🏞｜general-with-images

[smoky oak]

marry me, SDXL

[cyan snow]

wait, you got access to it?

[smoky oak]

We all have access to it in the bots bro

[oak osprey]

i wouldn't get attached to it, apparently the thing on clipdrop is going to be the release

[smoky oak]

wish I realized that they updated the bots sooner lmao

[oak osprey]

the bot version is a older research model

[smoky oak]

Wait, so we are gonna get the worse version from clipdrop?:/

[oak osprey]

well. for now

[smoky oak]

the version on there sucks ass in comparison

[oak osprey]

they're working on squishing the two models into one

see Emad's twitter

[smoky oak]

Why not just give us the one on the bot and not compromise with the bad quality of the clip drop one

[cyan snow]

the one on clipdrop is nothing like the model they used on the website, it's more similar to the early beta version.

[oak osprey]

i don't know honestly i said a couple weeks ago it was good enough

[smoky oak]

yeah

its nowhere near as good as the best of the 4 on the current SDXL bots here

[oak osprey]

i have no idea what they're testing / waiting for

yeah try to get a woman with freckles on clipdrop when it comes back, lol

[smoky oak]

Oh also, I won my first Pantheon :>

Mom asked for that gen haha

[cyan snow]

i was talking about this: https://stability.ai/blog/sdxl-09-stable-diffusion

[smoky oak]

yeah, that one is blegh

nowhere near as good as the one in this server

genuinely upsetting to see

[cyan snow]

the images they made with SDXL0.9 ARE WAY better than the ones from clipdrop, and the bot is somewhere in between

[smoky oak]

SDXL 0.9 is the one on clipdrop :/

[oak osprey]

and we can't tune the cfg on the bot or the clipdrop, so, for me, i find it very difficult to gradually tweak the prompt without going too far or not changing much at all

[smoky oak]

If we get images as high quality of these with little control, imagine what we can get when we can really rip into it and mold it how we need

[cyan snow]

i refuse to believe it is, the one on clipdrop looks like the early beta, not even remotely close to SDXL0.9

[smoky oak]

Some of these images genuinely blow MJ out of the water, and I am so here for it

[smoky oak]

Clipdrop on their own site say its SDXL 0.9, and they ARE Stability AI

[cyan snow]

they also said there was an issue with their inference, holdup il show you an example

[smoky oak]

If we get the one on clipdrop, then they massively disappointed after showing what's on the bots currently, cause its leagues better

but clipdrops whole site is lies and shitty experiences, so IDK

[cyan snow]

me and @oak osprey did a comparison, whatever they did on their website was messed up

[oak osprey]

the bot made the same broken shit by the way

multiple times

without a style tag i couldn't get a comparable result, and with the style tags i couldn't really change the result much, even with very different aspect

[cyan snow]

yeah, right? there is absolutely NO chance that what clipdrop and the bot used is SDXL0.9(the model they showed on their website).

not even close

[oak osprey]

i added just one term 'nebula' and it totally changed the output and then it was difficult to get the phoenix to show up as anything but an actual nebula, like a telescope would see

[smoky oak]

no seriously, the clipdrop one is terrible compared to the one in this server

[oak osprey]

clipdrop is worse yes

[smoky oak]

Trying to generate tigers was all deformed on the site, and washed out, and blurry and the upscaler is terrible

[oak osprey]

but both made bad stuff for me with fantasy prompts

clipdrop reminds me of Bing lmao

not in a good way

the same eyes tho

[cyan snow]

both are incomparable to SDXL0.9, idk what happened there

[oak osprey]

@smoky oak nothing tops that UncropXL picture from the family photo where it made the BOY pregnant

[smoky oak]

lmfao

[oak osprey]

look at this @cyan snow

also that's stephen merchant

[cyan snow]

did i have a stroke

[oak osprey]

i didn't ask for stephen merchant

[smoky oak]

where are you even getting these "incomparable" claims? imcomparable to what?

[oak osprey]

@smoky oak that's what SAI staff say but they don't really tell us much more so to me it's like blowing💨

[cyan snow]

the model they used on: https://stability.ai/blog/sdxl-09-stable-diffusion

the one on the right

[oak osprey]

all of the announcements behave as if clipdrop is cemented

[smoky oak]

looks exactly like images I have gotten out of the server bot lol

[oak osprey]

oh you can likely get those results from clipdrop if you click Generate 10 times but then all your credits are used

[smoky oak]

infact, i would say my wolf looks even better than theirs

[oak osprey]

your wolf is theirs 😄

it'd be kinda funny if they got the checkpoints all mixed up and cant tell which is which

[smoky oak]

Their example of an SDXL 0.9 wolf

[cyan snow]

idk man, if what they used on clipdrop is 0.9, im disappointed

[smoky oak]

the one I generated in this server

Using the SDXL bot ON this server

[oak osprey]

yours looks like a dog

[cyan snow]

the SDXL bot uses 4 models, one of which being 0.9(atleast what one of the devs said)

[smoky oak]

