#💬｜general-chat

1 messages · Page 40 of 1

vestal dew
#

Detail wise, isn't it better to generate a higher scale initial image? For example, if I do 512 x 512 I can get a detailed portrait, but if pull back to the full image, full body, details are lost. Does a higher resolution initial image address that?

#

You can change the resolution of the generated image

#

providing you have the VRAM to handle it

plush forge
#

Yes but then id have to know the dimensions

vestal dew
#

open it in a program that shows the details

plush forge
#

Is there no way to auto make it same as input?

#

Okok, ill taka that as a no

vestal dew
#

I use photoshop, but something like GIMP does it

plush forge
#

Ok but maybe there was an automatic option

astral goblet
#

Yeah there's no way yet for settings to automatically snap to the images dimensions

minor hatch
#

So can I use AI to blend the style of four different artworks I have made?

#

Pls dm to notify me habby agony

ripe girder
#

where i can generate

vestal dew
#

there is a PNG tab in A1111

#

If you load an image to that, then sent to img2img or txt2img, it will copy settings including the resolution

#

I don't think it copies/loads the model

#

I am loading now to check

ripe girder
#

how to generate here

minor hatch
#

Soo is this server used to create AI or is it another platform

vestal dew
#

eh, PNG doesn't give size

#

in windows, if you right click, go to DETAILS, you can see the resolution

minor hatch
#

I love being completely ignored🤦‍♀️

vestal dew
#

It is not used to create

#

you cannot create images here

#

png info sometimes sends resolution, sometimes doesn't

minor hatch
#

I made something

lilac reef
#

That is crazy good. Used or new and which brand?

lost marsh
lost marsh
#

the company selling the used gpu looks to be extremely reputable, plus my current 6900xt is also used, no issues whatsoever

rocky slate
#

howdy folks. anyone here that can spare a moment for a noob?

#

i want to learn to use controlnet and img2img with stable diffusion. is there a simple guide to this? does installing the model locally work as well as a paid option online?

vestal dew
#

youtube video?

lilac reef
lost marsh
#

this company renews it professionally, reapplying thermal paste or fixing fans. I checked before buying, anyways, if it ends up being damaged ill tell

rocky slate
#

during install is the consitent stream of "waiting for file to be created...." normal?

#

consistent*

lilac reef
lost marsh
lilac reef
#

never heard of them. Thanks

#

@lost marshPNY GeForce RTX 3090 24GB XLR8 Revel RGB | Fast Ship, Gift Quality!
$749.95

lost marsh
#

yeah thats on ebay

#

its 50$ more on amazon

#

i just needed more control over delivery

lilac reef
#

Well, I just spent 60 in gift cards on Amazon for my new 32gb of ram

#

damn

#

I haven't used ebay is oh, my maybe 6 years

#

They seem to be sold out on Amazon

lost marsh
#

oof i bought one last night

lilac reef
#

have a link?

lost marsh
#

nop, that amazon link is the only one

lilac reef
#

3090ti FE 1800 on amazon. LOL

#

I didn't see an amazon link

lost marsh
#

oh, the specs on pny cards are pretty much the FE cards

#

the one you found on amazon is it

#

theres no other link

lilac reef
#

yeah, PNY is who does their pro cards

lost marsh
#

i looked around plenty yesterday

lilac reef
#

Yeah, I just did a general search and nothing less than 1200 renewed so must be OOS now

wintry stream
crude willow
#

@olive sentinel dont put a command in pow submissions

#

there are actual channels to do that in, you are just typing useless text

olive sentinel
#

ok

waxen pasture
#

Dreambot is currently gone

crude willow
#

oh I thought they brought it back

waxen pasture
#

No not that I know of

wintry stream
wild steppe
#

hey @fiery dagger

#

what do you need help with?

waxen pasture
wild steppe
waxen pasture
storm shard
lusty verge
#

How are you doing

vestal dew
#

How YOU doin'?

wide hare
#

does anyone have a list of good models?

stable pawn
fervent thunder
#

How you doin

fervent thunder
wide hare
wide hare
wide hare
vestal dew
#

there was an extension I had...perfect or something like that...I deleted, but I want to get it back

lost marsh
#

made such a good image but my character's arms got amputated sob

stable pawn
#

One of my favorites is "Deliberate", it can do nudes but keep porn to a minimum so it rather safe.

wide hare
fervent thunder
#

well, I love some good realistic look, and analogdiffusion was one of my personal recent hits.
Loved also all the disney craze with ClassicDisney diffusion
My personal preference stays with Elden Ring model
And I'm a little biased since I just shared mostly models of my friends but I have awesome friends, just wanted to disclose.
I did a lot myself, but don't really like my personal style lol
No, to be honest, my top 3 :
realistic look : Analog diffusion, redshift diffusion
Not realistic : Elden ring
https://huggingface.co/wavymulder/Analog-Diffusion
https://huggingface.co/nitrosocke/elden-ring-diffusion
https://huggingface.co/nitrosocke/redshift-diffusion

#

I can retrieve links of any of those if you are interested

vestal dew
#

do you use any extensions...I wish I could remember the name.

#

ah, it's a negative. Badprompt

wide hare
#

Thanks!

fervent thunder
#

no problem

#

I'll plug mine since we are there lol

wide hare
#

ill also try it thanks!

fervent thunder
wide hare
#

what is your gpu?

fervent thunder
#

I have 2, the training GPU is a 3090TI

wide hare
#

damn

#

i have a 3070TI and tried creating models and it wouldn't let me coz I don't have enough Vram

fervent thunder
#

you have how much ? 8?

wide hare
#

yeah

fervent thunder
#

yeah, on that you don't have all options, but some are available still, I think it's TI training that is at 8? I'm not sure to be honest, I've been out of Internet for about 2 months, I came back yesterday

vestal dew
#

Lora too, no?

fervent thunder
#

maybe ? not sure, what I've told since I came back on LORA was from people that work on it, but I didn't try it yet

#

I may test it for the PoW, not sure, I need to see if I do one or not

hardy iron
#

Can I download stable diffusion on MacBook m2?

wide hare
#

although loras are still possible

fervent thunder
#

LORA feels like a godsend honestly. It's almost as good as finetuning for real, but is so small in comparison, and has the big benefit to just be an embedding, so you're able to mix and match with others too

#

it's quite incredible in comparison to dreabooth, in versatility

rain seal
#

I'm really new to Stable Diffusion and I haven't installed yet, how do you recommend learning it and setting it up? Note: My GPU and CPU can run it.

wide hare
fervent thunder
# rain seal I'm really new to Stable Diffusion and I haven't installed yet, how do you recom...

for setting it up, you can find some help on #🤝｜tech-support if needed be (check the pins too). There are different ways to use it, with more or less ease and features. Most require some minor technical knowledge to install for now, even if it gets easier by the day, it's not as easy as installing Word for example. Then for learning how to use, mainly practice, and the community here, as well as the doc of the tool you'll choose.
The currently most popular tools are known as AUTOMATIC1111 WebUI and InvokeAI.

rain seal
rain seal
fervent thunder
#

no problem, and don't hesitate to ask, the sharing spirit is quite strong around here

wide hare
fervent thunder
#

also, I like to plug this when new people come into the server and participate

#

please, do check the #⭐｜pow-info and if you feel like it, try and post a picture next week once you get acustumed to SD and feel like you can do something for next week's theme !

#

nothing to loose, some prizes to gain, and a community that thrives even more and learns every day

rain seal
lilac reef
#

I have been training on illuminati_diffusion_v1.0 a lot lately.

fervent thunder
#

hey general !

lilac reef
#

Heyas

fervent thunder
#

happy to see you, I didn't since I came back ❤️

#

I missed you tbh

#

I know we never talked a lot but I loved your art

lilac reef
#

Ahhh, thanks. A lot has changed over these last few months and it is very dead lately. hours and hours can go back between postings

fervent thunder
#

yeah I kind of saw that... general with images was dying already when I took a break

lilac reef
#

2.1 same now

fervent thunder
#

I intend to make it alive again, or at least try

#

for sur though, it's been out for a while now

vestal dew
#

seems like nothing happening with 2.1

lilac reef
#

great. I took all of January learning why I couldn't train 2.1 with well over 100 hours until I finally figured it out

#

LORAS, and Deltas

fervent thunder
#

you figured it out ?

