#💬｜general-chat

1 messages · Page 27 of 1

old marten
#

rule #1 - never use 2.x

#

its because someone dun goofed with how certain images were catalogued

#

so celebrities, artists, nsfw got messed up

#

1.5 is what sane people use except for niche situations

rain fulcrum
#

got it.
Thanks for the clarification

old marten
#

check the "tile" box

astral goblet
#

you know, gonna move this post to offtopic lol

crisp zinc
#

will Linux provide better it/s compared to windows?

astral goblet
#

but its not an optimization on it's own and windows generally works as goood

crisp zinc
#

Thanks

visual osprey
#

Hi, Guys. How are you?

Any ideas of a prompt that will show the OUTSIDE view of a room (from the window) ? Every resulting image I get is from the inside of the house, pointing at the window 😢

cursive gate
#

When it comes to fine tuning a model on different concepts with 2.1 and DreamBooth, can you train the model on different steps for each concept or does it have to be universal for all data?

grim aspen
mystic lintel
#

Heya guys hope you guys are doin fine here, wanted to ask a question , is there any other alternatives for clip interrogator?

fervent thunder
#

?

vivid herald
#

yeah looks like it's down globally

rain fulcrum
#

Uff

astral goblet
#

uhoh. conspiracy forces assemble

#

queue a huge section of the community to start theory crafting. in reality, it's probably 1 of 10000 possible scaling issues that popular file hosts often run inot, but by the end of today there will be people trying to cobble together a kickstarter / patreon for new crypto hosting or something

#

their host provider probablyw asn't aware of the traffic volume they'd really be pulling in when they sold these guys an unlimited plan for the first 2 years lol

cloud granite
short tapir
#

What are the chances of the AI becoming self aware?

#

They did a prompt for basically a supernatural cowboy. But instead of giving me images like that I only got one failed image where it was like all static and looked like a face in the static screaming.

astral goblet
short tapir
rain fulcrum
#

does stable diffusion uses VAE?

warm junco
rain fulcrum
#

thanks

#

I'm having problems with vae

#

that gives me a hint to solve the problem

signal nest
signal nest
# rain fulcrum does stable diffusion uses VAE?

this is an important point regarding the legal debate. It only uses a simple VAE as shortcut. The actual image generation is happening through diffusion (that happens to work in the VAE's space). VAE image generators can be used as compressors, because you can distribute the latent space code for specific images. SD can't.

short tapir
astral goblet
#

scifi about digital brain clones is awesome. are they self aware? maybe. there's one of those systems in neuromancer and i think it's my favorite depiction. robocop isn't quite there, but is like, the wetware of a brain digitally interfaced. thats cool too.

signal nest
#

in my view a functioning digital clone brain would be self aware. but this heads into philosophy , unprovables

astral goblet
#

yup. i am thinking that too. there's nothing "magical" about conciousness and it can be reverse engineered. sure is cool stuff, but it's part of reality and that means we can build it

signal nest
#

right it follows if you reject the idea of a supernatural soul. its "just" information flow. however, it's a lot of the right kind of information flow. we still dont quite know how to set it up, and it would be seriously expensive to do

astral goblet
#

towel, the most useful thing in the entire galaxy, isn't even an emoji. NOR is it planned as an emoji. i'm appalled at this discovery

#

actually i don't reject the notiion ofa soul, just that it's "super natural". if it exists it exists within natural laws. conciousness could very well be a field we haven't devised a test for yet. what i reject is that any phenomenon act outside of natural laws, whatever those may be

signal nest
#

IMO i think we've seen neural nets do enough already that it''s likely the understood aspects of physics are enough to do it, given enough connections (but i'm speculating, till its done no one really knows)

astral goblet
#

i think that machine awareness will happen before we understand why it happens

#

i'm really inclined towards the idea that intelligence emerges naturally from enough connections

signal nest
#

diffusion models and these lanuage models are doing feedback, so yeah.. interesting times..

astral goblet
#

what a time to be alive!

old marten
#

stop propagating this crap

signal nest
#

i've been well aware that plain backprop learning can't do the kind of interactive learning needed for AGI. but we see chatGPT able to absorb information from natural language exchange during one conversation, clearly they are layering enough temporary state for that. And diffusion models are sending a large amount of state through a net connecting spatial and verbal ideas. Diffusion models have also broken this limitation of "1:1" input/output mappings

old marten
#

I hate going on youtube just to see the annoying ass clickbait "oMG aI bECamE SentIENt!!1!"

#

100M views by idiots who dont understand how complex the human brain is

#

Give it another 100 years and maybe

signal nest
old marten
#

not with a bunch of 1s and 0s

signal nest
old marten
#

Being conscious is kinda hard to define tho. The AI can say it's conscious, but what if it's programmed to do so?

#

They would have to be 100% autonomous and have free will

#

which may not ever happen, ever

signal nest
#

and to be clear I dont think either SD or ChatGPT would do it. they're too static. But, they allude to capabilities that if wired up just slightly differntly, and scaled up sufficiently far.. i'd say it looks like AGI is within reach now

old marten
#

So at what point would we say it's conscious?

#

It writes its own code?

#

What personality would it take? Where would it get it from?

#

It would have to be as random as a human brain is

signal nest
old marten
#

no

signal nest
#

i'#ll be hard to define the exact boundary

old marten
#

well, I disagree at least is what I mean

#

we don't know if animals are conscious

#

to begin with... so yeah... its complicated

astral goblet
signal nest
brisk bloom
#

can you add denoising strength to a filename when saving?

astral goblet
#

i personally don't believe in magic

old marten
#

Im not saying magic

astral goblet
#

it can all be quantified and engineered

old marten
#

Im saying 1s and 0s is not enough

astral goblet
#

a deep learned neural net is a littel more than 1s and 0s

signal nest
old marten
#

So then, my question, at what point does it become conscious

#

is it gradual

#

is it suddenly

#

more calculating power = more consciousness?

#

I personally don't think so

astral goblet
#

thats what i mean. it'll happen before we even understand why it's happening, or if it was even happening at all

signal nest
astral goblet
#

one day we'll just be like "okay these have been concious for some time now, but we don't know why"

old marten
#

eh... if we don't know why... we wouldn't know how to make them in the first place.. thats my thought

astral goblet
#

and mine is it's a natural emergence of complex information systems

old marten
#

but maybe you're right. I think for the most part folks dont truly understand neural networks to the core rn

signal nest
old marten
#

but what I'm saying is that there's a hard gap

forest apex
old marten
#

if you put together all the GPU's in the world

signal nest
#

the hard gap is the scaling and figuring out the right arrangement

old marten
#

and make a neural network out of every single bit of information in the world, it still wouldn't be conscious

limber summit
astral goblet
#

deep learned networks have absolutely started throwing tendrals over that gap. we're even beginning to understand intelligence to a point that we're reclassifying many life forms as sentient

old marten
#

ur mistaking calculating power with consciousness

#

neural networks make calculations

#

tons and tons

#

but still might never be conscious... "might"

astral goblet
#

there's no real definition of consciousnes. it's just a state we use to describe part of the human condition. biologically, it's all just computations

old marten
#

Consciousness is a true sense of identity

astral goblet
#

unless you believe in some sort of magical power to it all

old marten
#

A computer doesn't know what it is unless it is taught, it doesn't know where it exists, it's not autonomous, it runs only with other people's instructions

signal nest
#

well i can't prove this , but as I dont beleive conscioussness has any supernatural component, and what we see in the natural world is predictable through computation (at most scales), then IMO it follows consciousness is indeed 'enough calculating power' , wired up correctly. Both are difficult in practice. The human brain has a huge number of synapses, connected in 3D.

old marten
#

we make slower and less powerful computations but we have a deeper understanding of the world than just 1s and 0s. at least that's my take on it

#

Well even if it's gradual, you can't teach a dog to communicate

#

in language

signal nest
#

it would take 16 petabytes to describe the human brain's connections. We're not using nets this big

old marten
#

u cant even teach chimps to use language, sign language monkeys are a fraud

signal nest
#

software is more maleable than chimps

old marten
#

so u can make a computer as smart as a dog but is it really conscious at that point if it can't use language?