It was prompted as a wolfdog/direwolf, to be fair

[cyan snow]

so, yeah that makes sense that the good images you got from the bot were 0.9

[proud dagger]

Do bear in mind that the 0.9 model is a research release intended for researchers/devs to start mucking with the architecture, not a final ready-to-go model for normal users

[oak osprey]

the bot seems to use two models now but idk

[smoky oak]

also, their example for this image is pretty bad on the site, especially when I saw people in here generating images like this:

[oak osprey]

normal users pay for api access to it soon

[cyan snow]

i can make WAY better cosmic jars using my 1.5 finetune, holdup let me make some

[smoky oak]

lets see you

[oak osprey]

he can

his model can also do elder magic, and sdxl cannot

sdxl can't do zippers either but i don't think any models outside of controlnet can

they look so weird. it's a very distinct pattern, almost like you can use it to identify AI images

[cyan snow]

bro

[smoky oak]

I mean, its cool, but I would not say better personally

also, jar in jar

did you ask for that? lol

[oak osprey]

i wish that canvas look were not there

[wispy nest]

no love for GTX 10XX cards

[smoky oak]

now generate one ontop of a table in a library

[oak osprey]

sdxl starts to fall apart when you get deep into composition like that too

[smoky oak]

it can't just do one jar, and it can't do a jar in a specific scene, where SDXL can

or at least not reliably

[wispy nest]

is it already 100% confirmed that it wont run on 10XX cards

[smoky oak]

yeah see, it doesn't work

what you are making is dope, but not what is being asked for

[cyan snow]

those are books next to it =/

[smoky oak]

When you ask SDXL for a galaxy inside a jar on a table inside a library, it delivers just that

an actual library, not a couple books lol

[oak osprey]

with blurry as shit background always

it's what MJ does too

[smoky oak]

I asked for bokeh, cause thats how cameras work lol

[oak osprey]

ask for flat zoom, or narrow aperture, or whatever fstop value you want, it don't do it

[cyan snow]

.

[oak osprey]

https://www.wildernessshots.com/wp-content/uploads/2014/07/f-stop-depth-of-field.jpg

[smoky oak]

That one is a little better

[cyan snow]

same prompt as you mentioned

[oak osprey]

right side = narrow aperture = no bokeh, just full scene in focus, and sdxl won't do it

[smoky oak]

Yeah, that one looks a bit better

[oak osprey]

the one cityscape i was able to get in focus took a lot of prompt mangling and it had the same destroyed looking buildings once they were 'sharp'

[smoky oak]

it at least can kinda sorta pass for a library

[oak osprey]

that's kinda overly harsh

[cyan snow]

but yeah, my goal is not to outdo SDXL, i was just saying that the version of SDXL on this server can't reach the same lever of detail as 1.5 finetunes, but im sure the one on their presentation sure can

[smoky oak]

I mean, it beats the shit out of a lot of 1.5 finetunes, but its also a general purpose base model that can do most things better than some of the best single subject focused finetunes

I get what you are saying, but its a little crazy to me

It beats 1.5 handily in realism

thats for sure

@oak ospreyI do get what you mean with it not being able to not have background blur, but I don't mind, as thats my preference for shooting, and is the most accurate for dimly lit scenes

I am sure that is something we will be able to do when we have better control over it

[cyan snow]

but yeah, i can guarantee you, that after i will get my hands on the base SDXL1.0 model, i would make a model that demolishes MJ

[oak osprey]

maybe, i'm just saying it's clever, because it matches human aesthetic preferences and covers up a shitton of flaws

[smoky oak]

As a quick sidenote, to remind of where we came from, this is the difference between 1.5 base and SDXL 0.9 base lmao

Exact same prompt, so yeah

[smoky oak]

hell, it already demolishes MJ in realism (not too hard to do TBH, I can do thate ven with a 1.5 model haha)

[cyan snow]

also, here you go, in a library.

[smoky oak]

I can only imagine how amazing this model will get, cause fuck 1.5 and 2.1 base are terrible without finetuning lol

[oak osprey]

the noise in the wolf's fur pissses me off. i know i can't get rid of it either, but i wish they could

[smoky oak]

Where? thats what they look like when photographed

[smoky oak]

I get what you are doing, but its pretty clear it is struggling hard to put just universe in the jar, letalone just one jar

[oak osprey]

man, this "that's how cameras work" and "that's how they look when photographed" i live in tofino where wolves are everywhere

i don't need a photo of a wolf

[smoky oak]

So you want it to look like real life? Thats kinda hard when all we can feed it are photos... not real life lmao

[smoky oak]

I do like the look of these, I will say

[oak osprey]

here is how a photo looks though. there is none of the uniform pattern in the fur like AI does

[smoky oak]

yeah, when you get close enought o see every hair

[oak osprey]

you don't see it? i see it when i'm not even zoomed in

i've been trying to eliminate it for a while now

[smoky oak]

real photograph of a wolf at a comparable resolution

[oak osprey]

that's a tiny res

[smoky oak]

like yeah, I get what you mean

[oak osprey]

1024x1024 is the SDXL output

[smoky oak]

yeah, but the head is only like 512x512 of the image

[cyan snow]

added 'house' to the prompt, this is a great prompt

[oak osprey]

hm, ok, fair

[smoky oak]

and that image is is 500x600

I get what you are saying tho

but the moire from the training data is likely a factor in that as well

[oak osprey]

894x596 was my other img tho

which has more detail than the 1024x1024 wolf face 😦

it just feels to me like a really good 2.1 finetune

[smoky oak]

can you send the whole image?

