🏞｜general-with-images

[oak osprey]

baseline 1.5 is some of the greatest stuff in the genre

the deformed outputs it makes are just top in class

favour for what though, family portraits?

[dense tapir]

here is ++ SDE

[wispy nest]

everything*, well maybe not photo stuff, but I don't think either 1.5, nor 2.1 has good results there :(

[oak osprey]

wireless handlebars

[oak osprey]

bing's photo generator seems to be the absolute best so far for photorealism

nails the 1980s kodachrome look too

[wispy nest]

I heard that googles ai is the best for everything, but I've never seen their results because "top secret" >:(

[sterile temple]

every model looks good when they only show cherry picked images

[dense tapir]

I have been testing a few other 1.5 models but I can't show any of them as they all came out nude

[oak osprey]

something off but i don't know what

[sterile temple]

they must have mixed in some porn models at some point

[dense tapir]

I guess so :/

[wispy nest]

nude, you say :P

[dense tapir]

yes, in your face and smothered

OMG, even realisticvision they are nude

[oak osprey]

what the bing

people are definitely cherry-picking their results lmfao

[wispy nest]

if this was the standard results, then I'd never need anything else when it comes to art. This is exactly the type of style I love when it comes to color, detail sharpness, etc. Or rather, the one I think is more than enough :D

[oak osprey]

ive always wanted a big family, with lots of miniature versions of me running around

[dense tapir]

Too plain for me

I'll use disney

I am going back to 2.1

[dark flare]

What is your dog's name? He's magnificent

[dense tapir]

LOL, eve MJ 1.4 my image is nude.

[wispy nest]

I just want the ai images to not look like ai images, I want them to look like they were made by a person. And from my own opinion, only the painting style would fit in a professional consumer product. I.E., no weird fingers, very little weirdness in filters, no noise, no copies, etc. Like, if you saw this in X, would you be happy with it?

Might be a tall order, yes :P

[oak osprey]

Michael

he has a masters in physics

there's an account in my local shoutout group that has an image of a 80-something year old woman, and she's always posting sus pictures and i'm thinking it's an AI account

she said this is her dog

[dark flare]

heheheee ........

He may be self-aware

[dense tapir]

Even with nude in the neg it would not pass clyde

[oak osprey]

he has human ears in another img

[dark flare]

I made a bunch of stuff with Firefly (Adobe) yesterday, they looked like actual photos. Of REAL people, not supermodels....

[dark flare]

He-Man's girlfriend?

[dense tapir]

LOL

[dense tapir]

well, back on 2.1 cause I could not make that non nude in 1.5

[oak osprey]

this Hermione is too magical

[wispy nest]

that's great! that's what I want, for someone to not see the image with any errors, or if they have errors, then it should be because of the style, like "big heads" should still look like it fits the style, etc.

But there are so, so few images that can fit those criterias. for now, but maybe not forever :D

not even the image I posted would pass as the hair on one side is wrong and the patch below it is weird.

[dark flare]

So, I know you guys are DYING to know ... the results of my training were MISERABLE, so bad I even deleted the images I generated because they were just so WTF.

And I never delete anything.

It led me to thinking, maybe the lora file(s) weren't loading properly .. .. . .. . so now I guess I'll do a different tutorial because mine was confusing??

[oak osprey]

@wispy nest but even when it's bad, it's good

[dark flare]

Keep in mind, a year ago, AI images were SHIT

Now they actually look like something

[oak osprey]

yeah go try Imagen

[wispy nest]

I don't want to be "that critic" so I'll say that as long as you're happy then it's enough. But if I were to try and be "objective" then I'd not want to be objective anymore :(

[wispy nest]

I know, I was there, in spirit mostly :P

[oak osprey]

i want AI to make impossible things because i can go use a camera to take photos

[wispy nest]

I want the ai to be able to make everything possible :D

[oak osprey]

i don't, marketing as an industry is already so toxic, i'm not looking forward to what these tools will enable them to do

its one of those things whose knock-on effects are hard to escape

even if you somehow manage to avoid ads and never see one - as if you can even know that - no one else is avoiding them, and your interactions are all impacted by that marketing

[dark flare]

Social Dilemma, anyone?

[oak osprey]

clones C1 through C5, HOMEWORK TIME

[dense tapir]

I've fallen down this rabbit hole and can't find my way back

[oak osprey]

ah, there we go

my fine-tuning on this model really did fix its ability to make faces reliably but like i never thought i'd say this, but i need more pictures of children to put into the training data i guess lmao

[oak osprey]

or maybe not, idk

[wispy nest]

Hehe, whenever I see a image I see them as if they were made to look as good as possible. I'm kinda boring when it comes to art actually :(

[oak osprey]

you can change

[wispy nest]

change back you mean? ;P

[oak osprey]

imagine receiving this pillow as a gift and it weren't your family

that's the AI hellscape future we live in

[wispy nest]

making something silly is easy, but making something perfect is harder, for me at least. That's why I slowly moved toward making "as perfect images as possible" :P

[oak osprey]

children of the corn 🌽

[oak osprey]

ok mr easy pants

show me the silly

Silly Wonka and his Silly Factory

[dark flare]

Hey I just realized there's a #🔧｜finetune channel so I'm gonna do my bellyaching there... 😉

[oak osprey]

i see nothing wrong with this image

the daughter's name is Little Arty. Arty Facted

[wispy nest]

I opened my older stuff folder and my computer almost died showing the 700,000 images :P

[oak osprey]

can you imagine being this close to your family

hive mind

really makes ya think

[wispy nest]

this is one of my silly images :P

[oak osprey]

that's not silly, that's epic

[wispy nest]

how about this then: cat origami?

[oak osprey]

where are they walking from

^ this would be sent to like the FBI as a faked proof of life to show someone is still alive when it's just photoshopped old family trip photos into new places

[wispy nest]

and this one is like in the middle of november :P

[wispy nest]

or maybe this? too little catnip

[oak osprey]

i have some silly photos of jimmy fallon smoking through spongebob after making him into a bong

[wispy nest]

last one before I need to head out!
So many old memories of weirdness 2022 was weird :P

[oak osprey]

night

[wispy nest]

👋

[wispy ether]

Raw edit

[wispy ether]

Waltuh

[wispy ether]

Tf

[oak osprey]

0 cfg

[wispy ether]

Chaos from skylanders

https://tenor.com/view/popcorn-eating-eating-popcorn-kaos-skylanders-gif-23497337

[dense tapir]

Look a lot like him

[viral crescent]

hihi

ahhhhh! it's baby apocalypse

[wispy ether]

Is the processing even finished

[sterile temple]

110%

[wispy ether]

What

[sterile temple]

whomst

[wispy ether]

?