#

I never managed to train 2.1, it made me desperate on training at some point tbh

#

my results were just so bad compared to 1.5

lilac reef
#

yeah, fp16 don't go over 100 for dim/alpha or the higher you go the less it learns because fp16 lacks the range of bf16

#

at 128 it learned zero

fervent thunder
#

100 for DIM, what do you mean ?

lilac reef
#

you need to know Lora to understand dim/alpha.

fervent thunder
#

so you need to train bf16?

#

ok

lilac reef
#

if you have it yes

fervent thunder
#

I need to learn lora yet

lilac reef
#

my card, nor colab's T4 has it

fervent thunder
#

ok nice though then, I mainly used everydream, and I think it got compatible with bf16

#

I need to update still

#

I haven't updated any training thing yet

lilac reef
#

as long as your hardware has it then yes, use it

fervent thunder
#

only invoke and auto on the other computer

#

I should be able to use it yeah, I just need to reduce a little batch size I think

#

I'm on 24GB

lilac reef
#

I think it is so dead because it is what, 3 months now since anything for SD?

fervent thunder
#

I haven't checked the news, but it seemed a little dry, yeah...

#

I know it's chatGPT that got me to AI game a little too

#

I hope we see something on text generation from Stability too

lilac reef
#

Lora is really cool, btw, and blows TI embeddings away...most times

fervent thunder
#

(just a reminder that I'm just a community mod and not teasing anything)

fervent thunder
#

lots of other upsides though, like size or merging capabilities

lilac reef
#

in the training food chain DB/LoRA/HN/Ti

fervent thunder
#

and did you try trainging multiple network types on the same concept/dataset to activate them all at once ? I was thinking of doing real scale tests on this, comparing HN, TI, LORA and DB

lilac reef
#

still people have issues doing styles in DB

vestal dew
#

We need an AI that operates Stable Diffusion

fervent thunder
vestal dew
#

an AI to handle and operate other AI's

lilac reef
#

I honestly have had no need of multi-concepts but it looks like it is easy to do

fervent thunder
#

I was talking about similar concept

#

having trained DB on like, Zelda

#

and a TI and a LORA and an HN

#

and use all of those at once

#

maybe overkill

#

lol

#

UBERZELDA

lilac reef
#

I really wish HN had the optimizations the others do, but for some reason it just never came. I mean 7h to train one on 15gig where I can only use a BS of 1?

#

Yes, LoRA + TI on even the same images give some really different, and fantastic, results

fervent thunder
#

ok good to know

lilac reef
#

though I have found some tis that lora blows up with.

fervent thunder
#

or to confirm at least

#

problem with testing this is, it's quite subjective

#

being 100% objective and rating things is hard when it gets close

lilac reef
#

what is bad about LoRA is that for 2.x Auto has never fixed it so we can use them natively meaning we have to use a script.

fervent thunder
#

ho

#

so no 2.X native lora, good to know

lilac reef
#

Auto can't seem to get the keys to match

#

that means I can't make a negative lora which blows as I was really digging negative prompt ti trainings like my bettermaker.

sleek raven
#

guys how can SD2.0 do upscale of image? It's only trained for max 768x768. How can it make images that's double that size?

waxen pasture
sleek raven
#

what?

waxen pasture
#

are you using auto1111 webui?

sleek raven
#

yes

#

I am asking how can such crazy thing be possible

#

not how can I do it

#

sorry for the confusion

waxen pasture
#

Oh

sleek raven
#

it's trained for 512x512 or 768x768, but it can upscale images beyond that resolution. That sounds crazy.

waxen pasture
#

Well from my understanding, it's basically doing larger images by splitting it up into segments and diffusing those seperately, which is why you can get weird duplication in large images.

lilac reef
#

with overlaps

sleek raven
#

does it mean during training, I need to give my model close of images?

waxen pasture
#

No

sleek raven
#

like if there was a rabbit picture, do I need to do close ups on its paw, face, and such?

waxen pasture
#

Not really

sleek raven
#

then where did it learn to fill up the details from?

rocky slate
waxen pasture
sleek raven
#

I didn't get that. But I guess it means the model is treating it like smaller section of work it needs to do and does what it needs to do just like it did for lower resolution stuff

waxen pasture
#

pretty much

sleek raven
#

SD model is about 5 GB when sitting in VRAM

#

I am wondering why they don't make bigger model?

#

bigger is like, more powerful right?

waxen pasture
sleek raven
#

I have a question

#

I tried to train "man" but I realized whenever it hears "man" there is a high chance that it will draw "mansion" a house. Do I have to use embedding or dreambooth? "man" seems like a concept that 2.0 model doesn't understand

trail ledge
#

make your own

#

if the model you're using don't have one

waxen pasture
sleek raven
#

picture of man

#

man

#

yes

#

single and multiple

#

picture of man then it draws a house

waxen pasture
#

try picture of a man, person, standing, brown hair, smiling
and in the negative house, mansion, architecture

sleek raven
#

man, also draws a house sometimes

waxen pasture
#

Your prompts are not detailed enough. SD needs more to work off of.

sleek raven
#

ok yes it stopped rendering house lol

trail ledge
#

if you don't give it negative prompts, it'll get creative

sleek raven
#

I see

#

I am studying embedding rn

#

if I make an embedding styleHell, whenever I use styleHell, I am using two tokens in the prompt right? the "Number of vectors per token" setting does not affect how many token is used when I use that embedded prompt, right?

#

the max [Number of vectors per token] is 75. If I had thousands of pictures, setting it to 75 will work with embedding to create an AMAZING embedding for that concept? Am I understanding it right?

#

ohhh

#

Number of vectors per token: the size of embedding. The larger this value, the more information about subject you can fit into the embedding, but also the more words it will take away from your prompt allowance. With stable diffusion, you have a limit of 75 tokens in the prompt.

#

I get it now that's why people were calling it embedding is just a clustered prompts

fervent thunder
#

token count is ... once again, and I'm sorry for it,.... not that simple

#

more tokens doens't mean more quality

#

for one, more tokens in your embedding means less tokens available for the prompt itself, and that can reduce quality

sleek raven
#

I see

fervent thunder
#

and secondly, more tokens in the training itself won't mean more quality either, after a point.
From what I had tested, with the old version of TI, there was some benefit, especially going from 1 to 2, or 2 to 3, ... but after 12? no more changed, not enough to justify the tokens

#

it may have changed with the new TI though

#

and if I was to describe a TI, it's no more than a specially trained talky walky to talk to your model, to access the weights you NEED, the ones you want. it doens't add concept, it access it better

sleek raven
#

Then is the best way to teach a new concept dreambooth really?
If I want to teach "catdog" a hybrid animal of cat and dog -something really cute- and (let's say) I have pictures of this fantasy animal, then is just running dreambooth the best way to do it? ( I am asking because I thought dreambooth can't teach a new concept)

fervent thunder
#

dreambooth can teach a new concept, for sure. TI can do, to a certain point. in your example, cats and dogs are close enough, and there are some cat/dogs in the artworks that SD was trained on, so TI should work correctly

trail ledge
#

well, you're assuming catdog, as a concept, doesn't already exist

fervent thunder
#

but if it was more abusrd, outside of what SD knows at all, then yes, LORA, Dreambooth, or even hypernetworks, would be necessary

#

(HN would be the worst quality for this though)

storm shard
trail ledge
#

TI is relatively easy to make and doesn't take a lot of vram relative to the other methods, so make one using, say, 20 images

fervent thunder
#

yep

trail ledge
#

also, 1000s of images isn't going to make it better, since you'll be challenged to caption all the pictures correctly

fervent thunder
#

I'm working on a training guide right now, and here is what I put for TI : TextInversion doesn’t modify the model directly. As such, it’s the least potent of the lot, but you can use multiple Text Inversion at once, and they only activate when you prompt for them. They are very light to share, and building a library of those is quite easy.

trail ledge
#

20 well captioned inputs is better than 1000 badly captioned inputs

trail ledge
fervent thunder
#

merging = blending ?

trail ledge
#

I love me some (face_embed1: face_embed2: 0.6)

#

I don't know the right term for it

fervent thunder
#

ho ok

#

you use both at once in corelation

#

didn't know about that

trail ledge
#

yeah, been making virtual people using different embeddings 🙂

sleek raven
#

Do you need to caption when you make embeddings as well?

trail ledge
#

yes

fervent thunder
#

depends on the method but usually yes

sleek raven
#

I thought you just need to define what embedding will be like 1girl and just add bunch of pictures with one girl in in it (using anim community as an example)

fervent thunder
#

there is an option to use the concept as name instead of the caption

sleek raven
fervent thunder
#

that's what I explained to you earlier, and yes, but really this is one use case. TI can be a lot more versatile