[cyan snow]

MJ would beg to be able to make images like that

[oak osprey]

https://m.media-amazon.com/images/I/71PWVGslWdL._AC_UF894,1000_QL80_.jpg

it IS

lol

i could use UncropXL 🤣

oh wait it's down

fuck

i was so ready to have a good laugh

[smoky oak]

yeah, thats all that res in this section of the face lmao

like come on man

[cyan snow]

the FBI showed up at their door because their tool made a little boy pregnant

[smoky oak]

thats like 200x300 on the generated image

actually, wait

[oak osprey]

i'm talking about the pattern tho

i mean, make a wolf even closer up in sdxl

fill dat frame, bb

[smoky oak]

not sure how to haha

[oak osprey]

i was super impressed with the tigers i made but i still saw that criss cross applesauce in the edges of the fur

i wish it could be fixed, because it's distressing and makes it feel like a waste of time to try and fine-tune the details

like, if i knew i could fix them with VAE training, that's what i'll do. but seems to be a question still even for SAI how to do it

[cyan snow]

also, y'all should see that AI_sponge thing, shit's wild

[oak osprey]

you know another thing, the text embeds for 2.1

they mega improve fine details. and HOW?

how do i get that merged into the fucking text encoder itself?

[smoky oak]

Man I hate when your generate prompts get deleted

[cyan snow]

that's with SDXL?

[oak osprey]

that's actually infuriating yeah

[smoky oak]

tried to gen an image, bot ate it

[oak osprey]

2.1

[smoky oak]

no shot lol

tho for 2.1, that does look dope

[oak osprey]

the skin looks a little too fake for SDXL tbh but the prompts people use to get these 2.1 images leave it with little other room than plastic unreal engine crap

[smoky oak]

still has massive artifacts for compression, but thats a 2.1 fault as a whole

[cyan snow]

yeah

[oak osprey]

that's because they're using text embeds on base 2.1-v

they're not using embeds on a fine-tune

[smoky oak]

the whole image looks like somebody ran it through a severe denoiser or upscaler or something

[cyan snow]

2.1 is like the crazy uncle of 1.5, i swear

[smoky oak]

stuff like this is why I don't do 2.1

[oak osprey]

hires fix

[smoky oak]

compression and distortions, and artifacts

[smoky oak]

oh man, is that what it does on 2.1?

[oak osprey]

no idea, i don't use it

[smoky oak]

it looks like they did something else to this image

its all... smeary

maybe that IS the TI

[smoky oak]

like this

[oak osprey]

that's the prompt editing i think tbh

[cyan snow]

it made a little village =]

[smoky oak]

ah maybe

[oak osprey]

it is so odd how prompt editing works

OH YES

you even showed it

the cartoony people it makes

[oak osprey]

i don't have text embeds on my bot, but i used the same prompt to show that image to you earlier

i didn't use prompt editing

[smoky oak]

Also, this section for the same part on the SDXL image is only 340x216 lol

So yeah, 894x596 is a wee bit more res to work with lol

well over 4x the pixels

[oak osprey]

another win for real photos

though you might get eaten/attacked

win for AI images

[smoky oak]

its 7.25x the resoltuion of the SDXL image

[oak osprey]

dont get all nerdy on us now

[smoky oak]

I math when I math! >:C

[oak osprey]

math is scary

[smoky oak]

lol

yeah, so 73k pixels, vs 532k pixels lol

[cyan snow]

bro what are you talking about, we talked about nuclear energy n shit a few days ago

[smoky oak]

oh lawd

[oak osprey]

realism

[smoky oak]

fuck the model on clip drop is so bad

its actually tragic lmao

[oak osprey]

that's your future, dude

:)

the arthritis is kickin innnnn

[cyan snow]

dude this is gotta be like top 5 AI hands

[oak osprey]

how do i make a photo of someone sticking their tongue out

on sdxl i tried and i could not

[smoky oak]

*Cthulhuritis

[oak osprey]

i tried to make a woman stick her tongue out and then i tried to get LSD blotter on her hand and instead, she became LSD blotter art

[strong sonnet]

does anyone have dreambooth discord link?

[cyan snow]

would you drink this

[oak osprey]

you mean EveryDream2?

[strong sonnet]

no, dreambooth(the tool to finetune/train)

[oak osprey]

you're on it

#🔧｜finetune

maybe you mean the huggingface discord. otherwise i have no idea

@smoky oak fucking ranciDXL #1100170365604483202 message

that's actually impressively bad.

[strong sonnet]

i want to take a look at their server

https://youtu.be/usgqmQ0Mq7g

you aren't aware of dreambooth?