[smoky oak]

inept

[wispy ether]

What

[viral crescent]

so Im looking for a decent femboy checkpoint any thoughts

[wispy ether]

https://tenor.com/view/bruh-gif-22596911

Civitai.com

[viral crescent]

no

all Lora's not checkpoints

[smoky oak]

Lots of anime models should be able to do femboys just fine, and the furry model I use can do them exceptionally well lol

[sterile temple]

couldnt you just prompt for it from most models?

[smoky oak]

yeah, should be able to pretty easily

[viral crescent]

I still seem to be getting Boobas

[sterile temple]

increase the weight for male

[smoky oak]

probably just bad prompting

[viral crescent]

Hmmm ok I'll keep trying

[dark flare]

Dang! I thought my 19,000 images was a lot...

[sterile temple]

show offs, with their fancy GPU doo dads 😄

[dark flare]

Well, Bitcoin is pointless now, gotta do something with 'em 😉

[sterile temple]

where did these guys learn how to ride a horse?

[dark flare]

hehehe

[dark flare]

Verra nass, how much?

OK, so ... I think I don't care about lora because I have 12 GB .... yes???? I can has regular Dreambooth?

[sterile temple]

you can use a lora with any compatible checkpoint

you can also create a lora from a checkpoint

[dark flare]

Well, I have a civitai model, and I just wanna add my face to it

There are sooooo many ways and 🙈

Is the result of kohya_ss different from the Dreambooth tab in a1111?

Scratch that, it seems the extension won't install/activate properly for me 😉

[sterile temple]

I've only tried dreambooth a few months ago on a colab notebook, and recently a colab notebook for a lora trainer that was really simple to use. I don't think I could do any training on my laptop

[dark flare]

Have you trained much or not really?

[sterile temple]

not really. I would like to play around with it, but it seems like a lot to learn 😄

[dark flare]

SEEMS THAT WAY FO' SHO'

[sterile temple]

I've been playing around with mixing models, recently trying the blockweight merge type of mixing

looks like you can connect a colab notebook to a local runtime on your computer

[dark flare]

Yeah, I could always run the old colab from last November but I wanna do it like the cool kids are doing now 😛 locally

[sterile temple]

I think I might try training a lora on my 6GB card. Looks like it might be possible, even if it's slow

could always run overnight while I sleep

[dark flare]

So yeah .... I didn't realize (until a couple hours ago) that lora was for low-VRAM pipples, and I shouldn't have wasted my time on it in the first place 😆

[sinful mirage]

the hair lmao

[marble wasp]

thats beautiful

seriously beautiful

i dont even know how something ike this is possible with SD

[sinful mirage]

ah i should have said it was midjourney, havent tried it in nearly a year and thought id try it lol lot of fun, i put it through sd controlnet tile tho for extra detail and resolution, they work really well together

[marble wasp]

ahh

[wispy nest]

19,000 is also a lot :)

[sterile temple]

what you talking bout son. course SD can do something just as beautiful 😄

[marble wasp]

his was better

lol

[sterile temple]

no, I'm delusional and I think mine is so much better!

prompting intensifies

Yeaaaaaaaaaaarghhh!!!

[wispy ether]

With dreambooth you cant train lora

[north monolith]

Glad people are talking about this, as I am just trying to install the Dreambooth extension, as there are plenty of videos showing that you can use it to train Lora's

But I'm having install issues with it, so considering kohya instead

Yep just checked you can use the Dreambooth extension to create Lora models https://www.youtube.com/watch?v=gw2XQ8HKTAI&ab_channel=NerdyRodent

[mystic marlin]

Random forest nymph? Still messing with Loopback Scaler.

[dark flare]

Yeah but apparently I don't care about loras in the first place. (I think. Once I get DB running we'll see if I get memory errors)

[marble wasp]

SD has problems with faces at larger than a close up shot for me

[sterile temple]

how close?

[marble wasp]

if you imagine a medium wide shot of a person, to where their body is in frame, it seems to have facial distortions that i have been trying to paint and inpaint out. then i added prompts about sharp facial features in different ways and it still renders it so weird. I got frustrated and closed the program

lol rage quit

[sterile temple]

Have you tried the after detailer extension? That's pretty good at fixing faces

[marble wasp]

im not aware of it

i'll look into it

[wispy ether]

https://tenor.com/view/galaxy-brain-gif-25912556

[wispy ether]

[wispy ether]

https://imgsli.com/MTgzNTAw

[wispy ether]

my vram when upscaling 512 to 2048

[sterile temple]

you should submit a daily #🔆｜dailies

[shut sinew]

First time playing around with ADetailer, kinda epic

[sterile temple]

Yeah, it's really useful

[shut sinew]

Getting somewhere

[dense tapir]

@smoky oak I am reading a few different things about getting the AMD 7K cards working (since no official ROCm support is there yet) and saw this - "Currently, SDP attention unfortunately doesn’t help with RX 7900 XTX’s iteration speed and is memory hungry. It’s better to switch back to old optimization and stay tuned." Ahem, for me, and you, that was the exact same case. Apparently, as the devs said, it really is just for a 3090/4090.

@smoky oak

[latent atlas]

https://www.instagram.com/p/Cs_V7LTIDt6/?igshid=MmJiY2I4NDBkZg==

veiled.vision.studio

Capsule Corps 🌑

Instagram

[oak osprey]

i would definitely take her here for glamour shots

when CFG is around 2.0-5.0 you get more poopy looking flowers like that, and the subjects might be floating off the ground

[shut sinew]

I did some macro shots of flowers before, no open fields tho

[oak osprey]

https://github.com/damian0815/compel/issues/35#issuecomment-1573214871

so the author of Compel also agrees that the prompt embeds are an issue for whatever reason in 2.1 but has a workaround in development

[oak osprey]

bannapple

lol they should fix that

[wispy ether]

[wispy nest]

was this the thing you talked about earlier?