#

it's slower to teach, it tries to respect more the rest of the model, ... it's regularisation data, and given your example, and the training curve for learning how to train a model, I though skipping it was best for you, and I suggested it too (when I told you to skip class/regularisation)

sleek raven
#

I think I need to start a project. My computer is going to be here tomorrow. So hopefully I can start it tomorrow.
If I want to just really make good ramen pictures. Pork belly ramen, vegi ramen, egg ramen, egg ramen on bamboo table, leaf ramen, pizza ramen, cheese ramen, carrot ramen. Should I use embedding to teach what "ramen" is? or Use dreambooth?

trail ledge
#

start with just using good prompts, both positive and negative

#

"ramen" is already in a model, I'm sure

#

if you really can't get what you want, then go with textual inversion next

#

and possibly LoRA

#

imo Dreambooth should be last, as it is the least flexible and take up the most space

#

but, if you really want to teach it something that really isn't in the models, then sure, go dreambooth

#

but you may find there are a lot of things that are in there already

warm junco
#

In many Anime models there is Ramen 😋

sleek raven
#

it can generate ramen with good promps positive and negative. But quality of ramen is very bad

#

I want top of top photorealistic super delicious looking ramen!

trail wing
#

gatekeeping ramen with AI

sleek raven
#

look at the images I generated

#

ramens the current 2.0 generate looks like spaghettis or soup, also looks very disgusting

#

it added mini tomatoes in the bowl as well

#

AI DOES NOT!!! know what ramen is

warm junco
#

Yea then train a lora xD

#

Super Ramen 2.0

sleek raven
#

bot that's REALLY good at generating ramen yes

fervent thunder
sleek raven
#

I do have a powerful computer(soon), I think I can do Dreambooth. In such case, should I do dreambooth or still go with LORA?

fervent thunder
lucid lynx
#

Hi all. Can someone suggest a guide for installing a model that can create AI avatars like lensa?

west sandal
#

hola soy nueva en esta paghina

#

y apenas estoy aprendiendo de esta pagina

fervent thunder
viscid mica
#

hello im looking for help i cant find webi-user.bat this is my next step can anyone help me

vestal dew
#

look in the stable-diffusion-webui directory wherever you installed it

lucid lynx
vestal dew
#

models can't do that on their own

#

to make SD make your character, you have to train something like a LORA

#

seen?

sleek raven
#

when captioning, I noticed I use lots of "a", "the" , and "and"

#

also lots of ","

#

commas

#

do I need to refrain myself from using these as they will be tokenized and recognized by the SD?

fervent thunder
wanton vigil
#

I'm training a Lora right now. I have 28 images, how many steps would I need to name my file?

fervent thunder
#

I have low experience with LORA, but with dreambooth, I trained for 100 steps per pictures first, then I started doing checkpoints every few thousands steps, adding up to 100% more steps to the initial 100 steps. it depends how fast it learns, and it requires you to test checkpoints and compare results, but it was the best way to get good trainings from my experiences

wanton vigil
#

Alright, thanks

lime mulch
#

Could someone define what the various "Sampling methods" selections mean?
Like, how they affect the image generation or if there are certain sampling methods that are better than others?

#

So far, I've had some decent luck with DDIM, but idk what it actually means lol.

trail ledge
#

There's no easy answer. In general Euler A and DDIM are considered the fastest with decent results at 20 steps, so if you want to experiment, they're a good go to, but not usually considered the best for final images

#

different sampling have different strengths and weaknesses. Some take more steps to come up with good images

lime mulch
#

I've also had a little bit of success with DPM++ 2M Karras.

trail ledge
#

best is to use the X/Y/Z Plot script to generate a grid of samplers vs. steps for the type of images you're going for

lime mulch
#

If there's one that gens better, but i just need to increase steps for, I'd def love to hear it. With a 3090, I'm blessed with 30 second batches of half a dozen images, so the time is thankfully not an issue.

trail ledge
#

like I said, it's more style dependent

#

so anime may work better with one, vs dreamy, vs 3D, vs photorealistic

#

if you have a 3090, just run the X/Y/Z Plot and decide for yourself

lime mulch
#

umm...

trail ledge
#

best way to get the sampler that works best for the style you're going for

lime mulch
#

I have no idea what the xyz plot is or how it works xD

#

Generally, I go for hyperrealism, but also trying to learn how to do anime now.

trail ledge
#

it's a bit out of date, since back then there was only X/Y, and the UI has changed, but the general idea should hold

lime mulch
#

Hmm, guess that's another one that doesn't work for me. =\

#

Like, I can't do the hires fix, and I can't do more than 6 for batch size, or it freezes up or otherwise just doesn't run.

vestal dew
#

not enough vram?

lime mulch
#

24 gigs enough? :p

vestal dew
#

ah

#

amdor nvidia

lime mulch
#

amd caps out at 16gigs for the 6900xt / 6950xt, it's the 3090 that I have.

#

I did hear that AMD cards have a hard time with SD, which is unfortunate.

fervent thunder
vestal dew
#

What does it take to do hires fix then?

#

I know people do upscaling ...do they have more than 24 VRAM?

trail ledge
#

hires fix works with less, vram just constrains the final size

lime mulch
#

Most likely I've just got some kind of issue with my install or something. :p

last merlin
#

Hires fix sometimes make the human face looks better any idea why

vestal dew
#

mine errors when I try hires fix

#

GTX 970 crap card

fathom thistle
#

did they add anything new to stable diffusion? cause i no longer could generate my anime type of images from Anything model even when i added in the model and vae

grave copper
#

what model/vae would i use for general images

#

something similar to midjourney

trail ledge
#

there's Openjourney

grave copper
#

openjourney?

#

mhm

#

and what vae along with it

trail ledge
#

i'd just use the default vae-ft-mse-840000-ema-pruned

static leaf
#

hello, do you know any way to add a person to an image via stable diffusion?

grave copper
#

ty

trail ledge
grave copper
#

and how do i download anything off of huggingface

#

anyone

trail ledge
#

that's kind of open ended question, but... click on the arrow?

grave copper
#

what arrow

#

and should i download openjourney or oj-v2

trail ledge
fathom thistle
#

any stable diffusion expert here? i cannot seem to apply my models to generate the image style i want from the current version of stable diffusion

#

also previous version i can only use 1 model, now i need 3 model to generate image?

#

nvm, that was checkpoint merger option

#

nvm, i found a save settings button after i select the vae, no issue now

shrewd arrow
#

hi, guys, is there a channel for PC hardware talking? I wanna build a PC specificly for AI art, what CPU is good for it

#

I already know RTX a4000 is a good GPU card , but its release date is 2021, in case buying a used GPU card , im considering RTXa4000 vs the new 4070ti,

past edge
#

so 3090 or 4090 if u a baller. dont even bother with anything else. amd is a naughty word around here too

#

If u gonna go for professional graphics card then go for a big daddy 80gb vram card otherwise ur better off with 3090. Any smaller card and the tradeoff for a card that is no good for gaming is not worth it.

shrewd arrow
#

20+GB vram card are about 2000$ , far beyond my budget, sorry , me not a baller

past edge
#

wut. 3090 is around $700usd rn. lol

#

that was way back in 2021

shrewd arrow
#

really， I will check that. but some article says RTX a4000 serie have raytrace and tensor core , don't know what, but sounds cool for AI art

past edge
#

no u have the wrong info. any card that starts with A is a professional card which I don't recommend.

#

All 3000 series have raytracing

shrewd arrow
#

im searching information for a CPU to go with RTX a4000

past edge
#

3090 is best value rn

past edge
#

make that 11th gen

shrewd arrow
#

why you don't recomend a profesional graphic card?

past edge
#

because the only ones that offer more than 24gb cost thousands and thousand and thousands.

#

The 3090 is only $700 and comes with a generous 24gb

shrewd arrow
#

they say profesional card are good for 3D and AI calculation

past edge
#

not by much.

#

the ones that you will find for cheap are old generation ones which arent worth it because they are only good for AI etc. not for gaming

tough bough
#

They are only good if you are using multi-GPU setups.

past edge
#

3090 is good for AI and gaming and everything else as well

shrewd arrow
#

only good when using muti-GPU, ? you mean the RTX a 4000 is only good when use many of it ? like rendering factory?

past edge
#

pretty much

#

Only good reason to spend on pro cards are for boatloads of vram which is worth more than gold per gigabyte

#

but certainly not worth 10 grand. im happy wif my 24 giggybites

astral goblet
past edge
#

nah G. used cards are where its at

astral goblet
#

i wanted a 3090 but couldn't find one that wasn't shipped from across canada way above msrp. so i compromised and got what was in stoock at msrp, a 4080. not as much vram but still super good value

past edge
#

any card at msrp rn is good value imo. But yeh, VRAM is the name of the game when it comes to AI

astral goblet
#

7900xtx is on par with 3090ti for gen speed apparently too

past edge
#

amd lmfao.