[oak osprey]

dreambooth is a research paper

[smoky oak]

WHAT HAPPENED LMAO

[oak osprey]

i don' tknow it deleted my gen request too to cover its tracks

[smoky oak]

can the server please fix this already, FFS

Every time, makes a mod ticket you can't access

[oak osprey]

oh nooo fuck man i don't even want to play this game anymore

[smoky oak]

brought it up to mods several times now

some sick fuck is trying to generate porn in bot 1

[oak osprey]

very realistic, much wow

[smoky oak]

tried to report it, but it made a dead channel

[oak osprey]

looks like regularization data from a good fine-tune

it sucks the image is so weird because the fur looks awesome lmao

100% would pet and die

[smoky oak]

how are we supposed to be able to report peoples gross gens in bot channels?

[oak osprey]

still has an odd distribution but could just be freshly washed

hamburger button -> apps

[smoky oak]

oh, they brought that back?

[oak osprey]

yes

[smoky oak]

looks like

that was gone for a while

nice

[oak osprey]

you had to restart discord lmao

your discord got so old it killed the commands

[smoky oak]

no I asked the mods, and they said that they just dropped it for the ⚠️ react

I asked as soon as I saw it gone

I used it all the time

The SDXL bots feel a little like they are uhhh... goofed today lmao

[oak osprey]

ah

it's probably the 0.9 version now

[smoky oak]

I swear, if they release the shitty clipdrop model instead of the amazing one in the SDXL bot, I am gonna be so mad lmao

ok, this one looks a little better

[oak osprey]

my theory is that the clipdrop one is without the refiner model

they're trying to eliminate the need for the 2nd model

[smoky oak]

Well they REALLY need it lmao

[oak osprey]

yeah i agree, if that's what it is

but that thing's vram use is like 16gb-21gb too

[smoky oak]

if thats the case, then SDXL will be DOA

[strong sonnet]

jaochim said that it has a discord server

[oak osprey]

it'll be great for my discord bot

[smoky oak]

Yeah, but most people will blow it off like 2.1

[oak osprey]

i can see bluewillow charging for it and calling it v5

people will just go to kandinsky

[static tusk]

my attempts at a wolf

[smoky oak]

walph

[oak osprey]

this one's weird

i tried to correct the colour but i'm not good at that lmao

[split rover]

we are testing a lot of things, some are not great but the data is useful

[smoky oak]

ah, so THATS why the images look way worse today, at least thats reassuring haha

still crossing my fingers that we don't end up with that clip drop base model as the actual release

[split rover]

yea haha, we debated on pulling some of the worst things but left them because the comparisons are valuable and then we can be done once and for all

[split rover]

they are running different inference settings to the bot. Dont go off that for what can do solo

[oak osprey]

i'm not going to use it anymore tbh that image of the wolf with its skin missing i received earler was a no-no

[proud dagger]

some of the options are just actually broken and i've been saying there's no use but The Decision Maker ™️ really wants to inhale data

[smoky oak]

Yeah, still gonna be massively disappointing if we get the clipdrop version, and not the excellent SDXL bot one

[smoky oak]

fair enough

[split rover]

They are the same model underneath (most of the time), sampling settings is very important though as the community knows by now haha. Plus we are actively working to improve it further before 1.0 🙂

[proud dagger]

remember that half the point of the bot is just to determine which of our options is best

this is literally... how we decide what to go with here

so, yeah, you'll get the best lol

[smoky oak]

Fair enough, though if we are losing the refiner model, and thats why the one in the bot looked so damn good, then thats gonna be a massive amount of bad press when you dangle something exceptional, and deliver something passable

Just my feedback

if we end up with that exceptional model that was in yesterday along with being able to use it on actually accessible hardware, then SDXL is gonna be fucking incredibly amazing

[strong sonnet]

@proud dagger would we have the option to finetune on sdxl1.0 which is releasing in mid-july?

[smoky oak]

holding out for that, cause I am exceedingly hyped after testing for myself, though that clipdrop model is extremely concerning haha

[proud dagger]

yes once you have it locally you can train all ya want

[smoky oak]

that is amazing

and thats saying that tools will be available for finetuning on day one then?

if so, massive W for SDXL

[proud dagger]

Our internal tools for training are or will be public (not sure if they're in sgm currently or not) and we have contact with devs of some public training tools to make sure those are ready, so yes

[split rover]

Oh we def are aware. The refiner has very nice detailing ability but we would like to achieve near if not the same performance out of the single model. The bot constantly runs with it varied between on and off. Right now though when approaching this research release we wanted to push it as far as we could and those two together pushed it much higher for the time being (just like how people love high-res fix and things of the sort)

[smoky oak]

Interesting... Thats interesting, and also concerning, but I guess only time will tell

[oak osprey]

still waiting for deepfloyd stage 3 and tuning tools for that

[split rover]

Ideally though a single unet can achieve the same if not greater performance, we are already considering improvements from here

[smoky oak]

oh yeah, so what was the hyped release date for today?

