxlY If they do this on their 700 USD MOBO I dunno so much about buying their 4090. :/
oak osprey
so i had a friend run that prompt through MJ and it completely misses the Uranium Core
2.1 is so prompt sensitive if you just change "red-hot melting" to "icy tundra melting" this is how it comes out
@hasty nova your model continues to impress
Michael from GTA5 Photoreal
trying to get one of the characters in the marketing material to become real
he went all greasy and wet
dense tapir
SD did a good job then as I have yet to see anyone in Marketing that isn't.
sterile laurel
Lol there’s a whale in the background
oak osprey
yeah i changed the prompt up wildly for that last one
pumped it through the normal sd2.1 Img2Img pipeline instead of CTU
just tragic what this thing does to Wooley Mammoths
oak osprey
so generate them 
i tried super hard to generate a floppy disk for you but it just doesn't know what that is
wispy nest
@oak osprey I tried generating floppa, no floppa comes out smh
oak osprey
have you tried Crammy Hamilman?
hasty nova
<@800501037793280030> your model continues to impress
Thank you! Looks great.
oak osprey
Thank you! Looks great.
stunning lack of wooley mammoths in the dataset
takes a lot of work to get one lmao
hasty nova
stunning lack of wooley mammoths in the dataset
There were in fact none.
oak osprey
that would explain it
here's what failure looks like
that's suppsed to be James May and Richard Hammond, lmao
oh but this one is kinda badass. i'd play that game
fresh hound
oak osprey
@fresh hound your img inspired me to recreate a 1980s TV show look with Pedro Pascal
you have disappointed pedro
i made him a "medical judge" so he can throw your ass into the clink
hearty karma
Jon Pertwee?
oak osprey
how did i go for so long without using SD2.1
it's like, the fucking best
but that ain't matthew broderick 
fresh hound
Jon Pertwee?
yes!
wispy ether
he did img2img
he made a colored bad apple using the frames and putting it into img2img
oak osprey
i literally don't care or want it to get them 'right', that's not what this is best used for. i don't need NSFW generated by AI either. there's literally anything you could want to see in Google Images already.
2.1 in this context is far superior. look at this Danny DeVito with glasses
oak osprey
it can even do two separate people without combining their features lmao
wispy ether
he looks chinese
wispy ether
you just said the n word
oak osprey
did i actually spell it wrong? that would explain this
naw that's how ya spell it
apparently inpainting just lets you fix stuff easily and it looks great with no effort
oak osprey
try with my whole prompt plz
photograph of portrait of Arnold Schwarzenegger in 1985 film, technicolor medical drama, surgical ward, korean war, MASH
oak osprey
whatever you like to do
wispy ether
oak osprey
that hurts to look at and makes me feel super fat
wispy ether
whatever you like to do
bro became chris pratt
oak osprey
that's way more acceptable than my result though LMAO
wispy ether
tiny thistle
oak osprey
tiny thistle
idk
wispy ether
young josef stalin
oak osprey
asian john candy. i wasn't kidding about 2.1's superiority
i was joking that it looks like danny though
to me, if a model knows real celebrities, it's okay but i rather it just generate imaginary people
oak osprey
try Karras sigmas, Sen
oak osprey
clipdrop is SDXL, right?
tiny thistle
is this correct?
https://stability.ai/
oak osprey
por que?
donde es por soy no los dos directemonte soya montanya
oh, you back-merged base model in?
16gb what, the 16gb model file?
yeah i coulda told ya that. it's non-EMA so there's more flexibility for training
you make an EMA version for distribution
improves inference by a ton
whatever base you use, should be non-EMA, that's all i'm sayin
when you're done you mean?
oh, i think that's for exporting it. but i've never used that
wispy ether
oak osprey
oh, fine-tuning apparently is okay with EMA but if you want to add whole new concepts to the model you should use the full model
yeah and if you use a lower learning rate with larger number of steps, eg. longer runtime for training
it does better without overfitting
jesus christ
wispy ether
jeez
oak osprey
that was it constantly paging in and out of your memory
get 128G RAM
Training the text encoder in addition to the UNet has a big impact on quality. Our best results were obtained using a combination of text encoder fine-tuning, low LR, and a suitable number of steps. However, fine-tuning the text encoder requires more memory, so a GPU with at least 24 GB of RAM is ideal. Using techniques like 8-bit Adam, fp16 training or gradient accumulation, it is possible to train on 16 GB GPUs like the ones provided by Google Colab or Kaggle.
timid heron
looks good
I've always just combined. only difference if you want to remove duplication like when making an inpainting model for example
overlapping occurs when merging and they just cancel each other out
timid heron
always working on the next model
do you see the same thing in my preview images? I always use (detailed skin:1.3) cause I love skin textures
timid heron
when prompting, think about what words the trainers would have used for that image.
wispy ether
timid heron
and put that in the negative if you want less of it. totally a duh comment. but have you tried blemishes in the negative?