[oak osprey]

nope, automatic doesn't have the thing i mentioned

i doubt it will because it uses Diffusers, they'll have to pull the code out of there and essentially pull it into bits and pieces to jam into A1111

[wispy nest]

I thought it was this thing :/

[oak osprey]

oh that apparently is the same thing yeah

the Mixture of Diffusers one

i see Multidiffusion is now essentially Mixture diffusion. he updated it last week to import their method and uses it instead of his own

the model_hijack inside the file there is the work i was talking about that'd need to be done

[wispy nest]

can't get it to work and need to read why, but I'm too stuck with the "neutral prompt" thingie I also got at the same time :D

[oak osprey]

well like we found yesterday, mixdiff is going to take hours to make something awesome

[oak osprey]

i'm kind of jealous at the way A1111 diffusion implementations can just be thrown in there without having to update the entire pipeline workflow, but it requires more insight into what the hell the tools are actually doing: https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/blob/main/tile_methods/mixtureofdiffusers.py

eg. this model hijack essentially reimplements the noise scheduling of the model inside automatic's hooks, which are done that way so that they can be passed around between different extensions

what diffusers would need is just one line of code:

pipeline = StableDiffusionCanvasPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", scheduler=scheduler, use_auth_token=True).to("cuda:0")

the thing about the A1111 version is that because everything is exploded and reimplemented for each extension/pipeline, you can do stuff that's not supported by diffusers, like, set the method used for CFG rescaling

Diffusers likely just needs 4 or 5 lines of code added to allow that, though

ComfyUI did roughly the same kind of thing A1111 did, they pull bits out of underlying libraries and reimplement them into their own class wrappers so that they have more control over the pipeline's inputs and outputs, which makes a lot of sense for that tool, but still, i wish that were more baked-in to Diffusers

[wispy nest]

well, I got it working, now to understand just how to work it :P

[oak osprey]

that's most of the time you'll spend lmao

[wispy nest]

was my first random prompt result :P

[wispy ether]

ai takes over the world

[dense tapir]

Seems the 4090 is being impacted by Nvidia stopping production on the chips and moving the facilities over to Hopper production has begun. Went on pcpartspicker and a lot of cards from manufacturers are not available.

Here comes the mining days again.

[oak osprey]

no one is buying them, why would they keep putting them out

i got mine 🙂

[wispy nest]

[oak osprey]

my model isn't trained for anime but it has interesting results nonetheless

[oak osprey]

@crimson dawn

come to the dark side

touch the sky, install gentooooo

[smoky oak]

Very interesting

[cyan snow]

this is unsettling, i just typed negetive prompt: earth, humanity

[smoky oak]

Oh my haha

I have played around with messing with negative and positive prompt switches for a bit, and it can do some fun stuff haha

[dense tapir]

It explains a lot

[cyan snow]

my educated guess is that it finds the concept least similar to the negative prompt

usually there is also a positive prompt, but when there is only negative it can make nightmare fuel

[smoky oak]

I believe the way it works in actual implementation is just that it tries to lower the triggering of tokens associated with those words

[cyan snow]

it made this again

i guess what ever that is it's the least simmilar thing to humanity.. lol

[smoky oak]

It's not triggering exact opposite, it's just triggereding everything but humanity

[cyan snow]

then why is it consistantly making similar stuff

[smoky oak]

Because it's consistently triggering everything but

[cyan snow]

huh

reminds me of that LOAB thing from a while back

[wild sorrel]

Your description reminded me "Sundered" game, fight against humanity... (yea, that's an actual boss and it doesn't look great too)

here's humanity for a reference

beautiful thing on right side

[cyan snow]

nah this is Ht'uimg'aeng'ytagyeh

[wild sorrel]

same thing, just a different language

[cyan snow]

this gives me backrooms vibes, idk

[smoky oak]

My grandmother just asked me to print her a poster of one of my gen tests I did 😅

I don't even know how to print a poster lol

[oak osprey]

fax it lmao

@smoky oak if you wonder why i get pissy at the shitty samplers in Automatic it's because i get poor reviews on Civit because it doesn't work with new, fixed, 2.1 models

[smoky oak]

Just add a disclaimer that it's not compatible with auto/Vlad/Ana

[oak osprey]

the thing is, i did get it working alright in there but i never fuckin remember which sampler it needs

that was still just mediocre results in comparison to a normal clean pipeline

i'll add that though

[smoky oak]

I have seen the results you have gotten compared to what I can get, and they are considerably better for sure, but not nearly enough to make me use something else for "better" samplers

[oak osprey]

yeah i wouldn't expect anyone to

i'm not sure how to phrase an issue and i don't have the energy to sit there defending it to the maintainers, who are unpleasant people already

i know they'll ask for more info and debug logs and, rightfully so, but, i'm tired all the time lmao

maintainers

[oak osprey]

u havin' a strok thar bruv?

i think you meant unpleasant, but that's not how i'd approach them with the issue

https://github.com/damian0815/compel/issues/35 this is how i approach cooperative resolutions for problems

[smoky oak]

My bad, I read your message wrong

[oak osprey]

it happens

[smoky oak]

I'm already super tired as it is 😅

[oak osprey]

i remember you had a headache yesterday, is that still a thing

[smoky oak]

Didn't have one today, luckily

[oak osprey]

that rules out anyeurism and brain hemmorhage

[smoky oak]

I do seem to get them when I sleep for very long amounts of time. I slept for 11 today, yesterday was closer to 15

My whole body in general is really going through it with my fibro

[oak osprey]

sleep is dumb like that

[smoky oak]

It is a flare based disorder, after all

[oak osprey]

not enough sleep? headache. too much sleep? headache.

[smoky oak]

I was doing very good for the last few months, but it just came in with full force a few days ago, and is not letting me go

[oak osprey]

u need lex luthor's latest product

he's partnering with snoop

[smoky oak]

I don't smoke, never will haha

More for everybody else tho lol

[oak osprey]

fucking awesome broken text encoder on a new test run

that's supposed to be a child

it's a dog horse shoe wearing monstrosity

he shops at cabtankro too

celebs are so relatable

[smoky oak]

@dense tapir have you seen SnapFusion?

Very interesting stuff, wanna get my hands on it

[dense tapir]

SnapFusion?

https://snap-research.github.io/SnapFusion/ Interesting.

[smoky oak]

Yeah, so its a diffusion method that runs on phones in a couple seconds, and apprently it has a massively more efficient and higher quality yeild than SD

very curious to see what comes of it

Imagine if we could get this new tech finetuned like SD and running on high end PC hardware

you could to images per second on a 3060ti more than likely lol

[dense tapir]

Well, I watched it and I await it on my PC

[oak osprey]

demon cat

looks like it'll produce the kind of deformed shit that i love

oh, they distilled the unet

joe penna was saying they're thinking of doing that with SDXL but that it is devastating to the results

[oak osprey]

working on a realism model 🥺

be gentle

@smoky oak how do you like that so far, even though she has no dangerous features visible

eg hands or face

i'm not sure how to prompt for that without getting something scary

[wispy nest]

I use words like ashamed and groan
wait, ahhh! you mean as in what to write as prompts! ;P

[oak osprey]

oh you

[wispy nest]

just use what you're feeling in the prompt! :D

[oak osprey]

this guy is a good omen if you see him during training

[wispy nest]

*now

[oak osprey]

in case you didn't know

training an image model is like fuckin around with a Ouija board

you can communicate with the dead

[wispy nest]

when I trained/made embeddings my test prompt was If it looks like a duck, swims like a duck, and quacks like a duck, then it probably is a duck :P

[oak osprey]

the robot pouring coffee sometimes looks like it wishes for a better life

like, what life lies beyond the doors of this coffee shop, for me?

none, because you're a coffee pot

[wispy nest]

taxes, and lots of politics

[oak osprey]

you think it'd run for orifice?