[proud dagger]

We have some really good leads for bringing it down to just one model with the full quality. This is still very much an active work-in-progress in that regard rn

[split rover]

thats on DF team 🤷

[cyan snow]

it looks undercooked, like it didn't finish diffusing

[proud dagger]

The number fairy Emad was counting to yesterday's announcement

[smoky oak]

Is that projected for release in Mid July, or is that for a later relase suggestion?

[smoky oak]

What?

[split rover]

Yep was open sign up for research access starting yesterday along with model details

[smoky oak]

Seriously lmao

That seems misleading and distasteful IMO

[oak osprey]

yeah idk why the excessive hype machine stuff

[smoky oak]

95% of SD reddit was swearing SDXL was out today, now they are all gonna be pissed lmao

[oak osprey]

you all do it, it's not just emad

[smoky oak]

||Emad is kinda a goon for hype and frankly stupid claims and stuff||

[split rover]

Its how the other SDs went too lol, research access for a tad, api access monday, then open wide soon after

[smoky oak]

API access Monday you say?

interesting

[oak osprey]

better look at the pricing there 😛

[smoky oak]

Yeah, is it a reddit level cost? lmao

[oak osprey]

$200 for 5000 gens or so? i forget. i thought it was rather steep

[cyan snow]

after the full release next month, will it be 2 .ckpt files, or an entirely new format?

[oak osprey]

it's basically $0.50 per image

[split rover]

safetensors ideally

[smoky oak]

You know, even if we do end up with the massively underwhelming clipdrop model, at least SAI seems to give a shit this time about finetuning, rather than just leaving SDXL to rot like 2.1

[smoky oak]

wow, that is certifiable yikes lmao

Turn my PC into a webserver and let people gen SDXL images for $0.10 an image lol

[cyan snow]

and requirements stay the same? there was a lot of confusion about the requirements

or higher VRAM

[smoky oak]

The requirements are just fully made up at this point, no actual data on them

[oak osprey]

5000 credits for $200, .5 credits per image and .2 credits to upscale it

[proud dagger]

The other thing that happened yesterday is our research team gave I think three separate talks? at CVPR 2023

[split rover]

Currently we got it running on 8gbs locally (20xx cards). I see @proud dagger coming in with details haha

[smoky oak]

from some people saying 2048x2048 just fine on 8GB, to then being corrected saying "it got bigger"

[proud dagger]

(about all the research papers they're publishing rn)

[cyan snow]

https://tenor.com/view/noice-nice-click-gif-8843762

[oak osprey]

no matter the case, it doesn't do 2048x2048 'just fine', deformations and dupes

[cyan snow]

man, i will make such great models with the new model

[smoky oak]

We have yet to see that, though we know the model can gen well under its base res, which is exciting

[oak osprey]

i thought it did 2048x based on the post from emad

naw, they've told us repeatedly in the sdxl channel

[split rover]

we can get some 1920x1080s out of it nicely but yea, its a 1024 model haha, all the same issues going up above

[proud dagger]

VRAM usage is 8 to 12 GiB to run in full or near full at basically any res.
It can go lower with offloading as usual

Basically if you have a modern nvidia card 12GB+, you're 100% golden. 8GiB+ or a modern AMD card, you're probably fine to run it well.

[smoky oak]

interesting, I have a 10GB card, which is reassuring

[proud dagger]

4GiB+, you're dependent on how well offloading implementations work

[smoky oak]

ok, so 4GB is still potential, that is also reassuring

[oak osprey]

2.1 seems to work on 1024x1024 too

[cyan snow]

idk about AMD, they are kinda trailing behind Nvidia at this point

[smoky oak]

hell, I am willing to offload for the level of quality that SDXL bot offers

[smoky oak]

massively so

[proud dagger]

10-series or otherwise very small, uh... in current revision it's not going to go well, but there might be community improvements to get it working if enough people are wanting it

[smoky oak]

AHA

thats why my images looked cooked lmao

[cyan snow]

been there, did that =/

[smoky oak]

was genning at 10 steps lmao

oh man, look so much better at 15 lmao

[proud dagger]

(Remember that when SDv1 dropped it required a 3090... and then it required 12gigs... and now it doesn't even need 4)

[smoky oak]

Very true

[split rover]

I swear running SDv1 feels like can it run doom now haha

[oak osprey]

you been workin on that a while eh

[smoky oak]

also, been curious about the prospect of FP8 support?

LLM models are down to 4bit quantized with very little degredation now, any sights on that sort of thing at the moment?

[oak osprey]

"very little degradation" is lul

get them writing code

[split rover]

Nows your time to shine @proud dagger haha

[proud dagger]

doesn't output a proper image natively but with highres fix, attention scaling, or other tricks, it can

[smoky oak]

They can't even do that at 32bit lmao

[proud dagger]

owo.
It's been very much looked into internally, we've just been busy on other tasks

[oak osprey]

they can but you need to know what you're doing to ask for stuff

[proud dagger]

I've been wanting to hardfocus on it but it was eating too much of my time

[smoky oak]

vicuna 30b at 4bit quantized only sees like 5% coherency drop from 32 bit IIRC

[proud dagger]

I want SDXL-4bit

[oak osprey]

measured how 😛

[cyan snow]

i love this https://www.youtube.com/watch?v=nTfqQv28Yx8

[smoky oak]

Should be possible

[proud dagger]

and I think we can get it. Just gotta put in the work.