wispy ether
hearty karma
guys do yall know if i could copy and paste my web ui to a different storage drive and have it still working?
wispy ether
guys do yall know if i could copy and paste my web ui to a different storage dri...
yes
i tried
timid heron
could be an unintended side-effect of another prompt. I seen some crazy things happen. I kept getting ocean and water when I didn't want it. figured it out. was coming from my "volumetric lighting" prompt
hearty karma
nep am i going to have to fix anything after i move it?
wispy ether
wispy nest
That's me
hearty karma
waiting for sd to install is pain
digital hornet
gaunt heron
timid heron
try a different vae that isn't as strong. which one are you using?
it is, but I prefer to leave the flexibility to use other VAEs as needed so I don't bake it in
oak osprey
it is, but I prefer to leave the flexibility to use other VAEs as needed so I do...
no worries, people like me pull the model apart and put it back into its separate components 
these single 180GiB files are like blah
stone cipher
Dark souls, Jack Black
oak osprey
fuck yeah, peter is back
hearty karma
i think my SD will load up extremely fast now :D
hearty karma
i put my sd onto an ssd
hearty karma
what?
oak osprey
ssd is ez
oak osprey
make it load faster from ssd than it does by default 
that is impressive
i used to work for a storage appliance company and we had a "feature" that would literally just scan and touch / read all of the most important files after the system reboots, to keep them in the mem cache
something like that could be done for your desktop 😄
there's a Linux kernel module called btier that runs block tiering between different storage devices, eg. all your most-recently used 1M chunks end up on the fast storage, and after being inactive for a while, they are moved to the slower tier. and you can have numerous tiers, and it just keeps cascading to slower and slower equipment.
stone cipher
oak osprey
we should send Stable Diffusion 2.1 to space with the next Voyager style probes
have the instructions on how to use it embedded on a golden platter, like we did with the original probes
just let them do unguided diffusion on it and see all of the weird sights of humans on earth
no prompt, lmao
hearty karma
thats a lot of cars
oak osprey
i think it's a small number of the total cars that exist, however
hearty karma
well yeah..
oak osprey
this isn't AI, i just finished converting my truck to a dually today
edgy shard
hi guys i am frustrated and need help. i'm trying to do some inpaint sketching and for some reason it's not drawing the inpaint sketch area only with what i ask in the prompt.
calm canopy
might not help but what happens if you generate with original instead of fill?
edgy shard
i get a lot of noise and blurriness
goal is to fix a hand and add some nail polish
focus area
result
blurry image with no attention to the area i inpainted
timid heron
yes
probably, but I'm making tools for artists. they may want a flatter color palette for color grading post processing. so I leave that up to them.
edgy shard
here's an example of inpaint masked, fill, only masked,
i mean that's a totally different image
hearty karma
kind quartz
noticed you are using dpm adaptive, what about eulerA 80-100 steps?
edgy shard
euler isn't for hands right
kind quartz
not you 😄 @hearty karma
but it is suggested for inpainting @edgy shard
edgy shard
ok so uh, what are the settings to inpaint sketch correctly? inpaint masked, fill, only masked changes the entire image
edgy shard
ok im mad, i'm about to table flip and walk away. what are the proper settings to adjust inpainted selections
?
the entire image changes
yeah it does i am getting a change. i want the hand to change
wut
hearty karma
your never going to get a solid hand when its that teeny
edgy shard
here is what i touched up in photoshop
but it's clipped
so i masked over that to fix it
is that not how it works?
hearty karma
yes but you have a pretty large image
edgy shard
yeah
hearty karma
with a tiny hand that your trying to fix
edgy shard
alright
hearty karma
edgy shard
ok happy with what i have now, going to do a 2x upscale
oak osprey
💀
hearty karma
💀
what?
i think something has gone wrong with my ctu and uu
64...
the total number of steps keeps going down
first use it was 700+ second time 200+ now 60+
what is going on
wait
does the denoising settings change how many steps it uses?
and i am on the web ui
no i dont remember 🥹
oak osprey
that looks awful lmao
reminds me of spray-tan lines
like she went to Rudy Guiliani's
oh you could just take secret pics of her jk jk
https://en.wikipedia.org/wiki/USS_Callister
getting a real USS Callister vibe here
The episode follows Robert Daly (Jesse Plemons), a reclusive but gifted programmer and co-founder of a popular massively multiplayer online game who is bitter over the lack of recognition of his position from his coworkers. He takes out his frustrations by simulating a Star Trek–like space adventure within the game, using his co-workers' DNA to create sentient digital clones of them. Acting as the captain of the USS Callister starship, Daly is able to order his co-workers around, bend them to his will, and mistreat them if they get out of line. When Daly brings newly hired Nanette Cole (Cristin Milioti) into his game, she encourages the other clones to revolt against Daly.