[wispy nest]

yes, both kinds

[oak osprey]

sing the orifice song

hey that guitar is hella in-tact

so close to looking correct

[wispy nest]

when I try and learn how to get X I make like 2000 images with the word or words at different places and focus :P

[oak osprey]

well i'm training quite a few thousand steps and each checkpoint makes like 13-15 images across a surface area of concepts eg. man, woman, child, robot, etc

would take too long to do all of those that you suggest 😛

[wispy nest]

and I wouldn't suggest you do either because most of the images are really, REALLY boring, but I still save them all for some weird reason. I'm kinda a horder when it comes to computers, or what's in them :P

[oak osprey]

i have used the same prompts on the same seed so many times. i have seen this douche bag in so many crappy knight costumes

[wispy nest]

like when trying to understand how the ai places hands :P

[oak osprey]

this is a first tho i've not seen this kid looking like a kodachrome image

see what i mean? lol

[wispy nest]

you start noticing similarities in people the ai makes after a while. There's not that many version of "a standard face" it makes without extra stuff

I add stuff like evil, sinister, etc before the "person prompt" to make it slightly different

[oak osprey]

its paintings are beginning to become a lot more painted

[wispy nest]

but is he a real wizrd?

[oak osprey]

you tell me

[vocal cloud]

Pretty sure I asked for soy. That looks like whole milk.

[oak osprey]

you can tell the text encoder fully froze now

the changes between checkpoints are pretty minimal

lol SD 1.5

2.1

@smoky oak my model is being used now to fix DDIM

by huggingface

[smoky oak]

Cool

[crimson dawn]

people blame mining often for the high demand of GPUs, but those guys saturated their market very early. GPU purchases from them only ever was to replace their old hardware and sell it off . Gamers who combed ebay and weren't afraid of mining cards were fine.

Some other demand came from data centers. even back then. Cloud gaming isn't that huge of a market, but those companies were buying up all the hardware. Stadia was ready for primetime and they expected it to be huge. It wasn't but they had the hardware ready to go. Yes that means clusters of GPUS in server racks. ML research has been going on for years and has gotten to this point because research labs have been flexing gpu clusters. they've probably even been buying up all the old mining hardware as it's replaced.

the real silicon demand eater is the military industrial complex. A whole side of the market that consumers don't have any idea about. They're taking in so much silicon and that's why there's a manufacturing shortage and companies like nvidia got to strategize to fill demand

if you're running a business and intend to suceed, you don't leave money on the table whenever you can

[sterile temple]

also 1.5 lol!

[smoky oak]

That looks sick

[dense tapir]

No, I saw it myself in 2020/2021 people going to BestBuy and leaving with basket/cart loads and some entire damn pallets of GPUs.

[crimson dawn]

runway ML made a really good base with 1.5. one of the best so far probably

Bestbuy is like, small beans compared to bulk orderers. I'm surprised you saw that. most bb managers would just sell the entire palette before it came in

[dense tapir]

Reason I never upgraded I couldn't get to them fast enough

[wispy nest]

I was also trying to make cars yesterday :D

[dense tapir]

they did here. first come first serve.

[oak osprey]

SD can do cars so well. it's actually one of the major biases when you do unconditional guidance. but looks like an atomic bomb dropped moments earlier

[crimson dawn]

Yeah. Consumer stores were only getting 5 at a time in some places. It was just high demand in all marktes

[dense tapir]

That shit was happening all over the USA with many pics, and vids, of people doing it

[oak osprey]

brick

[dense tapir]

Theft was up too with shipments just POOF

[crimson dawn]

miners were just one demand source that got the lions share of the blame, but hardly even a lions share of the use

[oak osprey]

it works like the human eye in many ways because when you're not looking at it, it's fine, but when you zoom in, it's like the eye didn't focus, and you zoom in on the garbage the optic nerve shows in the peripheral

[crimson dawn]

hardware scalpers always suck. i'm not against reselling gear you own, especially if people want to pay higher than msrp. but if you're buying it JUST to flip it, that's scuzzy

buying 10 to sell them? maaan

[oak osprey]

just curious if microcentre falls under that rule because it kinda sounds like it does lmao

[dense tapir]

well, it was a complete combo that happened and Nvidia, as I said at the time, along with AMD, was causing a lot of the issues with scarcity. We now know that was fact that we later had pretty much confirmed.

[wispy nest]

and that doesn't happen in 2.1? :P

[oak osprey]

oh it does but less

[dense tapir]

Now we have Nvidia, and AMD, trying to continue that gravy train.

[oak osprey]

drunk camera is drunk

[wispy nest]

less is not a number ;P

[crimson dawn]

apple is manufacturing their own silicon these days too and driving the profit margins on it higher too

amd an nvidia just rise with the tides

well nvidia rising faster than some. up there with apple suddenly wtf

[dense tapir]

No card in 2023 should cost what they are going for. I mean AMD or Nvidia especially with how much they have been gutted.

They love Apple and wish to emulate Apple's business model.

[crimson dawn]

well given market conditions, the unprecedented demand for the hardware, they do cost that much. they could charge more and they'd still sell out

[oak osprey]

just setting the record straight that 1.5 is still hot garbage lmao

[crimson dawn]

microsoft is getting into making hardawre for themselves because they see it as the only way to survive

[oak osprey]

all AI images are just garbage noise being turned into projections

[crimson dawn]

amazone is starting to make their own silicon too

[dense tapir]

Of course they could sell the mother fuckers for 10k each and still sell out but that DOES NOT make it right just greed.

[oak osprey]

to me, deepfloyd is the least garbage

[dense tapir]

Oh, shit Amazon silicon. ffs

[crimson dawn]

demamd for more hardware isn't going down and supply will only increase with more mega factories, which are hard to increase

[oak osprey]

i don't use it much til stage3 is out

[wispy nest]

yeah, I've heard the same from all the 30 people who use 2.1 mouahaha >:3

[wispy nest]

that's not how turbo works! :P

or is it 🤔

[oak osprey]

it's marketing material i made for a friend

[crimson dawn]

i dont know enough to disagre, but it looks right

turbo go brrrr

[oak osprey]

trust me thats how it works

theres pipes

[wispy nest]

you had me at me :P

[oak osprey]

honestly deepfloyd is incredible, it's just inaccessible

[dense tapir]

I am probably going to go AMD as I have seen the numbers. I want an Nvidia but I simply am not going to fellatio Huang to get one.

[crimson dawn]

turbos have pipes. internet has pipes. turbos have pipes. get it?

[wispy nest]

this is what turbo gave me :P

[sterile temple]

I work in the automotive industry and can say that's exactly how turbo's work

[crimson dawn]

i'd call it convoluted too. i've got it running locally once, but couldn't recreate that

[oak osprey]

want the prompt?

