💬｜general-chat

[old gull]

ah okay, then its definitely my AMD card for sure if youre saying it does, i guess i'll settle for 512x512 for now

[compact mantle]

how could i resize image from 768 to 4k size
the extra upscale only made it with more details but not the image size

[wide juniper]

I already have stable diffusion 1.5. How do I get 2.1? Hugging face has so many files

[forest apex]

@wide juniper follow the instructions from https://stable-diffusion-art.com/install-stable-diffusion-2-1/

[supple nymph]

Thanks!!

[boreal cosmos]

can a Quadro M6000 run Stable Diffusion?

[cloud granite]

Yes

[boreal cosmos]

ty, buying now

[cloud granite]

Ok

[boreal cosmos]

its better to have more VRAM right?

so thats why I'm getting an old M6000 with 24GB Vram

[cloud granite]

Yes it's better to have more VRAM, so it is possible to generate more images at once.

[boreal cosmos]

also if I read this right, I can almost halve the dreambooth training time (with a batch size of 2) if I have more than 21 GB VRAM?
https://www.reddit.com/r/StableDiffusion/comments/xp0a70/dreambooth_able_to_run_on_18gb_vram_now/

[tribal verge]

hi
Is there a site like lexica.art where people post their work with promtoms?

[lilac reef]

Yes, and why vram has always been the first driving factor for ML/AI. rather have a 48GB card at half the speed than a 16 at twice the speed.

[lilac reef]

Swapping in and out of the card is a heavy toll and PC mem is VERY slow in comparison plus the swap operation is slow.

[forest apex]

In the chat history someone had an M4000 and said it was fairly slow

[lilac reef]

In comparison to what? A6000? 4090?

[forest apex]

#🤝｜tech-support message

[lilac reef]

8GB GPU

[forest apex]

In comparison to a potato with a PCI connection... I dunno just wanted him to make sure

[lilac reef]

If it has at least 24 GB of ram then go for it if it has less get a RTX 3x or 4x card.

I think the M5000, or it was 6000 is a 1080 basically

[old marten]

Anyone here gotten the CPUAllocation memory error? Not cuda... CPU this time... good lord.....

[lilac reef]

Oh, and make sure it isn't like the K80 where it is 2x12GB not a full 24gb so is worthless.

[old marten]

Just when I had fixed my cuda issues, now its this. I once again suspect it's a software not hardware issue but im stumped.

[hearty vortex]

What is the best practice to show and automatically update a separate output image window on the second monitor with 1111?

[boreal cosmos]

I'm getting dis https://www.newegg.com/pny-vcqm6000-24gb-pb/p/2VV-000K-000Y7

[hearty vortex]

I just got a secondhand 3090. Hopefully not abused, but it has 2 years EVGA warranty...

[forest apex]

@boreal cosmos check out https://www.reddit.com/r/StableDiffusion/comments/xdp0fb/question_about_gpus_and_memory/

[slate crescent]

VRAM is the bottleneck for most AI applications across the current consumer product stack (and this seems highly unlikely to change anytime soon), but reallocate resources enough and at some point the bottleneck shifts elsewhere

like AFAIK, for instance if all you care to do is generate a single image, 12GB vs 24GB will be irrelevant to the speed; more VRAM can only assist when it comes to high batch sizes, and is primarily of interest for training

[boreal cosmos]

thats for generating images

I want to train stuff too

[lilac reef]

Better to get a 3090 (not TI) than that M6000 as it will smoke the M6000

[slate crescent]

the M6000 is a Maxwell chip, comparable to a 980

it would be an unwise purchase

[old marten]

why get a quadro over an rtx 3090? just curious.

[slate crescent]

all that VRAM and nowhere to go

[forest apex]

I have a 1060, it's not half bad, a 1080 8GB would be nice, but a used 3060 goes for cheap nowadays

[old marten]

If ur gonna go big and go home, get a 40GB A100

[boreal cosmos]

cheaper

[old marten]

Or if ur ass pimping, get that insane 80GB Tesla 🤤

[lilac reef]

I prefer owning my kidneys and they remain in me.

[hearty vortex]

Why not Ti? I wanted to get 3090 Ti because 3090 may have memory overheating issue which was solved with Ti.

[old marten]

Uh its 850 I got my 3090 for 800

[boreal cosmos]

wah

where u got it for 800

[old marten]

used

[boreal cosmos]

where

[slate crescent]

I am unsure that the Maxwell family (2016!) is capable for doing much in ML

[old marten]

offerup

local is better. no taxes

[forest apex]

reddit hardwareswap with notifications because people are thirsty for GPU's

[old marten]

I also sold my 3080TI for $600 there

so essentially just paid $200 for an upgrade of 12gb vram

[lilac reef]

ti has less mem

[boreal cosmos]

oh wao $730

https://offerup.com/item/detail/86b78ce1-0846-3854-ad57-dfdceed46485?q=RTX+3090

[old marten]

see

[boreal cosmos]

thanks!

[old marten]

just beware cuz offerup is fulla scammers

[hearty vortex]

Both 24GB...

[old marten]

if they have an account that was recently created and 0 reviews, dont trust em

[boreal cosmos]

thank you guys so much, I would have bought an extremely crap card otherwise

[slate crescent]

yeah, the M6000 is CUDA capability version 5.2

I dunno what the min req for the current toolchains in practice are but that's... certainly a new record for being low

[forest apex]

a potato with a PCI adapter

[old marten]

its cuz its an old card

back then 24gb was insane

[slate crescent]

the RTX 3000 series is spec 8.6

[lilac reef]

Yes, it was something else I was looking at. 3090ti is the better of the two as it has more cores, is faster, etc...

[slate crescent]

like Maxwell doesn't even have tensor cores at all

[hearty vortex]

Yeah, but Ti is about $250 more than 3090 on Ebay... I ended up pull the trigger on 3090. I may have to undervolt it to protect mem.

[old marten]

Sadly the Ti is also rarer so ppl still get away with scalping it. I too wanted the Ti but oh well

Undervolting it to protect it is a myth btw

[hearty vortex]

Really?

[old marten]

Even overclocked as I have it it works well

And electricity bills are no biggie, even at max workload I will spend 9c/hr

tops

[hearty vortex]

A good news.

[old marten]

Even LTT who did a youtube video about whether or not you should buy miner cards was not able to find any damage on those that were running 24/7

Not saying it can't happen, but it's not as common as ppl think

[lilac reef]

My 1-6- is suffering after being used for deepfakes 2018-2020, and other hard stuff. replaced tim/pads/fans already once, so I was saving for the 4060 to only get gut punched by nvidia. Now I have no idea and I do not trust the used market for anything so I am stuck as my pleasure is in training not just seeing images.

[hearty vortex]

Yes, I was trying to find an Ebay seller who should be obviously a gamer based on his past transaction + he has the original receipt.

[slate crescent]

I ran into a couple guys who had mining cards die after a month, so I nope'd out of that

[hearty vortex]

Ah...

[old marten]

Damn

But did they buy just one card or did they buy a bunch and just a few of em died?

Plus what exactly happens to em that they die?

[slate crescent]

no idea what the actual statistical odds are, or what tricks one could use to protect your chances (how to spot a bad card, etc)

these were one-off cases

totally anecdotal

it was just enough to trigger my "don't buy a used car from the north" warning bells

[hearty vortex]

That's why I got EVGA card. EVGA has by far the most generous warranty policy. Too bad they stopped making card.

[old marten]

I'm pretty sure the guy whom I bought it from didn't use it for mining, you can tell if they are selling multiple ones at a time, especially if u look at their sales history

I doubt someone would mine on just one RTX 3090. Plus he let me into his apartment to test it. I did the same with my buyer and its obvious from my tiny ass apartment I cant do mining here lol

[hearty vortex]

The seller sold his 2090 in 2021. I thought its a good sign.

[slate crescent]

tbqh the single best bet is if the person is selling a 3090 because they got a 4090--it's a perfect and satisfactory explanation that there is nothing faulty with the old card (just perhaps the human's judgement)

[astral goblet]

evga's warranty was good if you were in america. in canada they typically hated dealing with support cases and it was like getting a teenager to do a chore

[slate crescent]

it's very important to understand why someone selling a used product is selling it--it's not an invasive or unreasonable question

[astral goblet]

as a canadian, i stopped buying evga products a long time ago. thats not THEIR problem though. it's kind of the industry wide problem

[lilac reef]

The only way I buy used is never, and second way is if I know you, we hang out together a lot so I know how you treat your stuff. Since I don't hang out with people, and I don't know anyone irl (I am a loner) I buy new only.

EVGA is no longer a partner so for me Nvidia is dead to me since I do not like any of their other partners.

[astral goblet]

gpus are a device i'll often buy used and trust fine. especially flag ship gpus. the mining issue with bad products on the used market was just a blip, before miningtrepreneurs figured out that overclocking is less economical than buying another card

[old marten]

well think of it this way. If a GPU runs in a testing util like Kombustor, it's already proof that it is 90% issue-free

[hearty vortex]

I'll start mining with my used 3090 and make some money for my next new card, haha.

[astral goblet]

mining's a dead game

[old marten]

If you do a physical inspection, thats another 5% guarantee. The other 5% chance is what you can't see, which is wear and tear from mining, which doesnt always happen

[lilac reef]

mining is dead to never come back

[hearty vortex]

I hope so!

[slate crescent]

God bless.