"I have 3D Printing nipples, Greg. Want to milk me?" 🧓🏽
wait, that's not the line
raven ivy
hey
oak osprey
i was looking at your outputs and knew it was maxed out
hearty karma
thanks for reminding me about that
oak osprey
uh yeah
depends how much you need
if it's really already a clear image, go low
.3
hearty karma
instruction manuals are for nerds :-)
oak osprey
if you're actually upscaling, you might want .6, .7., .9. 1.0 is like, JUST THROW ALL THE NOISE AT IT. it completely replaces all the details of the image at 1.0.
it will then probably do nothing
hearty karma
are you guys speaking about the denoise strength thing?
oak osprey
no
hearty karma
no because you arent gaining details anymore then
oak osprey
yeah you need img2img to do something and depending on which img2img pipeline, that strength matters more or less
oak osprey
the non-CN img2img is more wild with strength and needs lower values
it replaced the helicopter with a fish at a high strength level
but that's why CTU was made. to keep the image and replace the details
i don't even use UU, it's not needed. CTU on its own is crazy. the Ultimate Upscaler, i had GPT4 explain the code. it's doing band pass and blends by making the padded edges transparent on the tiles before recombining them. it's not a great method, it shows up in the histogram
hearty karma
ultimate upscaler definitely helps
oak osprey
but not needed
dense tapir
Over the last week I have seen this more and more.
dense tapir
Probably getting ready for the 7950
oak osprey
when that drops, the 7900 and 7950 will drop in price more
dense tapir
Personally I don't give three shits about the W7900. Let's talk about NON business related cards.
oak osprey
i'm going to keep talking about the W7900 because it's 64G of VRAM in a single package for far cheaper than Nvidia would do the same
dense tapir
Like people think when the Titan comes down all consumer grade cards follow which has been historally BS.
oak osprey
it's a $4k GPU for 64GB VRAM
dense tapir
Yeah, they are in their own realm. Consumer, Professional, and Business.
oak osprey
HBM2 memory
dense tapir
I sure in the shit don't.
oak osprey
yes, but the ones who do, will stop buying your precious 7900XTX and 7950s don't matter to them
dense tapir
Aye
dense tapir
We don't like to listen...to you, not that we don't like listening. When you come back to reality then we can talk about it.
oak osprey
you can't even argue what i'm saying on its merit. you have to turn into an insulting person instead, and discuss whether i'm detached from reality. that is the sign of someone who lost.
hearty karma
oak osprey
if i were going to buy a 2nd GPU for my workload, i would likely be looking at **two ** AMD 7900 24GB cards.
dense tapir
damn, I get that much just rolling drunks each night.
oak osprey
every non-W7900 i buy, is two 24GB GPUs taken off the market
ergo, when the W7900 drops, more 24GB GPUs are on the market.
dense tapir
SMH
oak osprey
keep that reading comprehension score low, Sen
hearty karma
if i were going to buy a 2nd GPU for my workload, i would likely be looking at *...
ok... but like 99% of the population dont have the money TO EVEN OWN TWO
let alone even own one...
dense tapir
Shrek muddled the brain cells.
oak osprey
where did i say anyone else would be buying two? that you would want to buy two?
this is what i mean. you guys can't read
i said i will be buying two, which means two fewer GPUs for you all to buy. do you know how supply and demand work?
like talking to a brick wall
dense tapir
You hold yourself in too high regard
oak osprey
because you all will, as you say, never buy a W7900, you are not in competition with people buying those
hearty karma
oak osprey
ergo, when the W7900 is available, your buying market opens up
dense tapir
Well, I know with Nvidia you can't use consumer grade cards in datacenters and stuff. If you do then don't get caught.
They lock you in. WIth AMD I don't know.
smoky oak
What the hell did I just come back to
dense tapir
At least consumers have told Nvidia to screw off and Jensen just got a 10% reduction in pay due to his BS and people are not buying cards.
Of course he still makes 21m per year though. :/
loses 2.1ish million and that is like losing a dime to most people
oak osprey
yeah right you say that but he said on twitter he can no longer open a fresh pair of underwear every day. he used to throw them out at lunchtime and switch to a 2nd pair before the day was up.
now he has to do laundry
like everybody else.
smoky oak
What in the actual hell is going on in this chat
calm canopy
hell
smoky oak
I leave for half a day and come back to some of the stupidest takes and analogies I've ever seen
oak osprey
@smoky oak idk, i was explaining how people who can make use of the 64G will stop buying 24G cards and they went AWOL
smoky oak
Also, side note, now that Peter is no longer here, I'm working on taking the solo title of realism expert in the server
oak osprey
peter came back, mf
smoky oak
I'm working on modern and vintage realism now
oak osprey
you missed so much
smoky oak
He came back? After that whole dramatic ordeal about saying how he needed to delete every message he's ever sent because he's deleting his account?
smoky oak
You're kidding me right?
oak osprey
no
smoky oak
Great, now that I am messing around with vintage/analog realism, he's going to take every single image generation I make as a personal threat and challenge again
hearty karma
smoky oak
I don't know why Peter has to always take everything that everybody does as a personal attack against him, so I'm just going to keep him blocked so I don't have to pay attention to him.