[wispy nest]

good to have that cleared out by someone who knows! :P

nah, the way I write prompts I wouldn't be able to use it in the same way I'm used to. If it doesn't include some nice prompt words I haven't tried yet that is :D

[dense tapir]

I hate how Nvidia made an exclusive deal with Best Buy for their cards and BestBuy never has any PLUS is going bankrupt and closing stores to try and stop it. Any 4090 FEs you see in the wild are people who are reselling them for 500 over MSRP

[oak osprey]

i think it took me like 8 tries to get that image the way i wanted

[crimson dawn]

i typically build 5 year machines. last machine was all amd. i ran linux on it because i was sick of installing drivers so the cpu cores wouldn't park anytime i updated, or having my gpu suddenly decide to run with half lanes for no reason. the drivers on linux just work as they should on amd hardware. the windows side is like a constant struggle.

[oak osprey]

which i considered pretty decent for the result i got

[crimson dawn]

this machine i just finished is intel/nvidia. dawn of a 5 year journey

[oak osprey]

linux has better GPU drivers for nvidia and amd nowadays

nvidia's linux stack is completely open source now

it's kind of wild

[wispy nest]

I get what I want from the ai every time*
*does not include the 2000 test images I tried first

[oak osprey]

the source code for their driver was leaked and like 3 months later, it was open sourced

[crimson dawn]

So i've heard! When i tried to install my preferred garuda (iknowiknow bloat but i like features 😛 ) i couldn't get the nvidia stuff to recognize so i went windows instead. Might try to dual boot here eventually, but for now at least i just want it to run

still got my steamdeck running steam os so i can be like "i run arch" legit still

[oak osprey]

idk waht the fuck garuda is lol

[crimson dawn]

bloated arch with all the fixens

[oak osprey]

i use gentoo on my laptop which has amd and nvidia graphics both

i have to maintain something like 38 kernel patches to make it work because of lenovo's shitting the bed

they keep promising a firmware fix is coming

[crimson dawn]

gentoo on a lenova laptop these days? at least you've got that

okay that's actually a good flex

[wispy nest]

things I've yet seen the ai be able to make are any sort of long talons/claws, or blades like a cyber mantis or something for character's arms :P
Very none suspiciously specific thing, I know!

[sterile temple]

sounds very metal gear solid

[crimson dawn]

tons of these images are available. train.

[smoky oak]

The anthro model I use can do claws and tallons, but of course on Anthros lol

[smoky oak]

You know, I could probably do it

Maybe I will

[wispy nest]

I'd not be fat anylonger if it was that easy…mostly not easy because me too lazy :P

[crimson dawn]

with cyberpunk 2077 photomode you culd get them on 100s of characters and settings

[oak osprey]

perfection

[crimson dawn]

literally is that easy. you can eat while you do it.

[wispy nest]

I want sane results without having to do any work! Is that too much to ask? ;P

[oak osprey]

yes

[wispy nest]

😭

[smoky oak]

Honestly, love the honesty lol

[wispy nest]

why can't I get everything I want without effort? I must be cursed or something >:(

[oak osprey]

cursed like these guitar strings

[crimson dawn]

i love when things are automagical

[wispy nest]

ahh, another wizard from the school of auto-magic

[oak osprey]

i'm training this model with offset noise and terminal SNR together

[wispy nest]

another fellow wizard*

[smoky oak]

It's givinnngggg.... Jack Horner from the last wish lol

[oak osprey]

ALL THE NOISE

offset noise = more noise, terminal snr = end with zero noise

BATTLE

[crimson dawn]

that reminds me. im want to train this gobbly eye aesthetic

[oak osprey]

just use img2img through 1.5

it'll make your training data you desire

[crimson dawn]

heh ! yeah or undertrain someone

[oak osprey]

i think overtraining more likely to do that head-kicked-by-mule look

[crimson dawn]

no but i was thinking like, more comical. like take 100 photos of people and add random weird eyes to them.

[oak osprey]

use a googly eye filter from snapchat hahaha

[smoky oak]

That could actually probably work lol

[oak osprey]

or a face swap and do everyone as steve buscemi at 50% strength

[crimson dawn]

how can i make a name out of "head kicked by mule syndrome" that. i feel like there's opportunity to use ass here

[oak osprey]

ass... head... ass-head syndrome

[crimson dawn]

there's potential

feel like oppenheimer before the bomb got built

[oak osprey]

makes me think of the semi truck in Dumb & Dumber

[crimson dawn]

welp, cya later

[wispy nest]

this need to become the industry standard

[oak osprey]

https://www.youtube.com/watch?v=8ZZTY5jJSWU

ngl i didn't think it'd be hard to get that

it wasn't easy

[crimson dawn]

That's cool

once i unfocus the blindspot tho...

still cool but...

[wispy nest]

why do that thing(?) remind me of abe's odyssey?!

[oak osprey]

yea it wouldnt be my chosen image but none others have come out with any kind of long weaponry

[crimson dawn]

(i didn't actually leave i was just quotting a fave from d&d)

[wispy nest]

but yeah, having a image with claws etc can happen, I just wish there was a way to have great results 99% of the time :(

(great as in what I find is great :P)

[oak osprey]

it doesn't match the desire but it's really cool still 😁

[crimson dawn]

just freeze yourself till sd is awesome like that

[wispy nest]

no need to freeze myself, sleeping is one of my specialties, I can fall a sleep before finishing a whole sente—

[oak osprey]

😅 📢 🧓🏽

[sterile temple]

looks like the training images for swords and the like are absolute garbage

mostly low quality mmo screenshots or alibaba store stuff

and a lot of cropped swords too

[crimson dawn]

https://civitai.com/models/68668/pecha-swords-generator this is an awesome looking model speaking of swords. saw it the other day. probably not what you had in mind though

[sterile temple]

hypercars vs sword

[oak osprey]

that explains why the wizard looks like he's from wish

it's alibabaaaa

[crimson dawn]

swords have been around a lot lot lot longer than hyper cars. more images of shit swords because shit swords exist and are widely photographed.

archeological reasons or something dumb like that. get it outta the ai!

