🏞｜general-with-images

[smoky oak]

Interesting

Very curious to see how that works in practice

[glossy herald]

it feels around the same kind of uscale I get from their UI

(a new feature they added recently, when downloading a pic it lets you upscale it)

[smoky oak]

So basically not that big of a deal?

[glossy herald]

it was quite clean

and 0.2 credits

so quite ok in my opinion

[smoky oak]

Interesting, interesting

[glossy herald]

only x2 when I tried it though

[smoky oak]

Ah, that suggests it was the low quality version in that case

I have about 40 minutes to kill

Let me take a look and see

oh god, its its own own huge checkpoint

hmm

ohhhh, so you use it as an img2img model it seems

hmm

Yeah, I am way less interested in this

bummer

oh well

Real-Gan also seems really old... I am not sure what this is all about

oh.

Real ESRGAN is just R-ESRGAN we have had for months

I am genuinely confused what the point of that announcement was for-

I wish 2.1 wouldn't have been such a flop

this is base model?

if so, those look good for base model

Not bad for pure base model

sure, there are lot of artifacts, but the fact that they don't look terrible is a huge step up form normal base 2.1 lol

It still has some base model issues, but for base model that looks damn good honestly

Very nice

imagine how good it would look with a good model if it already looks that solid

getting base models to not look terrible really is a feat of its own lol

[neon crystal]

Reptilian Aliens in India Series 👽

[smoky oak]

It's not a subjective thing. It was objectively because I was using JPG for my process, and it caused the issues seen

JPG works fine for results, but not for upscale

As you can see, when you use JPG for the upscale process, it causes artifacts

vs PNG

I only found out that JPG was the issue around that time

[kind quartz]

looking for best carrot tree 😄

[smoky oak]

@kind quartzfirst try result lol

[kind quartz]

i got them many 😄
it is simply funny

[smoky oak]

I like messing with absurd things sometimes haha

Yeah sure but like, I'd like you take a picture of a snow leopard in the middle of a desert lol. Reality is boring and limited, AI isn't

I'll hold my thoughts lol

[wispy nest]

I mean, why is it okay for you to pay for a service, but not okay to use your own gpu?
You can do blood or whatever if you install it locally

[smoky oak]

I have nothing to add to that

[wispy nest]

• you'll get alot more tools to play with...not just gen

[smoky oak]

I think this one came out good

[kind quartz]

i have him on desert but still there is snow under leopard

[wispy nest]

Because you're losing alot of potential?
You can't do alot with it.

[smoky oak]

I mean, if you wanna spend money for ease of use, thats valid, but it is far more limited than just running locally

[wispy nest]

and idk how much you can actually generate before you'll reach the price of a gpu spent on it...

like 3060 or something

[smoky oak]

you'd reach that pretty fact

But mine looks good, so why the hate?

Its a little pretentious lol

@kind quartz

Just messing around

Sandy Snepard

@paper salmonYou didn't answer what is so wrong with my flower image. Why does it matter if its real or not if it looks good?

So if it did, it would be better?

[kind quartz]

@smoky oak rather spoilers it, problem i have is i got prompt in filename

[wispy nest]

I can tell those fingers are real

[smoky oak]

My point is, if I didn't tell you that it was made up, and you just saw it and thought it was real, what's the big deal?

Because as of right now, you're only taking my word that it's generated, you don't actually know if it's generated or not lol

I guarantee you the only reason you feel that way is because I told you it was generated, and you're just taking me at that face value.

I could have said it was a photograph, and you probably would have believed it too

You have a huge misunderstanding between the power of a base model versus a fine tune model

SDXL is a phenomenal base model, however a very well tuned 1.5 model can still handily beat it in several places. The only true excitement of SDXL is the fact that we will be able to fine tune it to become even better than 1.5 is now

Also, you're not able to use any form of fine tune or specialized model on dream studio, which in and of itself is a huge drawback

[wispy nest]

Dreamstudio is limited compared to local install tools, thats what I meant.
You get lots more options to play with , extentions, tools...

[smoky oak]

Models, LoRA's, textual inversions, dream booths, control net, deforum, pix2pix, megabatch processing, dedicated upscale workflows, unlimited scripting capabilities, and so much more

[wispy nest]

yea I couldn't be bothered to list 😄

[smoky oak]

Just to clarify, I'm not saying that there's anything wrong at all with using dream studio, it's a fantastic service, with a reasonable price, and makes a lot of sense for people who do not have the hardware to run stable diffusion on their own. However, it is by no means even remotely as capable or feature packed as a local stable diffusion installation

[cursive otter]

#imege smart boy corton profile picture

[smoky oak]

Imege

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[smoky oak]

Also, by the way, that flower I sent is a real photograph, not a generation.

So congratulations, you just downplayed the artistry in a photograph, because somebody said it was generated... Your judgment of that image was completely based off of the way that somebody presented it to you

This is a real photograph that I took with a real DSLR, with my real hands, in reality. But because I even so much as suggested (didn't even state) that there was a potential that it was generated, you had a problem with it@paper salmon

[cursive otter]

Smart boy corton profile picture

[smoky oak]

You can't generate images in the server

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[cursive otter]

#1072220168534642768 smart boy corton profile picture

[smoky oak]

You cannot generate images in the server.

[cursive otter]

Whare i genrate image

[smoky oak]

There is no way built into this server to generate images.

Read the messages I sent multiple times now, they explain how to do it

I'll send it again

[cursive otter]

Ok

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[smoky oak]

sighhhhh

You get one little taste of an extremely limited beta preview model, and now suddenly you're the king of generations

[wispy nest]

To be fair...we'll get SDXL on our local install later on too...soooo....

[smoky oak]

You do not support the developers. They do not make money off dream studio. You pay the electricity cost of what it takes to run their servers

Thats fair

The community by creating models and adaptations for people to use for free, as well as guides and probably one of the highest levels of support in this server, to name a few

oh and also

1.5 can do it just as well

SO ANYWAYS

oh come on man lmao

You had to buy a device that can use the web to access discord. Not exactly free!

It is free, as in no additional cost

have a decent PC already? Its free to run

What?

up until this point, I have been light hearted. I could really tear into all of your SDXL gens and point out the issues with them, but I am not like that

Dreamstudio costs wayyy more to run than local install

WAYYYY more

Also, here, hes happy now

Also, we have access to additions dream studio has no support for, which is another massive bonus to local running

Also, fun fact, a lot of native americans did not smile in photographs throughout history as it is thought in many native american cultures to be capable of stealing emotion or even the soul of the beings pictured

ok, so SDXL's big thing is it can do "Female Native Americans?"

Thats the selling point?

proceeds to send images with tons of issues

I see

Yes, we get it, its a cool little base model, but SD 1.5 can do all of these things too lmao

[open edge]

an akita inu breed dog mixed with a dachshund breed

[smoky oak]

@paper salmonI get what you are saying, but all of these images are just showing that you probably never messed with a properly tuned model

These are very "ok" images

. . .

again, we know SDXL can do decent looking images

[wispy nest]

We all know SDXL is a good base model, yea, that's why we're all waiting for it and 3.0 base model to release so we can finetune it and use more tools with it

I just don't get why is it okay for you spend money on credits, but not on your gpu

[smoky oak]

1.5 can do just as good, if not better lol

Are you done yet?@paper salmon

guess not

[wispy nest]

yea there only thing there old models can't do was text so far, everything else wasn't that impressive

[smoky oak]

@paper salmonI know things that SDXL can't do

here is a SD 1.5 image

good luck recreating it in SDXL

[wispy nest]

oh...right....I know...

[smoky oak]

Noise offset is just one of the many things SDXL can't do

Keep coping bro

Here, I have an idea

Try and generate toothless from how to train your dragon

see how well that works

I'll wait

Or an Opossum

generate an Opossum, I wanna see it

[wispy nest]

some older gens...

[smoky oak]

Or a blue person from Avatar, lets see that too

Genuinely the only thing SDXL can do 100% over 1.5 is text, and for that I will agree, SDXL slaps

but then again, we have access to controlnet which means we can make infinitely better text anyways

[wispy nest]

I think SDXL failed my "ripped, torn" clothes test too

[smoky oak]

Alright, lets get this over with

[wispy nest]

I'm not doing gens specifically to prove something on my 1050ti lol, maybe when I'll figure out collab situation...

[smoky oak]

There, happy now?

can we move on to something of substance

@paper salmon Let me know when you get SDXL to generate toothless from HTTYD, an opossum, or a blue person from Avatar. I'll wait

The truth is, there are genuinely things that SDXL can do better than most 1.5 models, you just haven't shown a single one of them lol

Oh, and here is a rule of thirds image of 2 native american women, for kicks

what lmfao

oh that's right, you have to pay to gen... Damn 😦

Anyways, I am sure that was a cover up to hide the fact that SDXL can't do any of those things I listed

oh crap, BRB

[wispy nest]

???

Didn't you just ask for it?

until few days ago I was using google's hardware for gen

well, for like 99% of my gens

was doing quick tests on my own gpu

I think if I paid for as as for dreamstudio I'd probably could buy 3060 already for spent $ 😄

[smoky oak]

Well let's see, my PC cost $1000 6 years ago, and I've generated 98,000 images... So how does that work out? Also, running the math, my GPU uses about 190 watts when generating for 1 hour straight. I can do about 1 image every 2 seconds when speed generating, which is 18,000 images an hour. Let's say that I were to do that consistently for 1 hour, that would be 200 Watts, each day for an entire month, which would come out to about 6,000 Watts, which based off of where I live would come out to around $0.40 a month...

Which one you convert that down turns into about $0.0001 per image.

Also, when factoring in the cost of my hardware, it should also be noted that my computer has made me well over $3,000 in the time that I've had it, so that cost is negligible

So yeah, running the software that is completely free on hardware that I've had for multiple years prior, works pretty good for me

So yeah, safe to say that I've generated a hell of a lot more, for a hell of a lot less

And that's also while having access to all of the exact same models that you have access to, as well as literal thousands that you don't, as well as dozens of different features that you don't...

So yeah, I think that works out pretty reasonably

I'd love to see your response where you try and figure out in what way I've managed to pay more than the nearly 10 cents an image you're paying for

Oh also, I should also note that we are planning on getting solar panel sometime soon, of which the electricity for my generations will quite literally be entirely free

[kind quartz]

wiser will retreat 🙂

[smoky oak]

My mother and I, we just moved into a new house

If we're talking about screens, two of my screens were hand-me-downs, one of them was a birthday present, my GPU was a 3060 TI of which I got for $400 MSRP, my CPU is a 7700K of which I got for $160, my motherboard is an Asus tough Mark 2 Z270, of which I got for $100, I have 32 GB of RAM that I purchased for about $110, and a 1 TB and VME SSD, as well as several others that have absolutely nothing to do with the function of my computer, but rather other things that make me money...

So yeah

I do live in the USA, yes

Oh yeah, there are lots of countries where graphics cards are outrageously expensive

Oh Jesus, I didn't realize it was THAT bad

Wow...

@paper salmon anything else to add?

[wispy nest]

3060ti in Russia is basically equivalent of middle class average salary , even higher a bit

[smoky oak]

The 3060 TI I did not get 6 years ago, I purchased it two years ago, everything else is 6 years old

I have a 3060 TI founders edition that I got for MSRP launch price on December 24th 2020. It launched at $399, and I bought it for $399

1.5 is not out of date software

whatever man, I have been engaging in this conversation cause I find your constant subject changing, goal post moving, and coping to be quite entertaining, but this is just causing me physical pain now lol

quick reminder that nobody here said SDXL isn't good, you were just the only one adamant that it's "better than 1.5"

which is entirely false, BTW

Question, are you Native American?

Also, you, posting an image with the same chieftain head dress

But yeah, are you Native American?

Oh for sure, 1.5 base is ass, but 1.5 fine tuned is far more capable, diverse, and feature rich than SDXL at the moment

SDXL is a fine model, a good one even as I am very excited for it, but its just not better than finetuned V1.5 in its current form

If somebody is saying something is objectively better than another thing, then they themselves open it up to comparison

The base is, but the finetunes are nowhere near as good as 1.5 on average unfortunately. Which is why so many people are worried to see if SDXL is finetunable, or if its gonna be a flop like 2.x

Right, so then why are you worrying about the cultural representation of a culture you are not a part of? If we wanna get into that

Their claim was SDXL base is better than 1.5 tuned, so they are the ones that opened that comparison

Their whole thing is how 1.5 in any form isn't as good as good as SDXL, which is just false lol

[wispy nest]

Idk how you'd call something outdated if most of SD community using it

[smoky oak]

right lol

You are exhausting

1.5 can do owls

Its an owl in general. I had it on hand

@paper salmonwhat now?

Guess they have no comment about my owl

oh well

I love the scene in this

So you will waste credits to generate things I have already shown, but none of the things I suggested. Right. Right

[surreal thistle]

Playground made these for me.

[wispy nest]

would you guys please give it a rest already. i pop my head in here every 10-15 minutes and you've been arguing about stuff for the last 2 hours. just agree to disagree or mute each other ffs

[surreal thistle]

Owls are cool.

[smoky oak]

You know what, fair point. This is like talking to a wall and is going nowhere anyways. I have been amused with how far they cope up until this point, but I get how its gotten old for some

[smoky oak]

I'll end it with this image, where the owl is looking at the camera like was suggested, rather than flying like was not suggested

I now have a new animal to play around with, which is nice, cause I have never generated Great Grey owls before

at least something came out of this immense waste of time lol

At least mine has eyes

ANYWAYS

I DIGRESS

[surreal thistle]

Have some girls.

[smoky oak]

the FEMOIDS

[surreal thistle]

They do look a bit like robots tho.

Because of the skin.

I still like them anyway.

[quick helm]

#🏞｜general-with-images 3 beautoful women wearing a white oversize hoodies by the beach and rocks by the sea in Sweden, windy weather. Full body image. Natural light. during golden hour, taken with a DSLR camera, using a 35mm prime lens with cinematic lighting.

how to generate?

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[surreal thistle]

Why are her eyes like that?

They’re not meant to be like that.

That’s the problem.

Yeah, gimme a sec.

[smoky oak]

...

I am just gonna act like I didn't just read that message lmao

[surreal thistle]

I mostly want more of this.

But there’s something unnatural about the skin.

[kind quartz]

if wrong eyes use extras tab gpfgan or codeformer, it will correct it

[surreal thistle]

What?

[kind quartz]

if your result has wrong eyes or so you can automaticaly correct it in extras tab in webui

[surreal thistle]

Playground can do it as well.

[wispy spindle]

If you're going to critique, please at least be respectful

[surreal thistle]

Close up on face, tropical impressionist painting. Wide open lips. E-girl makeup, blonde hair. Crop top with no shoulders, diamond necklace. Impressionist painting of a teenage young girl model. Nordic face. Tropical scenery, pale skin. 19 year old girl. Blushing. Russian model girl, the face of a model. Pino Daeni, Daniel F Gerhartz. Unbelievable lighting and breathtaking atmospheric scenery. Water reflections, fog. Palm trees in the background, smiling, cute model girl with beautiful reflective shimmering eyes and soft pale skin. Richard S. Johnson, perfect composition, beautiful detailed intricate insanely detailed octane render trending on artstation, 8 k artistic photography, photorealistic concept art, soft natural volumetric cinematic perfect light, chiaroscuro, award - winning photograph, masterpiece, oil on canvas, raphael, caravaggio, greg rutkowski, beeple, beksinski, giger

[smoky oak]

oh those are upsetting lmao

I hate how good those look lmao

[surreal thistle]

These as well?

[wispy spindle]

and then you attached a whole group to it needlessly. You've been warned yesterday so please keep it respectful. Thanks

[surreal thistle]

The skin looks too fake, is there a word that can make them look less like dolls and more like actual people?

Maybe “imperfect skin” or something?

[glossy herald]

hum .... sorry

[wispy spindle]

"textured skin" or "skin pores"

[glossy herald]

I didn't see you in there

[smoky oak]

The artists that you tagged are more than likely averaging out for that smooth look

I use tags like "textured skin, realistic skin, skin pores" and such

[surreal thistle]

Ohh.

[smoky oak]

They do, as I use them, but chances are terms you are using are canceling them out

its a balancing act after all

uhhh... No?

[surreal thistle]

The smoothness does result in an interesting design style, ngl.

They look more like people out of dreams.

[smoky oak]

ohhh, but that also depends on different models and such

[surreal thistle]

I don’t really care about realism that much, but I do want my girls to not look like dolls.

[smoky oak]

fair haha

[wispy nest]

so...after few tests on cardos animated..it actually looks like a mix of realistic + art \ drawing
Which turns out pretty weird o_O

Not sure if I'd use it, I'd prefer something a bit more specific...
Either closer to art or closer to realistic

[smoky oak]

for whatever reason, this woman reminds me of Frida Kahlo

[surreal thistle]

A bit, yes.

[smoky oak]

oh this is a very interesting inbetween

[wispy nest]

yea

[surreal thistle]

Big ol’ tohongerekoogers.

[smoky oak]

I WAS THINKING IT TOO LMAO

[surreal thistle]

I don’t really care about cleavage in my pictures, but I’m not against it, haha.

[wispy nest]

prompt didn't have it, it's just happens with ripped cloth...pretty often

[smoky oak]

Some Dobonhonkeros
A real set of dohoonkabhankoloos
Them big ol' bonkhonagahoogs!

[surreal thistle]

Humongous hungolomghononoloughongous.

[smoky oak]

HOW DO YOU REMEMBER THAT

I had to go through the whole video lmao

[surreal thistle]

My memory can be both a help and a nuisance.

[smoky oak]

Apparently lmao

I have a photographic memory for some things, but they are niche lol

[surreal thistle]

I think here the skin looks a bit better.

Still too fake though.

[smoky oak]

Yeah, thats looking much better IMO

[surreal thistle]

This is the same prompt, lmao.

I guess some turned out better than others.

These are perfect though.

Imagine an anthropomorphic owl in a tuxedo running up to you with an iron stick.

[smoky oak]

I have some anthro owls too

[surreal thistle]

Show them.

[surreal thistle]

Cool.

Also Dezgo can do anime pretty well.

Here the poreless skin actually makes sense.

It still needs work when it comes to fingers though.

[smoky oak]

I was messing around with owlin character design for my friends DND campaign

Also side note. My synesthesia has been really really aggressive these last few days, and this section of this song has had me in a synesthetic choke hold

Tickles my brain meat in the right way I suppose lol

[surreal thistle]

I’ve been listening to dub all the time lately.

Mad Professor, King Tubby and the rest.

[smoky oak]

There are just so many genres and scenes of bass music its impossible to run out of new and interesting songs

[surreal thistle]

For some reason I hate this music as much as I hate modern dubstep.

[smoky oak]

I mean, if you hate modern dubstep, of course you'd hate future dubstep lol

[surreal thistle]

Past dubstep was cool.

[smoky oak]

Curious to see what you think about this

thats your ping in there lol

[surreal thistle]

Not my cup of tea.

[smoky oak]

This is another new generation shapeless melodic drop. These are the sounds that really set off my audio visual reactions cause they are so organic I guess you could say

[surreal thistle]

If I listen to electronic music, I pick techno.

[smoky oak]

oh yeah, much much older stuff

also jesus why is that so loud

bout blew my eardrums out lmao

[surreal thistle]

Weird.

[surreal thistle]

I listen to it at a medium volume.

[smoky oak]

Little bit of modern bass house vibes

[wispy nest]

I'm mostly on dark trap lately...

[smoky oak]

I may have some songs if you wanna go more intense lol

oh, do you like to headbang by chance?

I have a song that puts me in a neck brace every time lol

[wispy nest]

If it's actual dark trap, like istasha, kiraw, corpse - go ahead

[surreal thistle]

This is a good mix of ska and dub.

[smoky oak]

Oh, this is very much complextro adjacent. Very nice

[smoky oak]

oh no, I was meaning like way more intense, less words, more anger lol

[surreal thistle]

Synesthesia is when you hear colors and stuff?

[smoky oak]

for me, yes

[surreal thistle]

Cool.

[smoky oak]

I have multi facet audiocentric synesthesia

[surreal thistle]

I don’t have synesthesia, but rather a very rich imagination.

[smoky oak]

Can be emotions, sensations, colors, shapes, all sorts of things

some songs affect me, others don't

[wispy nest]

hmmmm...anger sounds like what I might want but eh, I think I'm good on it for now, waiting for new album to drop

[smoky oak]

I have a few tearout tracks, which are about as intense as it gets before you end up at deathsetp, which is just straight up noise

nothing makes me smile more than CRUNCH

Oh, I do have a bit of color trap lol

its a very fun mix of styles

[surreal thistle]

This is a good dub.

[smoky oak]

ohhhh, you mean dub dub

ok, I se what you mean lmao

[surreal thistle]

Also, why does it sound worse in Discord?

[smoky oak]

I wonder if you would like New Amsterdam... It is very much old dub inspired with more complex sound design and a little more energy

[surreal thistle]

Mmm.

[smoky oak]

0 clue, unless you have an EQ

[surreal thistle]

Maybe because I’m using headphones?

[smoky oak]

Namster Damster

[surreal thistle]

This is pretty atmospheric, I think.

[smoky oak]

Also, just noticed your pfp, big fan

[surreal thistle]

Yeah. Only if I had a purple role to go along with it.

[smoky oak]

MMMM loud crunchy warning

Angry sound, aggressive oscillations lol

[wispy nest]

One day she'll figure out how to hold sword

[surreal thistle]

That's really not my style. I like bass in my songs when it's not ripping through my ears.

[smoky oak]

totally understandable

my tastes have gotten more and more diverse and subsequently nasty over the years of my venturing into bass music as a whole

[surreal thistle]

Like here.

[smoky oak]

and also I have synesthesia, which can make the goofiest sounds amazing to me

[surreal thistle]

It sounds much better in real life, trust me.

[smoky oak]

I do not

I don't tend to listen to bands over solor artists

[surreal thistle]

I don't really listen to anything other than dub these days.

But at least I try to find something new every day to not get tired.

[smoky oak]

oh yeah, absolutely not my style of music. I can appreciate it tho

I am not a fan of most traditional instruments unless done in a unique/new way, which is rare nowdays

I am listening to it lol

[surreal thistle]

You said you like complex soundscapes?

[smoky oak]

I do, yes

specifically spectrally complex

[surreal thistle]

The song I'm about to share does have traditional instruments.

But they masterfully blend in with the dub.

[smoky oak]

Yeah, I am not a fan of face to face at all personally

Its too traditional to me, sounds like a lot of songs I have heard, and i just generally do not have good reactions to most traditional instruments now

[surreal thistle]

Ohh.

Kinda weird.

Because you can't go wrong with traditional instruments.

[smoky oak]

the more unique something is, the more I am inclined to have a powerful reaction

[surreal thistle]

The guitar, saxophone, piano or harmonica are a staple of the dub genre.

[smoky oak]

not available in my country :<

same with that one

Stupid region locking

[surreal thistle]

https://youtu.be/YbSzPHqAo20

Maybe you’ll like this one?

It has both traditional and digital instruments.

AND dub.

[smoky oak]

Its a fine song, not really much of a reaction from me

I have an extremely picky ear for mixing and mastering, and I am not a fan of the mixing on that song

I do like some of the digital elements

[surreal thistle]

So future bass gives you the best reactions.

[smoky oak]

Several different niche genres

future bass, color bass, future riddim, tearout, and many other more niche genres

[surreal thistle]

Have you by any chance heard of Boris Brejcha?

[smoky oak]

That sounds familiar

Yeah!

[surreal thistle]

It's like house and techno mushed together.

[smoky oak]

I have

I recognize his visuals

[surreal thistle]

https://youtu.be/qc5Mte1FKbY

This is a good song.

[smoky oak]

Techno has never been too much of my jam, personally, but I do like a little bit of house

[smoky oak]

already listening lol

I have heard this song before

[surreal thistle]

Good.

Too bad you don't like techno.

[smoky oak]

had somebody else show me his stuff a while back

[surreal thistle]

Too bad I don't like future riddim.

[smoky oak]

Yes, its playing

[surreal thistle]

I'm a fan of traditional dub.

[smoky oak]

The beginning instrumentation is gorgeous in this

however I am major not liking the phase relations in this song

Dare I say severely disliking

but that intro was very very pretty

I do like traditional instruments in that case

[surreal thistle]

I don't have synesthesia. However, my imagination can be so strong that I can visit different places with my mind.

[smoky oak]

I have never been a fan of conflicting phase relationships

I find it horribly distracting, and even agitating unfortunately

[surreal thistle]

Dub music transports me to Jamaica or Africa.

[wispy nest]

I'd post something too, but I think all of my songs include explicit lyrics lol

[smoky oak]

I'll give it a listen

trap is a subset form of modern bass music after all lol

I wasn't prepared for just how much I love melodic trap lol

I vibe with this much more

[surreal thistle]

I love how this conversation went completely away from AI.

[smoky oak]

Its got extremely clear and clean production, which is typically one of my biggest turn offs. If its badly produced, good luck getting em to like it

but damn that is crisp and clean

[surreal thistle]

So you like trap.

And house.

[smoky oak]

Some aspects

I like modern house

[surreal thistle]

You might like the music I make.

I have a Bandcamp.

Unavailable.

Weird how Brazilian music is not available in Russia.

Also not.

So do I drop my Bandcamp or...?

[wispy nest]

Russia is outcast now, welcome to the world after 24.02.2022

[smoky oak]

sorry, I am gonna be busy for a little bit

[surreal thistle]

Unavailable.

[smoky oak]

Potentially, I am a music producer myself

[surreal thistle]

I use GarageBand, Soundtrap and Groovepad mostly.

[smoky oak]

Well, I am more of a sound designer than a producer at the moment, but I am working on an album

[surreal thistle]

I want to switch to Ableton, but my laptop is potato.

[wispy nest]

https://www.youtube.com/watch?v=Qslc4Lxlrdk
This one isn't too harsh, I guess

[surreal thistle]

Unavailable.

[wispy nest]

everything else is marked "explicit" all over it lol

[surreal thistle]

It's all Brazilian, right?

Also unavailable for some reason.

Very weird.

I guess that's got something to do with me being in Russia at the moment.

But Samba de Janeiro is available.

Literally the most Brazilian song ever.

[wispy nest]

W.A.S.P. available for me, Russia too

same for cazuza

and Sozinho

[surreal thistle]

Maybe it's got something to do with Discord.

[wispy nest]

oh yea I opened them in a browser

[surreal thistle]

Barbara Palvin's face seems to work really well with AI.

[edgy shard]

i dunno why but that sounds like a dig.

"you've got a face that ai just loves."

[surreal thistle]

Emma Watson's face is for some reason very hard for AI.

But Benedict Cumberbatch's isn't.

[mighty flame]

Can someone help me use vector studio to make a sticker out of a real picture?

[wispy nest]

just based on training data

[surreal thistle]

So probably Barbara Palvin's face was used for training more?

Or is it just simpler to draw?

[wispy nest]

yea, probably just more images with different perspectives

[dense tapir]

@smoky oak https://www.youtube.com/watch?v=NDmfMlZ2QFo

[dense tapir]

Oh, yeah I am sold on Vlad's fork for sure. So much faster in everything on my old 1060.

[dense tapir]

6gb

I am now doing 50 steps ddim for less time than 20

No, but why?

Naw, I have no issues with Russian and is better than having the dumb yellow, blue flag

you see that on twitter bios, etc...

[dense tapir]

Have nothing against them but I 1000 percent do their current leader. Anyway this is geopolitical so this political discussion is over.

[rain gazelle]

whoa

Performance on Vlad's was the same for me.

Did I miss a tweak somewhere?

[ripe cedar]

I use ControlNet to generate stuff with good hands and then hi-res fix upscale (I know, it's bad) ends up messing up the hands and making everything else better. Is there a way which makes the hands ok?

[rain gazelle]

try lowering your denoise for the hi-res fix.

g'night

[smoky oak]

Howdy

I have awoken

[smoky oak]

Heh... I just found that high res fix is good if you use it right

the problem is not high res fix, its latent upscale for high res fix

[dense tapir]

@smoky oakLet me tell you I feel the power

[smoky oak]

What happened?

[dense tapir]

I ditched Auto

[smoky oak]

For?

[dense tapir]

Euler_A 20 steps 40s Vlad 46s Auto1111. 768x768 Vlads is superior speed wise and has it all setup for pytorch2

on a faster card 2x

people are fleeing auto now for vlad

[smoky oak]

Interesting, but hows its upscale support?

[dense tapir]

I dunno that is your department not mine. He does updates and PRs and auto is pretty much abandoned automatic1111

[smoky oak]

Intereasting, maybe I will try it

[dense tapir]

you should since it does it all for you

at least to see if you get a speed increase

I know I sure did

[smoky oak]

sure

wheres the documentation?

[dense tapir]

polus all the shit I use for a gen it just loads it kerpow and there

same as auto

I can help ya but not much help will be needed

[smoky oak]

I just need the link if possible

[dense tapir]

https://www.reddit.com/r/StableDiffusion/comments/12srusf/vladmandic_vs_automatic1111_vlads_ui_is_almost_2x/

[ripe cedar]

I was getting black images more often with Vlad with ControlNet depending on the settings, but it works better overall. Definitely recommend it

[smoky oak]

thank you

[dense tapir]

https://github.com/vladmandic/automatic

[serene rampart]

hello

[dense tapir]

he is merging/working with another well known fork. This is getting good now

[serene rampart]

anyone make nsfw?

[smoky oak]

I have, but thats not content for this server

[serene rampart]

can i dm you?

[smoky oak]

ohhh, Vlad is a fork of A1111

[ripe cedar]

Some consolidation in the UI space is good I think. Perhaps it means people working on it will join forces in the longer term over a single project

[smoky oak]

that is attractive

cause that means it shouls support high res fix and the scripts

HMMM

[dense tapir]

yes, so is another about to merge so the coming together of two fantastic forks

[ripe cedar]

Add a knife and you've got a full meal

[dense tapir]

yep

that top dude his fork is awesome too

[smoky oak]

So I just clone the repository in my folder of choice?

[dense tapir]

yep

[smoky oak]

alrighty, I will do that then

[dense tapir]

after a few minutes I am already at home but the first few was a bit of a shock just because new

vlad is strict on github and he answers fast (does he sleep? no).

btw, I learned my lesson today, as did another, that the github search is shit. Made us both out as if we were lazy and stupid

Now on I use google to search github

I am really wondering how fast your normal gen is compared to auto on that 3060ti

[smoky oak]

If Vlad is as good as you say, I'll happily switch

[smoky oak]

its downloading at the moment, I'll hit it as hard as I can :p

[dense tapir]

He always was good as I used him when 2.0 came and auto refused to make it work for three weeks

it was just a working shell of auto back then now just wow

I wish it were not on gradio but can't have it all

oh, at least I haven't OOM on vlad yet but on auto 5-6 times an hour at least

so bad I couldn't even change models without an OOM but not once on vlad so he fixed something

[smoky oak]

this is the cleanest installer I have ever seen

[dense tapir]

if you don't speed up after I gained about 12% something is way wrong

[smoky oak]

very clean and tidy

[dense tapir]

yep

[smoky oak]

I appreciate being able to see whats happening

[dense tapir]

precisely

a lot of extension are now defaulted in

I don't use controlnet so I disabled it

[ripe cedar]

Yeah it installs a lot of the bread and butter stuff but it's convenient that you can disable stuff you don't need.

[smoky oak]

Yeah, Is aw that on the page

[dense tapir]

WOW, alright hmm

[smoky oak]

Cause this is still A1111, I should be able to run this as my persistent web server still, which is super nice

[dense tapir]

--api is gone

[smoky oak]

I don't use --api

I use --listen --auth

[dense tapir]

yeah it was replaced but I don't use it either

[smoky oak]

I port forward the listen to my router and direct connect with IP, then enter my credentials to gain access

[dense tapir]

first up, webui-user.bat is not there

[smoky oak]

yeah, just webui.bat

[dense tapir]

you can grab it from auto and place it in the folder and use it

watch for the flags that changed or no longer exist

all of my optimizations went POOF because they are now in settings. Oh, and KILL --xformers. remove it

[smoky oak]

wait why?

I am not willing to give up xformers for its memory optimization

[dense tapir]

cause you are about to use the real Pytorch 2.0 and sdp

[smoky oak]

alright, if you say so

I'll test both

[dense tapir]

yeah, you can switch via settings if you wanty

[smoky oak]

oh, so you don't have to restart the whole web UI?

[dense tapir]

Look at all those optimizations, lol

smoke 'em if ya got 'em

[smoky oak]

I need to copy some models over, then I can compare

There we go

luckily my NVME drives FLY at model transfer lol

[dense tapir]

This sounds like what SHARK does but I couldn't get it to work

[smoky oak]

I just relaunched the bat and its reinstalling the whole programs?

whats up with that?

[dense tapir]

didn't for me

of course I have no idea what is in your .bat

[smoky oak]

ok, it looks like it fixed some previous warnings or something

[dense tapir]

yes

it really is good at it

[smoky oak]

hope it doesn't do that every time

[dense tapir]

naw

[smoky oak]

oh jesus, that black orange style is hideous lol

oh well, it will work for now

[dense tapir]

I don't mind it at all now

it grew on me

sort of like herpes

when the other guy merges expect a shit ton of themes

even for mobile devices

[smoky oak]

hmm...

Its slower

9.4it/s

but it feels more responsive

like the image pops up when done considerably faster

Should I try BF16?

[dense tapir]

yes

[smoky oak]

do I need to restart UI for those changes?

[dense tapir]

maybe, or at least I would for something that drastic

[smoky oak]

ok, and now its errors

can't pull the prompt from the last image I generated

whole think is broken now

well thats a fun first experience lol

[dense tapir]

what did you do?

[smoky oak]

restarted web UI

now I can't access any samplers

and now the images don't show up when done

fun

Guess I am restarting the whole program

[dense tapir]

Oh, I thought you did

[smoky oak]

ok, it reloaded last prompt now

9.7it/s

[dense tapir]

now switch to xformers via settings

[smoky oak]

Where do I enable that?

[dense tapir]

settings

[smoky oak]

lmao

yeah, but where?

its not in cuda settings

ah, found it

settings-stable diffusion, cross attention xformers

[dense tapir]

stable diffusion

click it

this is so fast for me now I can have image previews and not a single slow down

[smoky oak]

ok, so restarting the UI is useless, good to know

so settings changes are a full program restart

[dense tapir]

these are huge ass changes

[smoky oak]

not THAT big of a deal when I get it tuned

m'kay, this UI is getting on my nerse

[dense tapir]

you know, I haven't looked at my vram usage. I know xformers is 5.9gb

[smoky oak]

restarted and now its nothing but trace back errors

[dense tapir]

I don't use that on even auto for that reason I close the cli and start it over

[smoky oak]

Never had a single issue with web ui restart on auto

oh well

where does one put command line args?

[dense tapir]

in the webui-user.bat

[smoky oak]

Not seeing any section for command line args

[dense tapir]

set COMMANDLINE_ARGS=--theme dark --lowvram --xformers --disable-safe-unpickle --port 9000

that is mine from auto

[smoky oak]

just paste that anywhere?

[dense tapir]

no

[smoky oak]

cause that isn't in mine

there is no set commandline_args in mine

only set I have in my whole bat

[smoky oak]

What the hell

oh wait, you copied over webui-bat over?

[dense tapir]

webui-user.bat I said just copy it over from auto

[smoky oak]

ahhh, alright

[dense tapir]

his didn't even come with one

[smoky oak]

ok so --xformers is gone

hmm

wonder why they nuked that

[dense tapir]

lives in settings now

[smoky oak]

weird, enabled it and no speed change

ok, vlad has a huge red flag for me

no insecure extension access

that is an absolute pain in the ass

So far I am not sold on this UI at all

its slower

@dense tapirany ideas why its considerably slower?

cause this isn't gonna get me to switch

gonna head to head compare for cerain

[dense tapir]

No idea because not a single person has said it was slower only faster

[smoky oak]

alright, for this test it took 11.79 seconds

time to compare directly to a1111

[dense tapir]

For me I can't use xformers it gives me an OOM

probably some setting I missed

yep, missed one

[smoky oak]

oh hohoho wait

Vlad is WAY faster in this test-

@dense tapirI am very confused

A1111 is faster on paper

but the actual time is slower

I get 10.5it/s in A1111, vs 9.7 in Vlad, but VLAD was way faster in 10 batch

[dense tapir]

Yep

that is Pytorch2 and his coding

[smoky oak]

ok and now its not as fast

I don't know

so weird

so A1111 is faster in single image

2.22s vs 2.34s

but in 10 batch its

15.85/11.79

and it seems to be strictly down to post processing time

doing 8 x 50 samples to make for sure certain

8x50 took 33.34s in A1111

trying vlad now

34.1s in vlad

trying a second time to make sure it wasn't just a first start up slow gen

33.83s

hmm...

[dense tapir]

not sure what to tell you but all signs are pointing to automatic1111 being abandoned and eventually you will have to leave it. Why yours is slower beats me.

[smoky oak]

It seems close enough to not be a huge deal speed wise, but Vlad is a hell of a lot more touchy for me, thats for sure

[dense tapir]

I mean, hell, I even got faster.

oh, let's try a setting for you

[smoky oak]

alright

[dense tapir]

go to settings

[smoky oak]

oops

just a sec

[dense tapir]

cuda settingas

*settings

mimic that

[smoky oak]

I will in just a bit

[dense tapir]

k

[smoky oak]

I messed up

[dense tapir]

still, mimic that it will help if you use sdp

[smoky oak]

Way slower now

Expiramenting with some other settings

[dense tapir]

Damn, I had to ticket this

00:23:05-192311 INFO Installing package: xformers==0.0.17
00:23:07-055986 INFO Verifying requirements
00:23:07-099210 INFO No changes detected: Quick launch active
00:23:07-102213 INFO Running extension preloading
00:23:07-103214 INFO Server arguments: ['--theme', 'dark', '--lowvram', '--disable-safe-unpickle', '--port',
'9000']
No module 'xformers'. Proceeding without it.
Available models: D:\vlads_automatic\models\Stable-diffusion 90
Loading theme: black-orange

no wonder I OOM'd

[smoky oak]

Hmm

this is all so strange

different combinations of cross memory attention are slower or faster for me

[dense tapir]

yes

same for us all

sdp is supposed to be the fastest

[smoky oak]

just trying to find a combination that can match a1111's sopeed

[dense tapir]

it is what came with pytorch2

what you should be doing is going to vlad and asking why it would be slower cause the dude knows his shit

he isn't a script kiddie like auto

[smoky oak]

ok, SDP isn't too bad

its just slightly slower than A1111

trying each combination. Its annoying that I have to restart vlad every time, but oh well

@dense tapirbingo

10.29it/s in vlad

[dense tapir]

That's a start

[smoky oak]

Scaled dot prod/SDP for that

trying other combinations still

[dense tapir]

I blew mine up some how

[smoky oak]

oop

[dense tapir]

made it way slower

[smoky oak]

I got down to 2.2s/it lol

[dense tapir]

my speed, YAY

lol

[smoky oak]

ok, doggettx is a lot slower

[dense tapir]

yeah, old

[smoky oak]

yeesh

43 seconds

[dense tapir]

Might work on 3 digit nvidia cards

[smoky oak]

only 8.02it/s with dog

[dense tapir]

Maxwell

[smoky oak]

for shame

trying invoke

also slow

not as slow

but slow

8.7it/s

[dense tapir]

I am going to reinstall 😦

[smoky oak]

trying sub quad now

[dense tapir]

what this needs is a reset settings to default

not all at once just a section one

I wanted that on auto too

[smoky oak]

oh LAWD Sub quad is slowwww

6.6it/s

[dense tapir]

notice you have sliders for that

[smoky oak]

💀

[smoky oak]

that is a little too much effing around and finding out for me-

[dense tapir]

even vram threshold

[smoky oak]

trying split attention finally

[dense tapir]

that left me with a black square

[smoky oak]

So far scaled dot/SDP is the best

[dense tapir]

yes, it is the newest

[smoky oak]

weird how it just barely ties A1111 for me

split attention is the slowest

alrighty

6.1it/s

ok, so disabling split attention

kills everything lol

OOM everywhere

[dense tapir]

heh

I hope this fixes the shit I did

I doubt it as it says unloading xformers

rip out venv

I know pytorch2 does use less vram

they said it would and yeah

[smoky oak]

@dense tapirOk weird

so

hmmm

I changed some of your settings and my batch got decently faster

from 1.58it/s to 1.67it/s for 8 batch

but my single image went from 10.3it/s to 9.5it/s

how weird

yeah, so my batch went to 31s which is baster than A1111's 33.34

its this setting

on gets me higher single image, lower multi

off is higher multi, lower single

how weird

[dense tapir]

that is a big one for 4090 and 3090 but mostly for 4090

[smoky oak]

FP16 seems to be the tiniest bit faster than BF16

like 2% maybe

[dense tapir]

BF16 is FP16 on roids is why.

they are both 16 bits

whatever you do don't zap venv as I am going to have to start all over it appears

ModuleNotFoundError: No module named 'clip'

[smoky oak]

@dense tapirWhat is cuDNN?

cause its giving me some weird results

HOLY HELL

ok

what is this

it takes like 10 seconds to work, but then boom, 11.9it/s

it takes a long time to start, and a long time to process after

but while its generating it hits the highest speeds I have ever seen on my GPU

but once it does that once, its fine

but only at that batch size

this is so weird

if I go from 1 to 8, it has to do the whole process again-

[dense tapir]

cuDNN is your cuda stuff

[smoky oak]

1.65it/s 8 batch, 11.6it/s 1 batch

[dense tapir]

yeah, this is touchy af. No more just ripping out venv

[smoky oak]

1.98s for one image

what is this sorcery

[dense tapir]

wth?

[smoky oak]

its really slow to first use, but then its insanely fast

[dense tapir]

how did you manage that>

?

OH, OH

[smoky oak]

Enable cuDNN

[dense tapir]

did you click that compile shit?

[smoky oak]

it takes a long time to gen, but once it has, it flies

I have seen up to 11.9it/s with this setting on

[dense tapir]

well, you haven't seen anything yet if that compile will work

[smoky oak]

I am disabling this for the time being

gonna try to get compile to work

[dense tapir]

Not sure how it works as there is no compile button but might be for batch work so only temp

[smoky oak]

its doing something

the gen has paused at the start again like cuDNN did

CPU is at 15%

RAM is slowly increasing

perhaps this is it

[dense tapir]

yes

[smoky oak]

ok, its no faster on this setting

1.54it/s at 8 batch

[dense tapir]

SHARK will make a big model file that anytime you use that resolution and that model it uses it instead and flies faster than the wind

[smoky oak]

ok, this is behaving exactly like cuDNN, but not as fast

[dense tapir]

use together?

[smoky oak]

ok now its faster

1.65it/s on 8 batch

11.4it/s again

screw it

trying together lol

I mean honestly

having to wait a few minutes at the start for a 10% speed increase from then on is a win IMO

cause I keep my SD active for like a week at a time lol

[dense tapir]

I wish it would work like SHARK

[smoky oak]

together is still 11.4 it/s average

[dense tapir]

Yeah, experimental but there is hope

I love how fast it loads a model

this is how fast Linux loads it on auto

At least my speed is back. I am not sure what happened

happened after I let it do xformers

[smoky oak]

trying cudagraphs

[wispy nest]

gave Vlad a try thanks to you guys talking about it. loads my embedding -substantially- faster. which is important for me lol

[smoky oak]

ok, its dooin stuff

ok, cudagraphs is faster to get to the sped up result

about 7 seconds waited before gen

[dense tapir]

I use 5 locons all about 800mb and in auto it took over a min in vlad seconds

[smoky oak]

oof, but the post gen wait is huge here

come owwwnnn

[wispy nest]

my load time dropped from about ~3 minutes to less than 1. pretty happy with that alone. so long as the speed is comparable, i'm a convert

[smoky oak]

7 seconds pre, 50 seconds post

[dense tapir]

I already deleted auto

[smoky oak]

11.4 it/s cap again

I wonder if thats a CPU limit I'm hitting

[dense tapir]

you need to be watching your cuda gpu in task manager. If it is staying 100% then that is your hard limit

[smoky oak]

trying aot ts nvfuser

[smoky oak]

taskmanager reports 0 utilization in SD

[dense tapir]

to get to the cuda you have to use the drop down

[smoky oak]

does the same with render

[static tusk]

switch one of the graphs to cuda

[dense tapir]

default is 3d not cuda

[smoky oak]

ok, nvfuser was the best so far

only a few seconds wait before and after gen

but its only 10.4 it/s

which is basically my normal speed

also, cuda is working now

thank you <3

[dense tapir]

🙂

I wish we could make that the default in TM instead of 3d

[smoky oak]

I think I have opened vlad more today than A1111 ever lol

[dense tapir]

I got an error so maybe my card is finally dying.

[smoky oak]

I am using cuDNN and its suddenly working way faster

[dense tapir]

cuDNN gave me a weird error

I need to use it maybe

[smoky oak]

Wait a sec

I think I know what cuDNN benchmark is doing

[smoky oak]

it seems to run the whole process, cache it, and then keep it in memory to speed it up

WOO 1.71it/s batch size 8

[dense tapir]

I don't do batches so which was for speeding up singles?

[smoky oak]

oh shoot, i forgot

[dense tapir]

batch is super slow even with the benchmark on

[smoky oak]

ahhh

its the uhh

torch memory

that speeds up single

[dense tapir]

this is hideous

10 times slower for batch 5

7-8 times so faster 1 by 1

[smoky oak]

Time to burn rubber

ok so torch mem with cuDNN is the fastest

but it has a really long start and end pause

but once you do that one, it flies

[dense tapir]

never worked for me

[smoky oak]

Now we're flying

batch size 8

Alright, these are my final settings I am sticking with

[dense tapir]

I wish

so you turned on benchmarking and

[smoky oak]

yes, it takes a long time before and after, but its the fastest of both

[dense tapir]

didn't do anything for me

[smoky oak]

weird

[dense tapir]

must be some tensor core thing

[smoky oak]

yeah maybe

oh wow, something with my settings is hammering my VRAM

Dreamstudio is the current option

Oh isn't lexica like extremely bad or something?

Yeah, lexcia was in a huge legal battle

They both appear to run basic SD models with meh results

My recommendation for free Ai is actually https://dream.ai/create

it does way better than the two you just said

Mage space's result

Wombo dreams result

Mage space

Dream

its also faster

and has tons of styles

its a solid service honestly, its what i came from

its not as good as actual SD, but its a lot better than lexica and mage space

[silk tendon]

Now everyone is equal

All hail communism

[mental cipher]

im new to SD and im trying something basic and i cant seem to get the result, im not trying to change her face just make her vibrant, i know somethings wrong with my prompt or settings just could use some advice from the pros

[silk tendon]

Are you using a vae?

[mental cipher]

whats that? probably not

[kind quartz]

how about in promt type: Make her vibrant

[silk tendon]

Look up berry mix vae

[mental cipher]

i added vibrant

[kind quartz]

i meant only make her vibrant without anything

[smoky oak]

just use a photo editor-

[mental cipher]

like gimp? or photoshop?

[smoky oak]

yeah, something

[still bay]

How to generate images

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[mental cipher]

i dont know how to do it in gimp or photoshop either lol newbie over here

[dense tapir]

@smoky oakI toggled those on and it is sitting here doing something as my card is revved up for a min now but no its

Dang, still sitting here

[smoky oak]

@wispy nestComparison between Mage, Dream, and dedicated SD

[dense tapir]

100% utilized now it comes

[smoky oak]

cuDNN takes a while

it has to go through the entire generation first, then actually do it, then cache it a third time

[dense tapir]

almost 5m

[smoky oak]

but you only have to do it once, then that res and batch is cached

unless you restart SD, then it has to do it again

[dense tapir]

5m exactly

[smoky oak]

but if all you do is square gens, the time saving could be worth it eventually

[dense tapir]

I am still waiting for it to appear

sheesh, must be a tensor core thing

still waiting

I guess this will be a 10+ min dealio

so, same rez and same model it remembers?

[smoky oak]

A photograph of a woman in a mexican dancing dress, smiling, tan skin, curly black hair

Mage, dream, SD

SD has a bit of a quality jump haha

[smoky oak]

yeah

[dense tapir]

This is not to be used for non TC cards unless you had a ton to gen with

[smoky oak]

batch, size, and model

[dense tapir]

SHARK does this but saves the file as a model so it just uses it the next time

[smoky oak]

the faster the GPU, the easier it is to use

also, both together is WAY slower

just cuDNN isn't that slow

[dense tapir]

you said to do both 😦

[smoky oak]

that is the fastest overall gen speed for single and batch, but it takes a longgg time to load

if you just want single then its just memory like I said before

[dense tapir]

Well, just one or the other didn't do anything for me

[smoky oak]

then I am not sure in that case

[dense tapir]

My card is still revving along

100%