For now I'm working on my first style LoRA, and I will be further refining my analog realism gens
wispy nest
so you are saying the 7950xtx will be faster than a 4090 for sd?
smoky oak
@oak osprey did he come back with the same account? Or is he on a new account?
smoky oak
so you are saying the 7950xtx will be faster than a 4090 for sd?
Theoretically, it very much could be
AMD has been rolling out significantly better AI drivers for their GPUs recently, and from what I've seen from leaks, the The 7950 XTX is faster in compute than the 4090
So it very much could be better at AI
oak osprey
<@636706883859906562> did he come back with the same account? Or is he on a new ...
same acct
wispy nest
hmm maybe ill grab that if its only going to be 1200$ or so
oak osprey
I don't know why Peter has to always take everything that everybody does as a pe...
he doesn't speak english
hearty karma
whos peter?
oak osprey
hmm maybe ill grab that if its only going to be 1200$ or so
and just like that, the future price of the 7900XTX goes down
smoky oak
he doesn't speak english
He does speak English, but he just always has a problem with what everybody does. Even when people tell him that it's not how he takes it, he insists that it is
oak osprey
He does speak English, but he just always has a problem with what everybody does...
he could be an AI
did you ever consider that? hm?
maybe his images are so good because he's AGI v1.0?
smoky oak
Now you're just being annoying 
oak osprey
Now you're just being annoying <:Glare:1073220046207340604>
idk why you have to personally attack everyone who does anything 😭
smoky oak
It was a joke :/
📢 🧓🏽
smoky oak
But I am not joking about how he is very combative with everything
I only started doing analog realism cause I thought he was gone and wasn't gonna give me problems for it :/
Anyways, i am just gonna mess with analog realism, IDC if he gives me problems for it
Anyways, what are you all up to today?
hearty karma
hearty karma
nah
smoky oak
ah, all good
hearty karma
is just meant to be robit
oak osprey
lmao
oak osprey
But I am not joking about how he is very combative with everything
I only start...
i got you, fam. you do you
smoky oak
man, look at that detail
wait what-
wait
no way
ok, nevermind lol
I thought this image came from Mature Male Mix lmao, but it looks like its just a graphical bug, its actually from photo real lol
timid heron
the recommended vae doesn't count anymore since you've changed the model. try a weaker vae like anythingv3 or ClearVAE that will let you crank up the CFG some. other ways to let you turn up the CFG scale are adding more tokens to the prompt or increasing the samples. But if you don't need more CFG, just let it flow. You can type in a number instead of using the .5 increments to give you more precision. Also I do have a kofi link on my civitai pages, thanks in advance 😍
smoky oak
@timid heronNice to meet you, I have been putting your model through hell and back, and have found that it has been the best photo real model I have had the pleasure to use
timid heron
<@112952647442501632>Nice to meet you, I have been putting your model through he...
thanks ❤️
smoky oak
fantastic work, I hope to upload a large collection of images to show off its capabilities at 4k and above!
oak osprey
technically it's not 4k
smoky oak
I don't know how many times I have to remind people that they are screenshots 
oak osprey
you'd have like 15 of those guys if so
smoky oak
It's a screenshot
oak osprey
it's upscaled 😛
smoky oak
Yeah, but upskilled 4K is still 4K...
oak osprey
yes, but it's independent of the model. those images are 512x768 native, unless the model can do 4k native without repeating etc
smoky oak
What are you talking about? You need a good model in order to upscale
There are lots of models out there that can't hold up at those resolutions
oak osprey
i upscale with base 1.5, or a good blended model, and both produce amazing results. the base model matters for producing a good starting point. but pretty much all of the 1.5 and 2.x models will do great work at smaller resolutions
and the upscaler does its thing on total trash too
remember the CTU example code upscales a 64x64 puppy to 1024x1024
using base 1.5
smoky oak
By that logic then you are suggesting that any model can upscale to an infinite resolution, which is not the case. Each model is limited based off of various factors
An anime model is not going to be able to upscale a 4K photorealistic portrait, that's a limitation of the model
oak osprey
they literally can. i looked in depth at the Ultimate Upscaler code yesterday. it is using multiple blending techniques and one of which is simply making the overlapping portions of the tiles transparent so that they can overlay on each other
smoky oak
Yes, in theory that's how it works, but in practice there's a practical limit to how high resolution you can go
oak osprey
the main limitation for how many tiles you can stitch together is your system memory, but that's resolved by paging out to the page file
smoky oak
The better the model, the better the composition understanding and token strength, the higher the resolution you can go before it starts to create high resolution blur
hasty nova
oak osprey
oh that just means you need to do intermediate upscaling steps to a middle ground
smoky oak
Nice
oak osprey
that stops the blur from occurring
smoky oak
oh that just means you need to do intermediate upscaling steps to a middle groun...