[forest apex]

but the chancellor is still on the brink :/

[astral goblet]

mining will keep going for hobbyist people who want to maintain chains for their own goals, but theres no economic incentive for people to mine public chains anymore

[old marten]

Guarantee the 50 series GPUs are gonna be way better for consumers

The 40 series were made while mining was still alive

[astral goblet]

the economics of mining shifted HARD

monumental shift

[old marten]

I doubt the 5090 will msrp for more than 1400

[slate crescent]

oh I would easily take that bet

[astral goblet]

the market is still bearing high prices. the 5090 will likely be 1700

[old marten]

nah G, 4070Ti flopped, 4080 flopped harder

4090s sold out only because of artificial scarcity

they are waaaay overpriced for tiny ass 12gb vram

[astral goblet]

machine learning is the driving force of gpu sales now. gaming is still the backseat audience

[slate crescent]

the only way the 5090 ends up lower is if the product stack shifts (they make room for a Titan or otherwise superior SKU)

[astral goblet]

"they sold out" yeah sounds like a huge problem for nvidia...

this is what you call wishful thinking

[slate crescent]

uhhhhhh at the halo, extending into datacenter GPUs, sure

[hearty vortex]

Cript did not change the world, but SD and Chat DPT certainly ddid.

[slate crescent]

the gaming market for desktop GPUs is still the vast, vast majority

[astral goblet]

you underestimate how many people are entering into the home machine learning space

[slate crescent]

naw I definitely don't

[astral goblet]

i've got 80 year olds asking me to build thema modern pc for training ai on

[slate crescent]

it's a lot, but the number of people who play video games is a stunning % of all of society

[astral goblet]

this tech has penetrated the masses very rapidly

[slate crescent]

SD 1.5 has been downloaded on Hugging face 1.25 mil times--that's a lot, incredibly impressive

[astral goblet]

i dont think the 4070 was a flop either. i mean, people bought it and it's only a few months out. you can't find one for sale on the west coast right now. weird way to convince me it was a total flop. the 70 series will likely target gamers while higher end cards are going to get a lot more expensive

[slate crescent]

but a small sliver of the desktop GPU install base

[hearty vortex]

Nvidia will release 16GB or 24GB 4060? Finger Crossed...

[astral goblet]

why would prices come down? the market is bearing the high prices fine

[old marten]

rofl in ur dreams

[astral goblet]

not likely. that will eat their high end product sales. if they stop selling then maybe they'll bring the gaming targeted cards a 24gb

[old marten]

not even the 4080 got 16gb m8

[astral goblet]

tehres no game that needs more than 10 on the market today. not even a 4k game

i've got a 4080 and it has 16gb

i didn't buy this card for gaming either. for gaming i would've just got an amd

[slate crescent]

the bigger issue is the memory controller on the 4000 series is speced for a certain width/speed relative to what it is feeding

you can't just, double the ram

you need twice as many lanes to carry that data

you are asking for bigger silicon

[astral goblet]

can't we just install the ram doubler?

https://en.wikipedia.org/wiki/SoftRAM (old scams are still fun)

[slate crescent]

it is conceivable that nVidia could consider a 24 GB lower end product regardless, committing to the engineering requirements--more niche products HAVE been made, including high VRAM cards. But as it has been pointed out before and even here, the extent to which it might cannibalize higher product sales is a big question

if it holds its own, it results in less 4090 sales; if it doesn't, bad press. lose-lose?

[astral goblet]

when gamers need 24gb, there'll be midrange cards with 24gb

[slate crescent]

keep in mind the 4090 itself already has this relationship to its big brother(s)

[astral goblet]

yup. the 4090 is going to be limited so that the a9000, whatever they'll call the lovelace pro line, can fly

[river maple]

price discrimination is evil

[slate crescent]

mate you don't want to live a world without price discrimination

flying would be so expensive

and basically all silicon would be way more pricy than the consumer market we enjoy

[astral goblet]

economics is a tough subject and i often try to remind myself it's all going to be evil through one lens or the other

[river maple]

the differences are practical, not artificial there

flying is expensive

[astral goblet]

morally, economics is a tough one to come to rest on

[river maple]

but consider how much it actually costs to manufacture many consumer products?

[old marten]

how much does a 3090 cost to manufacture?

[hearty vortex]

The market share of 24GB cards is most probably less than 1% of total sales of GPU, but because of ST, it will be much higher than 1% this year.

[old marten]

fuck it how much does a 4090 cost to make

[slate crescent]

pretty much all silicon products are overwhelmingly R&D

[river maple]

probably a few hundred at most. toss on an extra couple hundred for R&D if you really want to. then add retail.

yeah, I'm thinking RTX4090 should be $1k at most

[slate crescent]

"if you really want to" lol

[astral goblet]

how do you factor in the costs of building r&d facilities and years of investment towards a development goal? and those goals still need to be achieved so new r&d facilities will needd building sooner than later, and they'll cost

tech especially has so many aux costs that don't apply to the manufacturing process

[old marten]

at least the answer wasnt $10 for a 4090 as I was expecting. lol

[slate crescent]

the silicon itself, in terms of price quoted from TSMC is.... quite low

even the A10X GPUs, which are HUGE, record breakingly expensive, will be around $100 tops

[astral goblet]

and of course, theres always market forces and what the market is willing to bear. businesses aren't in business so they get to leave money on the table.

[river maple]

please see https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2023

companies currently make more than enough revenue

[river maple]

the trick is you buy in mega-bulk

[slate crescent]

if you are especially uncharitable you could construe the price of most intel CPUs to be ~$5 or so

[astral goblet]

i'm not at all saying they're in dire straits, to be sure. there's jsut a ton more consideration in pricing a product than material cost is what i'm saying

[slate crescent]

this entire industry is basically pure R&D; sand is not expensive my dudes

[astral goblet]

i like seeing when tech revenues are that high, typically get injected back into r&d and then me as an end user benefits greatly down the road

[old marten]

nah dude thats wishful thinking

[astral goblet]

its when middlemen companies like amazon get the lead that makes me woozy. what's with middlemen? what do they bring?

[old marten]

nvidia is obviously milking the gpu shortage (or was) with these wackass 40 series prices

remember when they first announced the 30 series

the 3090 was supposed to msrp for like $600

[slate crescent]

ironically amazon became king by being a single middleman that could replace 3-5

as a consumer trend, amazon is the middleman killer

[astral goblet]

it's how r&d has worked for decades though. You think we got down to 14nm circuits because these companies sold their chips for a reasonable margin over material cost?

[old marten]

Then the shortage happened and they were like "whoooops! Did we say $600? We meant $1200! Sowwy!"

Well they initially thought 600 was more than reasonable

so its not like a 100% inflation happened in the span of a year

They obviously are tacking way more than is fair

[astral goblet]

the shortage wasn't because of supply issues. that was a demand issue. market demand suddenly skyrocketed as people other than gamers realised this silicon was valuable

[old marten]

rofl no it happened bc of covid and chip shortages due to supply chain issues

on top of the crypto explosion which thank god died down

[astral goblet]

remember too that a lot of this factory output is taken by the mlitary industrial complex and all sorts of other industry. tmsc doesn't just serve gamers

[slate crescent]

in the context of GPUs it's not even a demand thing (overall demand as a sector is actually down) as much it is segment redefinition

[astral goblet]

thats not obvious though. that's just what you want

[slate crescent]

used to, a 970 was a high end card and a 980 was super high end

[astral goblet]

you're not the entire market, so remind yourself that the entire market are still bearing these prices

[old marten]

it is obvious when you consider their initial prices

they know ppl will still buy em even at crazy prices

except it didnt work this time with 4080 and 4070ti

[slate crescent]

now an XX80 is super-mega-plus high end, and the 4090 is we-don't-even-have-words-for-this-in-the-consumer-sector-end

the 4060 is a conventional high-end card!

[astral goblet]

material cost isn't where tech gets its value from. its much more abstract then that

[signal nest]

i get the impression it was close to 50/50 . supply chain hit other types of chip. i had no problem getting a prebuilt with a GPU.

[slate crescent]

the supply side was absolutely crunched

silicon orders were impossible to get

[astral goblet]

4090 is a 4k raytracing targeted card in the consumer sector. for gaming, its meant to give a 4k experience for spiderman or cyberpunk

[old marten]

dlss

[astral goblet]

there is absolutley market space for it with the games available to gamers

[slate crescent]

exactly, it's a newly defined sector

erm, market tier

[old marten]

its only the snobs who want native 4k that the card is for, aside from us who use it for AI etc

[astral goblet]

but.. no.. there are words for it

[old marten]

otherwise dlss already allows 4k at 120+

on 3090

[astral goblet]

highest end gaming products are always for the snobs. and they're gamer snobs too so theyr'e harmless

[old marten]

that is very very true

rich ppl will always spend money on the stupidest things

thats why we have 8k tvs now

[slate crescent]

I meant words in the traditional low/mid/high end product tiers, as defined by consumer needs and games that require them to run

[astral goblet]

when i was a kid, the guy who had the sega channel was a total snob. luckily i was his friend so i played a ton of sega channel

[old marten]

even tho they look exactly the same as 4k

[astral goblet]

ultra tier. i'm sticking to it.

[slate crescent]

sure, that works

[astral goblet]

/killerinstinctultraannouncer

[slate crescent]

it's just something way beyond what say the RTX 980 was doing

[astral goblet]

also it is called Ultra HD

GTX, they hadn't jumped the shark to the Raytracing marketing line yet

[slate crescent]

oh, right

yeah all this raytracing stuff is pretty silly imo

[astral goblet]

(i think raytracing in games is a gimmick still, but it sure sells! and its driving the adoption of cool tech like tensor cores so right on. go RT)

[slate crescent]

but, whatever, people always poo-poo new tech

[old marten]

it is not

rtx is definitely noticeable

especially in games like minecraft

[slate crescent]

right, like, better this than blockchain rofl, whatever gets you there

[astral goblet]

yeah it's noticeable yup. but its a gimmick

you can use shaders for a lot of the same effects in minecraft

[slate crescent]

my experience with the RTX ultra stuff (Portal, etc) has been "uh, that's a pretty piece of glass, I guess?"