Yep.

[smoky oak]

That one score that measures the coherency, I always forget the name of it

[proud dagger]

as soon as i can get away with borrowing some h100s im going to see if i can just shove the model into the new fp8 tensor cores

[smoky oak]

Could be amazing

Also, I heard that H100's are actually fast now for AI, right?

[umbral fern]

ended up making a car thingie

[smoky oak]

Cause man they were really struggling at launch

[proud dagger]

btw re LLM quantizating: we're seeing the first half-decent 2 bit quantizations rn.

I wish SD was as easy to quantize as LLMs are >.>

[split rover]

they got curves ok... all models are beautiful

[smoky oak]

H100 was getting its ass handed to it by A100's for a lot of tasks

My friend is a head Ai dev for unity, and they kept them all on A100's instead of upgrading, cause they cited big performance drops in what they needed

Now what those specific tasks were, IDK lol

but he said the A100 as like 40-80% faster

but I heard they have been fixed with drivers now

[oak osprey]

not all are. well, not for all use cases. look at Bark by suno

they trained a language model directly on audio

[smoky oak]

is BARK considered an LLM?

[oak osprey]

yes

[smoky oak]

hmm, interesting

but its not capable of conversation... I feel like there should be a distinction there

[oak osprey]

they ensure their models are compatible with released embeds

oh you can flip a flag on their inference code and it starts hallucinating like LLaMA

it's an autocomplete style transformer

[smoky oak]

ohhh, neat, didn't know that

[cyan snow]

WHY DID YOU PUT THE BOOKS IN THE JAR

[oak osprey]

basically it gets a history prompt, which is a numpy array of waveform data, WAV audio. and it continues the voice semantics and follows the prompt as guidance for how to form the audio

[smoky oak]

bad TE moment haha

[proud dagger]

I more meant that LLM code is essentially just one matrix multiply operation, repeated with different weights 50 times.
Whereas SD takes in text, spins it, flips it, bops it, puts it in a unet, grows it, shrinks it, flips it norms bops it drops it, takes the result, does a backflip, does the macarena, and then 20 gigs later there's your image

so much more code to write

[umbral fern]

spin it, flip it, bop it, pull it, shake it, shrink it, grow it, kick it, punch it, suplex it, do the nae-nae and boom, It's finished.

[cyan snow]

yeah, SD text reminds me of my handwriting

[oak osprey]

how are textual inversions capable of fixing fine details that this fine-tune hasn't improved yet? i can't seem to hit the quality of a textual inversion even with extensive fine-tuning of the TE and the unet separately/together

[languid pagoda]

oh i like them jars

model progress looking like a blacklight party lol

55%

[oak osprey]

what the hell is happening

[languid pagoda]

you'd think those strong genes would carry over

but the children have no mustaches

tragic

[oak osprey]

i've got true darks and lights again, yay

[supple jolt]

Someone know a website that good with creating bedrooms?

[oak osprey]

@cyan snow early stages of youthelder sybdrome

[wispy nest]

look what i hath done

[iron field]

Just starting to learn SD. Been practicing with Frank Frazetta and John William Waterhouse

[shut sinew]

[sick pagoda]

soon....

[oak osprey]

https://www.youtube.com/watch?v=k5JfV6rqajg

[smoky oak]

owo?

[sterile kiln]

👍 💯

[sterile kiln]

@oak osprey

[oak osprey]

[sick pagoda]

Nice name, u interested in horror?

[oak osprey]

spear wand

thomulas chong

[oak osprey]

sigourney will take that thing down

don't tempt fate, bucko

queen of diamonds is still figuring her damn shit out

the ones with 'photo'

seem to be the last to improve

portrait works great, photo not so much

[wispy nest]

For me, photo is more "scene" prompt while portrait is a "area" prompt 🤔🤣

[smoky oak]

Yeah, its half of my brand haha

[sick pagoda]

Ever tried one of my terror checkpoint or Nightmar shaper?

Because if you love horror I'm already your friend :D

[smoky oak]

I have never tried a horror checkpoint, might need to soon

[sick pagoda]

Here a brief example of terror vs nightmare

[oak osprey]

barking up the wrong tree with those

try males

[sick pagoda]

https://tenor.com/view/discord-gif-26012800

[oak osprey]

closer

[sick pagoda]

Nah

[wispy ether]

u sent that already

[oak osprey]

alllso this discord is pretty tight on the 'nsfw' rules and the amount of skin in that example is

[sick pagoda]

I forgor

[wispy ether]

he forgor

[sick pagoda]

I:m in too many servers go remember

[smoky oak]

These do literally nothing to show off the important parts of the models lmao

[wispy ether]

he sent it in 100 servers

[sick pagoda]

https://cdn.discordapp.com/emojis/979767138350944316.png?v=1&size=48&quality=lossless

[cyan snow]

this how i make the jars in some of my photos

[wispy ether]

1 man 1 jar

[oak osprey]

make him emaciated

[cyan snow]

hell no

[smoky oak]

@sick pagodaGonna download and try your model, see if it's good or not

[cyan snow]

also, can anyone get something similar to this out of SDXL?

we already tried with a different concept

[wispy ether]

what is arnold doing

[sick pagoda]

Also try the neg prompt suggested if you don't have 1 already

[wispy ether]

1 man 1 jar

no matter what

[sick pagoda]

What model are you downloading?