#

nty.

astral goblet
#

yeah me neither, but their new rdna3 cards actually have specialized cores finally i heard

past edge
#

I mean if u like banging ur head against a wall using linux then yeh go for it

#

Ima stick with good ol windows n' raytracing. team green for life

astral goblet
#

while i had amd hw i was an arch user lol

#

i often jump to new os's because i do enjoy figuring things out and knowing the whole field. weird i guess

vestal dew
#

4070ti msrp over 800...yet they can be purchased for under 700 at times

past edge
#

that is very very true. I should have added a disclaimer

#

4070, 4070ti and 4060 are absolute dogwater same goes for 4080 12gb

shrewd arrow
#

CUDA , do you mean, the reason why 3090 is so cool for AI is because 24G vram, and those tensor numbers and AI core numbers are not important. ??

past edge
#

yep. Idk about that tensor stuff

astral goblet
#

with xformers i've spread out a lot with large batches for training. it aint so bad at all

vestal dew
#

yes, but 4090 is a similar price to 3090, also 24 GB, so why buy a 3090?

astral goblet
#

xformers hasn't even been released for cuda 12 yet afaik

past edge
astral goblet
#

tried building it but , was banging my head on the wall too long there

past edge
#

Last time I shopped for a 3090 I got one for $700... lowest I could find for 4090 was $1650...

shrewd arrow
vestal dew
#

YES, Im talking price in stores

#

not price from private sellers

astral goblet
#

i wanted a 3090ti but i also was prepared to get a gpu that day

#

not another day

past edge
#

Ugh tell me about it. I wanted a 3090ti so bad but found nothing but indian scammers and cheapskates

#

but hey at least my ftw 3090 has pretty rgb. lol

astral goblet
#

i got the 16gb

vestal dew
#

on amazon, 3090 is over 1000

past edge
#

16 is still pretty good

astral goblet
#

i think all the 12s were pulled

past edge
#

nobody buys gpus from amazon

#

amazon is the worst scalper of all

vestal dew
#

where do you buy them?

past edge
#

newegg if I want a new one, otherwise 2nd hand market

#

best buy too but theyre always sold out

astral goblet
#

you're gonna get garbage teir tested hardware from amazone

#

cheap bootleg msi sticker on it

vestal dew
#

Newegg, 3090 over 1000

past edge
#

let alone that, they overprice all hardware

#

let me check hold on.

vestal dew
#

I'd like to know because I'd love a 3090 or 4070 TI around 700

astral goblet
#

cool frank tutor. what point are you making? that it's still substantially less than the 4090 and you're very wrong still?

past edge
#

....wtf? Ur right theyre hella expensive rn. Why this price increase all of a sudden

vestal dew
#

found a 4070 TI BB open box 650, but it wouldn't let me complete the purchase. Then someone else bought it

astral goblet
#

market demand heh. twe want the cards

#

we need the cards

past edge
#

dayum. well im happy to be a smart shopper and snagged one at a decent price.

vestal dew
#

see, the 3090 in stores all over 1000

past edge
#

if u look hard enough u can find a decent one from someone desperate to sell

astral goblet
#

you can find scores for sure

#

they come up rare

past edge
#

the problem is that im seeing them at 1800+... like wtf?

vestal dew
#

I found some on discord, but already sold at 800

past edge
#

only a moron would pay that much for a 30 series... damn its sad

vestal dew
#

yes, a few over 2000

past edge
#

shame on newegg

vestal dew
#

amazon, newegg, ebay

#

all over the place

past edge
#

ebay ive seen many go for low

#

700-850 easy

#

used of course

#

and ebay has seller protection so if they sell u bullshit u can get ur money back

vestal dew
#

3090's on Ebay refurb and used around 800. New over 1000

#

BB had a 4070ti at 650 open box

past edge
#

lmfao thats because 4070ti is a joke

#

I think even the 3060 outperforms it

vestal dew
#

no

past edge
#

idk man I dont phuk with 40 gen that whole shitshow is a scam

vestal dew
#

In benchmark on SD, it performs only under the 4090

past edge
#

the only 40 gen I would be interested are 4090 and the hopefully upcoming 4090ti

#

everything else 40 gen is shady af

lilac reef
#

anyone know of any embedding that will change the background? All my backgrounds are orange/yellow and I can't change them.

vestal dew
#

faster than 3090 although cant do as much training and large images

#

I'd jump on a 3090 around 700

#

cant find it

past edge
#

oh it took me a lot of bargaining

lilac reef
#

yeah

past edge
#

so yeh ull have to sweat to get one

lilac reef
#

3090ti 800 I would grab

past edge
#

offerup and facebook marketplace are good places to look for

vestal dew
#

yes general

astral goblet
#

i'd build a workhorse pc out of the 3090 if i saw it that low. ina heart beat. and game on my 40

vestal dew
#

although money is tight, so Im trying to get 1500 or less on the build. 100 is significant

astral goblet
#

shit games so hard

lilac reef
#

needs a reputable seller though else get cards half dead who spent their life mining

vestal dew
#

right, I don't want that

past edge
#

well I had a perfectly good gaming pc with a 3080 but got so into AI that I had to sell and buy a 3090

#

just for that juicy sweet extra vram. I hate cuda errors

vestal dew
#

the guy who bought the 4070TI at 650 also has two 3090's he has never used

#

but he doesn't want to sell them

past edge
#

I seriously doubt it sold for that low.

lilac reef
#

I was saving and Nvidia greed hit. In 2y my computer comes up for its 9 year replacement. I just saw what B650 entry level cost now. OMFG

past edge
#

receipts or it didnt happen lol

vestal dew
#

It id, BB had open box

astral goblet
#

miners usually would undervolt their hardware to save electricity once they found out pushing them past factory limits was inneffiicient and they earned more per watt that way. only cards around rx 480 and early gtx 980s would be firmware hacked to go harder.

vestal dew
#

My computer is 10 years old

lilac reef
#

Since when is entry level 200+ for mobos and 600 for gpus? Since 2023

astral goblet
#

the wave of mining horses today should be healthy stock

vestal dew
#

My life remaining is short, and I'd like to be able to use the SD with things like controlnet in the meantime

past edge
#

exactly i saw some dude selling his 3090s stack at like 700 each but I didnt buy from him just cuz I dont trust miners

lilac reef
#

My final upgrade arrives on Tue which is another 16gb of mem. 32gb of mem and done until this beast is 9-10 years old

past edge
#

controlnet is a gamechanger. fuggin luv that shit

#

cant wait for a true text-to-video engine

#

ima be all over that mf

astral goblet
vestal dew
#

is something wrong with the EVGA 3090?\

astral goblet
#

a controlnet trained for next frame generation should be doable, just gotta do it.. ugh

lilac reef
#

Yeah, no damn mining cards. Look, pristine never used. Take the heatsink off and the chips are being painted.

past edge
vestal dew
#

I see a listing for NEW 549.99

past edge
#

link?

vestal dew
past edge
#

keep in mind theres a shitload of indian/nigerian scammers out there

#

wtf

#

that looks legit

#

damn I wish I coulda found that when I was shopping for one. oh well

#

I say go for it and if its bs just ask ebay for a refund

vestal dew
#

8 items sold 100% positive feedback

past edge
#

theres also only 2 available

#

go for it bro

#

u wont find a better price

lilac reef
#

I dunno, it just seems so wrong to be forced to buy cards from 2 years ago for so damn much money.

vestal dew
#

seems too sketchy

past edge
#

sounds like a scam but u dont lose nothing by trying

#

I would even record myself opening the box in case I have to submit a dispute

#

u can also check to see if the serial number from the box matches the one on the card

vestal dew
#

also, it says 7 days before it ships...extended handling time

astral goblet
#

some of my best deals are from american ebay, ship it to a boat dock in seattle, it comes across to canada, i broker it myself pay the bit of tax and bongo bango

#

hunting for deals in canada is stupid

vestal dew
#

Seems like a scam

lilac reef
#

No thank you

past edge
#

that is very very strange indeed

#

there is nothing quite like buying in person.

vestal dew
#

all similar: Duct Tape, kitty litter, mop bucket, RTX 3090

past edge
#

just tell em hey can I see it running on kombustor for a bit. if they say no just move on.

lilac reef
#

LOL

vestal dew
#

how would you know it's really running?

past edge
#

simple. most gaming pcs have tempered glass. so if theres more than one GPU in there u would see it. and if its running on integrated graphics, it would look laggy as shit on the heavy tests

lilac reef
#

The worst part about modern gpus are the fans.

vestal dew
#

this is like me selling the stuff some scammer bought on my Amazon account, except that stuff is legit

lilac reef
#

why don't they use full sized fans? I mean they already take up 3 slots so reduce the heatsink and put a nice 120mm quiet fan on it. Nope, cause that would be too good

vestal dew
#

the scammer bought about 18000 worth of stuff, same stuff over and over, on my account. Paid for it some way. Amazon canceled some orders, delivered some which I wouldn't accept - I should have accepted them and sold them - and I have about 700 worth of stuff that was delivered

past edge
#

would one fan be quieter than 3 small fans tho?

lilac reef
#

yes

vestal dew
#

yes, it could spin slower

past edge
#

I used to have my AC running all the time so I didnt notice, but now that it's quiet in here I can see how loud it is even at rest.

lilac reef
#

more flow too

vestal dew
#

like one of those whole house fans

#

they spin very slowly

#

we need to adapt some bug to absorb heat, put out cold. Then a bug could live in the computer, eating the heat

lilac reef
#

I had this happen twice where people ordered on Amazon Apple shit and had it sent to me. I caught them by sheer luck. From then on I order, as I just did, and remove my card from the wallet.

vestal dew
#

that's the odd thing

past edge
#

damn thats fucked up

vestal dew
#

they didn't use my payment info

#

they just used my account, ordered the same things over and over until my account was locked

#

27 electric fireplace TV stands at $540 each

past edge
#

and they paid for it?

lilac reef
#

they used mine in the wallet. I mean the one I remember were 800 USD Apple Air Buds. wth? Why order that to have it sent to me?

vestal dew
#

7 packages of 4 smart plugs - I have those to sell now

past edge
#

oh they probably used a stolen cc

vestal dew
#

2 packages of 4 recessed LED light fixtures - I have those

#

Amazon doesn't want the items I have back

past edge
#

i'll take em. lol

vestal dew
#

I should have accepted the tv stands - 17,000

#

I am selling them

past edge
#

oh wait. they require hard wiring. nah nty

lilac reef
#

Why can't someone order a 4090 and send to me?

#

I didn't order this. No? NO! Keep it we will refund the card

past edge
#

I do need a tv stand tho lol

vestal dew
#

I have to pay about 7000 to fix my house foundation

#

I had money saved for laser surgery for my eyes - that's going to pay for the foundation

lilac reef
#

For what I do I really need the 4090 but no way in hell

vestal dew
#

yes, that's the card to have

astral goblet
#

training

vestal dew
#

but way too much

lilac reef
#

yep

astral goblet
#

also, generating datasets for training

past edge
#

wait until 50 series cards come out

vestal dew
#

people say they still get CUDA errors with 24 giggity gigs

lilac reef
#

1200 for the 4090 (the real price) it would be mine

past edge
vestal dew
#

ah

past edge
#

but some of them are

vestal dew
#

what else causes the CUDA errors

astral goblet
#

you can push vram out of bounds with 24 still. ez. 40 would be better

past edge
#

SD is unoptimized as hell so sometimes it can spike how much vram it asks for

astral goblet
#

xformers helps a lot

past edge
#

tbh idk, but I know 100% for a fact it does

lilac reef
#

I just read a 48gig TITAN is about to drop

astral goblet
#

with 16 i can hit an 80 batch

past edge
#

those cards are always $6000+

lilac reef
vestal dew
#

spiking to over 24 is still a VRAM issue thought, no?

lilac reef
#

Only 1?

astral goblet
#

good luck fitting it into a case

past edge
#

by torch

vestal dew
#

8000 for an A100 WITH 40 gIGS

astral goblet
#

i got a trcik out of the a1 wiki for that

past edge
#

so lets say u have 20gb of vram left, and u need 8gb for a job only, if for some reason there is not a fragment of at least 8gb available it will give u an error

lilac reef
#

I have been saying the day is fast approaching the GPU will be in its own case, with its own PSU requiring a separate household circuit attached to your pc by the newly revised NVLINK 3000

vestal dew
vestal dew
past edge
#

but going from 12gb to 24gb I can totally say the cuda errors are far rarer than ever

#

any error usually now is due to poor coding with dreambooth etc

vestal dew
#

they can sell a GPU with its own little nuclear power generator

#

for the card and fans

astral goblet
#

set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.9,max_split_size_mb:512 in the bat and while initial memory usage climbs faster, longer run it is lower overall. more useful for larger vrams

past edge
#

lol. they dont use as much power as people claim

#

a microwave uses 3-4 times more than any card

#

make that 5-6 lol

vestal dew
#

can SD work with a microwave?

lilac reef
#

If on Windows in Automatic1111 put this in the startup.bat (linux I forget what/where)

set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:24

vestal dew
#

to maybe bake an image?

past edge
#

because they keep changing the goddamn torch versions or whatever stupid dependency gets updated

#

so what works one day may not work next week

lilac reef
#

512 ceased I had to set it to 24

#

I also set the threshold lower

vestal dew
#

if I ever get 24 GB, I'll try it

lilac reef
#

I will say this on my 6gb since Jan's big update to automatic I no longer can use my old prompts as some of them were 2048x1024

#

he screwed with the mem shit I read

past edge
#

bro I have a 3090 and I dont even think I can even do that much rez lol

#

unless its upscaled or sum shit

vestal dew
#

I was going to buy the heck out of that 4070 TI open box at BB

#

12 giggity gigs

past edge
#

nah G 12 gigs aint shit

#

24 or nothin

lilac reef
#

No, you can but use lowvram or medvram it shoved it into cpu ram

vestal dew
#

better than the 4 I have now

past edge
#

I really hope the 50 series feature 36gb vram. yuuuuummmy

vestal dew
#

when are those coming out?

lilac reef
#

2y

past edge
#

yeh sadly a long time from now

#

but tbh, our biggest model is 768x768 so even with that ur good with 24gb

vestal dew
#

I want to buy soon

past edge
#

any more than that is just for futureproofing

lilac reef
#

not even a true 768x768

vestal dew
#

If I could find the right price local, I'd go buy it

past edge
#

Cuz once that 1024x1024 model drops... thats when big vram comes into play

lilac reef
#

I do know that before 2.0 dropped the roadmap was 2-768 and 3 1024

astral goblet
lilac reef
vestal dew
#

I thought to get good details on a full body image, a bigger initial generate is needed than 512x768

past edge
#

nah even 512x512 gens already look insanely good.

#

the only issue is the lack of skin details

lilac reef
#

not really going to go larger than 1024 square emad said

past edge
#

yeh there is really no need.

#

at 1024 you could get every single blackhead in a person's face

#

every last pore accounted for. lol

lilac reef
#

well, there is but not until we are all on 8k/16k

vestal dew
#

when I do 512 x 768 with full body, the face goes kind of blurry, soft, details lost, and body shape changes - oddly thin

past edge
#

nah G. some 512s already have good detail. 1024 will be all we would ever need

past edge
vestal dew
#

Im talking features like eyes, mouth, etc. A good, clear face

past edge
#

768 does offer more detail but everyone knows that model sucks ass for different reasons

lilac reef
vestal dew
#

so starting with a larger initial generate doesn't fix it?

past edge
#

nope.

vestal dew
#

Well, I can do half and half then

past edge
#

from what I understand, every single image was trained on 512

clever dove
#

HI! What is the best model to train LORA. People models, Realism focus. ?

vestal dew
#

Generate the portrait, then add the legs with outpaint

#

which I am doing now

past edge
#

so it can only have as much detail as u can fit in a 512 image

vestal dew
#

takes a long time on my GTX 970

astral goblet
#

controlnet helps control scene composition so much more. SAG is coming too. and A&E

vestal dew
#

but when finished, the details are good

lilac reef
#

Just remember this 8k is about what all the old films from Hollywood were shot at then 32k and 70k for Imax.

vestal dew
#

I cant run controlnet...cuda errors on my crap card

past edge
#

I do mostly face portraits so I notice the lack of details the most

vestal dew
#

yes, but a face portrait has more detail in the face than a full body image does

clever dove
vestal dew
#

I want to do action

#

walking, running, fighting

past edge
vestal dew
#

but when I try, the character changes, the face gets soft, loses details, body shape changes, everything loses details

astral goblet
past edge
#

controlnet by far is the most powerful img2img tool released

clever dove
#

any idea what to swap f222 for Lora?

vestal dew
#

however, outpaint works well

lilac reef
astral goblet
#

its not just img2img either. txt2img with random inputs is where i get the funnest results

vestal dew
#

just slow for me. And when I get a new system, provided I do before I die, then I will have controlnety

past edge
#

how do u use control net with txt2img

lilac reef
#

Dang, I wish I could tell SD "blue background" and it would obey but nope.