[old marten]

reflections are a good example too, u can only get that with rtx

[astral goblet]

blockchain was fun for a little while until bitconnect and other ponzi stock bros raided the whole scene and built it into a disaster area. a very quick little while.

[old marten]

RE:3 and 2 remaster make good use of them

subtle yet add a good touch of realism

[astral goblet]

i think theres still some potential for novel coool uses of blockchain, but we're not there

[peak portal]

Hey question, i'm rendering a batch of a PNG sequence from a video. Is there a way to make each frame a little more similar so the video is a little more cohesive vs. now where every frame is img2img from scratch? Would that be the looping script?

[slate crescent]

yeah, the well is pretty tainted by bad actors and conmen now though--basically anyone investing time and engineering into blockchain tech can be reliably guessed to have stupid or malicious intentions

[old marten]

no. for now we are stuck with that weird acid trip look for videos

[astral goblet]

you can fake reflections a number of ways, but they're not very needed for a game still. it's a tacked on gimmick and it sells cards. hurray.

i know all about how raytracing works. I've done courses during my time earning a degree on many facets of computer graphics, including optics. i'm no expert, more of a jack of all, master of none

[slate crescent]

there was a guy on reddit today talking about this + showing off a proof of concept, but said he was brute forcing it inefficiently

[peak portal]

Ok got it, thank you both! I did find EbSynth which kind of accomplishes this

haven't tried yet though just saw a youtube video

any experience with it?

[slate crescent]

personally haven't touched video; anyone?

[astral goblet]

ethereum foundation, that vitalik clown, he's a cool guy i think. he's building the tech for the market 10 years from now, not what it is now.

the ethereum company themselves are the ones that murdered the economy by moving to proof of stake right? so that's a big indicator they're not part of that poisoned well

[peak portal]

@slate crescent the workflow I saw the guy do was - export png sequence and run through stable diffusion, pick the frame he like best as the "style reference". Than use that photo as the reference in EbSynth and run through the raw png sequence in ebsynth with the chosen frame as reference

I'm going to try it today, will report back

[old marten]

Do post results if u get em

I wanna see how good/bad it looks

the technology is not quite there yet imo, but its nice to see the baby steps

[astral goblet]

new version of the nvidia studio driver came out today

[forest apex]

#🎥｜animation has some more coherent animations, and someone just posted a tutorial, and there are people getting results on reddit

[slate crescent]

I mean, "currency" as an application of blockchain is pretty dubious, even if PoS is magnitudes less insane than PoW (which is just dystopic absurdity)

[forest apex]

@peak portal you'll have to scroll up a bit

[peak portal]

do you remember the name of tutorial?

[old marten]

I mean to be fair, bitcoin is already ruined bc of uncle sam

U cant buy bitcoin anywhere without uncle sam snooping by and asking for bank documents and shit

I legit got banned out of coinbase for sending crypto to a darkweb address

like why tf do they care. its my money

[peak portal]

and FWIW i'm trying to do it with video frames, not generations from text so not sure if that makes a difference

[forest apex]

@peak portal I'll check but enigmatic_e on youtube gets some of of the best img2img renders

[peak portal]

@forest apex awesome, will check him out. Thanks!

[astral goblet]

it's less currency being staked, and the intrinsic value of that token. The eth token's value is pegged to the scale of the network. if the network is bigger, more eth is needed to fuel it. That intention all along was that the eth would be the energy economy of the network.

it's definately a lot different than pow, but that was never meant to be the final block chain design. it was just used becuse it worked (heh). but its inefficient by design

eth was never meant to be a daily currency like its treated by stock bros

[slate crescent]

for the question "well, who should be the first, original holders of our currency be?" the answer "whichever Chinese farm burns the most electricity" might be the single worst imaginable

[forest apex]

@peak portal this is for img2img transfer, it's similar to what enigmatic_e does https://www.youtube.com/watch?v=Z6pR7IB9in0

[peak portal]

thanks @forest apex and everyone for the info!

[slate crescent]

good luck Tim!

[astral goblet]

yup. proof of work began expiring the day that optimized asics were developed for that purpose. the gpu economic clusterfuck just highlighted that it's bad for economic forces and not great for democratizing the network. that's why memory intensive algorithms were developed, to shore up defenses against asic mining farms

[slate crescent]

it was an environmental and geopolitical disaster, just a trainwreck all around

our poor GPU market was like a tertiary consequence in the constellation of bad PoW outcomes

[river maple]

I'm ok with it since I've profited off Bitcoin

[astral goblet]

well, the environmental disaster part is questionable. while mining the ethereum and bitcoin networks was energy intensive, it's nothing compared to the carbon created by industry or big tech data centers. while it was smart to correct the inefficiencies of public blockchains, i think the media circus on that was a gigantic lobbied deflection

[signal nest]

the story that drove crypto was the instability of asset bubbles caused by central banking.. i watched those boom bust spikes of crypto prices and wondered how anyone could possibly think this was somehow a cure

[astral goblet]

no one has got as mad about youtube as they have mining farms

[slate crescent]

the key is that this is purely makework

youtube is a literal service that provides humans with happiness

[astral goblet]

right, but the actual scope of the impact was very insignificant still

[slate crescent]

like of course it's ok for goods and services to consume resources lol

[astral goblet]

people aren't mad about the actual big environmental impacts, and that's why these deflections work so well

pay no attention to the man behind the curtain

i'm sure there were a lot of stock bros who caught happiness from blockchains. they're humans too

[slate crescent]

I agree in part that the question of magnitude is important, but cypto minining's power consumption was not a drop in the ocean by any means

[astral goblet]

no, it really was. proportionally it was very innefficient yes, but total? it was a drop in the ocean. These claims of "it spends more electricity than some countries" are hyperbole nonsense designed to deflect attention

[slate crescent]

like the total power consumption of mining was, you know the stats, bigger than most countries

[astral goblet]

yup, that claim

right on time

[slate crescent]

right

[signal nest]

power consumption relative to actual service provided, basically. It's just counting other work. There's an argument that the finance world also wastes a lot of energy but it is supporting people to do it along the way

[astral goblet]

the power consumption of aws is more than some countries

[slate crescent]

right... because it's.... providing services...

UPS spends a ton of gas delivering my packages

[signal nest]

"the internet" consumes more energy than aviation, surprisingly. but its doing a lot of visible work

[slate crescent]

yup

[astral goblet]

is it though? like can we honestly say most of what runs on aws is needed? or is it all an environmental drain?

I'm not trying to justify blockchhains here. i'm trying to point out that the "environmental concerns" are flaccid when you step them up. people don't actually care about carbon output. just when it's the bandwagon

the real carbon impacts seem to operate without hinderence

[slate crescent]

I do think this is fair, but it's not irrational for people to jump on impact that is transparently needless, resource consumption for resource consumption's sake

like if you town has a propane burning competition, it's reasonable to be upset at this as wasteful even if it is 0.000000001% of the world's carbon footprint

the ratio of value to cost is near 0

watching youtube is an extremely carbon-efficient means of human entertainment

I agree that if one actually cares about the environment (or carbon footprints specifically), one's attitude should be "ok sure, let's stop this wasteful nonsense--and then keep caring about these even bigger things too, even if they might require some sacrifice"

[signal nest]

right this is how i justify it. Before telecoms & computers , people would need to physically travel and/or use more resources for real world entertainment. The big example used is going to a concert vs having music downloaded. I made driving games for a living, some would say "waste of time" , but it's surely less carbon footprint (& resource use) than actual motorsport - and directly enjoyable by more people

[slate crescent]

what studio tars?

[signal nest]

long gone.

[slate crescent]

ah

people do tend to get hung on on weirdly misplaced environmental concerns when it comes to technology

I remember a tour of an environmental engineering company once, people doing good work as far as I could tell, and some goober asked if it was wasteful that their secretary had two monitors

and the president of this company is having to explain to this college kid how the productivity efficiency gains of a human are, in fact, relevant

[signal nest]

(2 is the sweetspot apparently , not sure how that study changes with various sizes)

[mental sparrow]

hey guys, im really terrible at figuring out big repositories of code. i want to interpolate between two clip latents and generate images from it. how do i do this?

[slate crescent]

I am such a glutton in this regard; I am "down to 2" now, because one of them is a Neo G9--and I want more lol

I 100% utilized all 4 when I had 4 1080ps

if you worked in games, you can probably understand

[signal nest]

(hah yeah. it bugged me that you get a bezel in the centre, so i've usualy wanted 3 , but i was well aware from the amount of head movement that 3 widescreens is excessive. I'm guessing those ultrawides could do the job of 2 regular monitors.)

[slate crescent]

I am finding the G9 really great in Unity and visual studio, but noooot quite as good at having 2 monitors

[signal nest]

right definitely . exporter->debugger-> game for example

[slate crescent]

exactly

I am currently trying a setup with the screen divided into fifths, in which unity and VS each take up 3/5s, overlapping in the middle

[signal nest]

right when you have things fullscreened in each , its more comfortable

[slate crescent]

so both get all the space they want, and I can mostly do back-and-forth cross-referencing as needed

[slate crescent]

fullscreen is pretty universally useless in 32:9 though

maybe Unity/Unreal or "timeline" based programs would be fine

[signal nest]

right a superwide wants to be partitioned, i've never used one. I know most OS's have window managers that can tile these days

[slate crescent]

yes, I have found Windows Powertools to be very sufficient and even default Win11 behavior to be surprisingly robust

it is something I'm still learning into muscle memory, though

[signal nest]

i'm using "2.5" perhaps because I watch stuff on an ipad.

part of it is that a certain amount of head/eye movement saves you time & mental effort actually switching windows. it's possible VR could save on displays, if those ever got comfortable and high res enough

[slate crescent]

I feel like with VR the concept is obviously there, but it has just so far to go to make it meet those criteria

those two both form an "impossible triangle" of sorts with battery

[signal nest]

yeah i haven't bothered with it at all, vs other computing kit which I know I can use all day.

[river maple]

VR ----- AI
 \       /
  \     /
 Batteries

The triangle of impossible to perfect technologies

[signal nest]

VR needn't be portable but i think 'they' believe it would have to be to be ubiquitous. and yes the displays aren't even there yet.

[slate crescent]

I mean that battery (life) is directly at odds with either comfort or fidelity

[signal nest]

impossible triangle - trilema right? "battery life, comfort, fidelity - pick 2"

[slate crescent]

yeah, a common (and fun) logical pattern

this is more about tradeoffs than a strict trilemma, but it works

I mean I agree that Meta seems far more opposed (to a power cord having any place in their ideal future) that necessary, but I do think a cord would absolutely fall under comfort

most people prefer wireless headphones, and while I'm a weird duck that doesn't I still get very annoyed with the cord and take them off often

I would be immensely annoyed if doing so turned my screens off!

but yeah, popping the stack, part of the frustration with crypto and it's hype is a sort of boy-who-cried-wolf effect; it's harder to get people to take actual imminently-revolutionary advances in tech seriously when a lot of people just got burned by the magical funni-money

it doesn't help that the loudest flavors of the AI hype megaphones are so inflammatory

[astral goblet]

i too am thrown by win11's changes. tabs in my console? what is this sorcery

[slate crescent]

win11 has been 80% positive surprises, 10% annoying things to fix, and 10% fixes I'm still waiting on

overall better than I expected and good

[signal nest]

yeah i'm hearing this aLOT , people conflating NFT hype with AI art hype , "its the same people" (no it isnt, i loathed NFT's contributing to the GPU shortage holding AI art back)

[slate crescent]

yeah, this pops up in politics all the time--subgroups and opposing internal factions are everywhere in human activity, and constantly lumped together by outsiders driving by

[astral goblet]

yesterday i couldn't get gta loaded because onedrive held an ancient version of rockstar's folder it's it's cloud documents, and rockstar launcher was trying to load from that

why even one drive?

go away!

mostly cool surprises though

this nvidia studio driver fixes a bunch of adobe software instabilities, hopefully that means stability propegation to my weird cuda crashes that always happen. not memory errors. just scripts going haywire and leading to cuda failing. sampling.py always seems to be the tail end of the errors

[slate crescent]

I am hopeful that the overall software stack surrounding consumer ML will really stabalize in the next 2-3 months

any update on distilled diffusion?

[slim relic]

I really dont understand how to run other models on stable diffusion. Can anyone help please?

[slate crescent]

"other run"?

[slim relic]

jeez my english

[slate crescent]

ah

are you talking about using Automatic1111, or what interface? a web service?

[slim relic]

Automatic1111

and I want to use Orangemixs and AnythingV3

[slate crescent]

you can put other models besides SD (1.5, 2.1, etc) in the appropriate folder there in the project

it will appear in the drop down menu at the very top of the UI

specifically, put them in the models/Stable-diffusion folder

[slim relic]

lemme try

[slim relic]

"other models" means a folder, or a file in the folder?

[slate crescent]

a model is a single .ckpt or .safetensors file, generally 2-4 GB in size

SD 2.0+ models require a small .yaml file in addition, to tell auto1111 how to load them correctly last I checked

[slim relic]

I see I see

then what are the other folders for?

[slate crescent]

the other folders within /models/ are for the other (more minor) AI processes (outside of image generation) the GUI supports in some way, like the ability to run deepbooru to generate booru tags for an image

the "Stable-diffusion" folder could plausibly be renamed "image inference", but it's not inaccurate to say that all of these models are types of Stable Diffusion

[slim relic]

I'm learning a lot today

thx so much mate, great help

El Psy Congroo.

[slate crescent]

best of luck!

[graceful plover]

is it just me or does it get less creative when you generate higher resolution images compared to the native res

I always have that feeling when I try it

it's hard to say since the entire thing is inconsistent in general

but I feel like the results tend to be more boring and "vague" looking

[muted tangle]

i find it makes images more accurate if they happen to be in a native res

that the source was in

i.e. a landscape in a 16:9 format vs 768x768

[graceful plover]

idk if accurate is the right word but the results I get with 1024x1024 look less interesting and less rich than the ones that are 768x768
I just experimented with it, let me post it

posted an example in the other channel

[charred mesa]

Is deepfloyd even going to be easy to run on regular computers xd

[graceful plover]

I'm hoping we'll eventually get text models that can run on regular PCs

Apparently GPT like stuff needs 100GB of VRAM+

The 4090Ti might come with 48GB of VRAM, and it will definitely have NVLink so that's going to be a possibility

[charred mesa]

GPT3 alternatives like GPT-J and GPT-NEO-x require about +20 GB of VRAM lmao

and it's sad

[graceful plover]

are they as good

[charred mesa]

no idea

Probably not

[graceful plover]

though I only have a 3080 Ti so I can't run them either

[charred mesa]

You can try NEO-X at textsynth.com

[graceful plover]

That said my GPU seems to be just third place for SD performance, there's a tom's hardware article

[charred mesa]

wow

[graceful plover]

https://www.tomshardware.com/news/stable-diffusion-gpu-benchmarks

4000 series is slower for some reason

Of course they are only measuring consumer GPUs

but still

3090 Ti, 3090, 3080 Ti are the fastest in that order

Probably because 4000 series is still not perfectly optimized

[charred mesa]

xformers:

AItemplate:
TensorRT:

[graceful plover]

xformers is shown in the benchmark too

[viral vale]

I honestly fear what inpainting with... research models can do.. if you post a picture of yourself on the internet, you're at risk. Then again, this has been true since photoshop existed

[graceful plover]

we've had faceswap AI for a while

I'd say this has been true since photography existed

Photos can be manipulated

And I would argue AI editing is easy to spot in fact, compared to manual editing where you can make sure there are no artifacts

if someone is looking for artifacts the automated AI edits are more obvious than something painstakingly done by a human who knows what others will look for

[astral goblet]

likely will be released but not useable for most people. it could spark some basement engineer to somehow do magic and load the model onto home gpus

[glad jolt]

Likely, some people in the community will find a way, like the many optimizations added to stable diffusion. Give a community the source code to modify, and they will modify it.

Kinda like distilled stable diffusion which is supposed to be happening some time soon ™️

[slim relic]

what are the the typical suffix for vae ?

[swift snow]

Does anyone have a model for making cars?

[valid python]

"At first glance, Stability AI’s Emad Mostaque doesn’t seem like a man who’s going up against two of the biggest companies in the world — Microsoft and Google. But that’s exactly what he’s doing in the generative AI arms race.

OpenAI’s release of chatbot ChatGPT and image maker DALL-E have catalysed the biggest VC hype eruption in recent memory, capped off with Microsoft’s near-unprecedented proposed $10bn investment into the company.

But Mostaque — with all the swagger of a former hedge fund manager — doesn’t seem fazed.

“$10bn may seem like a lot — it’s actually still just the start of this,” he says casually.

Stability AI is no slouch when it comes to fundraising either. In October the two-year-old London-based company raised $101m from elite investors like Lightspeed and Coatue to scale its generative AI technology, which has been made famous by its open source image generator Stable Diffusion.

And when asked about how he plans to differentiate his company’s generative AI offering from the giants of Microsoft and Google, Mostaque doesn’t pull punches."

https://sifted.eu/articles/stable-diffusion-ai-emad-mostaque/

[old marten]

Thats a smart man.

ChatGPT-style engines are the future, 10bn are pennies compared to what this will become

Google searches will seem like someone using a yellow phone directory to look up businesses. And thats good. Because google searching stuff sucks.

2023 will mark the greatest leap for AI technology. The AI revolution is just starting, just like the industrial revolution once did

[still remnant]

“A percentage of people are simply unpleasant and weird, but that’s humanity” - Emad Mostaque

I wish this quote had existed for my highschool yearbook.

[silk tartan]

Can someone explain me how to use "/interrogate"? I get an error message when i try to use my PNG-image. I downloaded the png from here.

ERROR "no exif entries found"

[lilac reef]

What does it mean when blip repeats itself?

[lilac reef]

Precisely but raw power the 4090 is 56% faster than a 3090 ti. Once this is optimized we should see that amount of gain for AI/ML. If they had not knee capped it in other areas that the 3090 was not knee capped on the gains would have eaten into their business line of cards as 100+ percent faster would have easily happened.

4090 eats fp32 for lunch and is supposedly made for it from the hardware video I watched when the 4090 released. Damn hard finding anything that is AI/ML and not gaming related for cards.

[waxen pasture]

I'm so happy with how well my latest embedding was recieved on civitai :)
Got so many more downloads than expected

I just wished more people left a comment with some things they made, I always love seeing what people do with things I make

[fervent thunder]

isnt like gpt4 coming soon

and I really thought stable would drop something today. wrong again. shit

[old gull]

can someone explain model vs vae to me.. or a good place to read on what a vae is or how theyre made vs how models are made

[fervent thunder]

vae basically guides the model for better accuracy is how i understand it at least

[unborn phoenix]

Hi

[mental sparrow]

an autoencoder comprises of an encoder half and a decoder half
the encoder takes inputs and compresses them to latents
the decoder tries to reconstruct the the oirginal input from the latents

there are no constraints to the shape of the latent space in an autoencoder, as a result, small perturbations in the latent encoding of an input can result in unpredictable changes in the decoded reconstruction.

In order to make the decoding more stable, we have variational autoencoders which encourage the latent space to mimic a gaussian distribution. this makes it more robust and allows us to sample artificial latents from that gaussian distribution and decode them to reconstruct artificial inputs that mimic the training dataset.

VAE is just one structural component of the stable diffusion pipeline. We use the decoder part, specifically, to decode latents back into images.

i THINK this should be accurate but im still trying to figure it all out. apologies if I got something wrong.

[molten pecan]

anyone got a message on DM from an account in this server?

[forest apex]

@molten pecan saying what? a random person?

[waxen pasture]

Was it like "Join this server, hope this doesnt annoy you, yadda yadda yadda?"

Something with maybe blue in the name?

[steady arrow]

What's stopping SD from being as good as midjourney

will we get there..

[waxen pasture]

You can get results as good as if not better than midjourney if you do things right

[distant shore]

hi guys.. wanna ask something.. everytime sd launch locally there will be cmd box comeout.. but when i put my pc to sleep and turn back on.. that cmd bos was gone.. any idea how to close sd running locally completely?

[steady arrow]

I dont believe that for a second

Ive been tweaking and twerking sd since its first release and it is not capable of that

not even custom models, custom vaes, nothin can make it so

[fervent thunder]

I guess it's a lot to do with what you define as good as MJ, but for me I like my results better than what MJ can create...not that I have ever made a image on MJ so maybe I shouldn't really be saying anything :P

[steady arrow]

open task manager on windows and find it in the details tab and right click and click end task

[steady arrow]

give it a try you get 25 free minutes there to compare

[waxen pasture]

Well, it can get really close at least

May not be perfect

[steady arrow]

the truly best generations ever made of SD may come close to the average generation of MJ, but its incredibly rare

[fervent thunder]

thanks for the tip! Might need to try it sometime. But what I don't like about MJ is how blurry their images are. But that's also I believe just a "style enjoyment" issue :)

[waxen pasture]

But doesn't MJ have kind of a specific style, or is their model really diverse?

[steady arrow]

come #🏞｜general-with-images

[lilac mirage]

guys can someone help me pls? I've some problems using the Google colab to create a new prompt

I have problems in one of the lasts steps in the part of weights_dir

[lilac reef]

The Midjourney style is a thing and really I don't want the generator to have its own style as I consider that a bad bias.

[astral goblet]

lots of models have a very OVER TRAINED aesthetic. I dont think thats such a bad thing, though i think the future will be a base model you don't have to reload over and over, with textual inversion or hypernetworks for refinement

reloads are annoying. even if they're only 1min

[left knot]

you guys know that anti AI art fundraiser that was going on GoFundMe? the one where they wanted to lobby lawmakers for changing laws around copyright and AI art?

[lilac mirage]

can somebody help me with that please?

[left knot]

I think I can get their tax exempt status revoked, but I don't think that's a good thing to do as a human. so I might just let it lie

basically a month ago I initiated something that I knew would put them in violation of IRS laws, they failed to adhere to them and now they're at risk (if I report it). I think letting it go to court isn't a bad idea, as long as both sides are fighting it. Im sure Emad is dealing with it.

[lilac mirage]

when I left in blank the part of weights_dir it says error

what do I have to write there?

[left knot]

why aren't you using autos GUI?

https://github.com/camenduru/stable-diffusion-webui-colab

find the model you like

one click run, once ran, use the share urk

the stable diffusion 1.5 might be more recommended for a newer user

[fervent thunder]

I'm with you. People are obviously going to defend Stable here more, but I just dont see it myself. Im also not about to pay 30 let alone 60 for mj but the results speak for themselves

all that can change with a new model of course. Im waiting to be wowed by whatever they have on tap

[cerulean stirrup]

guys, how can I mix hair color, any keywords?

[fervent thunder]

did you try simply 'red and green hair' or whatever

roll the rng dice and you should get some results

[cerulean stirrup]

I have tried "white-purple hair" but it not work

[forest apex]

try [white:purple:.5]

[old marten]

it already is m8

[cerulean stirrup]

how about color at the end of the hair?

[lilac mirage]

guys I have these errors

with dreambooth

oh I can't send pictures

[fervent thunder]

you can in general with images or 2 point 1

or maybe you need to not have the leaf to share idk

[lilac reef]

Why is it that the trained image of an embedding is never the same I can get locally when the model, embedding, parameters, prompt, etc... are all the same?

Not even close.

[forest apex]

not the same how?

[lilac reef]

100% always a different pic.

I am talking about saving every 100 epochs and grab the emb and use it. What I get and what it got are never even close

[severe field]

ur so hot

can i lick your stomach acid

[forest apex]

totally not haram

[severe field]

did he eat a pork or blood product

[cerulean stirrup]

how can I have a img with more than 1 object? And how can I discuss what one of them were doing?

[severe field]

WHAT

[old gull]

is the artist name generator gone in the newest version?

[old marten]

yea there was some controversy around it so they removed it

[old gull]

huh.. alright ty

[short dome]

Hi guys. I'm new in this channel. Are the Dreambots channels free so anybody can request an image generation?

[elder cradle]

I don't think so, read the pinned messages, the artists weren't in the training data at the first place, it's just the switch from old OpenAI's CLIP to OpenCLIP

No artists were deliberately removed in 2.x versions

[deft kelp]

Hello I'm new here 😊

[jovial sky]

Dream studio not working for anyone else?

[magic citrus]

hey friends, a new post about image captioning for fine-tuning purposes. i think some of you might find it interesting - let me know what you think
https://followfoxai.substack.com/p/image-captioning-for-stable-diffusion

[old gull]

so can you just stick a pruned ckpt in the vae folder and treat it like a vae? or.. is that what a vae is? whats the difference between a checkpoint vae and a .vae.pt?

[stray gull]

did anything v3 repo got deleted?

[stray gull]

so using the name of the artist in prompt also doesn't work or they only removed the function to generate random artist name?

[old marten]

just the function

[stray gull]

got it

thx

[old marten]

Although on 2.x models artists don't work as well

as they do in 1.5

[fervent thunder]

on 2.x long prompts or short prompts?

[astral goblet]

the prompts that get the creation you want out of the model to the best ability

[fervent thunder]

Ohhh soo there's no matter if I pee standing or sitting because the goal is to get the waste liquids out of the system right?

[stray gull]

btw i am thinking of using these cloud gpu for hypernet training, i was wondering if you guys got any suggestions? (i just want to try it out so anything under 10$ would be good)

[astral goblet]

uh, thats actually a good analogy i think..

[lilac reef]

Hypernet is still a PITA but I love them and have trained a few.

[stray gull]

PITA?

[lilac reef]

Pain in the ass

[stray gull]

oh lol

[lilac reef]

a 16gb card is limited to 1 batch size and it is PAINFULLY slow. It acts as if no optimizations have been done to it.

lora, dreambooth, textual inversion all are much faster than it is AND require less memory except for Joe Penna

[stray gull]

i thought they were similar thing (i am a newbie)

[warm junco]

Yes like for the WD1.4 vae. It goes into the vae folder but its a ckpt

[stray gull]

so when i said hypernetwork i had all those things in mind

[lilac reef]

No, vastly different with DB being the best.

[stray gull]

So how much vram do i need with DreamBooth?

[lilac reef]

12 to 24gb depending on which one.

[old gull]

if i use sd_vae, can i just put it in the checkpoint folder instead and pick from that sd vae dropdown?

@warm junco

[lilac reef]

the 24gb one is the absolute best there is

[stray gull]

well i only have 4gb xD (1660 Ti) so i guess i have to use cloud gpu's

[lilac reef]

Joe said he could have made it 16gb but didn't need to.

1660/1650 has a horrible time with SD to begin with even for just working.

blame Nvidia

[warm junco]

No put it in the vae folder and you can select it with the drop down aswell

[stray gull]

yup i used to get black image output when i first used it

[lilac reef]

I am barely working on a 1060 and no training of anything locally

20x0 30x0 and 40x0 are all great for this but 10x0 and earlier is missing fp16

so I get different images than someone else with a 2/3/4 card

[old gull]

how do hypernetworks work? where do i put them? @warm junco

[stray gull]

in the models folder there should be hypernetwork folder (if not create one)

[lilac reef]

settings or like me I have a quick tab for them.

[old gull]

and how do you "activate" them ?

[lilac reef]

load them in via the drop down and that is it

[stray gull]

in Settings of webui

[lilac reef]

I hate this damn channel as it doesn't allow images where this is needed to help

[old gull]

is there something i can add to the userinterface to have a quick dropdown to switch hypernetworks like the sd_vae?

its probably sd_hypernetworks huh..

[lilac reef]

yes

[old gull]

huh.. its not showin up for me when i add that , apply, and reload

[astral goblet]

the newest version of webui changed hypernetwork behavior

you use them like TI embeds now, and select them through the ui button for extra networks

[wide juniper]

Trying out 2.1 but got this:
RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 768 but got size 1024 for tensor number 1 in the list.

Help please? Trying 768x768 dimensions doesn't work either

[warm junco]

Whats the error with x768?

[wide juniper]

RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 768 but got size 1024 for tensor number 1 in the list.
Time taken: 2.46sTorch active/reserved: 9312/12726 MiB, Sys VRAM: 15286/24576 MiB (62.2%)

I already set the size to 768 x 768 but the error remains the same

[warm junco]

Maybe you have the wrong .yaml file

Make sure its the v2-inference-v.yaml and rename it to match the model name

[wide juniper]

Here are the two files:
v2-1_768-ema-pruned.ckpt
v2-1_768-ema-pruned.yaml

[warm junco]

okay but did you Download the v2-inference-v or just the v2-inference ?

[wide juniper]

I copied and pasted the code to notepad and changed the fike to the .yaml file

[warm junco]

Okay thats the wrong way

You have to right click the file and "save as"

[wide juniper]

Oh. Thanks!

[warm junco]

Np! Had the same Problem weeks ago ^^

Only the save as worked. But if not i can send you the working file

[wide juniper]

Btw, does my automatic1111 update by itswelf, or do I have to git pull once in a while?

[warm junco]

That depends on your webui-user.bat file

[wide juniper]

Restarting webui atm but thanks for the offer

[warm junco]

There you can add git pull in line 7

Then it will look for updates and apply them at every Start

[wide juniper]

Here are the first few lines

@latent stump off

if not defined PYTHON (set PYTHON=python)
if not defined VENV_DIR (set VENV_DIR=venv)

set ERROR_REPORTING=FALSE

mkdir tmp 2>NUL

%PYTHON% -c "" >tmp/stdout.txt 2>tmp/stderr.txt
if %ERRORLEVEL% == 0 goto :start_venv
echo Couldn't launch python
goto :show_stdout_stderr

:start_venv
if [%VENV_DIR%] == [-] goto :skip_venv

[warm junco]

From what ?

Thats not the webui-user.bat

[wide juniper]

webui.bat

don't wanna spam up the whole chat

[warm junco]

You need the webui-user.bat not webui.bat

[wide juniper]

I see! Here it is

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

call webui.bat```

[warm junco]

What GPU do you have?

[wide juniper]

3090

My 2.1 is giving me blank outputs

black squares

[warm junco]

okay then edit the webui-user.bat like this:
`@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --autolaunch
git pull
call webui.bat`

restart and let it install xformers

i send you the working yaml from me in #🤝｜tech-support also we should switch there to not spam here

[fervent thunder]

lol chatgpt site broke?

[astral goblet]

git pull in the launch may cause issues and its smarter to do that as you understand what a fresh pull might change

they're good to have, but the nature of the changes may break your install, like today

there was a foundational update, torch 1.13.1+cu117 and xformers updates. i had trouble when i tried it out, but with a little hammering it came out and works great

[warm junco]

Yea works a bit faster for me

[astral goblet]

i've been trying to do a custom build of it myself for a couple weeks now. i still want to aim for cu118 support but this now is a treat. really stable finally, and so fast

[frank gate]

The update today broke my stuff too, but i just deleted thr whole venv folder and let it reinstall it

And that fixed it

Deleting venv seems to be the fix like 2/3 of the time

[steady arrow]

I got the 10$ plan but already in 4 days am 33% thru my credits lol

[vestal dew]

It gets expensive

better to do it locally

[stray gull]

can you elaborate ? i cant see any ui button for hypernetwork

[warm junco]

@stray gull Its the third button under the generate button

[stray gull]

oh

got it

[proud nova]

I am using a T600 NVIDIA Quadro GPU

[rough schooner]

Hey. Does someone know how to get hugginface diffusers run locally? I tried the installation guide on github and hugginface but it won't work

[warm junco]

You want to install table Diffusion?

[rough schooner]

yes i would like to learn how it works (better said: how to make a website with the stable diffusion) @warm junco

[vestal dew]

You run it on a computer, set it for remote access

[rough schooner]

But how? I can't get it to run. It is a bit overwelming for me to be honest

[vestal dew]

There are one click installers

[rough schooner]

rly?

but not web ui's?

just the "backend"?

[warm junco]

Wait i send you a tutorial link

[rough schooner]

Thank you

[vestal dew]

https://www.reddit.com/r/StableDiffusion/comments/zpansd/automatic1111s_stable_diffusion_webui_easy/

[warm junco]

@rough schooner start here, thats for a local Installation of Stable Diffusion Webui.
https://m.youtube.com/watch?v=VXEyhM3Djqg

[rough schooner]

thats a finished web ui. i would like to create one by myself

[vestal dew]

you have an Nvidia card?

[rough schooner]

yes

[vestal dew]

OK

[rough schooner]

the webui i am running right now

but thats to less to learn

XD

[vestal dew]

I see

[warm junco]

Ahh you want to make a webui by yourself?

[rough schooner]

yes

[vestal dew]

When you finish it, please PM me...I'd like to see it

[rough schooner]

i would like to understand how a frontend talks with a backend

backend = stable diffusion

[warm junco]

Would suggest to look at the githubs of Automatic1111 webui and invokeai

There you see the code

[rough schooner]

i did but i did not found any html file. i don't know where this web ui comes from XD

[warm junco]

But mostly its done with Gradio, thats the frontend

[vestal dew]

Does it mention who created it? Ask that person

[rough schooner]

i think this person has to much requests

[vestal dew]

You could pay me to explain it to you.

[rough schooner]

ah okey. is there no html?

[vestal dew]

Better yet, pay someone who understands it to explain it to you

[rough schooner]

if you really can do it i would pay.

[warm junco]

I dont know much about it. You should look at the Gradio site

[vestal dew]

I cannot, but start with the creator. Offer to pay.

[rough schooner]

probably that would work

is it gradio.app

[warm junco]

@rough schooner there you go:
https://www.gradio.app/docs/

[rough schooner]

but this should mean i have a console stable diffusion running?

[warm junco]

Yes

[rough schooner]

my stable diffusion won't run in the console

only the web ui form atomic1111 or what's his name

[warm junco]

You need the cli version

[rough schooner]

https://github.com/Stability-AI/stablediffusion not this?

[warm junco]

Idk

[rough schooner]

Okey. But thanks for trying me to help 🙂

[warm junco]

It runs in a console. It only opens in the webui

You can also use Automatic1111 Webui and delete the gradio stuff and make one yourself? xD

[rough schooner]

i didn't know about gradio

where is that located?

i thought it was some html or react app

https://ibb.co/nzKgzX0 thats the repo of automatic1111

[lavish current]

Well that's very annoying, looks like Anything V3 cloud model finally blew up and now I can't use any of the forks/mirrors

https://huggingface.co/Linaqruf/anything-v3.0

Hopefully it returns soon, that was the one model I used for everything, not even limited to just anime

Dang. Waifu is alright but still not nearly as good

Guess all I can do is wait now. Oh well

[proud nova]

https://web.archive.org/web/20230122174244/https://huggingface.co/Linaqruf/anything-v3.0

Try this, at least for now

[elder cradle]

Nah, the repo got deleted

The author stated that himself

You can try anyv4, or here's a pristine mirror https://huggingface.co/ckpt/anything-v3.0

[mental sparrow]

does anyone know how the automatic1111 implementation is so memory efficient?
when i try to load both the UNET and VAE onto the cuda device, my 4gb gpu ram blows up. this doesnt happen with automatic1111 since the webui runs (with a few hiccups here and there) pretty well on my laptop.

[lavish current]

Thank you guys

[proud nova]

Ur welcome

[fervent thunder]

does anyone know a good blog or news aggregator that focuses entirely on AI and related topics?

[fervent thunder]

it can run in multiple modes, full precision and half precision

the latter is smaller in ram

i guess itll be that

i think there was also a way to load up only parts of a modell or something i could be wrong tho

[mental sparrow]

alright ill check it out

[echo tulip]

Is there a method or solution to solve the hand issues? I tired to inpaint it and it still generate horrible looking hands. I have included the usual embeddings that reduce the chance of bad hands but it is still pretty hard.

[brisk bloom]

if its a picture I want to keep, I edit the hand shape in photoshop then do img2img and it usually completes the details

[proud nova]

Not yet. AI has problem with hands and will have for some time. But I guess it will be okay in a few months

The only solution is to hide hands behind back or in a pocket

[wraith sky]

when using automatic webui, when using models that require words such as mdjrny-v4 style in the prompt (for example the first openjourney model) does automatic need those special activation phrases/words in the prompt or is it smart enough to just use that model when selected via the dropdown?

[warm junco]

There is an extension where it selects the word when you switch

[bleak matrix]

Good morning, fellow friends! How are we today?

[bleak matrix]

Well. You can get some really good hands with SD

Posing them in certain positions does make it easier, for sure

Keeping the hands apart so helps

I've been working on prompts for beautiful hands but I'm not done with my research

But a lot of the time, iteration can help

I've found solutions for a lot of the issues. It's not 100% perfect, but it can mitigate things

[echo tulip]

you mean img2img? In txt2img I tried multiple different models, prompts and embeddings trying to create more accurate hands, however it is still very hard to consistently generate them, even after inpainting it's still quite difficult. Maybe we can only wait till SD figure out how to generate accurate hands

[echo tulip]

Really?

[bleak matrix]

Just with SD in general

Yes

[echo tulip]

is it with plugins or embeddings?

prompts usually are very limited

[bleak matrix]

Neither lol

[echo tulip]

hahahha models?

[bleak matrix]

Nope

[vestal dew]

what are the solutions?

[bleak matrix]

Part of the solutions are in the friendly guide

[vestal dew]

friendly guide?

[bleak matrix]

Though I haven't written the key words, to be fairrrr

https://docs.google.com/document/d/1aHJ9RBt_vlCwJQBVUUsb7VghKB-wynv7WGTxm9ozL1k/edit?usp=drivesdk

So, I'll just

"Share with the class" here

[echo tulip]

jesus you wrote all this 1/

!?

[bleak matrix]

Lol, you should see the 400+ pages I have written

Bahahaha

[echo tulip]

very impressive tho, how long did this took you to make?

[bleak matrix]

I don't know cuz I didn't count

[echo tulip]

please don't tell me chatgpt was involve ahhahahahah (just kidding)

[bleak matrix]

So, the thing with hands, especially up close, or with a lot of anything SD

[proud nova]

I see a writer potential here

[bleak matrix]

Is that you get that melding and that fingers turn into OMG MONSTERS!!!!

Oh, I am a writer.

Lol

[proud nova]

I knew it!

[bleak matrix]

Can't say my guides are all that lmao. But I am cuvjchhklj

So the simple solutions are to separate the hands, on a base level, regardless of the pose

And yes, you can still get good looking hands if they're together

[echo tulip]

so right now you are still working on adding the hand stuff into this document right?

[bleak matrix]

And closeup

Yeah

So I'm just saying it in chat lol

[echo tulip]

well i guess I will be looking forward to that, as there are not many useful tutorials out there.

[bleak matrix]

You can get a lot of beautiful women/stellar hands when you place them near the face

They look so good

Certain poses are just better than others

The thumb and fingers can be difficult to get in one go. Holding some objects are better than others, etc

[vestal dew]

So, where do we start? —------ – —------ —---- - - - —--------- – - – —-------------- – - —---------- -------------- – - - - - —--- - - - - - - —--- —- - - —---- - - - —--------- – - – —------ -------- – - —---- ---- ---- ------------ – - - - -

[mental sparrow]

@bleak matrix you could start a youtube channel with this much knowledge

maybe hack the next big thing in AI

[vestal dew]

Is that just filler, or are there supposed to be words there?

[bleak matrix]

I have one. That I never upload to 🤣

[bleak matrix]

There will be words. It's mainly for formatting because that's part of the length, and part because Google Docs is DUMB ON MOBILE.

You can't indent properly -_-

[vestal dew]

I see

[bleak matrix]

So, holding hands, cups, certain objects are better.

[mental sparrow]

@bleak matrix have you explored the codebase or is your document related to prompting only?

[vestal dew]

for me, the hands and feet are a horror story. The better the rest of the image, the more likely the hands and/or feet are to be mangled or hidden

[bleak matrix]

But like most things SD, hands are situational.

[echo tulip]

so basically use prompts to separate hands and poses that are holding something to get better results?

[bleak matrix]

I am a game developer--my sole focus isn't on prompting, but I don't have the time for technical support--my own research goes well beyond prompting lol

[bleak matrix]

Heh, oh, are they ever, lol, but like with the rest of the solutions

They are in a similar vein

Of problem solving

So, manicured hands, with long nails, they do very well

[mental sparrow]

how does one go about understanding the codebase and using it? i'm sick of trying to scrape together snippets of code and failing

[bleak matrix]

Your medium--photographt, watercolor, can also have some influence on how well it looks

[mental sparrow]

the automatic1111 webui repository among others is too intricate for me to understand :( but i want to

[vestal dew]

just tried adding "holding a cup in each hand"....left hand has 7 fingers, right hand off screen

[bleak matrix]

Yeah, so

[warm junco]

Try (peace sign:1.3) without holding

[bleak matrix]

Whee, Discord freaked out on me for a sec

[vestal dew]

next image...right hand inside cup (hidden), Left hand a fist with one big fat finger

I am trying it CS1o

[warm junco]

Worked well for my Cyborg girl

[vestal dew]

next image...lips look painted with lipstick - not happened with this prompt before adding holding cup - left hand is a thumb and one big finger, like a mitten

As Ive seen before, once it does something - in this case the painted lips - it keeps doing it

I generated about 300 images before this....no painted lips. Now every one has painted looking lips

[echo tulip]

it's just that I am confused why only the hands and feet's are the ones that have the most issues with? And everything else seems quite consistent.

[vestal dew]

Trying now with the peace sign

[vestal dew]

Could it be that a lot of the training with people was portraits?

[west grotto]

I am trying to train the AI on a model I made in Blender, but it seems to not be picking up the insect/multiarms/human torso anatomy correctly. When I use the embedding I made it starts off as a bird and then goes to an insect with no arms. It looks nothing like the model or what I'm going for. I made the embedding by taking multiple angled shots of the model and trained it at 100,000 steps.

[warm junco]

Its because on every image the hands are not the same.

[vestal dew]

Also, SD has trouble with depth and perspective. I wonder if using depthmaps helps

[west grotto]

Is their any way to make it work better?

[vestal dew]

Did you test it to see if it makes the thing when only the name is the prompt?

[echo tulip]

but same could be said with everything else, not every pose are the same, not every expressions are the same.

[bleak matrix]

(I am dreaming some examples for you guys)

[vestal dew]

first image with peace sign, right hand has fingers about double as long as they should be. Left hand has six stubby fingers with a super long finger pointing up from a knuckle

[warm junco]

Haha xD

Which model and have you enough negative tags?

[vestal dew]

Peace sign is adding extra long fingers on top of mangled hands

and it's still making the lips painted in every image

model is eldreths lucid mix

negative is "deformed hands, deformed feet"

[west grotto]

I just tried it and no such luck. Sometimes it's close but others it's a blurry blob. Does the checkpoint I have loaded make a difference in training?

[main sparrow]

Sometimes embeddings can't recreate what you want it to if the model doesn't already 'know' about it. Sounds like you might need dreambooth or lora training to be able to add new concepts to a model.

[vestal dew]

I did a textual inversion training. The name itself makes a perfect image. When I add it to anything else, I get boats in a canal

[main sparrow]

Embeddings really just hunt for tokens within the model to describe the images you put in. That's a wild oversimplification, but that's basically what it does and why thyer're so small.

[west grotto]

I'm very new to this and am just trying to get some art for a creature that isn't very common I guess. How do I do anything else to better this process?

[bleak matrix]

So, first, putting on object inbetween is a good way to keep the hands apart.

My prompt for this (using the bot) was

dream prompt:hands well, professional photography cfg_scale:9 number:4 negative_prompt:hands poor, fake nails, feet, toes, inflatable hands

[vestal dew]

using what bot?

[bleak matrix]

Here in the chat

[vestal dew]

I didn't know it makes images

seems hands and feet do better with txt2img than with img2img

[west grotto]

How would I do any of these processes? I'm very new to this and this is the first time trying to train anything.

[bleak matrix]

Sometimes adjusting the cfg higher and lower as you iterate can help.

Now, hands well and hands poor

Like hands drawn well, or hands drawn poor

The reason the hands is first

Is because I'm saying, "Hey, SD, these hands are SUPER important and this is what this sentence is about."

Mileage may vary, depending on your prompt

Well and poor are opposites

Just like square and sphere are

[vestal dew]

who would want poor?

[bleak matrix]

Poor is a negative prompt

[vestal dew]

ah

[bleak matrix]

(You can read the negative prompt section for more about square and sphere in the guide.)

Combining these together--positive and negative--can help

As I mentioned, manicured nails do very well

However, sometimes, they end up pointy

And glittery

Like crayons

So you can use manicured nails as a positive prompt in conjunction with model poses, blah blah

And you'll get some great stuff

Even with just the bot

[vestal dew]

I just tried hands drawn well and hands drawn poor negative...this is with img2img...hands are a horror

[bleak matrix]

But you can also use fake nails as a negative prompt

[vestal dew]

also, it's still doing lipstick lips...nothing about that in the prompt.

[bleak matrix]

laughs Yeah, it's a combination of settings and words

[vestal dew]

about 301 images...it started adding that, and now it's every one

[bleak matrix]

Lipstick lips?

Using fake nails as a negative prompt will lead to more natural looking nails.

[vestal dew]

Ues

With this prompt and image I generated about 300 images

I added "peace sign" , now your suggestions....now every image has lips that look like wearing lipstick

none of the earlier images had it

[bleak matrix]

Probably because of peace sign

Or what that might mean in relation to it

[vestal dew]

I got rid of peace sign

[bleak matrix]

I have a lot more prompts--this is just a sample

(Specifically more routed to the bot)

[vestal dew]

right

I suppose numberof iterations change things too

[bleak matrix]

(I do have ones not for the bot)

Yes

Significantly

[vestal dew]

sometimes need more, sometimes less depending on the model, prompt, settings

[bleak matrix]

Yeah

So, feet, toes, there's a lot of issues with those, some of which I won't get into, some of which are already in the guide

But using them as a negative prompt can sometimes cut down on that

Now, a very funny thing

That SD does with hands, besides length, and centipede fingers

Is that sometimes you get boneless fingers

Fingers with three fingers

[vestal dew]

yes, I've seen that

[bleak matrix]

Fingers with flat fingertips

[vestal dew]

hands with two sets of fingers

[bleak matrix]

Etc

So, besides poses

Is a negative prompt of cartoon, or inflatable hands

This is because of gloves or rubber gloves

[vestal dew]

have you worked with depthmaps. I wonder if that helps with limbs and hands/feet

[bleak matrix]

Here are some of my fave hands (couldn't find my fave--it's around here somewhere) from during the process of learning more about them.

Unfortunately, the Lupus fatigue is getting to me and I have to go now

These are all just the bot

The rest I will keep for the reveal

[desert bloom]

Without trying to start a debate on ethics, has there been any attempt to train a model from scratch with just images that the artists/photographers agreed for use for ai?

I havent heard of any group trying this. The next question being what is preventing people from trying this? (Not judgemental but just curious) like what are the barriers?

[warm junco]

I think the Barrier is to find every artist to ask when the images are uploaded anynomusly on image boards or other sides

You need a lot of pictures to train so no artist can give you like 100.000 images

[desert bloom]

There surely must be a good amount of artists/photographers willing to let their images used by ai, even if it's not enough images for a good model right now.

But technology keeps on improving and will need fewer images to understand concepts

[warm junco]

Yea when it improves maybe we need less images to train

But i dont think there are a lot of artists who say sure go ahead. When you ask them to train with their images

[desert bloom]

Still, you could always use the outputs of a simple model to improve it further

[upbeat birch]

They will be more willing if they get paid for their contributions.

Gotta pay the bills.

[desert bloom]

That's... actually not a terrible idea? Not in the "every image generated gets a royalty" thing but something more like "I'll pay you a small amount if you let your work accessible for ai use for ANYONE

Tho that's a pretty odd request and idk if that's even something you can buy from other ppl?

I'm just thinking that there seems to not have been much of an effort to organize images for who would be okay with ai use and who wouldnt. It's not legally necessary (and I hope it stays that way) but a lot of people are not willing to adopt or like ai for a problem that seems to have solutions, albeit cost prohibitive ones.

[cerulean stirrup]

when I'm trying to train my model

Exception training model: 'No executable batch size found, reached zero.'.

How can I fix this

[desert bloom]

I feel like we could make a effort to at least build a dataset, if not for making a model for right now, then for a future time when the technology will require less images to be trained

[old marten]

the real error is above that, look for another error

most times its because of cuda memory error

[stable pawn]

I use Automatic1111 so I see when my images are generated, they look good until the last step when it look like extra saturation and contrast is added, if I use loop-back the images are getting darker and darker for each step. Why is it so and and ideas I can test to avoid over saturated images?

[vestal wigeon]

I am new to stable diffusion

Which one shall I start using , SD 2 or SD 1.5 ?

I am a bit confused

[stable pawn]

I still use 1.5 as standard, but then added 2.1 as an option.

[vestal wigeon]

Which is the best way of using it? To use it locally by installing or is there any other way of using it?

[stable pawn]

I had some trouble with getting 2.1 to work and think it is not as good as Darkstorm2150 Photo realism https://civitai.com/user/darkstorm2150

I am self rather new and test, so I cant say "what is best".

[fervent thunder]

in my opinion, then 1.5 is better than 2.1 in everything but a few details; photo realism with slightly more details on a slightly larger default aspect ratio.

[stable pawn]

Yes, I forgot to say that, I seldom create "photo realism", I often use SD for concept art and to spit out ideas. for that 1.5 is great for me and what I often use.

[vestal wigeon]

I am a fan of photo realism, so which shall I pick?

[mortal inlet]

is there any plan to add paypal ? i stuck after using it for maybe about 30 renders ? ...it say<s free, i bought 2$ 10$ version but i dont have any credit now, i also dont have visa or mastercard. thanks

[fervent thunder]

photo realism as not a style of detailed painting but actually photos? then I'd say 2.1

[vestal wigeon]

@fervent thunder Yes actual photos

Thanks

[fervent thunder]

just be ready not to get the same results you see from others. Using the right prompts are an art in of itself :)

Anyone good with img2img? DM to make some money!

Have a logo and need variations

[vestal wigeon]

how about the F222 model? I heard its good for creating real life like human renders

just was scrolling across some articles but not sure

[cold siren]

isn't it quite old at this point

there are probably some newer & better models available for that

[mortal inlet]

i use blender, i rebuild all the AI generated pictures i made before, so phantastic to work !

[cold siren]

do you use the generated textures somehow?

[empty dawn]

how can one go about taking the photos i took of my city, and using the AI to turn it into a cyber city etc?

img2img correct?

• how could i do this with my photos which are taken with an iphone, therefore the pictures are not 1:1. its more in the shape of 16:9 (it is not actually 16:9 i believe, it is just not a squared 1:1 shape)

[vestal dew]

you can make images in the same ratio

photoshop or some other program like it can tell you the ratio it is

[vestal dew]

why don't you do it yourself?

[slim relic]

My outputs has really muted colors. Is there a way to improve this?

[main sparrow]

there's this weird concept of people paying others for what they don't have the time or expertise to do. I think it's called 'work'

[vestal dew]

wow

[main sparrow]

weird, I know. 🤣

[vestal dew]

When I work, it's doing things others can't do, either at all or as well as I can

[main sparrow]

There are going to be a bunch of menial tasks that AI artist can do to earn extra cash. Tweaking logos is one. many people don't have the time/expertise/desire to set up auto1111 or whatever to get their image worked. on. but it's totally worth paying people to do it.

[vestal dew]

I just did variations on a logo - came up with some interesting variations. However, I don't think any of them would be a final logo...just ideas

[main sparrow]

you'd be surprised

people are doing face training for dreambooth. easy task. getting 20 bucks in fiverr for it

[vestal dew]

ah

I should start doing that

I could make the money to buy my new system

[main sparrow]

well they were, anyway. I think there's a phone app that uploads your images to a CIA or foreign equivalent but gives you a cartoon back for like 2 bucks now.

[warm junco]

You need a vae for the model

[slim relic]

aight lemme try

thx

[vestal dew]

how would that be face training?

[main sparrow]

well they do the face training for you. but they also sell everything they can about your personal data to everyone they can.

[empty dawn]

can anyone link me to a working version of stable diffusion for AMD that includes img2img? ive currently got one but it only supports Txt2Img, and using img2img on normal diffusion slow AF cause it uses my CPU, 20 min per gen at that rate

[vestal dew]

ah

[empty dawn]

how can i stop img2img from changing the picture so heavily? im wanting to modify the image of my city i took, without completely changing the city

example in #🏞｜general-with-images \

[main sparrow]

turn down denoising strength

[empty dawn]

yea someone helped me ty

will batch size or batch count result in multiple generations out of one?

[empty flume]

can some one let me use his local host Stable diffusion link ?

i just want to try it, i have a weak pc

it will mean a lot for me

[empty dawn]

for txt2img?

if i can figure out how to set shark to public ill let you hop onto my gpu for a few minutes

[empty flume]

ye

[empty flume]

that will be a great minutes to spend

can you try this prompt:

A mysterious boy with snow-white hair reaches out boldly, grasping a gleaming golden sword from a colossal stone pedestal. As he holds the sword aloft, a brilliant light radiates from the blade, lighting up the ancient chamber in which he stands. The boy's expression is one of determination, his jaw set and eyes focused on the task ahead. He takes a deep breath and with one swift motion, the sword is his. The sword is said to be an ancient relic, passed down from ages past, and it is now in the hands of the mysterious boy.

[echo tulip]

for inpainting to remove extra body parts any tips on this?

[empty flume]

or this:
A mysterious boy with snow-white hair appears before an ancient chamber, illuminated only by the light of an intricate stone pedestal. He reaches out boldly with a determined expression, taking a firm grasp on a gleaming golden sword atop the pedestal. As he holds the sword aloft, a brilliant light radiates from the blade, illuminating the chamber and plunging the shadows into obscurity. It is said that the sword is an ancient relic, passed down from ages past, and with one swift motion, it is now in the hands of the mysterious boy. His gaze is resolute as he readies himself for the task ahead.

[empty dawn]

yes sorry i was changing out the propane canister on my heater

[pulsar canyon]

Yo

[lone siren]

Hey, absolute newcomer here. Can I run SD on my Ryzen 5600X + 32GB RAM + GTX 1060? Can it use my 1060 or do I need an RTX GPU to use it?

[empty dawn]

#🏞｜general-with-images

[still remnant]

I made a spreadsheet to help myself with this: https://docs.google.com/spreadsheets/d/17LXM_GgHib7W9dj88j7pqgenKsohNht-9f-CaV-yp9w/edit?usp=sharing

[rigid surge]

Is there a way to reset your stable diffusion settings in the A1111 app?

[old gull]

make a new directory and re git clone 😄

[rigid surge]

Ye honestly might have too, thanks

[old gull]

no idk, there might be.. sry i felt like being that guy..

[old gull]

this steps animation extension is pretty neat.. is that what people are using to make the cool animations ?

[viscid spoke]

Yeah it'll work fine on a 1060. Not the fastest but you shouldn't hit any hurdles

[old gull]

reasons to use '512-base-ema.ckpt' vs 'v2-1_512-ema-pruned.ckpt' as a vae ?what does pruining really do if anyone could explain please.

do you want pruned for generation, and full for when youre using it to do training?

[lofty sage]

hey

[rain fulcrum]

hello.
Does anyone know how to create lensa like portraits using stable diffusion?

i'm searching for a guide or something but I can't find it

i mean..
I could use a img2img generation on a photo of someone.
But usually I loose the face of the person in the process

[astral goblet]

the depth guided modes can add a lot of consistency, where a depth map is estimated and used as a mask, i haven't played much with it yet though so i'm not sure how to advise use of it. Results i've seen look great though!
low denoising values help too

[odd geode]

interesting

[lone siren]

I see many "flavors" of Stable Diffusion (versions, UIs, packages, etc), is there one that would fit better for someone mostly interested in creating artistic representation of pictures (making a picture look like a painting)?

[old gull]

the easiest way to use stable diffusion on your own terms is the automatic1111 webui, from there its about what models, embeddings, etc you use to get what art style you prefer

[rain fulcrum]

This vídeo may help:
https://www.youtube.com/watch?v=9Nu5tUl2zQw&ab_channel=OlivioSarikas

[old gull]

some artstyles the very basic SD models understand and you just need to include it in the prompt, such as "classical painting", "80s anime style", etc

if you want something more specific theres models trained on all kinds of art styles, such as valorant diffusion for the likeness of riot games valorant characters, theres disney models, etc

arcane diffusion was a petty cool one lol

[proud nova]

Question, Is there any way to use GPU and CPU performance for image generation at once?

[old marten]

nope

two different architectures

[proud nova]

Makes sense

[surreal kiln]

I can't post any, what happened?

[thin widget]

Is there a list of artists that consent to being in SD?

[astral goblet]

no

[signal nest]

communication between CPU and GPU would slow it down more than the CPU could help @proud nova

perhaps if you had a very powerful APU and a very weak GPU it could be worth it

[slim relic]

After I create an embedding, how can I use it to run images?

[old marten]

put the .pt file in the embeddings folder and add the keyword to ur prompt

[slim relic]

whats the keyword?

[grim aspen]

how would one have to promt to get a design from a round object like a soccer ball but as a quare flat texture?

[old marten]

whatever you trained the embedding with

[rain fulcrum]

Can someone explain me why 2.1v looks much worse than 1.5v?

same prompt and img to img generation:
1.5v https://media.discordapp.net/attachments/1066549893851783199/1067540213934604288/grid-0191.png?width=553&height=553
2.1v https://media.discordapp.net/attachments/1066549893851783199/1067544224146391171/grid-0192.png?width=553&height=553

[crisp zinc]

anyone try switching from linux to windows for performance imporvements?