[smoky oak]

I was just dicking around with sound design, and managed to make this nasty ass bass

Loud and intense warning lol

[oak osprey]

@smoky oak ever used Bark?

that prompt had something like [radio jingle] at the beginning

[smoky oak]

uhhh

interesting

[oak osprey]

it's a GPT language model trained on audio waveforms

so like CLIP is image-caption pairs, this was audio-caption pairs

[somber moth]

Noo

[oak osprey]

so you can make audio of people arguing at a basketball game

[somber moth]

Why not

[oak osprey]

eexactly

if you prompt it, Uh, hi, I'd like to order an oat milk latte it'll actually sound like she's in a coffee shop

it has basically, context, on top of the audio voice

[sterile temple]

I could use an oat milk latte 😴

[oak osprey]

my current fork of Bark is outdated, there's been a lot of developments since that sample was made, and it's higher quality now

[sterile temple]

trying inpaint sketch

[oak osprey]

my realism model training progress

[sterile temple]

img2img to clean up the smudgies

[wispy nest]

if you use auto1111's webui, then get the neutral prompter instead, it's awesome :D

[smoky oak]

I have messed with it, but I am not sure I understand too well how to use it as of now

[oak osprey]

i cant see its face but i know its better than the other leopard geckos i had

cute lil bugger

[wispy nest]

with the regional prompter where you can have the ADDCOMM and BREAK together with the neutral prompter like AND_PERP PROMPT_WORD_HERE:0.6 is a game changer in my opinion! Images are so much better with them!

[smoky oak]

do you have any info/documentation on how to use those sorts of things properly? I would love to learn

[wispy nest]

sure, I'll get the links and PM you them as I don't know if it's ok to share links here :)

[smoky oak]

It should be, as long as they aren't to problematic sources haha

[sterile temple]

please share. I found the examples in the readme confusing

[smoky oak]

same here

[wispy nest]

https://github.com/ljleb/sd-webui-neutral-prompt
https://github.com/hako-mikan/sd-webui-regional-prompter
these two, I tried them and now it feels like I never want to use anything else! :P

and I can answer some questions on how to use them if you want, though I have only used them since yesterday'ish so take that will all the salt you want :P

[smoky oak]

do you have any examples of things you have been able to do? I would love to get an idea of what kind of stuff can be achieved with them together

[sterile temple]

png info says a 1000 words 😄

[smoky oak]

that as well haha

[sterile temple]

I found it really fiddly, you either get overbaked images starting to look like something from 2.1, or the prompt after the AND_SALT/AND_PERP just replaces the first prompt

[smoky oak]

so they are saying to use them together, and you get much better results, so I am looking into that at the moment

[oak osprey]

what will you make

tell meee

[smoky oak]

I am not sure at the moment

[oak osprey]

did you get it working

[smoky oak]

ok, I see what they are doing here. Very interesting

[oak osprey]

that's like mixture diffusion

[smoky oak]

it is, this is only one half

[oak osprey]

it gives you a way to make vectors and inpaint areas with prompt for them

i dont get that but its cool

it actually looks easier than mixture diffusion

man, did you try that shit? impossible

[smoky oak]

ok, so apprently this isn't just location based multidiffusion

[smoky oak]

Yeah, I remember lmao

[oak osprey]

would be cool to put one of your subjects in a field or something with mixture diff

[wispy nest]

yeah, it takes some time to learn, but the results are so, soooo great, or so I think anyway :P

[oak osprey]

you can do that easily

three "a grassy field" prompts, as you want them to look. and the middle prompt, with your subject

[smoky oak]

lmfao

[oak osprey]

i haven't played with overlap depth

that old man is a mood

[sterile temple]

you could probably do that with after detailer

has anyone tried the tile diffusion regional prompter?

[oak osprey]

that's what i was talkin bout

i have some tests

the best i was able to do on my own with 3 fresh prompts using isometric view as a cheat

the demo image i could not reproduce

the thing that led me to understand, "this is not easy"

before fixing it so it works with 2.1

which i kinda like, tbh, it's like, deep metaphor or something

i wasn't quite happy with this result but also an attempt to use more than 3 images

it gets weird

[wispy nest]

yeah, but not for more than a day and I'm ecstatic for it! but that might be a incorrect feeling as I learn more from testing it :P

[oak osprey]

weird to see my wizard prompt become less magical

the other one still lookin like a painting tho, but that's in the prompt

[sterile temple]

the interface for the extension in A111 is more about overlapping regions, rather than tiles stitched together

[oak osprey]

this does both

i just am not using the other thing because it's a lil more complicated and i just wanted to see it working

eg. i provide just three prompts this way and overlapping pixel areas instead of vectors that describe a region and a prompt

[smoky oak]

Trying to use this regional prompter with Walach, and I keep getting this error on start, anybody have any ideas?

[oak osprey]

well the kid is definitely going places

if you actually try with the prompt tho it's pretty good

[sterile temple]

could be a module that exists in a newer version of A111. You said you haven't upgrade in a while?

[frosty swan]

☕

[smoky oak]

😶

If thats the case, then damn it, looks like I am not using this extension

[oak osprey]

just update your local version file

lie to it

ahahaha

"im new "

though yeah there's a bunch of new shit under the hood, why were you saying you don't want to update

other than torch2

[smoky oak]

torch 2 is the only reason really

[oak osprey]

is it actually a hard requirement

like for my stuff they're interchangeable

[smoky oak]

Its just worse

slower, uses more CPU, uses more VRAM

[oak osprey]

no i mean can't you run 1.13 on newer a1111

[smoky oak]

I have 0 desire to use torch 2

[smoky oak]

oh, I am pretty sure its a hard requirement

[oak osprey]

but there's no interface changes

to its api

[smoky oak]

no idea man

all I know is that torch 2 is objectively worse, and I am holding out cause I don't wanna lose perf and also lose resources

[oak osprey]

there's a few new methods to enable new speed stuff but that's so easy to avoid hitting if it's not there

wonder if SDXL is torch2 only

[smoky oak]

probably, and I would be fine with that as that is new tech

[sterile temple]

looks like it was first added in Feb, then updated in May

[smoky oak]

what was?

ohhh

great lmao

so it IS cause I have an older version of SD

[oak osprey]

1.5?

confused lmao

detail

[smoky oak]

so in order to use this I will need a newer version of A1111

[oak osprey]

that's the "shit under the hood" is the cfg callbacks

that's to work with terminal SNR too

[smoky oak]

so now I have to sacrifice perf and resources in order to use this new tool

[oak osprey]

there's probably flags / switches for sdwebui to bring that back somewhat

i wonder if it can use torch inductor at all

[smoky oak]

Just saw how old my SD is

pretty impressive

almost 4 months old

[oak osprey]

in 13 more months it will be born

[sterile temple]

just reading through a github thread about torch 2, looks like there was issues when it first came out with the nightly builds, and people found the performance much better if they built it locally. not sure if those issues have been addressed by now

[smoky oak]

I have tried ANA UI, VLAD UI and also Auto UI with torch 2, and they have all been worse performance with more VRAM and system RAM used, and more CPU usage

I guess I could give a second install of auto a test

I didn't care for Ana UI, and VLAD is a POS, so I'll try auto

I don't even remember how to install it lmao

its been that long

[oak osprey]

this dude's lit. i hope i look like that when i'm 40

[smoky oak]

ah yes, white hair at 40 😅

[oak osprey]

it's aleady started

there is no fighting it lmao

[smoky oak]