#

I even used weights but no dice

bright comet
#

Anyone was able to run Stable Diffusion on Apple MacBook M1 Pro ? OS Ventura?

lilac reef
clever dove
astral goblet
#

i've seen lots of that stuff. working on a few of my own pipeline ideas

clever dove
#

@astral goblet This is probably the only way it works with 100% accuracy

astral goblet
#

run into a lot of this here. "it can't be done" while whole time i been doing it

lilac reef
vestal dew
#

make a video of doing it, so others who think it can't be done can see how it can be done

lilac reef
#

They would just storm off, or throw more shit

clever dove
vestal dew
#

I don't know about others, but I know I say "it can't be done".

#

Not because I don't want it to, but because I've tried a lot of things but couldn't do it.

lilac reef
#

Well, I gotta conquer this damn SD as this shit with all this orange is pissing me off cause I can't remove it. STOP FIGHTING ME SD!

vestal dew
#

You show me it can, and how, and I won't storm off. I'll be very happy.

#

have you tried prompting for different lighting

lilac reef
#

yes

vestal dew
#

professional photography background?

lilac reef
#

though I am open to any suggestions

vestal dew
#

backdrop, single color background

lilac reef
#

using a ton of embeddings has helped but I just can't remove it entirely

vestal dew
#

always orange?

lilac reef
#

orange/yellow

vestal dew
#

with just a certain model?

lilac reef
#

it's the background

vestal dew
#

yes

#

I am doing right now "school yearbook"

#

because it gives a solid background color

lilac reef
#

whoa

vestal dew
#

it often puts a frame around it too

lilac reef
#

I can remove the frame

vestal dew
#

HS yearbook photo

#

or Jr High yearbook photo

lilac reef
#

well, then it would make them really young, right?

vestal dew
#

Both of those give a solid BG

#

However, I have nothing about lighting

lilac reef
#

it will change, I think, if that background can be shaded

#

solid blue would work

vestal dew
#

I see I get green, blue, brown, what looks like clouds

#

things that look like photographer backgrounds

lilac reef
#

Just throw "HS yearbook photo" at the end?

vestal dew
#

as soon as I mention color, the color of other things might change

#

I have it at the beginning

lilac reef
#

figured so, yeah

vestal dew
#

I see now the ones Im doing say "Junior High yearbook photo"

#

just because I was tutoring a Jr High kid, and it was yearbook day, and that gave me the idea for the photographer background

lilac reef
#

OMG, it is cursed

vestal dew
#

still orange?

lilac reef
#

Show you

vestal dew
#

then it's coming from something else

#

something the character is wearing, or some aspect of the prompt

lilac reef
#

it is due to the model

#

illuminati_diffusion

vestal dew
#

there's a person auraria is the one who helped me a lot with prompts

astral goblet
#

attend & excite might be whats needed. it lets you give attention to different tokens more

#

has a neat hugging space demo

vestal dew
#

thats why I asked about model

lilac reef
#

gives me great people/subjects but the backgrounds are all tinged to outright orange/yellow

astral goblet
lilac reef
#

o.O

astral goblet
#

you have to tell it what tokens you want to focus on and it kind of just works

vestal dew
#

there was a prompt style that did something similar

lilac reef
#

let's step back and do a simple prompt. Bradd Pitt

astral goblet
#

maybe coming to a1111. i think it uses a trained model

vestal dew
#

I tried it - separated the background from the foreground image

#

I dont remember what it was

lilac reef
#

How do I use that thing?

vestal dew
#

have you tried an AI background remover?

astral goblet
# lilac reef How do I use that thing?

type prompt, "blue brad pitt and orange leonardo dicaprio" then you'd have to say "2,3,6,7" as the attention index, i think. it'll tell you what tokens are what

lilac reef
#

I never got them to work all that well.

astral goblet
#

they use depth map estimation usually. just use the generated depth map to pull the bg out

vestal dew
#

I posted what removal.ai did - in general with images

lilac reef
#

I put in my prompt.

#

I need to kill the background during gen or that rim lighting will be there

#

that excite thing only made a picture

vestal dew
#

OK, I have to go.

lilac reef
#

alrighty, take care

vestal dew
#

I pray you get this resolved

lilac reef
#

thanks

#

well, attendandexcite has no bg but what did it do, and to what?

#

Yeah, I put any token indice and it has no bg

astral goblet
#

it helps guide the image towards the attention you want to give. like it is an aim towards solving "red hair" makes the whole image red problem

balmy jungle
#

has anyone made a model capable of true low-res pixel art?

fervent thunder
#

is there any way to convert a dreambooth character ckpt i trained on a google collab to a textual inversion? or a lora?

#

i see references on civitai where people have apparently extracted a style from a checkpoint and made it into a TI or Lora. unsure if something like that can be done with my dreambooth ckpt, but that's what i'm looking for

fervent thunder
# fervent thunder is there any way to convert a dreambooth character ckpt i trained on a google co...

nope. there are ways the other way around, you can merge certain things into a cktp, but not what you are asking.
But at the same time, you then say that it's possible, people have done that, so I'm not sure. a ckpt is a ckpt, made from dreambooth or not, and works the same way internally.
So I'm really not sure, they work by comparing each weight between the base cktp (1.5 or other, depending on what it was trained on) and the dreamboothed one maybe ?

fervent thunder
# fervent thunder nope. there are ways the other way around, you can merge certain things into a c...

it's possible they were just using the wrong terminology, i misunderstood what was meant, or it was a translation error. the exact words used on one example were

"The model was extracted from Elldreth's Dream Mix checkpoint elldreth"

it's probably easier for me to just train a new TI, but i had so much difficulty getting the character to stick that i would've preferred not to. maybe TI will be easier. thanks for the response though

#

part of the work is already done though, you have your dataset, you know it works, you can use it and train on it again on another method usually quite easily, especially since you already know how the concept reacts (not that good from what you were saying).
You can try and change a little the dataset, reduce or augment it, change captions, to see if it helps future trainings too. it's not an easy alchemy

fervent thunder
#

even still the ckpt i have i have to prompt illustration, comic book, etc and fill the negatives with all the other types EDIT: actually..never did think of trying to interrogate it to see if a specific artist style mimics it closely enough...gonna give that a shot

#

Как ботом пользватся бл

fervent thunder
#

it could be lots of things that make it struggle, but finding a good artist to use as reference like you did is a great help too, well done

#

one tip : I ask chatGPT a lot on this. I'm not very literate in terms of art, and don't know artists. chatGPT can understand what kind of style I look for and give suggestions

#

at least it sure did help me make some datasets already

fervent thunder
blissful flame
#

An interesting fact about a person

Your eyes have not grown. They are the same as when you were a baby.

sleek raven
#

How much of visual cocnept can a single vector represent?

#

Say I embeded subwayStyleSandwitch. There are many sandwitchs. I need a vector that can represent sandwitchs well enough

#

How can I estimate the vector size reqyired to represent such a concept? Try and fix?

static blade
#

👀

fervent thunder
fervent thunder
#

8 is a good start usually for me, whatever the subject

compact quiver
#

Can you put a cartoon image into SD and have it recreate it as a realistic person ?

fervent thunder
#

for sure yes

compact quiver
#

How can I do it?

#

and what prompt would I use to make it realistic,.. Do i just say, make photo realistic lol

fervent thunder
#

sorry back

#

well there are lots of avenues for this

#

you can do the simplest way, img2img. you use your cartoon as source image, prompt something like "Photo of a young bear with a red shirt holding a honey pot, by National geographic" and you put denoising around 0.5, you should start getting things

#

if you want if to feel even more photo realistic, you can also use a model specialised on this, like redshift and add the "redshift style" token to your prompt, it would add cool camera effects for ultra realistic look

#

then you want to also add negative prompt like "cartoon, drawing", ... to push those away too

#

lastly, you can call on our new friend ControlNet and go into txt2img, using your drawing as ControlNet source, and try the different control net modes

#

one thing I didn't say, based on your question : no, you don't just say "make photo realistic". you describe it as if it was realistic

fervent thunder
compact quiver
#

Ok cool. so I add my photo. and then describe my photo adding some prompts like photo realistic / raw etc. cool i'll try that - thanks.

fervent thunder
#

So, when do we get a 1024 base model? Really like to see a v1.5 1024x1024 base model.