That actually makes it worse
I've spent hundreds of hours finding the best way to go about this, I know how to actually get around that sort of thing, and that's not how to do it
oak osprey
well not on my bot, which i can flip many knobs for and change its noise level
you can completely nuke all of the detail from the image, and CTU will rebuild it.
smoky oak
Yeah, and you can also just give it a blank white image and it will be able to hallucinate information out of it, that's because it's a diffusion model, you can only hallucinate the information that the models trained on
oak osprey
a 1024x1024 image is literally 1024 times larger than a 64x64 image, and the default CTU model (v1.5 base pruned EMA) upscales that like a champ
smoky oak
You also don't need CTU in order to do that, you can do it with basic image to image
It's able to take a blurry image and hallucinate the information based off of what prompt you give it, and 1024x1024 is just barely out of the reach of its base resolution target, it better be able to do that, lmao
At that point it's not upscaling anything, it's just generating a whole new image that's very very loosely based off of a blurry blob you provided it with
oak osprey
yep and the way the UU works is by breaking it up into whatever number of tiles, and upscaling each to 1024x1024. it has to upscale to 1024. it can not work at other resolutions. it produces absolute trash. one of the dimensions has to be 1024, and the other is resized to keep the aspect ratio - that's in normal CTU, not the tile upscaler, which makes everything square tiles.
smoky oak
except you can use different aspect ratio tiles?
oak osprey
the magic of UU is that it has many different tile creation patterns and uses padding around the edges so that it processes beyond the boundaries of the tile, and any "crappy model" can be given more padding to have more context of what it's working on
smoky oak
if you try to generate a high res image of somebody with skin texture with a model that doesn't understand what skin texture is, its not gonna work
its not hard to understand
oak osprey
you can change the prompt to the img2img and help it do better, but yeah, there's always going to be something like that. but if it already has SOME skin detail, using lower noise strength will go from that rather than erase it (like 1.0 does) and replace with its own
using many iterative runs of a low noise strength did pretty well, especially when manipulating the CTU img2img prompt so that you can instruct it what you want more or less of.
smoky oak
Yeah, and that was the basis for my ultimate upscale guide V2, but it also has its own huge issues which have to be fixed by the workaround in my ultimate upscale guide revision 3...
oak osprey
idk why we're even arguing that, it just seems like you feel you're the domain expert on it and so anyone saying something that doesn't jive, is very wrong and stupid. all i was taking issue with is you saying this model has "great 4k abilities", when it's just photoreal at the normal ones and same deformations at higher resolutions. that was my only point. calling it a 4k model will piss people off, who think you mean on its own.
but you also tell me 2.1 sucks so like, opinions and all.
smoky oak
Nobody said it was a 4k model, I was saying I have goals to show that it is able to hold up at 4k and higher with upscaling, which many models I have used are not able to do, as they start to hallucinate details and start overbaking the same details.
Nobody said I am an expert, but I certainly have the most experience with ultimate upscale between the two of us, which is not something to be slept on
you suggesting that using iterative upscaling over and over again at lower upscale factors is literally the basis of my V2 upscale guide, which leads to the same issues I, and the people who used it, ran into.
It can only go so far before you have high res blur, which then needs the new steps introduced in my V3 guide revision in order to fix
oak osprey
well, i don't use A1111 or any of its scripts. whatever issues they have are likely their own.
smoky oak
but you also tell me 2.1 sucks so like, opinions and all.
Also, never said it sucks, but alright. What I DID say was I have yet to use a model that can match my 1.5 models I use without having to overcomplicate the process
oak osprey
well if you're using SD for NSFW then i could see why 2.1 won't work for you. it is likely resistent to Dreamboothing it on women you know to have them suck dicks
smoky oak
Sen, don't start with me, you told me ultimate upscale was trash for over a month (after you basically demanded I showed you my process) before you turned around and preached it as gospel lmao
so I was just faking my 4k gens back then? lmao
oak osprey
lmao the original img2img is incredible but for different reasons
smoky oak
oh come on, I RARELY get into arguments here lol
oak osprey
right, you and me get along and we don't even use the same scheduler sampler 
smoky oak
That won't work
having more than 1 tile at a time will not work, as it violates one of the principles that UU uses
oak osprey
i would recommend looking at how "enable_tiling" works in the VAE first, because it's almost the same damn thing and it should be possible just to upscale using tiling WITHIN the VAE
smoky oak
or at least it won't work in its current form. Using it with Cnet might be able to help around that limitation
oak osprey
what's weird is it still runs out of memory on my 80G with VAE tiling and trying to make an 8k image
so there's something there that still needs help, code-wise
like the state of the image needs to be moved to CPU after the tile is processed
but it stays on GPU and then it tries to reassemble and stitch them all and 💩
if this seems like ooga booga gibberish to you, maybe turn back now, or just have lower expectations of the outcome, but higher expectations of how much you will learn.
smoky oak
there is, I have been saying that the whole time lmao
oak osprey
if you can even point me to the location where this happens in the diffusers source, i will go through it with you
smoky oak
you can't have 2 tiles at the same time, cause they need to look at the padding between to line up
if they can't see each other, they can't have the same padding
but, that is where Cnet might be able to help
oak osprey
but, that is where Cnet might be able to help
cnet should just, imo, be baked into the VAE
but i'm honestly an idiot and i don't know what controlnet is doing
smoky oak
on the oppoisite side, sure
but they need to be one after the other when sharing an edge
oak osprey
it uses a circular tiling pattern in ultimate upscaler, making the image wrap around so that there's no edges to the image
smoky oak
that is the big limitation here, but cnet might be able to help there as it shows UU what the whole image looks like at the same time
oak osprey
that blows my mind tbh
it's very well written code and i wish it were a standalone library
but everyone is like "just use a1111"
you just need queuing and batching extensions in A1111 from what i hear
you can also write python scripts to automate the calling of it
there's some REST API it exposes and you can send a payload to like /generate/img2img
so you can create python unit tests for your models
hasty nova
I just realized I have been trying to do some controlnet tile upscale and I never hit enable on controlnet 💀
hasty nova
I just realized I have been trying to do some controlnet tile upscale and I neve...
the result wasn't that bad though
smoky oak
you can do queuing in A1111 without even using an extension, just open the web interface in multiple tabs and type in your prompts in each, then hit generate on each in succession and it will queue them one after the other
I have used that feature several times
it hasn't for me at least
all it does it tell it what to generate next, so its really just the seed, prompt, and settings
it does one gen, then does the next
My PC is a webserver, so I and my friends use it at the same time, so it just puts us in queue behind each other
oak osprey
sen nothing stops it from being both a library and a library that's used in A1111. i am going to say this once, because i believe it very strongly. open source's greatest benefit are the contributions that come from people who use it in ways you did not expect. they're running into issues you aren't hitting. and they're contributing improvements and fixes.
if the ultimate upscaler were abstracted into its own library, i could immediately begin a) using it, b) contributing fixes to it, c) contributing performance improvements because i know how to parallelise things and automate benchmarks so that we don't waste time clicking buttons
smoky oak
You can access my SD web UI anywhere in the world if you have my IP and proper port forwarding port
smoky oak
Maybe one day I will mess around with linux again, but my Volta ML tests were enough to deter me indefinitely as of now
7 hours of installing and bug fixes for a hyper specific ML API just to find out its slower than A1111
oak osprey
well, i will say this. i tried Ubuntu expecting it to be easy lately and it was a pain in the ass and i'm a professional kernel dev who's used Linux since 2005
they lock everyone into what their devs thinks works well, and then to get the newer versions of anything (or older versions, in NVIDIA's case), is impossible
smoky oak
I had to use ubuntu for Volta ML
oak osprey
gentoo has a different curve to it, you have a rough time getting everything set up but then it's amazing and flexible
smoky oak
man, if Volta ML worked reliably on my PC and this new GPU, I would be steamrolling 4090's in generation speed lmao
oak osprey
when i was 15 i spent all night installing gentoo and my computer wouldn't boot for 4 days
i was persistent af
smoky oak
I get that persistence with some things as well
its good and bad 😅
I did 1000's of upscale data collections with UU, only to have my method not be reliably viable cause of an issue with A1111 at its core
however, I was able to learn several small tricks that allow me to generate higher resolution images with less fuss
oak osprey
man, making a good dataset is an art form of its own
people are always like, sHreK, Y u No TrAiN oN a100 80G???? and i'm like, bcuz datasets, dumbass
smoky oak
I spent 500+ hours on that guide in the end only to find out that SD itself was the problem
oak osprey
yeah SD is a miracle it works tbh considering ALL THE FUCKING ISSUES with their ahem datasets
smoky oak
people are always like, sHreK, Y u No TrAiN oN a100 80G???? and i'm like, bcuz d...
Data sets are super easy with LoRA's, but I don't touch dream booths cause they are a pain in the ass and take way too much time to achieve decent results with training
oak osprey
this just keeps happening, too. am i the only one noticing that? like, LLaMA was trained on a subset of The Pile which was discovered to have so many problems, it is no wonder that thing is as batshit out of the box as it is
hasty nova
smoky oak
I can train a LoRA on somebody in less than 30 minutes and have it look excellent. You don't need dream booths for a single person lol
hasty nova
medieval peasant dude
oak osprey
well dream booths are incredible though, that's how Junglerally got DD working
hasty nova
anyway I gtg
smoky oak
well dream booths are incredible though, that's how Junglerally got DD working
Yeah, they are great for some things for sure, but for most things its just pointless/overbloated
oh, that reminds me
oak osprey
yeah LoRAs are like cheat mode.
smoky oak
@bold cave Did you ever finish your Amber heard dreambooth?