#

yeah but chimps brains are infinitely more capable than any computer

#

so, you would have to surpass even that

#

with bits and bytes

signal nest
#

langauge isn't the dividing line. what if someone was mute, or born ferral, and had ann accident damaging the vocal chords. you're doubting that they'd be conscious?

abstract ice
#

True

signal nest
#

I think its probably possible to build with todays tech, but too expensive to be worth doing - and we'd have to experiment alot because we still dont know the precise config.

#

we've good hints of different aspects of human consiousness in various nets (and software systems) IMO.. stable diffusion is showing some kind of imagination, for example. whilst deepmind's game playing software shows hints toward the big picture (i.e. experiment with a game and learn to win it)

abstract ice
#

I don't feel stable diffusion or chat gpt have any actual intelligence behind them

#

Like that Google engineer thinking the chatbot was sentient lol

signal nest
abstract ice
#

Yeah they have tool level intelligence

#

It's picking the highest probability word to write next based on its training etc

signal nest
#

what it does wont change session to session, its not updating its net based on the interactions.. thats one big difference

#

but imagine if it recorded each session and re-trained itself including that latest interaction in downtime.. like we have sleep & we dream (speculation that this is associated with long term memory formation)

astral goblet
#

really, a waterwheel mechanism or any mechanism that performs repetitive tasks, looms especialy, they're all "AI" in that we've taken our intelligence, and turned it into an artificial source of it

#

machine learning came along, where machines can change up how they work over time, because of software. that was pretty crazy

#

society is built with all sorts of cogs of artificial intelligence. specialized little capsules of intent all over the place

signal nest
#

hive minds , with a lot of mechanical components already

#

civilization already builds things and makes decisions that are way beyond individual comprehension

astral goblet
#

yeah so many little bits of knowledge built on top of each other

#

city of ember was a really great metaphor for it all

cerulean stirrup
#

what a different between training with dreambooth and without dreambooth?

timber dock
cerulean stirrup
#

Did civitai just crash?

astral goblet
#

might help to do a hard refresh if they changed hosts. hold shift and click refresh

fervent thunder
#

never heard of a hard refresh before. gonna do it all the time now 🎊

lone siren
#

I've been playing with SD, trying to make some photos look like the were painted but I still can't get a satisfying result. The resulting image is either too different from the original, or not stylized enough. I tried searching for tutorials but the ones I've found weren't what I was looking for.
Can someone point me to the right direction?

gloomy elm
#

What's the prefix for the ai?

vestal dew
#

prefix?

wide juniper
#

Does anyone know of any models that can do interesting dance poses?

vestal dew
#

I think for any kind of action poses, depthmap would make a big difference

#

You'd need a model trained for dance positions.

wide juniper
#

What’s depthmap? Related to the depth2img thing?

#

Any models train for dance positions?

#

I’ve been using vroid outputs but it’s tedious

vestal dew
#

yes

#

The SD can create a depthmap, and then when you use that to generate, it does a better job of making distances, perspectives

wide juniper
#

How does that help with action poses though?

vestal dew
#

because in action poses arms and legs are posed with distance and perspective

#

SD as it is has trouble with perspective - that's part of why it mangles hands and feet, adds extra limbs

#

I havent used depthmaps. My card can't handle it

#

I just posted "man ballet dancing" in general-with-images. It mangled his hands and feet - has troub le with the perspective of the limbs

scarlet falcon
#

why is everyone using sd 1.5?

#

no one is making 2.1 models lol.

vestal dew
#

they dont like what was removed for 2.1

scarlet falcon
#

boobies?

#

really?

wide juniper
vestal dew
#

adepth map lets SD see perspective

#

I havent tried them - I cant generate them on my PC. I am suggesting they can help with this

#

I cant give details. I cant use them on my PC - dont have details

wide juniper
#

Ok so like are you saying if I have a pose I find on google, I use depth2img instead

vestal dew
#

I think it would help, yes

#

did you see what I posted?

wide juniper
#

Yes

vestal dew
#

SD does a bad job with motion, with placement of limbs in perspective

dire dove
#

Hey guys, if I use a mask to inpaint the background of an image containing an object in the foreground, the resulting image always seem to have a variation of the unmasked foreground object being covered by the original foreground object. Anyway to fix this? Thanks

vestal dew
#

I cant get inpaint to work for anything

wide juniper
vestal dew
#

yes, but I don't know if it's a model

#

im going to try it

#

maybe I can use it if not generate one

#

ah, I already havbe iot....cant run it

dire dove
#

https://imgur.com/a/EzBxHiJ
This is what im talking about... i inpaint the background of the bird, but in the result image, the original bird is overlaying another slightly larger bird with strange feet

vestal dew
#

that happens to me when I try to have someone behind a foreground figure

mortal musk
#

Guys, how are you?

I have a challenge, anyone willing to help me solve it, I'll be happy to pay money for the help.

I own an agency that develops caricatures. my business is not doing very well these days as I pay real artists to do the drawings, including myself.
I'm thinking of changing the way I work and try to integrate artificial intelligence in the creation of these cartoons.

For that, I would need to understand a few things:

1- Is it possible for artificial intelligence to learn 100% of my style? My style is like this:
https://uploaddeimagens.com.br/imagens/_LVhVPA

2-Let's say I train a model with about 100,000 drawings in my style (yes, I have a database with that many drawings), because I have such a large number of drawings, does this help in learning? or from a certain number, the quantity becomes irrelevant?

3- In my job, I look at pictures of people and make drawings of those pictures in my style. Can artificial intelligence do this same process? In other words, after I train a.i. to learn my style, is it possible to upload a photo of a person and wait for the artificial intelligence to deliver the drawing of that photo as a result, but in my style?

4- If you, who are reading this now, know how to do this and know how to answer these questions, can you send me a dm to help me make this happen?
I know it's not complicated, I'm not new to the field, I just need direction. If you want to charge to teach me, I can pay.

Im just a cartoonist trying to find my place in this new era. (and pay my bills)

dm me!

vestal dew
#

Why don't you try it?

#

You could also train a model with the images of the client. Then have it generate an image of the client. Then use that generated image with the same settings : seed, prompt, etc (PNG tab) using your trained style to generate the caricature

#

how would that change your business model? You'd fire your employees, charge less?

vestal dew
#

he asks for help, but he doesn't respond

brisk bloom
fervent thunder
#

hello , i want to use a new model called anything 4.5 model , i am using google colab , how should i choose to use this model ? if i click the custom will it choose that ?

vestal dew
#

are you using UI?

#

copy it into models/stablediffusion, then update models in UI, pick it

opaque sky
#

Guys… best models/embeddings for Dragon Ball style?

forest apex
#

the dragon ball z model? on huggingface

forest orbit
#

I'm going to install SD and 1111 on my computer at work. Since it's for work, would it be better to install SD 2? Seems that everyone prefers 1.5 but I assume that its for the NSFW capabilities.

warm junco
forest orbit
swift snow
warm junco
#

It wont load without it

jade bramble
#

@ocean raven Hi

forest apex
#

rpilocator FTW

opaque pivot
#

hi everyone .. i am new here

#

can you tell me where is the best place to start learning how to create

#

thank u

forest apex
opaque pivot
#

thank you i will check

dim copper
#

I am making a game with stable diffusion 1.4, but I'm worried about it generating NSFW content accidentally. I'm not going to use any safety checkers as that takes too much time. It will be for adults so a bit of tasteful nudity is fine. Like if the prompt is "manga girl" or something. I heard 1.4 is better than 2.0. I don't think anything too graphic has ever appeared in my images. Basically, if someone hacked my game and put in a bad prompt will I go to jail? 😯 On the other hand one could type bad things into the Chrome Browser and similar things would appear I guess. But the difference is that the bad images might not actually exist but just be imagined by SD. What do you think?

#

BTW, I'm not worried about it generating Mickey Mouse as I can just put the word "parody" in the description and then it's fair use 😉

warm junco
#

Also i would recommend using SD 1.5 or 2.1 as model

dim copper
#

@warm junco Thanks, I'm not using a webui, It's a custom build of stable diffusion just using the weights so it's built right into the game. In a game every second counts so it depends how fast those filters are. I may end up just filter over the text input which might be quicker.

warm junco
#

Ah okay good idea

#

Yea but then use 1.5 or 2.1 for better results

dim copper
#

I haven't tried SD 2.1. Is it really that bad, or do you think it is acceptable quality?

warm junco
#

No the Bad model was 2.0 everyone was complaining about

dim copper
#

oh i see. So 2.1 is just like 1.5 but a little less nudity?

warm junco
warm junco
dim copper
#

I heard that 2.x they removed some of the NSFW content and copyright content. But other people were saying it doesn't produce as nice artistic pictures. So, I guess I'll have to just try it for myself! A few more GB to download I guess 😁

warm junco
#

So they reverted in 2.1

#

But i think they also put less artist and nfsw stuff in there

dim copper
#

Oh haha 😁 Yeah, nothing wrong with female bodies. #equality

warm junco
#

I would suggest you try 1.5 and 2.1 and compare ^^

#

For 2.1 you also need a .yaml file to get it work

worn trail
#

Any finetuned models or related for generating pixel art?

warm junco
worn trail
#

thanks ^^

severe field
worn trail
#

idk, Im not from the US

severe field
worn trail
#

?

severe field
#

do u consume energy

worn trail
#

What are you trying to accomplish here? Just get to the point lol

severe field
#

can you lick chocolate from my belly button

signal dome
#

wtf

proud nova
#

Dude wtf

fervent thunder
#

I'm sure this is a commonly asked question, but I couldn't find anything conclusive so I wanted to ask here.
How come various AI-Art sites default to Stable Diffusion 1.5 (or old versions in general) when the newer versions are available?

brazen fox
#

SD switched from CLIP to openCLIP for 2.x and it works way less with artist names and certain descriptors, and instead relies more on negative prompts for instance

severe field
vapid tangle
warm junco
#

with 2.1 you can generate 512x512 too

brazen fox
#

compared to 768 generations- even with the 512 base model it's weird

warm junco
brazen fox
#

there's nothing wrong with SD2- I love using it way more compared to 1.x- but 512 generations just aren't that good with it

fervent thunder
#

I know its upsets some, but to me 2.0/1 were just flops

#

I dont know the reasons, but they just arent as good/fun to work with as a lot of these 1.5 models

#

add in the longer gen time, its a pass for most

#

like it seemed to me stable had a nice head of steam going with 1.4/5 now theyre just flop era after flop era, but all it takes is one good model to change the convo

vivid lance
#

I think SD is just going to surprise everyone eventually

#

like atm its good, its pretty damn good with custom models and model mixing

still remnant
#

SD surprises me almost every time I hit "generate"

brazen fox
vivid lance
#

but I think it will eventually overtake MJ

fervent thunder
#

i never used greg or any other artist in my life, i do photo real 9/10

autumn compass
#

embeddings help a lot in 2.1.
A must, in my case

brazen fox
fervent thunder
#

yeah artistficially like I said, one good model and the convo changes

vivid lance
#

I use one artist consistently in prompts lol, 18th century painter William Henry Hunt

#

if I want a even modern digital painterly look

brazen fox
vivid lance
#

I go for him

warm junco
#

yea i realy dont use the official models at all since custom models became so good

still remnant
#

honestly, in the AI world, I'm just glad we have so many competitors, competing models helps secure future use, and means multiple solutions are being researched simultaneously.

sudden isle
#

I'm into John Singer Sargent myself

fervent thunder
#

yeah absolutely mielkman. speaking of that, has dalle ever dropped a new model since dalle 2?

vivid lance
#

I don't think a lot of artists realize they aren't doing much "new" its just modern looking because of the tools on the strokes and the subject matter

#

like if you wanted an "ethical" Public domain model, you could easily just style jack from old painters with a very simple photoshop script to prep the images in batch for a more digital look

severe field
#

so silly

vivid lance
#

:/

fervent thunder
#

what?

still remnant
severe field
#

will that be acceptable

vivid lance
#

I would still rather my toes remain unmolested XD

severe field
#

I never said thats what it was

still remnant
severe field
#

what if I just wanted some chocolate

fervent thunder
#

are you trying to gross the whole room out

severe field
#

on your toes specifically

#

the nerve of some people

fervent thunder
#

cant you block people on discord

#

i guess you can

lilac reef
#

I really sat down with kohya-trainer today and I can't make heads, nor tails, on it. I am 100% lost and the dreambooth extension is broken beyond repair for 2.x SD.

#

Seems Lora will die because there are no tools for it, and is excessively overly complicated in comparison to DB and TI/HN.

#

Doesn't help when content creators say it sucks.

bleak wolf
#

@lilac reef @brazen fox @orchid wadi @still remnant how is the situation with lawsuits? i was away for a week

brazen fox
#

afaik, little to no change

lilac reef
bleak wolf
#

ty

lilac reef
#

YW

#

Man, LoRA looks like it could really work but it is so broken, or this one koyha colab and it takes a rocket scientist to figure all the steps out.

#

I think it is so complex, and so hard to work with that most creators have shunned it, or had bad results (when it worked) that they tell their subs it sucks and just do Dreambooth.

#

It became broken, and so complex, when it released for all models and 2.x base.

trail wing
#

humans in 2.1 come out all slightly vertical stretched or am I tripping balls

lilac reef
#

Depends

trail wing
#

I get it at least half the time

lilac reef
#

were you 768 squared?

trail wing
#

yea

lilac reef
#

I never get that since I left 1.5 unless I change the resolution.

trail wing
#

it's odd, the proportions look slightly off, and you see it instantly in faces, but could me

lilac reef
#

No, not just you

#

People saw it far more in 2.0

#

stretched faces

trail wing
#

oh, I kind of skipped on 2.0 really

lilac reef
#

not all that different but 2.1 is far far better and I honestly do not get the issues. Now I will tell you if you just arrived to 2.x you MUST up your prompt game especially negative prompts

#

negative prompting is everything with 2.x

trail wing
#

yeah, I get to what I want, was pretty heavy on negs in 1.5 already, it's just when human faces are on or some full body humans, I see the slight stretch out

#

it's not all the time, but it's there a lot of the time

lilac reef
#

you might be missing a negative

trail wing
#

like a negative for x y ratio ?

lilac reef
#

yeah. bad proportions is one I see around

trail wing
#

gonna try that out see what happens, though am pretty happy on 1.5

fervent thunder
#

a

lilac reef
#

I left 1.x immediately the day 2.0 dropped and never looked back. I lost all my models and stuff but meh.

#

I prefer 768x768 and when 1024x1024 hits with 3.x I will drop 2.x immediately too.

#

I doubt we will go larger than that

brisk bloom
#

1024 will rock

lilac reef
#

yes

#

my card will take 2m per generation though

#

actually about 2.5m I timed it

brisk bloom
#

ouch

lilac reef
#

yeah, 20 steps Euler_a

trail wing
#

zie pain is real

lilac reef
#

my card doesn't have real fp16 so everything I see is 100% different than what someone else will. I train and what I get is not what it gets.

#

it being on 2000+ cards and colab

orchid wadi
trail wing
#

I think they got a fix out for 32 bit math on fp16 that may give you closer results to others, commit was done today

orchid wadi
trail wing
#

ok I think I get that, was doubting what I was seeing really

lilac reef
#

I own a 1060

lilac reef
trail wing
#

that's rough

lilac reef
#

yep, but there was a post on reddit about this showing how each card looks different with the same prompt.

#

my 1060 was blank, lol.

trail wing
#

heh

#

I'm hoping laptop 4080 come with 12 or 16 GB vram and go to that in some months

#

I'm on 3070 8 vram laptop, and it's on the edge of very usable for some stuff and useless for others

#

but there's always colab

lilac reef
#

Colab for LoRA is a dead duck

#

DB extension is just bad and it crashed last night as it was saving (lost connection in the webui).

#

I do have the koyha notebook but I am not a data scientist, and there is no help vids or docs so lost.

trail wing
#

haven't tried lora

lilac reef
#

I have done it all but it so wanting to sink in but can't

trail wing
#

to tell you the truth am trying to decide for the last couple of hours if I should update to latest auto web ui but can't find if TI is broken on it or no, or if it will create xformers crazyness at start and randomly throw me into a multi hour charade

lilac reef
#

16h for me one day straight and I never managed it. I hate xformers due to that as everything else was just working. In 1.5 I did a recompile 4 times with no issues.

trail wing
#

16 hours damn

lilac reef
#

yes, I went around and around and it was bad. Finally I got it

orchid wadi
lilac reef
#

Since you have a new card a manual compile is not needed so you should be okay

trail wing
orchid wadi
trail wing
#

heh

lilac reef
#

backups are for sucks, lol

orchid wadi
#

Also you can always manually download the earlier commits build as zip

lilac reef
#

run it rough, run it dry and hope

orchid wadi
#

from github

trail wing
#

ok, here we go then, wish me luck...

lilac reef
#

luck

trail wing
#

I can't tell...

orchid wadi
trail wing
#

in the same conda env

orchid wadi
#

Rename the main folder with like Folder2 and then use the other one as Folder

#

Updating auto shouldn't mess withyour python setup...

trail wing
#

well, had some venv screw ups

orchid wadi
#

If it does, well just reinstall python again. Done it a few times because of other software problems not that hard

#

venv is the local that programs thing

trail wing
#

I know, I didn't actually know if it could screw conda

orchid wadi
#

you should delete venv regularly for "clean install" if there are odd issues

trail wing
#

now I know

#

do you have the new torch and xformers update warning or did you update both ?

orchid wadi
#

Updated both and no issues

#

Just put it to do it's thing, went to take a shower came back and it was there

brisk bloom
#

same, both updated without issue

trail wing
#

did you guys update both at the same time or run one flag, then quit, then another update flag

brisk bloom
#

one at a time

orchid wadi
#

did both

#

same time

trail wing
#

ok, that settles it then, here we go

orchid wadi
#

Like what is the worst thing that can happen? You having to reinstall the whole repo? I mean like it takes few minutes but it isn't like your computer gets messed up like with some programs.

trail wing
#

or I could grow an extra arm, there's that

#

you never know

orchid wadi
#

Do say if that happens. As a welder an extra would be nice

trail wing
#

practicality above all

#

major hangups for me are not being used to python at all

#

I program in haxe, it's a simple packet manager and it uses MS compiler

#

this thing... the wheels, the ckpts that can have code in there

#

spooky

orchid wadi
#

Try labview and then youll appreciate everything in the world more.

#

Because fuck me is that a... awful thing

trail wing
#

lmao

#

dude, I used softice on windows 98 se

#

you're only 18 once

#

some stuff is great when you don't know what you don't know

orchid wadi
#

I regulatly use AutoCAD. It is as shit now as it was every point in the past.

#

Why the fuck do we use it?

trail wing
#

hehe

orchid wadi
#

Because it is industry standard?

#

Why?

#

Because everyone uses it?

trail wing
#

it was such a pain

orchid wadi
#

Why?

#

Because it is industry standard

trail wing
#

even the dos 3ds was better

#

ok, it's telling me I should update pip

orchid wadi
#

Modern software doesn't improve. It just stay at equal level of shitteness for modern needs

trail wing
#

but I can't remember if I got pip through conda...

#

it's from the venv... 22.2.1 -> 22.3.1

charred iris
#

Guys is it possible that windows defender just flagged protogen as a Trojan?

#

I ended up scanning my gui for other reasons and windows defender goes insane and starts reporting two Trojans inside the cptk file for photorealistic protogen

#

That I got off of civtai

#

I downloaded malwarebytes to get a second opinion and it's not detecting anything

orchid wadi
#

Don't download anything but safetensors

brisk bloom
#

if you do a checkpoint merge with 100 of file A and save as safetensor, would that remove any potential viruses?

trail wing
#

ok, updated, pip upgrade had to happen or it hang up xformers updating

#

weird

#

let's see if TI train is running...

forest aspen
#

anyone got any luck with pix2pix in a1111?

#

for some reason i can't get it to work properly..

trail wing
#

TI training is going on new update, yey

charred iris
#

ok so what it reported was this C:\SDGUI-1.9.0\Data\models\protogenX34Photorealism_1_ckpt.ckpt->archive/data/204

#

I think I actually did download this as a safetensors, I then converted it to cptk with stable diffusion gui

#

I did that because straight cptk files did not work for me for some reason and I was advised to convert them myself

#

and it reported it as Trojan:Win32/Sirefef!cfg

brisk bloom
#

ahh I think the model isnt downloaded yet, maybe that was the issue

lost surge
#

anyone have any recommendations on the best way to train a custom stable diffusion model on Mac?

trail wing
#

I tried pix2pix on hugginface space and it worked so good

#

I just typed: turn them to dogs... and BAM!

trail wing
charred iris
#

guys does anyone have an idea what the protogen thing could be?

trail wing
#

I was reading discussion yesterday and one dude said he was thinking about doing it and another dude comes in and says he is finishing it alreayd

trail wing
brisk bloom
#

yeah my net isnt that fast today. waiting for the 7gb file to finish

charred iris
trail wing
charred iris
#

good to know

trail wing
#

I wish it was less of a crap shoot and some day someone is going to slip something in there that is gonna hurt, but that's the best I can tell ya about it

alpine dust
#

I made ChatGPT make some shit, lol

#

Steps: 25
CFG: 8
Prompt: Wide-angle photo of a majestic dragon, soaring through the clouds, highly detailed, 8k, ornate, intricate, cinematic, dehazed, atmospheric, (oil painting:0.75), (splash art:0.75),(teal:0.2),(orange:0.2), (by Jeremy Mann:0.5), (by John Constable:0.1),(by El Greco:0.5),(acrylic paint:0.75), dark style picture, hard shadows and strong rim light, watercolor effect, watercolor (medium), traditional material, art by Juan Giminez and Atey Ghailan and Sachin Teng, flat color, raining, basic white background, style of Jordan Grimmer and Greg Rutkowski, crisp lines and color
Negative prompt: (frame), (frames), watermark, ugly, frame, frames, watermark, ugly

#

Its respond

#

Results are kinda impressive

charred iris
#

how are cptk files different from safetensor files?

fervent thunder
#

latter is more secure i think

#

from viruses

charred iris
#

have there already been instances of ppl getting infected by running models?

fervent thunder
#

wassup dude

charred iris
#

sorry hello

fervent thunder
#

and idk fandyus, just what ive heard people say about safetensors, supposedly more secure. cant hurt ya know

orchid wadi
# charred iris how are cptk files different from safetensor files?

CKPT can contain code that is executed. This isn't just for malicious purposes. But you can use CKPT to also carry speacial code relating to the model like instructions on how to use the model for a program or parameters.

Safetensors can not carry this kind of code. Downside being that it can not be used for models that would need such things, upside being that it is always safe to use.

#

However unless you work with things where you'd need such things as extra code in the model, then you know how to deal with the risks and you know where you get your files from.

#

Because CKPT is nothing but a funky zipfile. You can open it with 7zip or such if you want to look inside. It takes diffusers model and packs it in a certain order.

#

You can transform CKPT and diffusers with a script that knows how to set the files.

fervent thunder
#

dude this isnt some dipshit conervative discord

#

its for ai stable diffusion

#

so please fuck off to off topic if you need attention for that stuff

orchid wadi
#

I don't even know what is about...

#

I'm gonn assume something racist and USA related

fervent thunder
#

hes posting stupid conservative memes around the discord

#

yeah exactly

orchid wadi
#

Ah...

orchid wadi
#

It is a shitty meme all things consider... Also you could have upscaled it or smth.

#

Like if you gonna shitpost at least do high quality shitposting

#

I don't want you to really

#

I'm not from USA. Nor am I conservative. Nor do I care.

fervent thunder
#

he joined both here and discord two days ago, in other words hes just some alt trolling like a moron

#

too bad i never see mods here

orchid wadi
#

You can just like @ them.

fervent thunder
#

you can say whatever you want, im gonna use the block function on you

orchid wadi
#

Well I wouldn't say that what you posted is relevant nor of particular quality.

#

Like sure if we were talking about... conservative whatever in relation to SD or technologies or it was just in the current zeitgeist of the discussion. Yeah sure.

#

But just random shitpost linked from... Whatever cdn that is... Is trolling

#

Dude I'm finnish. The mustard race of shitposting. My bar is high.

#

You can't even shitpost properly.

trail wing
orchid wadi
#

gtfo amaterur

trail wing
#

hey, this new xformers version is it supposed to be just a little bit faster on it/s

#

I think I see it

brisk bloom
#

not perfect

orchid wadi
#

I got like a 0.5it/s more

trail wing
trail wing
brisk bloom
#

will try another pic

trail wing
#

hey at least it is working proper

orchid wadi
#

Like I don't really feel it when I run like bs of 9 and grids of 10x10 😄

trail wing
#

yeah well

orchid wadi
#

I put the thing to do its thing and go have a walk in the park near me. Then come back and go through the stuff it made.

trail wing
#

you sounded like that dude... yo dawg, we foudn out you like rims, so we put rims on ya rims dawg!

brisk bloom
#

I tried "remove glasses". only one side of the frame was removed. lol

trail wing
#

50% accuracy

#

safe and effective, ship it!

#

gonna try it out soon as TI train ends

brisk bloom
#

im sure it will improve over time

trail wing
#

yeah

#

the one on hugging face spaces was working great, but diff implementation maybe

#

nice to see it included so soon even

brisk bloom
#

yeah the more tools the better

#

sweet. I got one to work. changed a guys beard from white to black

grim glacier
#

Haven't been here in a while, is there a requests channel?

drowsy tendon
#

3.0 when?

still remnant
#

3.0 when we can finally deserve it.

lilac reef
old marten
#

3.0 when never. It will have to be fanmade or crowdfunded.

#

You realize midjourney and stable diffusion got sued into oblivion and you want them to release a new model?

trail wing
austere wasp
#

I'm trying to make conceptual music videos w ease! what AI program is best for that?

#

was thinking diffusion

lilac reef
fading sandal
#

Hey all, just curious on performance stats. I've got a 3090 TI, but I've got a pretty old rig. i7-6700k running on a 7 year old lenovo motherboard. I've got an m.2 that i'm running SD on and I'm up to date with all the latest xformers and all that stuff. I still see, on average, around 12-14 it/s. I've gotten as high as 18 once. Even if i'm doing a 20 step euler a with basic prompt, i'm not cracking 15. Am I being bottlenecked by my CPU? or maybe by my pcie 3 motherboard? This all doesn't seem likely, but I wanted to know if there is some optimization I'm missing or if a 3090 TI can really only pull 15 it/s. Thanks!

trail wing
lilac reef
trail wing
#

oh

#

xformers: 0.0.16rc425

#

release candidate, the mad man!

lilac reef
#

well, I was on 0.15 so yeah, 0.16 is way new

#

I can't get it to roll up so I can smoke it, so I have to pass.

#

Not wanting to be spanked for another 16 hours to end up defeated again.

trail wing
#

had trouble with pip not being latest, so had to update that for xformers to update

#

or it simply didn't

lilac reef
#

pip is an easy upgrade it is all the other stuff. this needs that so you upgrade that and it will not work with this so you update that now the previous thing says it isn't for it.

trail wing
#

it's whac a mole updating

lilac reef
#

yeah, too much for me as it is so interdependent.

#

we are supposed to get optimizations to speed all this up from emad but I lay odds it wants tensor cores which I don't have.

fading sandal
#

...i just upgraded pip

dim wedge
#

hey, does anyone know how to outpaint the other pictures (not only the first one in line)

old marten
#

How long does it take you to do a 512 x 512

#

CPU usage minimal, mine barely goes above 6% so I doubt it has anything to do with it, it probably isn't as slow as you think? I have a 3090 regular.

fading sandal
#

i feel like weirdly this has gotten ever so slightly worse. i was doing sub 2 second times pretty recently it felt like

#

i know this is picky, but i'm just like... is there some optimization i'm missing? i've got xformers and even that --opt-channelslast in the command line

golden trench
lilac reef
fading sandal
#

ah, ok, so i'm not loosing my mind

lilac reef
#

No, you aren't but I thought it was just me or my card.

#

my its have decreased from about 1.4s/it to 1.75s/it

fading sandal
#

ooof. i used to have a 1080, but upgraded to 3090ti for this stuff

lilac reef
#

I was saving for a 4060 and was so excited. 16GB msrp of 379 then Nvidia. oof

#

4060ti is a disgrace all around having even less performance than a 3060ti and less ram costing 200 more

fading sandal
#

yeah, i think the 3090ti might be the best overall card for SD. the only competitor is the 4090.

lilac reef
#

Well, ChatGPT said it was the 2080. Yeah, it needs to be updated, lol.

trail wing
fading sandal
#

tell me more

#

like i should be getting upwards of 20+ right?

lilac reef
#

If Nvidia continues with the insanity I will just continue as I am or drop SD/AI stuff like a hot potato.

trail wing
#

I'm on 11 +/- it/s on a 3070 , on a laptop

fading sandal
#

hrm

trail wing
#

at 512x512 euler a 20 steps

lilac reef
#

512 sq you should be smoking it

#

at least 17-18it/s

fading sandal
#

yeah, right. @trail wing any ideas on how i could debug this?

#

i was getting 17-18 at one point. that's as high as i ever got

trail wing
#

you're on latest commits and xformers and all the mumbo jambo ?

lilac reef
#

that is about normal for a 3090

fading sandal
#

i got all the blablablas installed

brisk bloom
#

my 3080 averages 14it/s with those settings

fading sandal
#

deleted venv folder yesterday and git pull

trail wing
#

damn, go to the discussions at auto1111 and search "3090" man

lilac reef
#

4090 (40x0) as bugs with SD and I mean training bugs and shit

trail wing
#

there's got to be something happening there

fading sandal
#

where's the automatic discussions? that a different server?

trail wing
#

on github

fading sandal
#

oh yeah, i already looked around there

trail wing
#

issues and discussions ? on both ?

fading sandal
#

oh, there's a gup-go-brrr discussion here. that seems liike the place

trail wing
#

heh proper named

lilac reef
#

If you train the new bug is with cross attention on a 4 thousand series card as it is giving crap. There is a bug report about it. Turn off the optimization and training goes back to working.

trail wing
#

that's a damn shame

#

next step is installing linux on an external ssd m.2 I got over there and checking those it/s go brrrr

#

just a little bit more brrr actually, but brrr still

#

every brrrr counts

bold wave
#

Hey all, if you could generate images from things you read in books, would you rather have a chrome plugin to generate from selected lines in pdf browser view? Or have a web ui to paste the desired text into a box and generate the image there

trail wing
#

can you hook it to a right click with send to option and then it goes ?

#

cause that sounds cozy

bold wave
#

Great idea

trail wing
#

but so does a non-linear keyframe editor for deforum

#

yet, here we are

#

matploting views away

bold wave
#

I mean, it would be great if it was added to kindle 😛

trail wing
#

yeah

forest apex
bold wave
#

I actually have not started anythin hehe, was just asking what would a user prefer

#

In case I would develop it, I think I would have to you gpt to grab context words from the paragraph or something

#

and at the end summarize everything in one sentence

forest apex
#

A friend sent me some Russian poetry earlier that looked really visual and like it'd make for good promptfood but the results were all over the place. describing some further context or finding the right model seemed needed

bold wave
#

would you care to share the poem

#

perhaps I can start prompting a bit

forest apex
#

Some kind of NLP or other interpretation would be needed, making automating this require more than pasting several prompts into the Prompts from textbox script

forest apex
#

Using something like Prompt Magic which adds keywords might help... And a context menu, for the UI of whsr you're proposing, would be easiest to me

sour lake
#

有中文吗

floral zealot
bold wave
forest apex
#

it is in the Dynamic Prompts extension for automatic1111

stable pawn
#

I need prompt help or suggestion. in SD 5.1 I write "cartoon female bunny character" and I get a cartoon character walking upright and that is what I want, but as fast I set a style or add ", pencil outline" or "Cross-Hatching" the character become a normal bunny. So why is it so that my style at the end over ride the "cartoon" in the beginning?

#

I am running SD locally with Automatic1111 WebUi.

waxen pasture
stable pawn
#

@waxen pasture Not exactly, but... Wow, thank you now I need that 😄

waxen pasture
fervent thunder
#

can anyone give me any tips on how to get less jumpy animations?

forest apex
#

@fervent thunder how are you animating?

fervent thunder
forest apex
#

are you using deforum? img2img? another script?

forest apex
fervent thunder
#

thanks

fervent thunder
willow garden
#

Hey guys, any tips to transform a photo into an anime drawing (akira or ghibli style) ? I mean, about the models to use and what the parameters

stable pawn
#

You know all those negative prompts we write, deformed face, deformed fingers, double head, ugly..." and so on. Take them and place them on the regular prompt and then the regular prompt in the negative. Now you will have a image of everything you hate in a character.

calm plover
#

Anthro or anthropomorphic can give objects or animals human qualities

silk fossil
#

Hey, all. I'm in Antarctica and I want to play with stable diffusion. How large is the file download?

forest apex
silk fossil
#

I can't download gigs of data, unfortunately.

forest apex
silk fossil
#

It'll have to be sneakernet. Any guide to installation on a portable hard drive?

#

I can have someone bring me one, mail isn't realistic right now.

forest apex
#

Hmm.

#

You might try asking on reddit, I think it would be kinda difficult

#

you'll also have to download things like pytorch, which is 2GB

#

Having a friend just install it on a machine like a laptop or something might work ;0 otherwise it'd probably be easier to use it remotely, like over TeamViewer or gradio

silk fossil
#

It pretty much has to be local.

#

We just don't have much internet.

#

It can work, it just has to be a portable install on an external drive

#

Or I need to locally host the assets

#

I must bring a new technology to a new continent!

warm junco
#

For a local install You need the files so you have to Download at least 5-7gb

#

But yea it can work of an external ssd

dim copper
#

I just had a realisation that every Stable Diffusion version cost a million dollars to train. So version 2.x could just be a very costly mistake. But would someone admit a mistake that cost a million dollars? 🤔

fossil mortar
#

Я пукнул

forest apex
#

this sort of progress isn't always linear... I don't think 2x was a mistake.

orchid wadi
#

2.x is better than 1.5 IMO

forest apex
#

Yea, for some things, it's not as bad as it seems at first, just different prompting

#

The midjourney embedding for 2x seems to do better than openjourney, for instance

forest apex
edgy cedar
#

Hi guys I’m new in this serve, Maybe u already talked about this, but I I would like to import an image and expand it, It’s possibile? I thought paste the browser link of my image like in Midjournay would work but it doesn’t

forest apex
old gull
#

anyone train a model on snowflakes yet??!?!

spark kite
#

The Stable Horde has provided the first batch of ratings to LAION! Over 50K images with 3+ ratings each! Total 250K ratings

horde_ratings=# select count(id), ratings_count from images where ratings_count > 0 and ratings_count < 6 group by ratings_count;
 count | ratings_count
-------+---------------
 73884 |             1
  7655 |             2
 58140 |             3
  3713 |             4
    73 |             5
(5 rows)
inner tapir
#

I've noticed Loras have become very popular options, and Hypernetworks seem to have dropped off, despite Loras being substantially larger files. Is there a quality difference between them? Something else? Just curious as a non-tech person

severe field
#

i fell off

orchid wadi
inner tapir
fervent thunder
#

where do i find embedding files?

#

any website recommendations?

#

also is it possible to use a Lora thing with SD 1.4?

orchid wadi
#

To summarise it simply:
Textual Inversion embedding, is a instruction how to find things from the model that recreate the concept.
DreamBooth is injecting training thing in to the model and replacing things in it.
HyperNetworks are additional networks of indorformations. However they must have an exit node, so they can only make on concept and it's variations and put that in to the space between "neurons" of the base model.
LORA is kinda like DB that is merged in instead of overwritten.

#

Like... LORA is kinda tricky in way because what it does is kinda funky.

#

LORA doesn't touch the parts of the model it doesn't need to touch.

inner tapir
orchid wadi
#

Mathematically it kinda like goes in reverse. HN/DB build up to make something, LORA tries to like.... Do the opposite? In a way

#

basically instead of find mathematical solution to represent the concept, they make a solution to represent the OPPOSITE. Then adjusts the model knowing what is the reverse

#

Like closest would a photo negative? Or like negative mould to make a tile with.

#

You use a negative to make a positive. But this stuff get strange because it isn't like that in a network space.

orchid wadi
#

Since LORA changes the model.

#

Same thing with TI and DB. Anything that CHANGES the model has to be done before anything relies on it.

#

So if you'd want to combine the lot you'd go: base model training > base model finetune > DB > LORA > HN > TI.

#

Not sure how they'll actually play along in reality but that is the theory.

fallow hinge
#

hi

orchid wadi
#

The greatest benefit of LORA is that it doesn't touch parts of the model that it doesn't need to.

#

Oh and LORA can work at like really low VRAM.

#

Well "LOW"

#

Compared to the others

inner tapir
orchid wadi
#

Stability are saying that they are releasing low-end user finetuning toolset of their own. Since the next model version is going to revolve around being even purer base model, that is even more finetunable.

#

Talking with Emad, he stated that the goal is to make like "0 bias" base model, then everyone trains in to it the things they want.

fallow hinge
#

where can i start these generation of images

orchid wadi
fallow hinge
#

is there any limit in dreaming images?

winter pike
#

unlimited

hollow crow
#

anybody have any reccomendations on what gpu to buy? im going to need it to run professionally but im not a pc expert

#

and do i need something else aswell?

warm junco
#

Photoshop, Blender, Unreal Engine, Office, Browsing, Ai Image Generation?

fallow hinge
hollow crow
#

so ai image generation

#

does the gpu and other components quality affect the quality of the results or just make it faster?

wide juniper
#

what are weights, as opposed to models?

warm junco
finite scaffold
#

HELP IT SAYS STABLE DIFFUSION MODEL FAILED TO LOAD, EXITING HELP!!!!!!!!!!!

warm junco
hollow crow
warm junco
hollow crow
#

my friend offered it to me

#

maybe it would be good for starting

severe field
hazy solstice
#

Pls why am I getting this error?

severe field
#

because something isnt working the way it should

hazy solstice
#

Please include at least one positively-weighted prompt

severe field
#

I never seen that error

hazy solstice
#

Oh am not allow to post image,

severe field
proud nova
#

Is generating images training the AI itself, or do I need to use "Train" option in the menu directly to actually train it? I am asking because my last results were much better compared with the first ones I made with the model

forest apex
#

@proud nova no you have to manually train

hollow crow
proud nova
wide juniper
#

what are weights, as opposed to models?

astral goblet
severe field
astral goblet
#

gross

wide juniper
hushed tangle
#

How can I do img2img on a cover art to get children bad drawing of it?

digital aspen
#

Is there any reason why i am getting a message request from someone advertising for another server?

fervent thunder
#

because youre on the internet and spam and scams are a thing

winter pike
vestal dew
vestal dew
digital aspen
digital aspen
# digital aspen Yup

I was in there long b4 i got the message. But getting these kind of messages is really annoying

warm junco
severe field
winter pike
#

very shady way of promoting their services

severe field
#

so unfortuntate

vestal dew
#

I got a warning on Fiver for making a suggestion to some seller on his page. He's an SEO expert, selling that service, but he has his own name misspelled on his profile.

proud nova
#

How fast do you generate ur images? I am reaching 4s/it

fervent thunder
winter pike
proud nova
orchid wadi
proud nova
#

Nice speed

lilac reef
vestal dew
#

they said it is against TOS to promote my services to other sellers

lilac reef
#

o.O

vestal dew
#

I guess the service is that I gave a suggestion to fix his profile

lilac reef
#

Ahhh, yeah not good.

#

lol

vestal dew
#

I was looking for an SEO expert on the cheap for a neighbor's side business web site. I noped off that page when I saw he had his name misspelled.

lilac reef
#

Maybe they thought you were asking to fix it for him.

narrow sluice
vestal dew
#

If he doesn't even have the attention to detail to make sure his profile is right, how can he be trusted with your site tags etc.

lilac reef
#

I hate SEO and consider it the work of some cosmic evil entity.

narrow sluice
#

So basically, I need to add 500 swears so people know it's my personal opinion and hyperbole?

#

SMH

waxen swallow
#

when is text-to-video being released?

narrow sluice
#

Yeah, eff that.

narrow sluice
#

It's like asking when a chicken is going to lay an egg.

vestal dew
#

text to video to generate a 10 second clip in 1000 hours?

lilac reef
waxen swallow
narrow sluice
#

Just broad timelines

still remnant
#

we shall receive text-to-video when we deserve text-to-video.

vestal dew
#

you get a lot of chickens so you get eggs every day

lilac reef
still remnant
vestal dew
winter pike
#

Fiverr lack of real moderation. I had to request a refund via paypal after days of trying to contact them, they replied instantly

narrow sluice
#

What was that programming analogy?
"Nine women can't make a baby in one month"

#

But yeah.

vestal dew
#

Yes

waxen swallow
#

I have an idea of a hack that we could use to train SD for text to video. It would require a lot of training data though. Break video into frames and train SD on the frame sequences.

vestal dew
#

If you give me nine women, can I make a baby in a month?

lilac reef
vestal dew
#

it is what it is

lilac reef
vestal dew
#

if someone makes porn, and thus I can learn the techniques he used to do that to do what I want, then it's good

narrow sluice
#

48 FPS means...oh.

lilac reef
#

Now clean the filth out of society by any means necessary then we might be worthy.

narrow sluice
#

Yeah, it's gonna suck.

still remnant
vestal dew
#

porn is about sex. Without sex, none of us would be here.

narrow sluice
lilac reef
narrow sluice
#

Weird edge case, but okay

vestal dew
#

Oh I see. I thought you mean "REAL WORLD" , like videos of real people having sex...deep fakes

lilac reef
#

CP is the biggest reason we are not worthy of such tech.

#

DF they had CP and R it was disgusting.

silk fossil
#

What the fuck did I just walk into?

waxen swallow
waxen swallow
vestal dew
#

couldn't some motion be short cut...pose A, pose B, interpolate between them

waxen swallow
wide juniper
#

Do wild card prompts work out-of-the-box, or do I still have to install extensions? If so, which extension is recommended?

silk fossil
#

Is the one-click install bundle from cmdr2 acceptable and safe?

narrow sluice
vestal dew
#

true

cloud granite
#

Guys, the official Midjourney Discord server is nearly reaching 10 million members!

old marten
#

given how people create 50 accounts so they can make more free images

#

mj is for normies anyways

cloud granite
#

So they don't actually have 10 million members, they probably have like 100,000 right?

old marten
#

probably

#

I would guess they have more than SD

#

but not by a margin of 9 million

cloud granite
#

I was all excited for the 10 million milestone but they are fake alt accounts.

#

There is no point of streaming if the members are all fake.

old marten
#

you have one weird youtube channel

cloud granite
old marten
#

Bizarre videos

cloud granite
old marten
#

You make videos about calculators?

cloud granite
old marten
#

Oh ok. That's pretty cool then I guess

cloud granite
old marten
#

is your pfp a roll of toilet paper? lol

cloud granite
uncut plinth
#

someone have problem with saving the image in the webui?

old marten
#

not really

proud nova
still remnant
#

got a commission from someone wanting a person hanging off a cliff holding on to someone's hand. Hands: the bane of AI Img gen.

#

I guess it's back to traditional tools for me!

winter pike
uncut plinth
uncut plinth
winter pike
#

my models were randomly changing too

uncut plinth
uncut plinth
trail wing
#

weird thing happening, had a safetensor loaded, changed to another, then tried to change back and no matter how many times I tried it kept loading another "fallback model" saying the one I wanted didn't exist

#

like wtf homey, I can see it in the directory

#

this is the second or third time this happens, no clue why

#

also, no clue why the "fallback model" was the same and where that is established either

#

maybe it was just alphabetical, it started with an A

#

but anyway, weirdness

dry trellis
#

I had that happen once...I had to move all models to a different folder except one... then load auto1111 then move the models back

silk fossil
#

So the large files are the data files or weight files?

warm junco
rare cairn
#

Hi sorry if this is the wrong place. I'm new to this. What is the largest data set I can download?

warm junco
#

the biggest datasets are all 7,5gb big

#

it depends on the model, you also can merge models to get a bigger dataset but 7,5 is the limit

silk fossil
#

Which datasets are necessary? I'm trying to do it with the one click installer but I'm in Antarctica and my connection is unstable, so I'm going to try to use another tool to download the largest datasets.

#

It will have to auto-resume.

warm junco
silk fossil
#

Might try that.

#

Yeah, that's doable.

warm junco
#

its the best you can get at this file size i guess

silk fossil
#

So I can run stable diffusion with just this small one? That'd be neat.

warm junco
silk fossil
#

Is there a guide to basic installation for use with those?

warm junco
silk fossil
#

antarctica needs memes, and i need images to generate those memes

#

very important

#

Unfortunately, watching videos here is very difficult

warm junco
#

oh okay

silk fossil
#

it is almost impossible.

#

(I really am in Antarctica)

warm junco
#

Then here the short Version:
Automatic Installation on Windows

Install Python 3.10.6, checking "Add Python to PATH"
Install git.

Create a folder on an SSD without spaces in the name. Go into the folder and right click, choose Git Bash here.
Copy this command:
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git

Place stable diffusion checkpoint (model.ckpt) in the models/Stable-diffusion directory
Run webui-user.bat from Windows Explorer as normal, non-administrator, user.

rare cairn
#

I saw there were pre-trained data sets that are several TB large that I could download? I just don't know which is the best. Any suggestions?

warm junco
#

Nope these are not for the users.

silk fossil
#

@warm junco tyvm

warm junco
silk fossil
#

I do.

warm junco
#

Perfect xD, we need penguin images here alot

rare cairn
#

what are those larger dataset for?

warm junco
#

these are the raw materials the algorithm "watched and learned" (Laion B made these large files). the endresults got compressed into the model files we have then

#

i think its like 1-3% of images are in the model of the whole raw data

silk fossil
#

the Adelies are still here, the Emperors should arrive...soon?

#

Skuas, we have plenty of.

warm junco
silk fossil
#

The most dangerous animal in Antarctica. They're HUGE angry seagulls and they will attack you if you have food.

warm junco
#

Ohh 😮

silk fossil
#

so a major use of stable diffusion here is going to be "a skua attacking a ___"

rare cairn
#

so you recommend merging the data sets? Is 7.5gb the max total that can be combined?

warm junco
rare cairn
#

Is there data sets you recommend? I reserved 10 TB for data sets because I assumed bigger would be better.

warm junco
#

@silk fossil then i have to say for photorealistic images like animals the SD1.5 model will be better than the 2gb Dreamlike

warm junco
rare cairn
#

Yes lol. But I guess you're saying 7.5gb is the max so I didn't really need all this space.

#

but yeah whatever you recommend

cloud granite
warm junco
trail wing
#

yeah, all of a sudden those 1TB ssd m.2 are getting smaller by the release

#

I'm like an addict at this point

#

and also, peeps should prune their models

brisk bloom
#

the largest model I have is the protogeninfinity at 8gb

warm junco
#

I have that too but tought its 7,5

#

Will check tomorrow

silk fossil
#

I remember the old deep dream AI thing, I actually liked that style.

#

With eyes everywhere and shit, that was dope.

#

Weirdly, my download is progressing.

warm junco
#

Okay Dreamlike is very good and not weird. The trigger word for the model is dreamlikeart

silk fossil
#

What do these datasets actually represent?

#

I know it's the data that was created when the model was trained, but what does that actually look like?

#

What is the "stuff" of the model's "knowledge"?

unkempt gull
#

So I just downloaded the protogen 3.4 and 2.2, but the safetensors have a different hash. Downloaded from civitai

#

any reason why that might be?

#

are the hashes only for ckpt? Pruned or not pruned? I can imagine the hashes would have to be different

frank gate
#

safe and ckpt will have different hashes for the same item

#

same for pruned and not

rare cairn
unkempt gull
# frank gate same for pruned and not

Right. That's what I figured. I'm also though unable to replicate the images shown on civitai for those protogens. Copied exactly, so the hashes thing got me wondering.