#

I don't think that'll happen with 1.5, since it's base trained on 512x512...

#

I think overall AI model training also might slow down until world will figure out what's legal and not, what libraries they can use and what not

tired brook
lofty parcel
#

Is LoRA single
She seems kinda cute ngl

fervent thunder
#

she sure does. At least, she got the whole AI sphere hot for her features

lofty parcel
#

Awooga

gentle otter
#

i like lora but she doesnt want to spend time with me

wild steppe
#

gotta take her out to dinner first

warm junco
#

i hope she dont get overtrained, then we need [abs:0.3]

wide juniper
#

When doing LoRA, when do I need to adjust my play Mumbai's? just wondering because the results have been crappy

last merlin
#

what is play Mumbai's

shell quail
#

hi guys

fervent thunder
#

Hey GranBlue !

karmic brook
true dawn
#

good morning yall!

potent fox
#

Just another great hint from me, especially when using multi controlnet: only change one thing a time. Being it a prompt change or some parameter. That finally saved me some trouble 😂

blissful pendant
#

so do i need to run controlnet on my own hard drive

fervent thunder
#

You can install it on a service you're using too

true dawn
#

Can't wait for ControlNet 2.1 768px

narrow herald
#

If I want to use a photo of someone to change them into a new person with controlnet, is it best to use text2img or img2img

true dawn
potent fox
#

My idea is to use controlnet to create poses for any character, and environments. Which works really good so far.

fervent thunder
potent fox
#

Controlnet gives you more control over the shape of what is seen in your image. The content itself relies on your prompt and / or base image.

fervent thunder
gentle walrus
#

so where is the prompts channels?

fervent thunder
#

the prompts channels ? I suspect you mean the bot that makes pictures ?

#

it's dead, sorry :/

surreal perchBOT
#
FAQ: I'm new here, how do I generate images ? Where is the bot ?

Welcome ! There is no bot currently to generate your images on discord. You may want to start by taking a look at the #1072220168534642768 channel. You can access Stable diffusion in different ways : 1️⃣ the official website, https://beta.dreamstudio.ai/. The easiest and fastest way to access Stable diffusion with 200 free credits. For any question on it, you can find help in the #1025467151206854736 channel. 2️⃣ Installing Stable diffusion on your computer. There are numerous projects that let you do that, and you will find help in the #🤝｜tech-support channel. 3️⃣ Running Stable diffusion in the cloud, through rented GPU services, using notebooks. You can find lots of them shared and discussed over in the #1011228442399883294 channel.

gentle walrus
#

what since when

waxen pasture
#

since then

fervent thunder
#

27/01

frozen quail
steep sage
#

Will there ever be 3.0 or 2.2

fervent thunder
#

Quick question, when new SD models get released such as 3.x, 4.x, 5.x etc, will the previous models always be available to use?

waxen pasture
fervent thunder
potent fox
fervent thunder
potent fox
#

You have to think from the view of controlnet, and what it 'sees' using the different models.

lusty mason
#

What the hell is stable digest

potent fox
#

Digest? Never heard of that.

lusty mason
#

I just got pinged about it

jolly delta
#

metaverse-first interviews dear god that is cringe

waxen pasture
#

There is a bug with the digest on the mod art images, if you try to zoom in to one of the lower two images it just opens the one above it

potent fox
#

Stable and Metaverse? Do i get this right?blush

vestal dew
signal dome
#

i am worried to make my clone and talk to me what if it comes to irl 💀

vestal dew
#

He Man and the Masters of the Metaverse

vestal dew
#

I have a story about a demon that possesses a doll. When released, it posseses a little boy...but maybe it should possess an avatar

potent fox
#

If if i can talk to my virtual girl, i would not be satisfied 😉

#

Because there is no physical interaction.

vestal dew
#

If I could talk with an AI girl, just imagine it, talk to some fake gal from south Peru

potent fox
#

So all the Metaverse stuff is just a hype that will never happen.

vestal dew
#

it already happened

#

where do you think you are?

acoustic tangle
#

the internet

vestal dew
#

I mean in what you consider the "real world"?

potent fox
#

Where should the Metaverse happen? Only in VR. And that is to be honest, not the complete thing that live has to offer.

vestal dew
#

In 1997 the Earth was struck by a giant asteroid

#

aliens uploaded the human consciousnesses into a virtual world - that's where we now are

potent fox
#

Okay, if you imagine your view of the world as a computer game...

vestal dew
#

oh that's only half of it

#

some human consciousnesses were placed into robots who terraformed Mars to accommodate human life.

#

those in charge then let some humans return to their bodies - this elite society now lives on Mars, served by robots who are occasionally memory wiped when they start remembering their human lives

potent fox
#

That is far beyond human capabilities. But a good story 🙂

vestal dew
#

it is

#

space aliens did it

#

dont you remember the news in the mid to late 90's about an object detected, heading to Earth from the direction of the sun?

potent fox
#

I recommend playing a game. In my case it is Cyberpunk 2077. There you will get some hints of what is happening and what could.

#

Because it has some parallels to our world: crime, cyberspace and transhumanism.

vestal dew
#

I just told you what's going on

potent fox
#

Then we are basically talking about the same thing.

vestal dew
#

partially

potent fox
#

What part is different?

mossy bridge
#

Guys I am thinking about coding an Prompt Creation Assistant, which helps you build prompts by asking questions to you and adding the according keywords to the prompt. Would you guys need that?

vestal dew
#

we're living in a virtual world while the robots serve humans living in the real world on Mars

vestal dew
potent fox
#

That is a scenario of the future. I call it utopic.

vestal dew
#

it's now

cedar quiver
#

So, uh, realistically speaking
Is SD's main model not just actively worse than some contemporaries and even its older model due to NSFW restrictions

#

because the intensive rebranding doesn't really seem like much but a way to distract from that severe handicap

potent fox
vestal dew
#

that is awful for useful prompts

cedar quiver
#

Doesn't Midjourney have a (fanmade) site where you have a lot of options for subject, framing, material etc. and you just click-to-add? Something similar can't be that obscure for SD

vestal dew
#

I have a tool that takes your prompt suggestion, suggests to you what AI is best to generate an image from that prompt\

potent fox
#

I only thought about using an AI to generate me a prompt, because i don't speak english natively.

vestal dew
#

give me a short prompt. I will type it into my tool for the suggested AI to generate the image

potent fox
#

In other words, ChatGPT helps me to optimize my prompts.

cedar quiver
mossy bridge
#

Typical conversation would maybe:
U: I want to create a person
B: Which gender does the Person have?
...
B: How is the Bodytype of the person?
B: What is the person wearing? (Adds typical Proffesion or Outfit Keyowords
B: What art style? (Adds Photorealism Keywords)
etc.

vestal dew
#

could be useful

waxen pasture
vestal dew
#

B: How is the Bodytype of the person? Answer: he was born without a spine. He's much like a jello.

waxen pasture
#

It's a standalone gradio ui, not an auto1111 extension

cedar quiver
#

I would love to give such an AI the most bizarre and unhelpful responses I could. Gender? None. Body type? Also none.

#

Ethereal sexless being.

waxen pasture
vestal dew
#

B: Which gender does the Person have? Answer: The pronoun is Swe

cedar quiver
vestal dew
#

B: What is the person wearing? Clothing styles are a social construct designed to separate people into conflicted groups, you bigot!

cedar quiver
#

"I am sorry, as an image model I cannot generate anything that would indicate the race, sex, nationality or ideology of an individual, in the case I might represent it slightly negatively."

vestal dew
#

you get a prompt that generates a grey blob

cedar quiver
#

Like. . ChatGPT, my guy, people never needed AI to be racist. But they do need AI to do things people wouldn't bother to do. ClerkLife

vestal dew
cedar quiver
#

When's the Huggingface project for webcam support to feed the AI interpretive dance as a prompt

vestal dew
#

we need an AI to operate Stable Diffusion and other AI's

cedar quiver
#

You think parentheses are hard? I hope you brought your flags, because this fork only takes prompts in SEMAPHORE!! evil

vestal dew
#

this one takes prompts through morse code generated by eye blinks

#

interesting images during dust storms

cedar quiver
#

not developing a third transparent eyelid to protect your eyes from dust

#

blank . o <(pathetic.)
🐫

errant yacht
#

what the current best colab? I just installed jupyter on my pc

#

something with the ability to remove nsfw disabling, because it's a huge pain

#

I was using NOP's a long time ago

wild steppe
potent fox
#

I just asked ChatGPT about gender-aware speaking in germany.