I have been waiting like a week
oak osprey
yeah everyone on my discord server has been waiting too
cant believe you guys didnt even know who she was
smoky oak
All this big talk about doing it better than me, and I finished in less than an hour, and you never submitted
if you say so, I'll believe it when I see it
oak osprey
i watched a 2.5 minutes long Dreambooths videoing and i'm an experts now
smoky oak
imagine having to transfer dreambooths, LoRA's could never
oak osprey
LOL feeling inadequate? i was making a joke about the fact i literally watched a 2.5 min video
smoky oak
who is this supposed to be?
oak osprey
you have to prompt the upscaler "fuzzy lady"
it's funny, do it
Snoop had an old dude in a scratchy looking jacket and i made that thing into a damn shag carpet
smoky oak
I look forward to you trying to beat my amber heard LoRA
oak osprey
spin up AWS Spot Instances at a low price per hour and run automated tests and email the results to yourself
idk Sytan's will probably win. it sucked the soul out of this room just by being pasted. that's really good
idk if you all know anything about her but she's a terrible fucking person lmao
smoky oak
oh yeah, we do lmao
smoky oak
idk Sytan's will probably win. it sucked the soul out of this room just by being...
so, its just a screenshotted image of her
oak osprey
you were like "who is she" 🤣
smoky oak
I wanted to see if Sen would hit me with the "it looks nothing like her"
like they always do lol
smoky oak
right lol
smoky oak
You know what, I have a better GPU now, I am gonna update my amber heard LoRA to be better. Lord knows I am not under any time constraint lol
smoky oak
@bold caveI have no doubt you can make a dreambooth that looks like her, but I DO doubt that you can gen her well enough to look realistic
if this is what we have to expect for realism, then I rest my case
even better
oak osprey
SD2.1 doesn't need upscaling ❤️
Cloaca ahahahaha
the prompt was cloaca nocturna
pretty close to the text and it nailed the cloaca
made me wonder if it's reproducing a book
oh, SD2.1 can do round glasses
that's "Danny DeVito"
new and improved
what the hell happened to her neck and forehead
smoky oak
plastic, is what happened
oak osprey
omg dude
why would you neg Uma
she's ma boi
WHAT
that's cold
that's so cold
you should have used Stevie Nicks as a neg
or Angelina Jolie
her eyes are damn near on the side of her head
me over here making NSFW Betty White images and you're all like, "omg make them younger"
the butter lady?
no one was surprised
smoky oak
@dense tapir My LoRA training is SO much faster lmao
like 90% faster on this GPU cause of more VRAM and a faster GPU in gneral
Went from 2it/s at bs 2 to 1.85 it/s at BS2
wispy nest
jesus people. i did a doubletake. i was like "don't have that many people blocked here...are they all chatting at once?"
smoky oak
@wispy nestHi combo :>
Did you see I got a new GPU? I can run vicuna 13 b now!
I tried it and it was janky AF lmao
wispy nest
i did. grats on that!
wispy nest
that is obscene. i paid nearly that much(290) for a 3060.
smoky oak
Highschooler bought it for his PC, but he got accepted to college in a different state, so he wanted to sell all his stuff to have a car in the new state, so it was one of those "I need this gone NOW" sorts of things
it is a beauty, super happy with it
this is it
wispy nest
it's been a while since i used a locally run llm now. i had to look which one i was using for conversations. i switched to TheBloke_gpt4-alpaca-lora-13B-GPTQ-4bit-128g for convos and stablevicuna 13b for storytelling
smoky oak
ah, I see
wispy nest
but again, i haven't ran them for at least a couple weeks now
basically since stablevicuna came out
smoky oak
I am still trying to use my model for the same uses I brought up before, so I will have to look for models for that
Vicuna 13b keeps going insane when i talk to it lol
"how are you doing today?"
covers eyes so I can hear you "Remember, I will be home later after tomorrow cause I need the chair for the new job I start yesterday" writes that down in your email
Stuff like that lol
static tusk
sounds like 6 fingers are typing
smoky oak
bruh lol
smoky oak
hasty nova
oak osprey
jesus people. i did a doubletake. i was like "don't have that many people blocke...
imagine flexing about how antisocial you are that Sytan is your only friend
hasty nova
wispy nest
@smoky oak you are in the states?
oak osprey
Dwayne Johsnons starring in "Dwayne Johnson 2: The Two Johnsons"
smoky oak
Yessir
smoky oak
Dwayne Johsnons starring in "Dwayne Johnson 2: The Two Johnsons"
what in the hell did you do to him lmao
wispy nest
I'm running Vulcan now and getting 3it/s opposed to 24s/it haha
hasty nova
I'm liking this prompt
oak osprey
it can kind of get ol' Dwayne Johnsons sometimes
wispy nest
I haven't used OfferUp in years
hasty nova
@oak osprey try it with DD, here is the prompt:
Positive: photograph, medieval peasant in a medieval village, portrait, close up, fantasy, cold lighting, dramatic, overcast, cloudy, dirty, melancholy, beautiful, mud
Negative: low quality, low res, messy, grainy, smooth, sand, big eyes, anime, fractured, cracked, wrinkles, makeup
wispy nest
Honestly think I'll get the 7950
wispy nest
I'm liking this prompt
You render in 512x512 and upscale?
oak osprey
my negatives are (deformed, distorted, disfigured:1.3), 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
hasty nova
You render in 512x512 and upscale?