[cyan snow]

bro wtf, there is not even a human in that photo

[cyan snow]

not even 1 jar

[wispy ether]

1 MAN 1 JAR!

great result tho

[cyan snow]

your sense of humor is even more fucked up than me, and i joked about nuclear reactors

[wispy ether]

i remember generating nuclear stations exploding from sd

[wispy ether]

1 ||jar|| 1 jar

[cyan snow]

was that the prompt for that gen? there is not even a jar

[oak osprey]

two jarheads

[wispy ether]

check it

should include it

in png info

[oak osprey]

what in the HP Lovecraft fuck did you train on

[wispy ether]

skulls

💀

[cyan snow]

your model is fucked, it didn't follow your prompt

[wispy ether]

what was my prompt

[wispy ether]

oh bruh

demon in new york destroying everything, realistic

[smoky oak]

@sick pagodaOh wait, just a sec

is your model 1.5 or 2.1?

[oak osprey]

those models can probably be repaired by reconnecting the trained unet to CLIP L

[wispy ether]

1.0

[oak osprey]

and throwing away the broken CLIP

[wispy ether]

what is CLIP

[oak osprey]

https://giphy.com/gifs/thelonelyisland-season-3-i-think-you-should-leave-itysl-gJH0l1ItLtaFdHNt3S

[oak osprey]

@smoky oak ahahahahahahahahaha it gets way better when i put midjourney into the negatives

[smoky oak]

Your model is looking a lot better as of late

[wispy ether]

isnt midjourney a paid website

[smoky oak]

Midjourney is a pretty meh overpriced service

[oak osprey]

thanks i really started spending a lot more money on training it ahahaha

[wispy ether]

150steps

[oak osprey]

the elongated neck lady is the 'photo' prompt starting to become tall aspect friendly

[wispy ether]

neck man

[oak osprey]

@smoky oak i thought you'd appreciate that excluding the midjourney image attributes allows me to clean up the output and make people look more realistic

[smoky oak]

it does haha

[oak osprey]

this is because every incoherent image caption i have from midjourney i just replace with the term "midjourney"

so any random MJ image becomes a viable negative prompt

you can also throw it in as a positive prompt though

that makes the output wild

[cyan snow]

also, can anyone get something like this out of SDXL?

[wispy ether]

150steps looks pretty realistic

[oak osprey]

probably, but it would take a lot of work. similarly, SDXL images, you cannot easily reproduce in 2.1 or 1.5

its just three totally different styles of prompting and they are incompatible. it's not to say you can't do the same stuff, but if you're wanting that, just stick with the model that does it well

[oak osprey]

that Sigourney Weaver image in the grand scheme of things has taken me about 3 weeks of fine-tuning on 2.1 and, i haven't gotten to play with that checkpoint in a very flexible way yet, but it's still not as easy to get those realistic people results as it is in SDXL

[chilly zenith]

So I've been trying to get Stable Diffusion to run locally using the AUTOMATION1111 WebUI and v1-5-pruned-emaonly.safetensors but the results so far have been absolutely terrible. I'm running it on my M1 Macbook Pro with 16GB ram. Is this expected output on a device like mine...? Default settings

[oak osprey]

@chilly zenith try a model like Artius v2.1

[wispy nest]

I've also noticed that it's quite hard to make something "similar" to someone elses art because it's not only about the image itself, but the style the person who made the first one has. Like, similar how? Color? details? realism? or all of it? etc, etc :P

[smoky oak]

That model is trash lol

[chilly zenith]

All right that makes sense lol

[oak osprey]

@smoky oak it's bad but he should be able to reproduce the reference images

[chilly zenith]

Thanks for the suggestion

[smoky oak]

you need to try finetuned models, those will give much better results

[oak osprey]

base 1.5 is mostly useful for making awful looking stuff to laugh at and scare your friends with.

[chilly zenith]

Hahah well I'm very glad to hear this. Thought I was going crazy with those amazing pictures you guys have been sharing here

And then my handicapped tiger drawn by a 4yo

[oak osprey]

i think we've all been there

[wispy ether]

apple sucks

[cyan snow]

yeah i mean, MJ can't compete with this stuff

[smoky oak]

A little reachy, but I get what you mean haha

[wispy ether]

https://tenor.com/view/mj-michael-jackson-king-of-pop-dance-gif-13036441

[smoky oak]

lol

[wispy ether]

"lol"

[oak osprey]

MJ can compete with that. FOR MONEY

[cyan snow]

eh

[oak osprey]

you're going to be pouring credits into MJ's wallet to reproduce any images you want, but if you're open to accepting stuff you didn't expect, well. then it's cheaper.