#

It correctly stated that is a source of confusion 😉

#

Or should i say he, or she?

#

How do you call that conversation?

Question: "How could i be aware of everyone's own personal gender definition and is there a chance to fulfill this approach by changing the language rules?"

Answer: "It can be challenging to be aware of everyone's personal gender identity and preferred language. Gender identity is a deeply personal and individual experience, and not everyone feels comfortable or safe sharing their gender identity with others.
However, one way to create a more inclusive and respectful environment is to use gender-neutral language and to ask people for their preferred pronouns or gender identity. This can help to create a culture of respect and openness, and signal to others that you are committed to creating an inclusive space.
Some examples of gender-neutral language include using the singular "they" instead of "he" or "she", using "folks" instead of "guys" or "ladies and gentlemen", and using gender-neutral titles like "Mx." instead of "Mr." or "Ms.".
In terms of changing language rules, there is a growing movement to create more gender-inclusive language in many languages around the world. While language change can take time and effort, it can be an important step in promoting inclusivity and respect for all individuals.
However, it is also important to recognize that language is just one aspect of creating an inclusive environment. In addition to using gender-neutral language, it is important to create policies and practices that support all individuals, regardless of their gender identity, and to challenge harmful stereotypes and norms around gender."

#

"How can humans procreate?"

"Humans can procreate through sexual reproduction, which involves the union of a sperm from a male with an egg from a female. During sexual intercourse, the sperm are ejaculated from the male's penis and travel up the female's reproductive system to fertilize an egg in the woman's ovaries or fallopian tubes. This fertilized egg then implants itself into the lining of the uterus and begins to develop into a fetus.

There are many factors that can affect human reproduction, including fertility, age, and health. In some cases, assisted reproductive technologies such as in vitro fertilization (IVF) may be used to help people who are having difficulty conceiving. However, sexual reproduction remains the most common way for humans to procreate."

"How many genders are needed to human procreation?"

"Human procreation requires two genders: male and female. This is because sexual reproduction involves the fusion of a sperm from a male with an egg from a female to create a zygote that develops into a fetus.

While there are a range of gender identities and expressions that exist beyond the male-female binary, sexual reproduction still requires the presence of both male and female reproductive systems. However, it is important to note that gender identity is not the same as biological sex, and that not all people who have male or female reproductive systems necessarily identify as exclusively male or female."

fervent thunder
#

I still can't believe "gender identity " is an actual thing and it becoming bigger and bigger...

#

wel, I'm not American , so yea...that's just weird for my culture

potent fox
#

It is not a thing. It has only been materialized by certain people. And they are not the affected ones.

fervent thunder
#

yea...

potent fox
#

So the question is, why are they doing this.

#

"Why are people promoting gender aware speach? Is there a basic motivation that has some realistic evidence?"

"People are promoting gender-aware speech to create a more inclusive and respectful society for all individuals, regardless of their gender identity. There is ample evidence that gender discrimination and marginalization continue to exist in many areas of society, including education, employment, healthcare, and housing. This discrimination can take many forms, including unequal pay, limited job opportunities, harassment and violence, and denial of basic rights and services.

Using gender-aware language can be one way to help address this discrimination and promote inclusivity. It signals to others that you are committed to creating a respectful and inclusive environment, and can help to reduce feelings of marginalization and exclusion for people who do not identify as exclusively male or female.

Additionally, research has shown that using gender-neutral language can have positive effects on people's attitudes and behavior towards gender, and can help to challenge harmful gender stereotypes and norms. By expanding our understanding of gender and recognizing its complexity, we can create a more open and accepting society where everyone is free to express their gender identity in the way that feels authentic to them."

fervent thunder
#

well, it answered the question...
But I think you wanted something more specific? gender identity origin?

potent fox
#

"How many gender identities are existing?"

"There is no definitive answer to how many gender identities exist, as gender identity is a deeply personal and individual experience that can vary widely from person to person. However, it is widely recognized that gender is not limited to a simple binary of male and female, and that there are many people who identify as non-binary or genderqueer.

Some people who identify as non-binary or genderqueer may experience their gender identity as a combination of male and female, as neither male nor female, or as a gender outside of the male-female binary entirely. Others may identify with terms like agender, bigender, androgynous, or genderfluid, among many others."

#

And now there comes seriously questionable stuff.

"Where does the gender definition originate at?"

"The origins of gender are complex and multifaceted, and can be traced back to a range of cultural, social, and historical factors.

In many societies, gender has traditionally been tied to biological sex, with people being assigned a gender identity at birth based on their genitalia. However, as our understanding of gender has evolved, we have come to recognize that gender is not solely determined by biological factors, but is also influenced by social and cultural factors, including norms, expectations, and stereotypes around gender roles and behavior.

Gender is also shaped by language, which reflects and reinforces societal norms and expectations around gender. The way we talk about gender can have a powerful impact on how people understand and experience their own gender identity, and can contribute to feelings of inclusion or exclusion.

Today, many people view gender as a complex and multifaceted aspect of identity that is shaped by a range of individual, social, and cultural factors. This understanding has led to a greater recognition and acceptance of the diversity of gender identities that exist beyond the traditional male-female binary, and a growing movement towards greater gender inclusivity and equity."

olive scaffold
#

Hello, could you tell me how I can generate infinitely, that is, not having to press the button every time

vestal dew
#

You can't

potent fox
vestal dew
#

You can do a batch size , and then have a text file with 100 prompts

#

100x100 = 10,000 images

#

Set batch count to 100. Script "prompt from file or text"...have a text file with however many prompts you want

#

each prompt will generate for the number in batch size

potent fox
#

Is there a way to batch txt2img? I would find it useful.

vestal dew
#

yes

olive scaffold
vestal dew
#

in a1111 there is a batch count in txt2img

balmy frost
#

guys , i have a problem , you know how can i use lora in google collab

vestal dew
#

there is also files from prompt or textbox in txt2img

potent fox
#

I don't have a batch tab in txt2img

vestal dew
#

I don't know if there's a limit - i have only tried with a file with 3 prompts with batch size set to 100 - the max - and it did 100 of each

#

are you in a1111?

potent fox
#

How you do that with the file? Please explain.

vestal dew
#

you write various prompts in a text file

#

set it in the script prompt from file or text box

#

You can also just write the files directly in the text box

#

this is in A111. I don't know about other UI's

potent fox
#

Yes, i am using sd-webui

vestal dew
#

then set batch size to whatever...the limit is 100

#

I don't know about that UI

#

in a11111111, it then generates batch size number of images for prompt after prompt in the text box or file

#

you can have the same prompt each one or different prompts

potent fox
#

I will try that in A111111111. Even though i tend to keep the prompt if it gives me good results.

vestal dew
#

The prompt is in the image

#

so later, you can load an image in PNG Info, and then transfer that to img2img or txt2img

#

it transfers the prompt used for the image

#

provided you haven't erased the meta data

potent fox
#

Img2img batch processing has the advantage that it can be used with ControlNet.

vestal dew
#

or if your UI doesn't do that, you can find a meta data viewer, view the prompt, transfer it

#

Im trying to figure out what GPU people have when they do video with SD

#

I copy prompts from the images I like, try them with different models, try with some changes etc

potent fox
#

I am thinking about different poses of a character. Can it be achieved by a prompt, or better using ControlNet.

vestal dew
#

Prompt has trouble with it, especially img2img

#

controlnet does best

#

I cant do controlnet - my card has only 4 gigsd

#

Iget CUDA errors

potent fox
#

I usually don't spend money for things i don't need. But my GFX card gives me so much fun 😍

vestal dew
#

I need to buy one: 3090 or 4070TI

brisk bloom
vestal dew
#

they seem to be rising in price daily now

#

you are kidding me

#

I didn't know that's there

#

there's your answer in A1111 - right click, generate forever

#

actually I, um knew that. I wanted to see if anyone else would mention it. Yeah, that's it.

#

when I did Generate Forever and asked for a time estimate, ChatGPT told me to go outside to fix the antenna

regal snow
#

Does anyone have any tips on integrating specific items into diffusion? I made a LORA self portrait and had a lot of success, but when it comes to trying to make a LORA of any object, it comes out a bit wacky regardless of my prompts and choice of model

vestal dew
#

maybe a lora + controlnet?