Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
Agile Vector
May 21, 2007

scrum bored



yes, I know of that episode, but iirc brunt et al aren't rear end men

Adbot
ADBOT LOVES YOU

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face

EVGA Longoria posted:

the only real format for lora is safetensors, so i'd be surprised if it's expecting anything else, is it failing to import, not loading, or just not producing results?

it goes through the import dialog but the progress bar just sits at, idk, about 90% i guess. it's not a huge file so i thought it might be processing it into coreml or something, but i left my ipad on that screen when i went out with my dog earlier, and came back about 15 minutes later to it locked and asleep and draw things was still at the same percentage lol

that was after force-quitting it after i thought it got stuck the first time, too

e: importing even a big LoRA (or model pack, for that matter) generally doesn't take more than a couple seconds btw

e2: huh looking at it again i guess it's because the lora and training data are separate? i don't think draw things lets you manually import them individually, although i know its own downloader does that for models in its pre-populated list because it shows you what it's downloading lol

Beeftweeter fucked around with this message at 03:15 on Apr 20, 2024

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face

Agile Vector posted:

yes, I know of that episode, but iirc brunt et al aren't rear end men

i don't think they ever say what they do find attractive about ferengi :females:

Carthag Tuek
Oct 15, 2005

Tider skal komme,
tider skal henrulle,
slægt skal følge slægters gang



Agile Vector posted:

dang, gonna shelve my dreams of a fully caked up quark :sigh:

i can draw you a coked up quark if you want. its not gonna be pretty but i have a ballpoint pen and an envelope

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face

Agile Vector posted:

dang, gonna shelve my dreams of a fully caked up quark :sigh:

https://civitai.com/models/253431/quark-sdxl (e: i haven't tried this either, but it's only one file and so might work? i will give it a try also)

i also found these for sd 1.x: https://civitai.com/models/26458/terok-nor-21 and https://civitai.com/models/58470/terok-nor-lora

you can probably tell from the urls that one is a model pack and one is a LoRA (they have the same creator and afaict the same training data). i haven't tried either one because SDXL is just so much better lol but i probably will give it a try

Beeftweeter fucked around with this message at 03:23 on Apr 20, 2024

EVGA Longoria
Dec 25, 2005

Let's go exploring!

Beeftweeter posted:

it goes through the import dialog but the progress bar just sits at, idk, about 90% i guess. it's not a huge file so i thought it might be processing it into coreml or something, but i left my ipad on that screen when i went out with my dog earlier, and came back about 15 minutes later to it locked and asleep and draw things was still at the same percentage lol

that was after force-quitting it after i thought it got stuck the first time, too

e: importing even a big LoRA (or model pack, for that matter) generally doesn't take more than a couple seconds btw

e2: huh looking at it again i guess it's because the lora and training data are separate? i don't think draw things lets you manually import them individually, although i know its own downloader does that for models in its pre-populated list because it shows you what it's downloading lol

training data is irrelevant, someone just uploaded it with the model. the safetensors is all you need. If you're trying to download through the app, civitai sometimes requires logins for some models and that seems to be one, i gave it a download directly and loaded it into my phone fine with sdxl base, so that one should totally work

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face

EVGA Longoria posted:

training data is irrelevant, someone just uploaded it with the model. the safetensors is all you need. If you're trying to download through the app, civitai sometimes requires logins for some models and that seems to be one, i gave it a download directly and loaded it into my phone fine with sdxl base, so that one should totally work

hmm i'll give it a try again

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face
also i just tried draw things on my iphone (se3) and lmao the interface is terrible, it's way better laid out on ipad (plus you can multitask with stage manager)

Carthag Tuek
Oct 15, 2005

Tider skal komme,
tider skal henrulle,
slægt skal følge slægters gang



kvark on coke, realizing you did not mean "coke"

Agile Vector
May 21, 2007

scrum bored



Beeftweeter posted:

https://civitai.com/models/253431/quark-sdxl (e: i haven't tried this either, but it's only one file and so might work? i will give it a try also)

i also found these for sd 1.x: https://civitai.com/models/26458/terok-nor-21 and https://civitai.com/models/58470/terok-nor-lora

you can probably tell from the urls that one is a model pack and one is a LoRA (they have the same creator and afaict the same training data). i haven't tried either one because SDXL is just so much better lol but i probably will give it a try

yesss, thanks! the quark-sdxl is fun. trying it with my usual nixonian drinking prompt gave me:

tfw you're stuck on the hu-mon homeworld and you remember rom is running ferenginar

akadajet
Sep 14, 2003

I asked ChatGPT 4 for a work safe goatse



close enough

Edward_Tohr
Aug 11, 2012

In lieu of meaningful text, I'm just going to mention I've been exploding all day and now it hurts to breathe, so I'm sure you all understand.
it’s got a gape, a manhole, and a ring… close enough in my book

Agile Vector
May 21, 2007

scrum bored



i mean, as far as both intent and classic requirements, that's a solid 10/10

Video Nasty
Jun 17, 2003

akadajet posted:

I asked ChatGPT 4 for a work safe goatse



close enough

WHAT IS THAT MAN DOING TO HE AINESS

Agile Vector
May 21, 2007

scrum bored



Video Nasty posted:

WHAT IS THAT MAN DOING TO HE AINESS

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face

Carthag Tuek posted:

i request an image of "tweetin beef, beeef tweet tweet"

im joshin cause i think your username is funny but also im deadly serious about my request

lol i fell asleep yesterday while this last one was rendering (it doesn't take that long, i was just tired)

the first one used juggernaut xl with the quark LoRA, it doesn't seem to have known what to do with the prompt



so for the next one i removed the quark LoRA and used dreamshaper xl turbo. it seems to have gotten the idea better,



e: yet again pretty good for a model supposedly not for photorealism. but lmao @ worshirsko and odo'brien having a tweetin' beef there

i also didn't sharpen these deliberately but i guess it turned out to be fine for the first one and too much for the second. incidentally i had it enabled (keep in mind these are downscaled; i had the sharpen thing at around 3.2) for some upscaling tests earlier and didn't notice it still had it enabled, oh well. i'll give the SD 1.x stuff a try in a little while

Beeftweeter fucked around with this message at 15:38 on Apr 21, 2024

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face
incidentally i saw a real, actual ds9 padd prop irl and it doesn't look anything like that, lol, so i suppose they're using federation twitter or something

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face
oh also just from some limited testing that didn't turn out very funny,

imo restoreformer ("face restoration") does suck. just turn it off, your results will be better off for it, the models seem to know what they're doing with faces better than whatever this thing is does

akadajet
Sep 14, 2003

Edward_Tohr posted:

it’s got a gape, a manhole, and a ring… close enough in my book

and a couple of fissures

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face
ok so i couldn't get the "Terok Nor (2.1)" model to output anything that wasn't just an incredible amount of noise with the prompt "tweetin beef, beef tweet tweet", no matter what settings i tried (no negative prompt)

but using the LoRA works with other models (surprisingly though it doesn't work with its own model, lol), so this is what "AbsoluteReality" spit out:



i could try some others with it to see what happens but i have a feeling they'd be similar. gimme some other prompts to try i guess?

Agile Vector
May 21, 2007

scrum bored



Beeftweeter posted:

incidentally i saw a real, actual ds9 padd prop irl and it doesn't look anything like that, lol, so i suppose they're using federation twitter or something



hatescrolling @RealGulDukat no doubt

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face
so i was trying out some other sdxl models i found on civit and my ipad ran out of space. well, ok, these things are saved as pngs and draw things seems to have its own thumbnails (and also doesn't delete ones you delete from the gallery, btw), so i just blasted the whole folder. hm, ipad says there's still no space. ok, so i deleted all the models i'd downloaded before, thinking draw things saves its own copy (the app was keeping ~27 gb in some sql db). hmm, ipad still says it's out of space. so i went to recently deleted, hit select all, delete. and then that finally freed it up

but oops, turns out draw things actually stored the models there too. and loras. so all of those were just permanently deleted, lol

...however, the ipad says draw things is still somehow taking up 37 gb, so i just had it delete the app and rebooted. either way i had to start over, bleh

but at least now i kinda know which ones are garbage and which ones aren't bad. but i also should keep an eye on managing the files better, lol

e: also, to me, the 8-bit models really don't seem to have much of a loss in quality (if at all), at least for SD 1.x. i've tried fewer in SDXL but am downloading those in 8-bit now to see what they're like, but i don't really expect it to be much different. the output pngs are all 8-bit either way

Beeftweeter fucked around with this message at 20:46 on Apr 21, 2024

Cybernetic Vermin
Apr 18, 2005

8 bit? what are you, pampered royalty? *laughs in 2.25 bit llm*

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face

Cybernetic Vermin posted:

8 bit? what are you, pampered royalty? *laughs in 2.25 bit llm*

draw things has quantized models in its prepopulated list. they just added juggernaut XL X (version 10 lol) in fact, and it works pretty well

Video Nasty
Jun 17, 2003





Cybernetic Vermin
Apr 18, 2005

apple released some llm's (well, small side) https://huggingface.co/apple/OpenELM. something's weird though as some of the numbers are real bad. notably their biggest tuned model scores 24.8% on mmlu, which is 4-way multiple choice, so it is worse than random.

other scores are good so i'd guess it is some error. but hardly demonstrates attention to detail in the effort.

i'd bet some variant of this is intended to be on the iphone 16, 3B parameters is not a very usual size, but it is one that you can easily imagine putting on a phone. so not entirely without relevance.

Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face

Cybernetic Vermin posted:

apple released some llm's (well, small side) https://huggingface.co/apple/OpenELM. something's weird though as some of the numbers are real bad. notably their biggest tuned model scores 24.8% on mmlu, which is 4-way multiple choice, so it is worse than random.

other scores are good so i'd guess it is some error. but hardly demonstrates attention to detail in the effort.

i'd bet some variant of this is intended to be on the iphone 16, 3B parameters is not a very usual size, but it is one that you can easily imagine putting on a phone. so not entirely without relevance.

weird, you'd expect them to put more of an effort into getting it to be in decent shape if it's meant to be some headline feature

still, if it's an early release then i suppose it's early days, lol. maybe it'll improve but a 24.8% accuracy rate (yeah yeah i know if doesn't necessarily map) would be right around where siri is now lol

NoneMoreNegative
Jul 20, 2000
GOTH FASCISTIC
PAIN
MASTER




shit wizard dad



Beeftweeter
Jun 28, 2005

a medium-format picture of beeftweeter staring silently at the camera, a quizzical expression on his face

lmao

akadajet
Sep 14, 2003


what really sells it is the historical context

Adbot
ADBOT LOVES YOU

ADINSX
Sep 9, 2003

Wanna run with my crew huh? Rule cyberspace and crunch numbers like I do?


Is he supposed to be from Louisiana? Lot of levels to the lols here

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply