🏞｜general-with-images

[smoky oak]

I am now in the mood to compare AI image gens lol

[dense tapir]

I want to see what this does if it finishes as I will regen to see

[smoky oak]

once you wait, you don't have to anymore, so might as well see

[dense tapir]

this has potential if they will save it and reuse it as SHARK does

the reason shark is so fast

still waiting, lol

maybe it froze up

still waiting

[smoky oak]

with both, a 512x512 gen took abotu 1 minute 50 seconds to cache for me

[dense tapir]

well, my cpu is 12% and gpu 100%

[smoky oak]

just like mine

[dense tapir]

Yeah, I am so wanting to stop this

25m and counting

[smoky oak]

Yeah, I'd say just cut your losses

its not like this is something you can afford to wait every time you gen

[dense tapir]

must be those TCs

more you have the better this will be

it just finished

[smoky oak]

oh it worked?

[dense tapir]

there it is

[smoky oak]

nice

might as well use it for now then

[dense tapir]

LOL, not a single bit of speed up

actually is slower

ffs

know what? something is bad

[smoky oak]

lmfao

I don't understand what keeps happening to you 😅

[dense tapir]

same prompt, same seed, different image

I am genning again

wth?

[smoky oak]

A photograph of a snowy forest, realism, mountains, sunset, pine trees, snow, pretty

Lexica, Mage, Dream, Stable Diffusion

[dense tapir]

Well, I may have to go back to auto, damn it

far too random for me

[kind quartz]

pitty that hand 😄

[dense tapir]

@smoky oak https://youtu.be/mtMZGdCjUwQ?t=359

I linked at the time for you

[smoky oak]

oh I did that a longgg time ago haha

[dense tapir]

Ahhh, orange is best

[smoky oak]

Hard disagree, but to each their own

[dense tapir]

anyway I am done with vlad as I can't handle this being at seed -1 all the time.

[glossy herald]

Yes CivitAI, in fact, I do love waking up to find your database so busted again that my profile comes back as 404
Thanks for such a nice welcome civit 🙂

[smoky oak]

That website is less stable than diffusion

[dense tapir]

overload is partially why

a victim of its own popularity

[smoky oak]

I wanna like vlad, but I am back on auto 😅

at least for now

[dense tapir]

I hate losing the speed gain

[smoky oak]

A pencil sketch of an old man, pencil, sketch, drawing, pencil sketch, pencil texture

Dream, Lexica, Mage, Stable Diffusion

I find it crazy that all of these services are based on SD, yet they have such a much lower quality on average compared to locally run

[dense tapir]

resources

[smoky oak]

I'm just surprised there isn't a service that offers high quality results at the trade off of taking a while

cause these results from the others were probably less than 10 second gens

[dense tapir]

some do but for pay I think

some were using colab free which is why we lost it

[smoky oak]

I did use high res fix for SD, but even without it

still the best for sure

and only 3.6 seconds for 512x768

which is the fastest of all of them

[dense tapir]

highres fix is the best one I read because it works in the latent space while the others work in the pixel space.

[smoky oak]

yeah, upscaleing in SD isn't a pixel thing, its a diffusion thing

[dense tapir]

I don't use it because it has always been far too slow for me

[smoky oak]

fair for sure

[wispy nest]

i assume you were using sdp attention? did you try using this option? sdp is non-deterministic, enabling that option makes it deterministic at a slight speed hit

[smoky oak]

I just recently found out that highres fix works really good, the only issue comes from the latent upscaler, which is what causes those huge composition changes

[dense tapir]

Well, I was going to as I read about that as the -- option but xformers is non deterministic as well only it minorly changes this thing is just like I did a seed -1

[smoky oak]

on my hardware, xformers is 100% reliable

for same seed gens

[dense tapir]

it is still 100% non deterministic

[dense tapir]

ut oh

[dense tapir]

Applying scaled dot product cross attention optimization (without memory efficient attention)

on vlad's I can't use (apparently others can't as well) xformers so let's try this and see

[smoky oak]

enabling xformers for me on vlad made 0 difference at all

[dense tapir]

fucked me over and from that video I see it did many others too

[dense tapir]

black box gen

[dense tapir]

way slower too

[wispy nest]

seems pretty deterministic to me

[dense tapir]

yep

I can't use it though as I get a black gen

[wispy nest]

you've got something set weird then lol

[glacial sand]

An Apocalypse where people are running for their lives.

[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!

[dense tapir]

no, 2.1 without xformers has this same issue

[kind quartz]

i cant run 2.1 without xformers, having 8GB VRAM

[smoky oak]

general Awareness doesn't have an RTX GPU, which causes issues with tons of optimizations for AI

[dense tapir]

if you tried you would gen a black box

yes

[kind quartz]

i have 3070

[wispy nest]

what? not sure what you're referring to

[dense tapir]

like AMD gpus could not do 2.1 due to this since xformers is not for them so they had to do full 32bit and hope they had enough mem

--no-half

1.5, and 2.0 does not have this issue.

[wispy nest]

sd 2.1 model?

[dense tapir]

SAI is aware of it and tried to force xformers on people

[dense tapir]

version 2.1 not model

[tired basin]

I think there is instructions out there on how to modify base a1111 to use torch 2

[smoky oak]

it sucks, we tried it over a month ago

[dense tapir]

there is and it was horrid

[smoky oak]

if there is an optimization, we have tried it lol

[tired basin]

no improvement with newer versions of xformers?

[dense tapir]

the problem is auto is MIA and needs to implement it whereas Vlad is present and did implement it

[smoky oak]

I even spent 7 hours installing VoltML through WSL and hypervisor, and it was for nothing

[smoky oak]

A1111 isn't made to benefit from its optimizations

it was about 20% slower for me on average

[tired basin]

^^ He thanks you for your dedication!

[dense tapir]

So, if I use pytorch 2.0 vlad I am forced to forever have random gens

[smoky oak]

https://tenor.com/view/foreva-loveandhiphop-beef-gif-4932175

[wispy nest]

what do your settings look like?

[tired basin]

I just hope a1111 is OK, and that he's just having a break

as in the person

[dense tapir]

yep, until I get a card I have enough mem to do --no-half

someone said he is back to playing dota 2

[tired basin]

I went down the paid paperspace route

[smoky oak]

wait what? --no-half shouldn't have anything to do with this

[dense tapir]

my god it does

[smoky oak]

I run at half, and my gens are rock solid reliable

running no half is just inefficient

[dense tapir]

you only half read what I said

[tired basin]

if you only have a 4gb or 6gb VRAM card, then loading a full fp32 model in is impossible

[smoky oak]

As you can see, half is pretty reliable in Vlad lol

[dense tapir]

I said, consider this like an AMD. They do not have xformers so they gen black boxes in 2.1. All over the net about that so the only other way to gen without xformers is --no-half

[tired basin]

Even some of the unpruned 1.5 models don't load (like one of the Protogens was 7.7gb)

[smoky oak]

ohhh, I see what you mean, alright

[wispy nest]

i'm not saying that's 100% not the case, but i will say there are a LOT of people online who are convinced they understand all the settings they are changing and they don't

[dense tapir]

end of line

[wispy nest]

in any case. thanks for discussing the vlad fork, i will definitely be using it

[smoky oak]

on my 3060ti it was basically the same speed, but with more issues

There were some things I could do to get it faster tho

but they were kinda annoying

[wispy nest]

on my 3060 12gb it's substantially faster in model load and iamge finalization, about the same gen speed

[smoky oak]

yeah, same here

[wispy nest]

and all the extensions i was using transferred over perfectly fine, so pure win for me

[smoky oak]

the gen was about the same, other parts were faster

[tired basin]

If anyone actually understood every setting that goes into a generation, their head would explode - there is so much to it!

[nimble timber]

Guys i have a problem installing automatic 111, where can I ask help?

[tired basin]

#🤝｜tech-support

[nimble timber]

thanks

[smoky oak]

Vlad doesn't offer anything for me at the moment to justify a switch

[tired basin]

The two things for me were the lack of the switch height/width button, and the png info missing as a seperate tab

[smoky oak]

oh I didn't even notice the lack of switch button

I use that all the time

[wispy nest]

i noticed this as well, but the arrow button does the same thing so i've just switched to that now. (the png info tab)

[dense tapir]

sdp determinsitic and no xformers for the first and the second is turning on --no-half (full precision) at about 4.92s/it. This is why AMD users said fuck 2.1.

I am regenning and I bet it doesn't change but at almost 5s per iteration I'll pass

[tired basin]

is that at 512x512 or 768x768

[dense tapir]

doesn't matter the rez only that it is 2.1 but this was 960x544

[smoky oak]

So I found out that the only issue with high res fix is using latent upscale

base

[tired basin]

Res affects speed significantly

[smoky oak]

Latent

[dense tapir]

well, no shit lol

[wispy nest]

i'm not sure what's going on under the hood with vlad...but i can tell you it does something much differently as far as memory management while genning goes. i can't run batch size 20 hires fix from 512x512 to 768x768 on auto1111, but i just got done running batch size 30 hires fix from 512x512 to 1024x1024 on vlad. i've got very little reason to do batches that high, but the fact that it can do it at all without OOM means it's doing something right

[smoky oak]

swin

[dense tapir]

WHOA, it changed

wth?

[tired basin]

but 5s/it at that res is ouchies..

[smoky oak]

indeed it did

[dense tapir]

seriously

why is this changing like this?

I have determinstic sdp on

[smoky oak]

no idea man

[dense tapir]

512x512

[smoky oak]

[dense tapir]

pretty steady 3.08-3.1

[wispy nest]

on a 1060? with what sampler?

[tired basin]

I was running a 1060...

--medvram and --xformers was giving 1.2it/s on 512x512 gens

I couldn't find a better optimisation combo

using Euler A

[dense tapir]

Euler_A, sorry

[smoky oak]

I made you a DDIM lover :>

I use DDIM 99% of the time

[dense tapir]

Yes, I switch between it and euler_a

[wispy nest]

both are great

[kind quartz]

how about uniPC? got similar result to Euler but can on realy low numbers of samples

[dense tapir]

unipc demands tensor cores

[wispy nest]

the faster model loading is a large part of why i'm gonna stick with vlad. a lot of times i'll start a project by checking how my topic is interpreted in each model with each sampler. takes a while, but gives me the best base to work with imo

[kind quartz]

@dense tapir oh o.k. didnt knows that

[dense tapir]

I wondered why it was crashing and was told that was why

turing, and higher there are no issues

[kind quartz]

it crashing ui from time to time, need restart

[dense tapir]

I seem to have blown out vlad's again

my speed is shit now

it is weird because there are settings that if you touch you can't untouch and have to start all over

happens to a lot where on the net they say to never save xformers

it got me now this got me

[wispy nest]

it seems i have to rescind my previous statement about speed being about the same on my 3060. i hadn't enabled bf16 and channels last yet on vlad. this is with DPM++ SDE Karras. first is stable diffusion with optimizations, second is vlad with optimizations. i'll definitely be sticking with vlad

sure, it's only a ~7% increase, but that'll add up

[smoky oak]

yeah, understandable

I can get about a 14% speed increase in vlad if I mess around fully, but its more annoying IMO

[dense tapir]

I have no idea but I saved the image and got the same image back

interesting

[wispy nest]

yeah...definitely switching lol

[smoky oak]

I probably will as well at some point, but I just don't like certain aspects of it

also it seems to be less VRAM efficient I believe, compared from my gens

but I did so many I could have misaligned the results

and there are wayyy more factors to vlad as well

[dense tapir]

512x512 euler_a 20

[wispy nest]

it seems to be substantially better at vram management from what i saw. as i said earlier...auto1111 i couldn't do batch size 20 512x512 to 768x768 hires. vlad let me do batch 30 from 512 to 1024

[dense tapir]

that is even going through a HN and a lycoris

I am not even running low, or medvram options now

[smoky oak]

That is an extremely over the top use case, but I supposed that could be the case

[wispy nest]

yeah, as i said. i don't have a reason to do that. i was just testing if i COULD. and it did it without OOM

[dense tapir]

one thing I hate is that in the sampler drop down I can't get rid of unipc or plms yet they are unchecked

[smoky oak]

I wish I could run A1111 and vlad at the same time rather than having to deal with the annoyance of switching between

[wispy nest]

that does seem to be the case, unfortunate

[dense tapir]

a bug I hope it nips. probably a really simple 1 value bug

[wispy nest]

probably

[wispy nest]

it doesn't seem to switch the "generate" button over like auto1111 does as well, but i rarely use that as an indicator anyways(and could have something to do with me changing settings for preview and progress bar, haven't messed around enough with the settings menu yet)

[dense tapir]

switch the gen button?

[wispy nest]

change it to generating or whatever it was on auto1111. i think it greyed it out as well? i honestly don't remember

[wispy nest]

ah yeah. that's what it did

i would prefer that it did something like that in vlad, but haven't yet found an option for that. probably because the stop and skip buttons are just always visible

[dense tapir]

That is vlad

[smoky oak]

muahaha

I have both running at the same time

:D

[wispy nest]

also: it seems it doesn't clear the annoying image at boot...so i may have to check the code and find a way to turn that off

this thing

[dense tapir]

yeah

[smoky oak]

ok so, in side by side

[dense tapir]

I saw it

[smoky oak]

A1111 is faster gen, Vlad is faster processing

[dense tapir]

For me both are faster

[smoky oak]

A1111 for 1.52it/s at 512x512x8 (total 13.74s)
Vlad got 1.44it/s 512x512x8 (total 12.88s)

So Vlad was faster in the end

trying single

[wispy nest]

it's made a few of them lol

[smoky oak]

A1111 won the single image dra race

A1111 9.74it/s - 2.11s
Vlad 9.01it/s - 2.25s

[dense tapir]

for me there is just no comparison Vlad is simply faster in everything on my 1060

[smoky oak]

I wonder why its not for me

I am gonna give them 50 samples to drag race them harder

[dense tapir]

I am baffled

[wispy nest]

ah, changing the default temporary directory will allow it to clear it. that's a relief

[wispy nest]

this was the case on my 3060 as well

[dense tapir]

sytan is a 3060ti

[wispy nest]

yeah, i know

[dense tapir]

seems so odd

[smoky oak]

A1111 10.5it/s - 5.20s
Vlad 10.1it/s - 5.44s

at 512x512 50 samples

They do produce identical results, which is cool to see

[dense tapir]

although I have seen about 3-5% of the comments saying it was slower

[smoky oak]

gonna try with high res fix

[dense tapir]

They never mentioned their card though

[wispy nest]

i've talked with Sytan quite a bit before 😛

[dense tapir]

No, I meant the 3-5% who said it was slower for them too

[wispy nest]

ah

[smoky oak]

Interesting

👁️

Vlad is faster at high res fix

[dense tapir]

btw, the memory management is out of this world with this as my vram is full so it is paging it in and out (I can tell) yet I do not have lowvram or medvram on

[wispy nest]

it definitely does something different with hi-res. i read that vlad has multidiffusion as a buiilt in extension, so i imagine that's what does it

[tired basin]

1060 may not be new enough to take advantage of something in the cuda library

[dense tapir]

vlad even has the video extension too

[smoky oak]

A1111 - 10.6it/s 512 - 1.92it/s 1024 - 13.63 total
Vlad - 10.05it/s 512 - 2.06it/s 1024 - 13.17

[smoky oak]

I have a 3060ti

[wispy nest]

i've personally only dabbled with it, but apparently it helps with memory for VAEs? i dunno

[smoky oak]

this is a very interesting result IMO

[wispy nest]

that may be why it doesn't OOM on vlad with high batch size while auto1111 dies

[dense tapir]

damn

[smoky oak]

512x768
A - 6.84it/s 7.94s
V - 6.6it/s 8.11s

very strange

[smoky oak]

let me try a full workflow run and see how they compare total time wise

[smoky oak]

oh. i haven't been checking VRAM

paying attention to that now

[dense tapir]

UniPC

It seems slow so what is UniPC's claim to fame?

It messed this up at 30 steps

[languid holly]

Hikari

[smoky oak]

Vlad 512x512x8 batch = 1.56it/s 11.95s 4810VRAM 3143RAM
A1111 512x512x8 batch = 1.64it/s 13.41s 5794VRAM 3114RAM

Vlad 512x512 to 1024x1024 = 10.15it/s 512 2.07it/s 1024 9.97s 8192VRAM 4436RAM
A1111 512x512 to 1024x1024 = 10.6it/s 512 1.96it/s 1024 10.35s 8192VRAM 4390RAM

so yes, Vlad is faster here

[dense tapir]

I am not seeing the lure of UniPC as the images are rather, well, ugly.

[smoky oak]

UniPC has just been worse and slower DDIM form my testing

[dense tapir]

Yeah, same

it has to have some reason in being

[smoky oak]

feel free to take a look at the benchmark numbers above I just posted

[dense tapir]

my batch is slow when I tried it yesterday

over 3m for 4

ddim 30

[smoky oak]

512 to 1280 is 17.05s on A1111, 15.73s on Vlad

both ran out of VRAM at 512 to 1536

drag racing 150 DDIM steps

A1111 10.55it/s 16.58s
Vlad 10.11it/s 17.05s

for 150 steps

trying 768x768 to 100 to see the higher res perf

interesting

Vlad 4.30it/s 24.24s
A1111 4.30it/s 24.22s

HMMMM

time for 1024x1024

to 50

Vlad 2.06it/s 25.47s
A1111 1.96it/s 27.61s

A1111 wins at low res, vlad wins at high res

how bizarre

[wispy nest]

ok. i may have asked too much of the hires fix lol

[smoky oak]

bruh lmao

@wispy nestso in high res fix Vlad uses less VRAM, but in batch, A1111 uses less

512x512x6 A111 7921 / Vlad 8104

A1111 used less system RAM as well

2853/2880

[smoky oak]

@dense tapir Oh god, Vlad sucks ass for ultimate upscale-

Identical settings-

also, it doesn't use the upscaler you tell it to use

no matter what, it uses the super slow swinIR

Which means A1111 took 11 seconds to upscale, and Vlad took 49

ok, something is wrong here

its upscaling 2 times

Hmmm @dense tapir

Fixed the double upscale :>

[sinful mirage]

👽

[smoky oak]

nice lol

[sinful mirage]

wasnt able to use hires fix much in auto, having lots of fun with it in vlad lol

[smoky oak]

unfortunately, ultimate upscaling seems to be really bad in Vlad

[sinful mirage]

havent gone past latent yet, hopefully its fixed by the time i get there

[kind quartz]

the fame of uniPC is that you can make 12 step and having quite good preview, and you can then render by Euler type of sampler not A version to know how final will look

[smoky oak]

oh, latent upscale sucks for high res fix. Try out ESRGAN_4x at a low denoise like .25, it works way better

(way better in meaning it doesn't change the image dramatically)

[smoky oak]

Its that what DDIM is for as well? You can get excellent results from DDIM in 15 steps, and its considerably faster than UniPC

[kind quartz]

i am not sure. I will send you 1 step

[smoky oak]

my gens I do are 15 steps in DDIM

[sinful mirage]

ah ill probably use that one for post process, i usually like the random detail, right one is latent

[smoky oak]

totally understandable

[sinful mirage]

eh pros to both, ty for the tip :)

[smoky oak]

I do tons of gens till I find one I like, and I don't like when high res fix changes it drastically haha

[sinful mirage]

ah ye i feel that, so many upscalers now, seems a few are pretty recent

[ripe cedar]

Where are the denoise settings for upscaling?

[sinful mirage]

in text2img -> highres, not sure if theres a post processing setting

birb

[ripe cedar]

I have an image already made in another folder, so I need to upload it into the UI before upscaling, so text2img won't work

[smoky oak]

Upscaling doesn't work that way

That is pixel upscaling, not generative upscaling

[ripe cedar]

I can't upscale an image I generated and saved in another folder? I have to do it after its generated within the UI?

[smoky oak]

if you wanna do diffused upscaling, you need to use img2img

[ripe cedar]

Ok, I'll play around with that. Its a noob thing so I'll learn. But, a different issue:

I used an add-on to make a mask as a test

[spiral hill]

how to create pictures?

[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!

[ripe cedar]

I used these settings. For some reason it works when I do 512x512 but when its the same res at 1920x1080 I get OOM error Edit: ok hang on, it keeps showing that but in another area in the add on section it shows the generations perfectly, which managed to change the hair color as I was trying.

[spiral hill]

։

[ripe cedar]

You have to download SD. There are lots of tutorials on youtube showing how to install

[smoky oak]

I sent a message literally right after you saying how-

[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!

[ripe cedar]

I used Segment Anything to more precisely select a mask of hair from the original generation. However, you can see it draws hair within the mask, rather than changing the color in a normal way. How to do it properly?

[smoky oak]

Does anybody happen to have a recommendation of where to start learning python? I have a potential job opportunity and it is paramount that I start learning python as soon as possible

[wispy nest]

personally i learned it just watching youtube tutorials

but i think the kind of tutorial you want depends on your background and context

if you know other languages you may just wanna read a syntax sheet

and get the quirks

if you're new to coding in general you need a tutorial that has all the basics

[smoky oak]

I know 0 code, but I am an extremely fast learner. I need to learn how to properly interface with diffusers, which my friend is telling me is extremely easy, and his place of employment is hiring, so he's offering me a sort of suggestion of what to learn to apply

[wispy nest]

python is very easy to learn because it has a very intuitive syntax

[smoky oak]

He said my experience with SD and image generation in general as well as model training is already a lot of positive things, and if I can prove myself useful with basic diffusers interfacing, I could land myself an extremely well paying job

[wispy nest]

it is like a gold rush right now

the last one

[wispy nest]

https://wiki.python.org/moin/BeginnersGuide/Programmers.

btw, AI told me to send you this

[smoky oak]

Thank you!

[dense tapir]

@smoky oakI am losing my patience with Vlad.

[smoky oak]

That was short lived-

[dense tapir]

https://github.com/vladmandic/automatic/issues/553

I gave everything I was given about the OOM issue but it apparently wasn't enough. I can't shit information I wasn't given.

[smoky oak]

That was a real rude response from him

[dense tapir]

Yes

[smoky oak]

big yikes IMO

[dense tapir]

Hence, my response. Seriously, he was given all the information I was given.

I have to hand it to the InvokeAI devs they were top notch with a bug as I gave them the info and they remoted in to my machine and saw it then fixed and a PR all in under an hour.

[smoky oak]

Thats the type of development we as a community need

[dense tapir]

I agree

[smoky oak]

I'm currently taking the very anxious dive into python

[dense tapir]

good luck. Python isn't hard it just demands formatting that is beyond stupid. If you never programmed before you shouldn't have an issue with it.

[smoky oak]

I haven't, luckily

I have heard from many people that python is quite easy to get into. Here is hoping. I have a once in a life time opportunity being hinged on me learning python

[dense tapir]

As I said it is super easy and why we have the majority we do in this community merely script kiddies. It was actually written to be just that so monkeys could use it, no offense. Make it super simple so you idea(s) could be tested. The rapid prototyping of the computer programming world.

[ripe cedar]

Might want to join a python discord as well if/when you get stuck, to speed up the learning process

[smoky oak]

Gotta love when you have a python error before you even start the tutorial

._.

[dense tapir]

Don't get frustrated and take your time.

[smoky oak]

good call, I'll have to take a look into that

I am doing a tutorial that does the typical "hello world" thing

and before I even type anything, I have errors lmao

[dense tapir]

probably that tab/space shit that bites me hard

[smoky oak]

no like, making the project threw errors lmao

I wasn't even able to type before it was already giving errors

[dense tapir]

coming from c++/C background we aren't that strict about whitespaces.

well, damn

no clue about that

[smoky oak]

._.

I click on debugger and I get an error saying it can't load lmfao

[surreal thistle]

Benedict Cumberbatch as Genghis Khan.

[wispy nest]

at this point i only want to see metamemes about the AIs internal representations

https://tenor.com/view/venom-tom-hardy-screaming-gif-11035878

[surreal thistle]

You mean how the AI sees itself?

[wispy nest]

no how the AI sees

because it is a clever way of data compression and we are extracting data from it by decompressing it computationally

and it probably sees similar to us, by abstraction of perceived patterns into components that in combination can recreate a much broader set of patterns than the originally perceived ones

it compresses the information but also expands the space of possible perceived patterns

https://tenor.com/view/think-emoji-thonk-meme-gif-11987870

[surreal thistle]

https://tenor.com/view/big-brain-gif-25037715

[wispy nest]

right, so we explore latent space with our clunky word input

but in the full latent space, there are more things to describe outside of the word input

because the way it is stored or compressed, the model itself, it allows for a much broader application because it is abstract

so how do we access that through just words, it is a very specific combination of words that opens access

holy shit just had a stroke writing that

[surreal thistle]

Get a life. /s

[wispy nest]

the more words we use ,the more precise the location in latent space we explore because it is like an anchor

or something, i'm not a datascientist

i'm just imagining things

[surreal thistle]

AI conjures images out of nothing, they didn't exist before.

So they can't be mapped onto a specific location in the latent space.

[wispy nest]

no it just abstracts the essence of the data we feed into it and generalizes it

and we explore the space of possible images based on the concepts it learned from the data

[surreal thistle]

Now that makes sense.

[wispy nest]

and the sentences we put into it, or edge data, or pose data, etc, it just serves to move around that abstract space

[surreal thistle]

你很聪明。

[wispy nest]

maybe some math genius can elaborate if there are sets in the output of these models that are not accessible via the input methods we have

[wispy nest]

barely

[surreal thistle]

Turing可能帮助你。

[wispy nest]

we have less than one year until AI has juiced out every smart person on the planet and we have nothing more to add to it

[wispy nest]

i will ask AI

maybe it knows

[worthy bobcat]

#image https://share.icloud.com/photos/0de4txk3W7RlKQ1WJvoMCw9Tw

[dense tapir]

@smoky oakVlad informed me that his fork uses more vram than auto's does.

[smoky oak]

ok, so I was RIGHT

it used more in all of my tests

[dense tapir]

He said that my 6gb and smaller are too old.

I am lucky, then, that sdp uses less mem but that is why xformers no longer works for people.

not sure how, or why, his would use more vram since he is a fork though.

[smoky oak]

no idea

[dense tapir]

he said people who said rez 1234 will now have to use 1233 (iow they will have to reduce their resolution).

if they wish to use xformers that is

I know I will not be submitting tickets henceforth, that is for sure.

[smoky oak]

hmm...

seems very weord

*weird

[dense tapir]

What is the default interrogator in auto?

I can't use this one in vlad

[smoky oak]

I don't recall

[dense tapir]

There is a new game in beta people are complaining about as it uses up to 19GB of vram on normal settings and on a 4090 can't even get 30fps at 1440p

as others have said 16gb is the new min and you must have DLSS/FSR to make the newer titles playable.

engine is Unreal Engine 4 I think he said

[smoky oak]

hmm

weird

[dense tapir]

devs who design are saying be prepared as it is what it is and expect these demands going forward as no one optimizes now

[smoky oak]

I know cyberpunk 2077 OD is extremely hard to run, but for good reason

[dense tapir]

we have gone back to the 1980s and 1990s where the mantra from software devs was "just buy better hardware".

leave it to their devices we will all be needing 128GB vram cards within 10 years

which isn't going to happen

[smoky oak]

I'm wondering why most people abandoned texture streaming

it works damn good, and didn't cost really any extra time or money

[dense tapir]

I don't know

I am still wondering why the industry just will not optimize now. I think laziness since they now have hardware to take up the slack so it will keep getting worse and worse

then they will hit a brick wall

[smoky oak]

The ones that do still optimize are not employed at major studios anymore it seems

although, there are still really good games being made, just not by the studios with big history in their names

[dense tapir]

agree

yep, I cannot use interrogate using the exact same model in Vlad's as it simply is OOM. In auto's no issue

I really need at least 8GB for vlads

[ripe cedar]

OOM is a problem when generating even a single 1920x1080 image with ControlNet in Vlad. Could make batches of those in auto. Also, black generations.

[dense tapir]

I only got the black gen with 2.1 and turning on deterministic SDP (no memory optimizations iow).

[ripe cedar]

The other major problem is that it gets errors using the extension segment anything if I have to edit a preview. I couldn't even install it in Auto though

[dense tapir]

Well, if this did not have pytorch2 with SDP I wouldn't even be able to use it for anything so all a moot point if that had happened.

the extension problem is not vlad's issue it is poor programming by the extension authors. For instance WD14 tagger had to have 1 line changed and one removed/commented out because of the way he wrote it.

Now I have an extension (SAG) that just doesn't work. I suspect it has to do with how it hooked into auto. Stuff that hooked probably ceased to work.

[dense tapir]

He said it worked on an older commit but on the latest it ceased to work. I let him know.

Hmmm, I had a 4090 check and this worked in vlad. Must just be vlad and my old card

might be a pytorch2 and old card issue or something /shrug

[sinful mirage]

blueberry

[dense tapir]

Why?

@smoky oak

[smoky oak]

peeng

I have no idea man

[dense tapir]

look above at the two images

see, wtf?!?

auto was never this random

[smoky oak]

I will say, when doing all of those owl gens, they actually did have very small variations in Vlad, even for me

which is weird cause in auto I have pixel perfect recreations

[dense tapir]

if they minorly change for you would explain why they drastically change for me

Got it.

[static tusk]

Guessing here, but its likely the variation seed - strength 1.3 ?

[dense tapir]

no the -1 variation seed

set it to 0

[static tusk]

in auto thats under the extras tick box

[dense tapir]

yeah, I never use it as it was for batch work and I never click the extra box for eons

that leaves just the SAG extension so I hope he can help figure out why it doesn't work even after I just reinstalled it

[tired basin]

you got a url for the SAG extension?

[dense tapir]

@tired basinhttps://github.com/ashen-sensored/sd_webui_SAG

[dense tapir]

Tat watermelon looks sus

[surreal thistle]

Some of my best girls today.

[surreal thistle]

At least give me some feedback, like what do you like and not like.

[surreal thistle]

@dense tapir

[dense tapir]

Not sure how to answer that

[surreal thistle]

I need to know which of my girls you like and which ones you don't.

Public opinion is important in all cases.

[echo terrace]

a cat

[dense tapir]

First one has more feminine features

[surreal thistle]

This one?

[dense tapir]

These two are the most feminine with the first holding the crown

[sinful mirage]

I saw some images made with sd that combined 2 prompts by having it do the first half of the render with one prompt and the 2nd with a separate keyword, does anyone know what this is called or how to do it? ty :)

[surreal thistle]

Yes, I can agree.

These two aren't as feminine.

[dense tapir]

No, they aren't.

@hasty nova

[surreal thistle]

These are though.

[dense tapir]

First one for me then 2 then 3

3 a bit more masculine

[surreal thistle]

Is there a combination of words to give them more feminine features?

Cause I'm sure I gave the AI enough information that they're supposed to be girls.

[dense tapir]

not sure

[surreal thistle]

They're still all fine though.

Tomboys are appreciated as well.

[sinful mirage]

found it ! I used "[dog | beautiful valley with trees]" in the prompt with simple additions like highly detailed and photorealism, it combines them in a different way instead of having a dog and a valley with trees

[surreal thistle]

That's basically all of the good ones.

[sinful mirage]

([portrait of a skull | beautiful valley with trees]:1.1), highly detailed, photorealism, concept art

[rain gazelle]

evenin'

[dense tapir]

Fem with a streak of tom boy you better look out in her

[hasty nova]

Looks good @dense tapir

Wait is that clown one dd?

[dense tapir]

which one?

[hasty nova]

I guess the last two you just posted

[dense tapir]

They all are

[hasty nova]

Damn

Those are things I didn't think it could do, are there embeddings?

[dense tapir]

Why I can't wait for V2

1 HN and 2 locon I made. I need to check them out with my old embeddings

I never leave home without my HN I made

[hasty nova]

I see, that's how you are getting that style

[dense tapir]

I guess

[hasty nova]

Looks awesome

[dense tapir]

I miss making embeddings

[dense tapir]

trying that last one with one of my embeddings

I don't think I made this one but can't remember

No, I didn't I remember now

[hasty nova]

Still looks cool

[dense tapir]

the thing about embeddings is you just never know how they react with different models

same, now, with lora/locon

This is my embedding

My embedding folder is a hot mess

[hasty nova]

None of my other embeddings are for 2.1 so I can't test them and I can't make them anymore

[sinful mirage]

i need some positive embeds, only have the neg ones

[dense tapir]

Why not?

[hasty nova]

Digital Diffusion used to be an embedding

Because xformers causing issues and such

[dense tapir]

then move over to pytorch2 and SDP

[hasty nova]

I am but I'm not sure that's working with training? I'll have to test it again, but I tried once and oom

[dense tapir]

I do not know but it should. The problem is the devs. Embedding inside Vlad is SDP

[hasty nova]

Oh, well then I might try again

[dense tapir]

Yeah, I could train in auto using xformers but so damn slow

[rain gazelle]

evenin'

[dense tapir]

AMD is about to follow in the Nvidia footsteps and do a stupid release for over priced 8gb card

7680 I think it is

You see no one out here gives one damn how much it cost them to make so if you can't make it for what people are willing to pay then don't make it. They are starting to wise up to that. A VERRRRRY hard pill to swallow for them.

using my wolverine embedding I made last year

back when all this was fun and people were making all kinds of embeddings on here

[sinful mirage]

ye i remember u from months ago, still havent gotten into any sort of training myself

[dense tapir]

Well, I miss those times it was damn fun for us all.

Now it just seems to be all cloud based, or going that way, with porn galore and if it isn't anime related no one cares.

I expect it to pick up with SDXL then fast die back down especially if it is going to have near to, or the the same, hardware reqs of that dumb deep floyd

We'll see though

[sinful mirage]

im at 8gb so im pretty much stuck here lol but i can do a lot with a little these days so who knows

the optimizations have come a far way lol like 5x

[dense tapir]

10-16gb for SDXL expected and 16-24gb for deep floyd

[sinful mirage]

i used it a lil and the img2img is pretty good, gotta try it more

[dense tapir]

well, I can't run it and I refuse to touch web based

[sinful mirage]

same other than some huggingface things to try the new chatbots, nothing too crazy so far

[dense tapir]

Here we go. Let's go to the circus

[hasty nova]

I'm with an 8g 3070 and I locally created the model @dense tapir is using right now.

And tons of embeddings that are of decent quality, one even has almost 9k downloads.

So yes, you can do a lot with a little.

[dense tapir]

You know in Vlad's I just can't get inpainting to work right. I dunno if there is some setting or what is up

[hasty nova]

What's wrong with it?

[dense tapir]

I use it to fix faces in auto but here I do the same and it doesn't work. really weird

[sinful mirage]

whats ur name on there? cant find it

[hasty nova]

On civitai it should be junglerally_

Just look up Adventure Diffusion, it's a 1.5 embed with a bunch of dls

The model general is using is my latest creation, a checkpoint (not the embedding) by the name of Digital Diffusion

[dense tapir]

it doesn't even match the colors

inpainting needs its own set of settings

[sinful mirage]

oo ye ive used adventure diffusion a bit and just got the clrs one today, ill have to try the model

[hasty nova]

Oh neat. Keep in mind the model is for 2.1 and not compatible with any of my embeddings.

[sinful mirage]

([advntr | clrs | skull]:1.1), paint splash, dripping ink, photorealistic, tropical colors, highly detailed, Greg Rutkowski

[dense tapir]

yeah, it is broken

changes the color and all

You can see it happening in the live preview

[dense tapir]

that prompt was "along came a spider"

I'd play that video game

Probably need a 48GB gpu though

in times of need one always has death

[sly gazelle]

prompt is at https://bashable.art/generate?generationId=6444d25748de44c55fc1f529

[ripe cedar]

Anyone use ComfyUI? It's pretty reliable, although nodes aren't for everyone and some functions are cumbersome

[dense tapir]

@ripe cedar I love nodes having used DaVinci Resolve a lot but comfyUI is far from comfy. I wish he would look at how resolve did them as they are more in the background if you want the power while giving you a basic interface at the same time.

[brazen anvil]

i use comfyui, i just use a1111 for inpainting or testing new things or tools

[ripe cedar]

I've been trying to combine ControlNet with GLIGEN (basically a regional prompter) but can't get the ControlNet node to connect, as the KSampler can only accept 1 positive input. If I connect it, thenit loses one of the GLIGEN nodes input. It's a lineart ControlNet with multiple characters; I'm trying to see if its possible to give each character specific colors on hair, clothes, etc.

[brazen anvil]

can you share me the workflow to see it in detail? I'm more of the type that likes to see things further apart and not so close together

for example,, this is my playground xD

[ripe cedar]

Like this? I think discord takes out metadata from images, and also removes json files

[tired basin]

Nothing about that looks simple/user friendly 😛

[ripe cedar]

Yeah the name is ironic. But I think for power users its actually really powerful. It takes time to set up but once its done, its possible to spam complex workflows because its already set up

Its also quite fast, and rarely bugs out

[brazen anvil]

Have you made images with multiple characters before? I think the easiest way would be using an AreaComposition, either in the "vanilla" way that comes in the wiki examples or using custom nodes, let me review Daveman's nodes and I'll tell you

https://comfyanonymous.github.io/ComfyUI_examples/area_composition/

[ripe cedar]

Yes. However, I wanted enough control to make consistent characters, which is why I used 3d and extracted line-art to use an input for ControlNet

The composition is easy to make with ControlNet; I just need a way to ensure color control on each thing like the hair, shoes, etc.

I was trying to use the regional prompter extension in Vlad, but it got errors

I also tried to use Segment Anything with GroundedDINO in Vlad and I got OOM errors and others

[brazen anvil]

you have this custom node? https://civitai.com/models/24537/comfyui-visual-area-conditioning-latent-composition

[ripe cedar]

I also tried to use instruct pix2pix and it changed the hair color to what I wanted at one point, but then changed the entire picture's color for other stuff

I didn't download it

[brazen anvil]

im having dinner atm, im going to check it in detail in 20 min and try gen something

[ripe cedar]

K ty

I'm having treatment at the doc tomorrow and it's late here so I'm off the PC. If I'm awake I'll see what's up in bed

If there's a way to use img2img with ControlNet for the colors, then that could work too. I can flat-shade the 3d characters and the result would be a direct overlay on top of the line-art (if need be) because the image is if the same scene, just rendered differently

For example, this image was made in Maya (nothing to do with SD). However, it's a combination of line art and color, each which can be its own layer. So if I could use that line-art and it's base flat colors, SD could add its own shadow, clothing folds, style, etc.

I used that character (on the right) to make a pose, then line-art, and then feed the line-art into ControlNet. But I just can't control the color with prompts alone. But I'm impressed with how ControlNet kept the hoodie strings, hairstyle, etc. I didn't clean up the line-art (lines were messed up because I didn't do the weight painting properly in 3d) so there are some messed up parts, but it's really good in what it does. What's why I only did color prompts, no compositional prompts.

[wispy nest]

for reference, that is exponential growth on a logarithmic scale, meaning it is double exponential

i'm sure that we will solve a lot of problems very fast

[wispy nest]

i'm testing a prompt engineer

so far pretty good images

[kind quartz]

cant get godzilla to work properly and sudently nothing but it.

[wispy nest]

what is your prompt?

maybe i can help

[kind quartz]

no thank you, i modified it, but should be part of name of image 🙂

[kind quartz]

she is going forward unstoppable as Terminator 😄

[dense tapir]

🙂

[edgy shard]

From this and me wanting my original character art design remade

To this, which took me quite a while but was WELL WORTH IT.

this took me about 3 hours? i'm not sure if i'm doing good yet or not lol

[distant cloak]

niceee!

[edgy shard]

i am losing my mind at how incredible this is

[distant cloak]

its magic fr

[edgy shard]

wow that's crazy

[distant cloak]

hehe thx

[edgy shard]

orangemix?

[distant cloak]

this is ghostmix

[edgy shard]

looks like it could be a merge

Huh.

[distant cloak]

u havent tried ghostmix??

[edgy shard]

no i am trying to go for more "anime movie" production quality visuals

lmao.

[distant cloak]

ahh. which model r u using atm?

[edgy shard]

counterfeit

freaking love counterfeit, whatever it sourced from is just nuts

i get huge consistency from it

[distant cloak]

it is pretty good. but imo I like MixProV4 better for anime

[edgy shard]

yeah i downloaded mixprov4

it was very aggressive

[distant cloak]

ya thats true

for your character are u specifying each color and clothing in your prompt?

[edgy shard]

i inpaint color.

[distant cloak]

im wondering how i should do it cuz it looks fun to try something similar

[edgy shard]

sure let me show you

[distant cloak]

ooohh wow

[edgy shard]

sec

[edgy shard]

https://www.twitch.tv/videos/1805649752

turn your volume down a bit i get loud (haven't hooked up my compressor yet)

[distant cloak]

wooow i didnt even know color inpainting was a thing! thx sm for showing me this

[edgy shard]

sure man :D

[distant cloak]

now i wonder if there is some way to "save" these colors and apply them automatically to another prompt

[edgy shard]

not sure, inpainting saves the headache and takes seconds

[distant cloak]

mhmhm

[edgy shard]

https://clips.twitch.tv/OnerousEmpathicPonyHassaanChop-NDJ4sW1YvgGLcipW one more

[thin roost]

Only missing the pointy party hat

[distant cloak]

Looool that reaction

[edgy shard]

LOL, to the side and a party favor

[thin roost]

I like the idea of a warrior character with a party hat/favor

[thin roost]

Gringotts stargate ?

[dense tapir]

Pretty much

[kind quartz]

would be fun to add some empty speech buble and fill it with anything 🙂

[smoky oak]

@dense tapir I'm making pretty rapid progress in python IMO

[dense tapir]

told ya

[smoky oak]

been doing it for about 2.5 hours now, I am already doing stuff like this

I made this, then googled cause I knew there was a more efficient way to do it

found the while command and boom

Pretty proud of that haha

[dense tapir]

a never ending loop while 1=1:

[smoky oak]

what do you mean?

[kind quartz]

if you try ask AI, it should help you even more

[dense tapir]

I mean it is always 1 therefore the loop never ends

[smoky oak]

I am not gonna rely on chat GPT for code. Even a beginner like me can see how inefficient it tends to be

[smoky oak]

it only loops if the number is outside 1000-2000. I already tested it

[dense tapir]

while 1 = 1: is a never ending loop

[smoky oak]

oh, I thought you were saying I made it wrong and it would never stop, my bad

yeah, its never ending while that value is outside of the range

[dense tapir]

It is used when another input will break it out

of course you add a sleep in it else it will bog your cpu down

doing a while 1 = 1

[wispy nest]

the magic is not just in the data

you can have infinite pictures about a pencil and would still know as much as if you had 10 pictures of pencils

[wispy nest]

i think the magic is about this https://arxiv.org/abs/2204.01437

[wispy nest]

https://www.semanticscholar.org/paper/Does-Deep-Learning-Learn-to-Abstract-A-Systematic-An-Lin/2ae807688d5bae0a7331992793be066b93d7655f

and this

A set of controlled experiments are conducted based on this framework, providing strong evidence that two probed pre-trained language models (PLMs), T5 and GPT2, have the abstraction capability

so we probably dont need massive amounts of data but sufficient amounts of data to find out how to create architectures that perform better at creating abstracts and translating them into concrete hallucinations from input vectors

[kind quartz]

They usualy have two rows of teeth

[wispy nest]

synthetic data

[pulsar jackal]

logo

[wispy nest]

novel architectures

we just need to make smarter models that need less data

and not just models, but also ways they can interface with themselves for even more accuracy

there is a project that managed to increase accuracy by allowing the model to reflect on outputs before sending them, evaluating them in a self-conversation

like a thought process

its not all that black and white

ofc real data is needed

and who cares about copyright to learn concepts

if people don't want others to learn anything from their content, they should not release them

cannot prohibit thinking and learning

[kind quartz]

AI and programing is very effective. Just one must have some level of programing ability, asking right question and as said, knows how to program

https://media.discordapp.net/attachments/613357739846139924/1101455639517085746/image.png?width=908&height=700

[ripe cedar]

You mean LAION? There was a post about this. Interesting discussion. Apparently, it shouldn't be too difficult to do: https://www.reddit.com/r/StableDiffusion/comments/1313939/an_indepth_look_at_locally_training_stable/

[tired basin]

Im guessing you didn't get to the comments - the claims made down in the comments were not the sourcing of the images, but the quality of the tagging/captions that dictates the "quality" of the database

and thus the eventual quality of the resulting model

You were discussing the database, not the technology...

Agreed, only those with access to the tech have the massive advantage, but in terms of image databases those are much more available

Dunno, but I'm guessing you're about to tell me....

There is already a hand pose database..

[kind quartz]

🙂

[ripe cedar]

I did. There was no consensus; the database in theory could be much smaller, and tagging could be done with AI now as opposed to not being possible before.

[smoky oak]

we are in desperate need of a closed model that can generate, and towards the last few steps run alterations or functions in order to correct things it itself sees as a problem. Like extra fingers, it could recognize there are too many, and inpaint it and then continue the gen process. I am sure that would take a let of a lot of compute and memory, but it could be cool

[smoky oak]

TOE WEEL TO TTHE

A YOU IT WETTHEY YOU OHE FOUBET

Truer words have never been spoken

brings a tear to my eye

[kind quartz]

it was because some arguing and any picture was needed 😄

[smoky oak]

👁️

💧

[kind quartz]

it is general with images, and wasnt needed. o.k.

[delicate sedge]

yo guys my friend is making a live 2d model and asked me to make the art for the character itself and i am not really sure how to make it, anyone got any ideas if there is anything i can add to the prompt or a lora or something that can help?

[smoky oak]

Is this supposed to be a response to me saying I didn't wanna waste my time with chat GPT for coding? Cause if so, that code it just made is massively more inefficient than what I made and I have liek 3 hours code experience lmao

it can code, but not as good as many people think

I was told and now realize that using chat GPT is a good way to get inefficient code

and learn inefficient workflows

[ripe cedar]

What do you need to add? There is Char Turner which helps make a character sheet

[kind quartz]

have you in your code things agains wrong input? 1000 included or not?

[smoky oak]

what?

[kind quartz]

what if somebody pressed M instead of number?

also < or <=?

[smoky oak]

the code chat GPT returned only asks 1 time

[kind quartz]

nothing is easier, just properly ask

[smoky oak]

mine keeps asking until you answer correctly

[kind quartz]

If you check code from AI, you can lear something as well.

[smoky oak]

it uses far more operations, and doesn't ask again for an input, I am not sure its better lol

it doesn't prompt a second input

[delicate sedge]

i mean the style itself and the details

also i dont need different angles

just her standing should be enough

https://nikke-db.pages.dev/v

some examples of the what the aim is

can this char turner make similar results?

[smoky oak]

I stand corrected, I see how its doing it now

[kind quartz]

because it has things like except ValueError for example

[ripe cedar]

I can't remember but I think it can do different styles https://civitai.com/models/3036/charturner-character-turnaround-helper

[delicate sedge]

alright i will see if it works

thanks

let me know if you got any other suggestions too

[ripe cedar]

You can find different but similar characters to the character you want. Then you can use the img2img function with high denoise settings; it can generate a similar but different character. You can use ControlNet with it to make the pose you want for it.

[delicate sedge]

oh good idea i will look into that

[ripe cedar]

You can make batches of them as well, and then pick the ones you like and upscale them

[kind quartz]

AI suggested some prompts, curious if example or always different 🙂

[neat echo]

#🏞｜general-with-images (full body visible), (thick thighs),ultra realistic 8k cg, masterpiece, pink bodysuit,shiny clothes,((ultra detailed background, delicate pattern, intricate detail)), formal, serious, queen, goddess, divine, (hair over one eye), absurdly long hair, very long hair, blonde hair, looking at viewer, nsfw, breasts out, lace, lace trim, lace-trimmed legwear, (displeased:1.5, anrgy, upset, head tilt, glowing red eye),white background, huge breasts, pussy peek, ocean

[kind quartz]

that is suggested to me

what is difference in prediction models?

[lament halo]

Cyberpunk guy

Hands still looking weird

[kind quartz]

@lament halo i got similar einstein 🙂

[lament halo]

Cool

Yeah I’m gonna try putting Einstein into something

[kind quartz]

einstein is well trained

[smoky oak]

uhhhh

what happened to /faq?

[kind quartz]

curious as well

[smoky oak]

tried to send it for the guy above, but all that comes up with table flip

hmmm

[kind quartz]

@stone cipher dont worry, same here with context

not sure if real like but like it 🙂

if no fingers, i cant say

i meant if is posible to see from outside to inside of spacesuit

i think this isnt real, because it reminds me one picture, where somebody looks over shoulder, but i can be wrong, one cant be sure

this is common problem, not sure how and if posible to solve it

planet on right and left side

yes got such pictures. But if you put something in front of horizon, it happen not sure how frequently

no i know have i been pawned 😄

This prompt work,
An astronaut floating in space, with a breathtaking view of Earth in the background
but those thing apears in midle ages as well as anywhere else.

curious if asking next time for prompt suggestion i get different words from AI

[open sun]

hey ppl, just wrote batch metadata remover script, It's easy to use, put images with gen data or any other metadata in "images" folder, run main.py script and it will clean all images and place them in "_processed" folder.

no plans for UI, if anyone chimes in and adds it, that'd be swell.

[kind quartz]

trying pseudo pixel art

[brazen anvil]

I tried different configurations to try to generate something before going to sleep but none worked for both characters at the same time, were you able to solve it?

[tiny wharf]

Does anyone know how to generate videos like these. Ive download sd locally and tried imgtoimg and interpolation with diff. Settings but unable to achieve something like this

[ripe cedar]

Yep, it was so simple! I already generated colored images from 3D (Maya), and overlayed generated line-art to keep the composition of the image. I simply used img2img+controlnet.

I actually posed the example on the anime chat here

[tiny wharf]

I meant videos like these where car evolves

[kind quartz]

Monkey with nicely enhanced M, love it

[crimson dawn]

make an embedding to fix it

[kind quartz]

thank you, dont take it negatively from me, i cant correct it (because i know nothing), but good there is a way.

[brazen anvil]

(masterpiece), (best quality), (official art, extremely detailed CG unity 8k wallpaper), (highly detailed), ((absurdres)), night, sky, a [tentacle moon|(red eye:1.2):.25], light particles, lunar light, looking at viewer.
It's not what I was looking for but I liked the result

[glossy herald]

A little favor/challenge I wanted to post to the artists around here.
I'm trying to finish/complete this oil painting my mother did and never finished.
I do get some results, but nothing really great...
I wondered if someone wanted to take a go at it, see what they manage to do. It's an important piece for me, and I've never managed to really make anything that satisfies me.
I have no specific criteria though, except staying close to the original style maybe ? To be honest, even a tribute is something that does feel good on this, but yeah, I was trying to "finish" it when I started on it

[smoky oak]

Oh whaatty

Deepfloyd?

That looks sick-

[outer spear]

Deepfloyd making some rather scary faces

[cyan shoal]

GREAT

[kind quartz]

where are you taking images IF?
Try text on perspective something, for example skyscraper 🙂

[kind quartz]

Faces are o.k. just flat fingers

[pseudo aspen]

Can someone help me with face heading using controlnet and openpose?

I want two characters looking at each other but I keep getting them looking at viewer

[outer spear]

prompt?

[pseudo aspen]

lyco:EnvyBetterHands-LoRA:1 , looking at another, masterpiece, sunlight, light shafts, best quality, perfect hands, nice hands, (perfect shadows, perfect lighting:1.3), (extremely beautiful and detailed eyes:1.4), 2girls, AND takao, ((standing)) schoolgirl_outfit, lora:Takao_Schoolgirl-1.0:1, looking at another, lying, smile, (golden eyes:1.4), AND lora:atagolora:1, atago, stunning_speedster, cute expression, perfect eyes, (golden eyes:1.4), light smile, ((standing)), looking at another

[outer spear]

sure is alot of stuff

did you trying putting "looking at viewer" in negative

[pseudo aspen]

I'll try, one sec

[outer spear]

I would start with a simple no other stuff prompt like "2girls, looking at each other" then at stuff on top

[oblique nimbus]

used your prompt

[brazen anvil]

to get something similar you should use darkSushi25D25D_v20 model and this config

[edgy shard]

i was a little surprised last night, an artifact generated with very clear text https://i.imgur.com/vqiljbI.png

[edgy shard]

then you show me this guiz lol

[glossy herald]

this is using Deep Floyd, that just got out

https://huggingface.co/spaces/DeepFloyd/IF

generated here

https://github.com/deep-floyd/IF

[smoky oak]

so deep floyd is gonna be a model we can download and run evenetually, right?

its different to SDXL?

[oblique nimbus]

I did it with https://dezgo.com/

[wispy nest]

was a lot more excited about deep floyd before i saw this

[dense tapir]

@marsh lily https://github.com/RadeonOpenCompute/ROCm/pull/2094

[dense tapir]

No Windows support and, of course, that page is full of hate.

It spilled over into a different post on github is the only reason I know.

[dense tapir]

What I know now about what is coming nothing less than 24GB will suffice for me. Let me know how it goes with training and speed etc...

SDXL I might be able to run at 256x256

[smoky oak]

I wonder if that stable vicuna model can be run on normal hardware or if its an unoptimized mess like stable LM 😅

[dense tapir]

SDXL on 8GB I am not sure will run SDXL even at 512x512

[wispy ether]

[smoky oak]

SDXL is 2.5x the size of 1.5, so no idea

but I was talking about the large language model

[dense tapir]

I know 16GB for 1024x1024 which is what I was telling people when 3.0 was announced. SDXL is just marketing speak for 3.0 they said would be 1024x1024 last year.

[smoky oak]

SDXL is not 3.0

[dense tapir]

I know you were but I am only concerned with SD.

[smoky oak]

3.0 is not as big as SDXL

I talked with devs about it in this chat before

[dense tapir]

Ahhh, cool but either way most people will stick with 1.5

[smoky oak]

SDXL is a bigger version of 2.x with new adaptations that are being tested to be implemented in the new and even more impressive release of 3.0 which is not as big

Thats what the devs told me at least

SDXL models are 2.5x as big as 1.5, cause they have 2.5x the params, however the dev I was talking about it with said there should be ways to optimized by pruning params for dedicated models

like if you want a realism model, you should be able to prune all the params not associated with that, is what they made it sound like

[dense tapir]

Aye

the problem is that has negs to it as well. No matter

[smoky oak]

Also, they did mention about the fact that SDXL/3.0 do use a drastically new text encoder

[lyric mica]

what

[smoky oak]

The way they talked about it SDXL is just a random offshoot of 2.x with an XL size parameter set

And SD 3.0 is supposed to take the benefits of that and then combine it with a much better data set as well to produce drastically even better results, from what they said

[smoky oak]

is this the new vicuna 13b stable LM one?

[dense tapir]

All I am seeing is more of a bitch to train. Sad

[lyric mica]

yeah I'm trying it here
https://chat.lmsys.org/

[wispy ether]

ok this one looks better

[crimson flower]

What was the fix?

[wispy ether]

3legs

nvm its 2

[dense tapir]

Depends on the extension. One of mine doesn't work, doesn't throw an error either so I have no idea what is wrong with it.

[crimson flower]

Asking about WD14 tagger.

[wispy ether]

:I

any ideas how to fix the sword dissappearing

or the hand

[dense tapir]

from pathlib import Path
from argparse import ArgumentParser

default_ddp_path = "C:/Users/omega/.cache/huggingface/hub/models--SmilingWolf--wd-v1-4-vit-tagger-v2/snapshots/1f3f3e8ae769634e31e1ef696df11ec37493e4f2"

def preload(parser: ArgumentParser):
# default deepdanbooru use different paths:
# models/deepbooru and models/torch_deepdanbooru
# https://github.com/AUTOMATIC1111/stable-diffusion-webui/commit/c81d440d876dfd2ab3560410f37442ef56fc6632

parser.add_argument(
    '--deepdanbooru-projects-path',
    type=str,
    help='Path to directory with DeepDanbooru project(s).',
    default=default_ddp_path
)

make your preload.py in the tagger folder this, but with your location for the hugging model. The dev, that guy is never going to fix tagger, hes abandonned it for real, so atleast the hard ref to the location is set...i think i might even move the model directly into the tagger folder.