LMFAO

wait lmao

ADATA needs to go back to primary school math lmao

[oak osprey]

this is me at 50

1600w is a nominal rating

[smoky oak]

ah yes, 4x 450 watts with 1600 watts

[oak osprey]

my solar panels are rated for 1980w but that's in nevada at certain time of year. i see more like 2100w here in canada

[smoky oak]

it needs 2 wall connectors lol

[oak osprey]

i am willing to bet it voids the warranty to do that lmao

[smoky oak]

cause it can supply more power than a single outlet can provide lol

[oak osprey]

it's like a, "just because we can, doesn't mean you shall"

oh wait

are they marketing it for that

[smoky oak]

I wonder how much it is lol

[smoky oak]

they themselves showed that it cando it, so

[oak osprey]

can != should

the Jurassic Park theorem

[smoky oak]

I wonder how much it is

[oak osprey]

axiom

well i have a 1600w psu and i don't fuckin remember how much it was but it's probably not that one

[smoky oak]

its not, for sure, this PSU is in a teir of its own

so apparently, they were able to keep it under 2250 watt load with no problems

which is like...

That is just ASKING for an explosion lmao

found the price

[oak osprey]

yeah but okay, so, a 747 jumbo jet famously can land on an aircraft carrier. once. it is rekt after that

[smoky oak]

$800 PSU moment

just get 2 PSU's lmao

get 2 1200 watt PSUs for less than half the price 😅

[sterile kiln]

How to burn capacitor

[smoky oak]

wait a sec

it has 1 wall plug

what

well that doesn't add up lmao

[oak osprey]

20 amp circuits are common now

the plugs are.. .not rated for that, but

YOLO

maybe they tested it in Europe

can get like 3000w out of an outlet

[smoky oak]

yeah, they better get ready to be sued when people try that on a 15 amp circuit lmao

[oak osprey]

breakers flip for that reason

well, unless you weld a penny in there

[smoky oak]

fair

[oak osprey]

they'll probably still get sued

[smoky oak]

its just like how people think that extension cables are dangerous, but they aren't lol

just have to use the right power rated ones lol

[oak osprey]

depends on the cable, the current

[smoky oak]

yeah, but that doesn't make them dangerous, that just makes you stupid lol

[oak osprey]

extension cords at the big box stores for outdoor use are a pretty dumb design

those are all pretty dangerous

underrated like mad

er, overrated

the bad one

one of the problems is that the same plug is used for everything

should be a separate plug type for higher current circuits, or at least keyed so there is one sideways

there is but it is up at like 220v

[sterile kiln]

Yeah, yolo, and in 6 month, you can send back your cards to warranty.

the most destructive for advanced electronics are undervoltages, an unstable voltage curve. which must largely be the case

[oak osprey]

well those liquid nitrogen tests are pretty destructive in the same way

you have to well-insulate things and corrosion still sets in even with petroleum jelly caking everything

electromigration is more hidden than that, and is invisible damage to the actual conductors

[smoky oak]

yeah, but that is not a common usage comparitively

[oak osprey]

yeah but eg. high voltage going into CPUs can be a spectrum from long term damage through electromigration or immediate explody like AMD had

same for PSUs

[smoky oak]

yeah, thats true

[oak osprey]

they just rode the razor, baby

[smoky oak]

had to recommend against Ryzen 7k to a family member because of the whole mobo manufacturer situation

so annoying

[oak osprey]

i mean i would recommend against it for other reasons, that one to me is pretty low on the list right now tbh

it's a solved problem, but others are potentially not. or, not yet. and after buying first gen threadripper i'm like, wait and see

the 1900x was a weird one, it got no attention and they basically ignored its existence and the 3700x was great. i like my 5800X3D but the 5800H has some firmware issues

we have EPYC systems that i'm abusing for ML stuff and they're pretty crap tbh, overhyped for a majority of use case unless that's like, extreme threading and PCIe I/O needs

trying to convince them to acquire some i9 systems but really something like an ASIC for Pytorch would be really cool

that single thread limit is really brutal

bitcoin miners used asics back in the day

probably the issue with pytorch on there is that everything changes so rapidly it'd probably be obsolete quickly

[oak osprey]

shits' gettin weird

the knight now has a kid

[smoky oak]

He wants the best single core perf for gaming, which is the 7800X3D, which is a highly effected CPU unfortunately

so he is holding off on the upgrade for now

[oak osprey]

the motherboard was at fault

asus

[smoky oak]

yeah, and its still an ongoing issue on some

Asus had to released 3 fixes to get it "working"

but its happening to other brands too

[oak osprey]

eh we'll see, the melting cables from nvidia ended up being mostly user error

[smoky oak]

User error for a poorly designed and effectively recalled product lol

[oak osprey]

had no issues with mine

[smoky oak]

Looks like the 5 top brands affected in order were Asus, MSI, gigabyte, ASRock and BIOSTAR for Ryzen burning

[oak osprey]

it sits at 500w for hours and hours

biostar what the hayull lmao

they're still alive?

[smoky oak]

Yeah, doing decently well

[oak osprey]

how have htey not been sued out of existence

[smoky oak]

I wish EVGA was still around

[oak osprey]

every garbage motherboard i ever had, biostar. and one gigabyte slot 1 CPU mobo that had a hairline crack

oh, an intel x79 board had broken clock source

shame on them, they fucking invented TSC

[oak osprey]

he got tired of Jensen making woosh sounds

i like to think the owner of EVGA (Anthony) saw the keynote in advance and was like

https://giphy.com/gifs/southparkgifs-3o6Ztb4BriSL47jEGs

[smoky oak]

hmm...

this pytorch 2 install isn't too much slower

about 0.7it/s slower

not too terrible

[oak osprey]

might not do same native res as before

i had to fiddle with it on my 8gb laptop

[smoky oak]

the idea of having to install all of my extensions and stuff again makes me wanna puke

[oak osprey]

can't you symlink them

windows does have it it's just limited

anyway im goin to bed

[smoky oak]

I have tried, but I have never been able to figure it out

its like 15 steps

[oak osprey]

hm :\

[smoky oak]

oh great, can't have both versions on at the same time either lol

new version goes out of its way to disconnect the old one from its webserver

[sterile temple]

have you switched this off?

[smoky oak]

I'm not sure, I'll check when I get back to my computer

[smoky oak]

where even is that?

windows settings is showing nothing

[sterile temple]

win 10

[smoky oak]

Ah, weird that you can't search for it

its off

[sterile temple]

that's good

[smoky oak]

Was hoping it was on so I could get better perf or something lik

*lol

[sterile temple]

mine was on. got a 10% performance improvement when I switched it off

😄

[smoky oak]

I remember when torch 2 was said to give you a big uplift in performance, so much for that lol

[sterile temple]

have you tried the various launch options like --opt-channelslast --opt-sub-quad-attention --opt-sdp-no-mem-attention?