#

Ah well, I probably have something wrong.

frank gate
#

could be you are doing it right

#

hardware itself can impact the generation

#

as can other settings

#

stuff like if xformers are on or not

old marten
#

1tb is the new 256GB

#

I don't fuck with anything less than 2 TB, in fact, I returned my motherboard when I found out it only had two M.2 slots.

#

From what I've seen, 2 TB is the sweet spot for M.2 in terms of price for storage. Anything else and it gets more expensive per gigabyte. So optimally you want to get 6-8tb in 2tb cards and a big ol' 6+tb hdd for storing old files.

ebon pewter
#

"You are running torch 1.12.1+cu113.
The program is tested to work with torch 1.13.1.
To reinstall the desired version, run with commandline flag --reinstall-torch.
Beware that this will cause a lot of large files to be downloaded." my auto1111 still working though. can i ignore the message?

merry inlet
#

anyone here running on 6000 series amd?

somber vapor
#

guys is there any difference between the dreamstudio website and dreambot on discord? and what about credits on discord, is it free or is there a way to connect our account on website?

plush haven
#

anyone know of any tutorials on using transfer learning to train an image classifier. I have 50k images to classify 😦

unkempt gull
vapid tangle
#

wait

#

dreamstudio is the official site right? in that case i do not know the anser to that

unkempt gull
merry inlet
#

@unkempt gull I am running ubuntu on a dedicated drive and have rocm

#

so no lol not supported but it works

somber vapor
astral goblet
#

So machine learning engineers have learned how to not just train a base model with random inputs, but rather give it a giant framework of ideas and concepts all encoded exactly how the want that, with a new programming language, and then refiine that new foundation base that is compiled exactly how they want it to be

vapid tangle
astral goblet
#

imagine you have an engineering problem and you have a number of solutions for various parts. you can encode those as key knowledge, then train that model from that base compilation.

it's going to become a pretty crazy world

silk fossil
#

this thing is so fucking cool.

empty barn
merry inlet
#

@empty barnso I have 8gigs vram 6600xt and I cannot make a single 768x768 on SD2.1

#

is that just not possible for me?

empty barn
merry inlet
#

@empty barnpython launch.py --no-half --medvram

#

if I don't have --no-half it won't load

empty barn
#

--no-half doubles your VRAM requirements.

merry inlet
#

I will try it again leaving it off and see what i get

empty barn
#

It should work, your GPU definitely has the capability to run FP16. It seems ROCm has issues with running it on not officially supported GPUs, sadly.

merry inlet
#

okay so it does load the weights etc and I can use the web ui but when I generate I get a lot of errors immediately

empty barn
#

What error?

merry inlet
#

RuntimeError: mat1 and mat2 must have the same dtype

#

Error completing request
Arguments: ('task(rb67kqf3dybeoiy)', 'cat', '', [], 20, 0, False, False, 1, 1, 7, -1.0, -1.0, 0, 0, 0, False, 512, 512, False, 0.7, 2, 'Latent', 0, 0, 0, 0, False, False, False, False, '', 1, '', 0, '', 0, '', True, False, False, False) {}
Traceback (most recent call last):

#

there are a lot of files inbetween those though

empty barn
#

We should probably switch to tech-support with this.

merry inlet
#

occam can you run any HIP commands? like to see gpu usage

#

Im not sure if this is the issue but I am running this on ubuntu 20.04 and using rocm 5.2.5 because I could not figure out how to get it to run on ubuntu 22

empty barn
#

Switch to the tech support channel, let's not clutter the general chat.

fervent thunder
#

hello , is it possible to use inpainting in batch mode

warm junco
warm junco
vestal dew
wide juniper
#

Anyone has any success using instructpix2pix to rotate a character?

vestal dew
#

I cant even figure out what it is or how to use it

ornate pilot
#

Random thought what if every time we prompt up a new picture and it’s all deformed and disfigured that thing we created actually pops into existence in the multiverse somewhere and looks at itself and realises how hideous it is and then jumps off the nearest escheresque building only to land back where it started over and over for eternity 🤔

vestal dew
#

That's not what happens

#

The thing already exists in the multiverse

#

SD does not work as claimed. It works by taking actual photos of things in different universes

#

Each time it does this, it weakens the space time fabric between universes

#

There is a race of people - Thumbalongs - on one universe who are working hard to get into our universe. They want to come here to steal our perfect thumbs.

#

It was they who created SD, got it propagated into our universe. We are unique in all the multiverse - our perfect thumbs

ornate pilot
#

Hmm interesting 🤔 maybe that’s what ufo’s are, it’s the AI from another multiverse stealing pictures of us to scam those higher dimensional beings thinking they are making stuff

vestal dew
#

likely so

#

Every time an artist creates an image here, every photo...it strengthens the fabric

#

But the AI is generating so many images that the fabric is weakening. Soon, the Thumbalongs will make it through.

#

Beware of people wearing gloves, hiding their hands, especially wearing mittens

#

people with club feet, missing limbs...these could all be Thumbalongs

#

their movement to break into our universe began, even though nobody there knew about our universe at the time, with the man who founded the continent where most Thumbalongs live.

#

Christopher Thumblongus discovered the continent. Chris had a thumb where his nose should be. This positioned his thumb perfectly for picking his nose. But he had no nose to pick

#

That frustrated him. Frustration led to anger, and anger led to the poorly defined side...or limb, or finger, or toe

ornate pilot
#

I’ve seen strange people in the park sometimes… who look like they are something else wearing people suits, just something not quite right about them they just scream fake human

vestal dew
#

Stephen King kind of had it with his Low Men in Yellow Coats

#

but he doesn't understand the whole picture

#

like an image from noise, this is becoming clear with the use of image generating AI

#

Thumblongus became obsessed with finding perfect thumbs. He wrote extensively on his theory that somewhere, in some "world", there are people with perfect thumbs. His writings inspired a movement, and eventually the AI technology came out of that.

#

The people propagating the technology have nobel intentions. They want to come to request thumbs, not take them.

ornate pilot
#

What if the multiverse is so big no matter what AI creates it already exists anyway and the amount of possible stuff that can exist out there is ♾️

vestal dew
#

however, preserved by some dark rituals, Thumblongus was preserved. Some say he still sleeps, others that he is already awake.

#

Understand, still sleeping or awake, sooner or later he will come for our thumbs.

#

it could be that some of those who first cross over - maybe already have crossed - are Thumblongus cultists or agents sent by him.

vestal dew
#

Elvis and JFK are not dead, BTW. They are hidden, awaiting the arrival of Thumblongus and his agents, preparing to defend us from the sinister plot to steal our thumbs.

#

Mark well the name Christopher Thumblongus, for even if you only now know it from my ramblings, soon you will know it well, and you will fear it.

#

Imagine what such a man, a man with a thumb where his face should be, would do to hide his identity if he wasn't ready to be known. What might he do to make people think of a face covering, something covering the nose, as normal?

ornate pilot
#

Me and my friends where coming home from fishing it was early sun hasn’t risen yet and we stopped at the traffic lights and along comes elvis park beside us looked exactly like him he was driving one of those classic le mans racing cars that had no roof and no paint just metal body work, lights turned green and he waved at us and he was gone

vestal dew
#

You see?

proud nova
#

This is so fun to read

timber dock
#

today I thought "I'm gonna actually use git instead of just downloading the zip"

#

bruh ls doesn't work on windows

vestal dew
#

In fourteen hundred ninety two, Thumblongus sailed the ocean blue. In fourteen hundred ninety three, he dreamed of taking the thumbs from you and me.

ornate pilot
#

My thumbs look longer than a normal thumb, I guess we expect the ideal thumb to be stubby?

vestal dew
#

try to generate an image of "man with thumb where nose should be" or something like that. It will be hard to get it.

#

That's because the Thumblongus cultists don't want us to know about Thumblongus

vestal dew
ornate pilot
#

There come a time perhaps when we can rewrite the human genome in a living person and like a negative prompt erase away features we don’t like and replace them with perfect ones, but how boring would that be ! Perfect people everywhere