No, 640x768 and hires fix 1.75
wispy nest
No, 640x768 and hires fix 1.75
Ahh ok I'll try that
hasty nova
I'm using a 2.1 768 model, so if you are on 1.5 it won't work very well
oak osprey
and that Dwayne is DD
smoky oak
it can kind of get ol' Dwayne Johnsons sometimes
why does this look nothing like him lmao
is that was 2.x thinks he looks like?
hasty nova
my negatives are `(deformed, distorted, disfigured:1.3), poorly drawn, bad anato...
I usually make my negatives as I go, start out simple then add what I need based on the style of the image
wispy nest
I tried 2.1 and 1.5 but 1.5 seems to be a little better to mess around with for me
oak osprey
is that was 2.x thinks he looks like?
its all my other prompt bits tbh
hasty nova
my base neg is low quality, low res, messy
wispy nest
it can kind of get ol' Dwayne Johnsons sometimes
Chinese rock
hasty nova
My PC is on top of my desk, in a corner blowing out to my closet (and its side panel was taken off for cooling reasons so it also doesn't get too hot in there)
Got it to actually show the background a bit more
smoky oak
ahhhh lol
wispy nest
My PC is on top of my desk, in a corner blowing out to my closet (and its side p...
Was thinking about venting mine outside lol
hasty nova
Was thinking about venting mine outside lol
not a bad idea, but it could be dangerous if you don't do it right
oak osprey
Korean War version of Dwayne Johnson
wispy nest
not a bad idea, but it could be dangerous if you don't do it right
Yeah think I need to just do a bigger radiator so the liquid traps more heat
oak osprey
man i can't tell if post-Korean War Dwayne is better
hasty nova
the hands look kinda funky but are actually not that bad, just in a weird position
oak osprey
just a low res quick one with 'dwayne johnson'
i love it at high resolutions
oh mi
oak osprey
can you imagine 3 jack blacks in one body? comedy powerhouse
oak osprey
LOL DO MATTHEW BRODERICK
hasty nova
LOL DO MATTHEW BRODERICK
don't know that one but I will run it through soon
oak osprey
it definitely knows the concept of a Matthew Broderick but it does not know who he is
like, try Ferris Bueller
hasty nova
OH ferris bueller?!?
hasty nova
And inspector gadget?
hasty nova
Ah yeah I know who he is nevermind
just didn't know the name
yeah it does not seem to know him at all
hasty nova
lmao jim carrey
rain gazelle
evenin'
oak osprey
that's pretty good i could see him putting makeup on to get there
rain gazelle
whatcha guys makin'?
oak osprey
trying to see how well Junglerally's model knows celebs
oak osprey
can do Reagan, Bush, Obama, ya know. becuz political deception is cool
i got a pretty good one of jimmy fallon but what the hell is he doing LMAO
hasty nova
oak osprey
IS THAT A NEEDLE
oak osprey
hasty nova
looks too cartoony
oak osprey
LMAO WHY SWORD
hasty nova
bro wtf
hasty nova
now my gens folder is full of these
oak osprey
Hugh Grant or Tony Blair?
i never realised Tony Blair would make a good Senator Palpatine
hasty nova
oak osprey
@hasty nova on your model, try <person> as a warrior priestess in middle earth photograph portrait cinematic epic 8k sharp focus
oak osprey
rey reyyyy
because he's a feeling horseish
i think ryan reynolds just has a very long face.
hasty nova
yeah no, look at the one above
oak osprey
dang, it knows joni mitchell
hasty nova
Seems to be able to do everyone's favorite
hasty nova
idk, I put Emma Watson because she is a fan favorite
oak osprey
if an AI model can't do EW people would just throw it in the trash.
hasty nova
oak osprey
what the FUCK happened to her SKULL
LOL
i thought you said nips
i was like .... uhm where
((((((massive areolas))))))
aw shit wrong window sorrry
that had nothing to do with this gen btw
Hildawg
nice
d33p
damn boi
u tryina get nooked by like da clinton foundation
LOL
young hillary clinton is a fucking hoss
weird idea of her but i'll take it
people are going to think i'm weird but they'll be right
ask their parents 🥁
that's definitely what an A100 is for
"Can you make me look young again?"
"For $26,000"
"Wow, what surgery is that?"
"It's an NVIDIA GPU"
go for it
so it seems like you wanna dreambooth junglerally's digital-diffusion
😄
put another 2.1 model into the world
i dare you
it's pretty competitive with SDXL
but we'll see
SD2.1 be like "i can fight them. 😤 "
no
ask JR when he's around tomorrow
static tusk