[smoky oak]

I have tried opt-SDP attention or however you put it

That basically doubles my performance

It's still slower than xformers on torch 1.3 tho

I will give those other ones a try though, I have not seen those last two

[sterile temple]

funny when it generates the camera mentioned

[smoky oak]

@sterile templeok, I tried them all and it was slower

gonna mess with individuals now

[sterile temple]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations

[smoky oak]

ok so, looks like SDP and channel last is as fast as 1.3 for me

I did have channel last on 1.3

[sterile temple]

cool

[smoky oak]

@sterile templeSeems to be faster with sub quad attention

14.73it/s to 15.18it/s

gonna drop channel is last and try just quad

14.5it/s without channel is last

SDP and no mem att is 14.95

[sterile temple]

bit faster than mine 😢

[smoky oak]

retested and it seems all of them on is the fastest lol

I had a YT video that was slowing gen perf lol

all of them on got mr 15.37it/s

so about .7it/s faster than 1.3

hmmm

wait a minute

[smoky oak]

@sterile templeIt looks like these optimizations severely lower batch perf

[sterile temple]

oh really, I don't usually use batches

[smoky oak]

its not as bad as I thought, but it is slower

1.3 is 2.49it/s, and 2.0.1 is 2.36it/s for 8x 512x512 images

ok, so pytorch 2 has slightly faster post processing, so it kinda offsets the slightly slower batch size perf

[sterile temple]

have you tried changing this power setting?

[smoky oak]

I have been on normal up until now

just set to max, lets see if it makes a difference

¯_(ツ)_/¯

seems to be basically the same

worth a try tho lol

[sterile temple]

--no-half-vae

[smoky oak]

that just makes things slower in general

[sterile temple]

have you replaced the cudnn dll files?

[smoky oak]

not that I know of

[sterile temple]

https://developer.download.nvidia.com/compute/redist/cudnn/v8.8.0/local_installers/11.8/

[smoky oak]

what am I supposed to do here?

[sterile temple]

download the windows installer. open it like a zip file

cudnn\libcudnn\bin\ there are several dll files

[smoky oak]

oh, is this just to update my cudnn?

[sterile temple]

yeah i think so, the version pytorch is using

copy them into this folder stable-diffusion-webui\venv\Lib\site-packages\torch\lib

[smoky oak]

Pretty sure I am on 11.8 as is

[sterile temple]

people say replacing these dlls gave them a performance boost

[smoky oak]

alright I guess

[sterile temple]

I just did it, but I don't have a 4090 and it seems about the same for me

[smoky oak]

wait, how do I open a .exe as a zip? never done that lol

ohhh, I see

[sterile temple]

I originally just copied the temp directory path it showed in the installer, once it extracted there, before it finished installing I just went to that directory

I have 7zip installed so there's an option in my context menu to open the archive with that

[smoky oak]

which one?

[sterile temple]

torch

then lib

[smoky oak]

ok, replaced them

lets see I guess :p

watch my perf eat ass lmao

[sterile temple]

hope it does something, or nothing at least

[smoky oak]

I did it on the new install, so no worries if it dies lol

seems to have made no difference

oh well

[sterile temple]

at you can say you tried it lol

[smoky oak]

the perf difference does seem to be ever so slightly in the favor of pytorch 2 with those optimizations you shared

I think I am just gonna update my base install so I don't have to deal with transferring things

I will copy the venv just in case, that should be enough, right?

[sterile temple]

backup that and the extensions folder

or just rename them

[smoky oak]

alright, YOLO

[sterile temple]

I'm on the dev branch myself, risky business. I did it to try out that tensorRT thing, turned out to be 40% slower on my card

[smoky oak]

oh you got that working?

👁️ 👄 👁️

[sterile temple]

yeah it was fairly straight forward

[wispy ether]

always use batch size 1

[wispy ether]

what card

[smoky oak]

I managed to get overwhelmed 😅

[wispy ether]

gtx card?

[sterile temple]

3060

[wispy ether]

same

let me try

https://github.com/AUTOMATIC1111/stable-diffusion-webui-tensorrt

this right

[sterile temple]

yes

[smoky oak]

oh righhhttt

I remember why I couldn't do it

cause NVIDIA lost my account lol

oh!

it worked this time!

[wispy ether]

hehe

[smoky oak]

I have to make a whole profile and BS lol

[wispy ether]

how do i instlal it tho

[sterile temple]

use the install from URL option

[wispy ether]

i did too

stderr: Traceback (most recent call last):
File "C:\Users\Lukag\Desktop\stable-diffusion-webui\extensions\stable-diffusion-webui-tensorrt\install.py", line 29, in <module>
if trt_paths:
NameError: name 'trt_paths' is not defined

btw

[smoky oak]

can somebody send me the tensor RT file or something? NVIDIA is cockblocking me

I don't have an organization/university name and URL

[sterile temple]

can you send files in discord?

[wispy ether]

https://developer.nvidia.com/nvidia-tensorrt-8x-download

the zip is 1.3gb

[smoky oak]

uhg, I will probably have to find it on reddit or something

[wispy ether]

i mean i can use a file host

@sterile templehow do i apply tensorrt

[smoky oak]

if I could download it myself, I would

[wispy ether]

i can send you via host file

[sterile temple]

I just followed the instructions on the github page

[wispy ether]

file host*

[wispy ether]

do any

like i did

[sterile temple]

If it's private use, just put your name in the company. That's what I usually do

[smoky oak]

I don't have a URL tho

[wispy ether]

chill

im uploading

[smoky oak]

lmao

I just put no.com lol

[strong sonnet]

I used minigpt4 to caption this image. I think the results are great.
no hires or control net, just prompts
https://www.reddit.com/r/Faces/comments/13yz53v/the_face_of_a_girl_whos_now_6_months_self_harm/

What do you think of it?

[wispy ether]

fake

[strong sonnet]

Sorry, I didn't get you

[smoky oak]

@strong sonnetI am mostly interested in why the hell somebody would use 150 steps lmao

[wispy ether]

https://tenor.com/view/alan-wake-npc-stuck-walking-game-gif-23946845

[strong sonnet]

for better detail

[smoky oak]

thats not how steps work

[strong sonnet]

So the deaatils are much clear rther than they just smoothing out

[smoky oak]

20 steps is plenty lol

there is almost no visible difference

hell, I do 15, and 10 for upscale lol

That is 15 base, 10 upscale

[wispy ether]

but guys

[smoky oak]

@sterile templeshould I do Tensor RT 8.6 GA or EA?

[wispy ether]

how do i install the tensorrt zip

i mean apply

is that fine