[oak osprey]

the efficacy of paying for MJ is dependent entirely on expectations of the user

[wispy nest]

oh, no, I'm also starting to notice the ai placing light sources in the image! 😱

[wispy nest]

hello Moes!

[wispy ether]

im going to moe lester you

[wispy nest]

I love it when women walks my direction…less so wielding knives...again :P

[wispy ether]

what the fuck

[oak osprey]

AI WOULD NEVER DO THAT

[wispy ether]

https://tenor.com/view/bro-ayo-wtf-bruh-moment-gif-21632234

[wispy ether]

ru sure

[oak osprey]

he is hunting the elder magic virus

he has a team of elder-youths on his side (not pictured)

[wispy nest]

I saw so many weapons before I finally learned why they always had one fused to their hands :P

[oak osprey]

@stone cipher i miss u

that was fast

if i'd known i would have tried sooner

@smoky oak peter is still alive

[oak osprey]

that's sweet

[wispy nest]

nooo! I was just about to goto work and now I will have to fight of the urge for ice cream as well!

[oak osprey]

the puerto rican day parade

[wispy nest]

I thought of you and the faces at a distance for some unrelated reason ;P

[oak osprey]

@smoky oak you know classifier free guidance?

and how to help it along while training

[oak osprey]

sir why do you need to see under my robes before i get into the theatre. why do you assume i am hiding other people under there, and snacks, and drinks? it is impossible

[weak sage]

Unexpected error: Processing could not begin, you may need to refresh the tab or restart the service. 0it [00:00, ?it/s]
Total progress: 0it [00:00, ?it/s]
Total progress: 0it [00:00, ?it/s]

I just restarted stablediffusion in both browser and console

what's the point of XY script existing if it just gives us more things to debug while we are debugging

no time fo that shiz

oh wow though, night and day, this new 2m sde thing is better than sliced bread

its as fast as 2m but as effective as sde, idk how thats possible

[oak osprey]

anyone ever seen a sweater sold like that?

i guess it just assumed to make a side by side profile lmfao

[solid gate]

Anyone knows how to solve this? It happens to me when I want to use the GPU for Roop.
This problem does not happen when I use the CPU.
I installed the toolkit and steps from github to use the GPU, but still the problem

[smoky oak]

@weak sageHow nice to see thou, how are thou doing?

[oak osprey]

man that's so cool

[weak sage]

i am well, thank thee

Hmmm, I need a GPU that supports a PNG with like, 50 times as many pixels

[smoky oak]

whatever you are running right now can do that

[weak sage]

are you talking about img2img tiles?

but like if i could gen at 10x resolution it could fix all these annoying face and hand issues

[smoky oak]

That would make them worse actually

[smoky oak]

just have to get better at prompting, and use models better suited models for stuff like that

and also inpainting

[weak sage]

you what, type "good hands" into the text box and magic happens? lol, best i can do is controlnet inpaint global harmonious and reroll 100 times, get one passable gen

[smoky oak]

I use models that do hands just fine, personally

[wispy ether]

Roop???

[oak osprey]

"mJ cAn Do hAnDs"

[smoky oak]

lmao

[oak osprey]

man it's like they don't fine-tune their text embeds at all

that halo effect

fuuuck that

i would rather use filtered LAION data for training

[wraith karma]

Generated it for my dnd character

[oak osprey]

slick

@smoky oak downloading only high-res laion images with camera models in their exif data

it's not bad

no wonder we can't do buildings

[fleet goblet]

Peekaboo.

v

Yeah, mangledmerge is ridiculously good.

Really REALLY good.

[fleet goblet]

I just tried out Artiusv21, and holy crap I can actually generate 1280x1280 images.

Watermelon.

Power.

\

It'd be hard to tell it's fake.

[fleet goblet]

Papers, papers...

[fleet goblet]

Hands.

[fleet goblet]

\

[wispy ether]

wowie

[wispy ether]

ok wtf

another vae

[wispy ether]

edited

[wispy ether]

try negative prompt bad geometry

[cyan snow]

why, what's wrong

[wispy ether]

maybe it makes that look better

[wispy ether]

did you use it there

[cyan snow]

idk man, looks fine to me =/

[wispy ether]

or try deformed geometry

[cyan snow]

bro, i don't need to, there is no need to get rid of something that's not there. thats not how negative prompts work.

[wispy ether]

wat

its meant to make it better formed

[wispy ether]

bruh

[wispy ether]

can you add "raytraced reflections"

[cyan snow]

why would i add a pretty good detail to the negative prompt?

[wispy ether]

not negative

why only 27 hires steps tho

id always use 0

[cyan snow]

why would you try copying my settings without my permission?

[wispy ether]

do i need your permissions

🗿

its open source

[cyan snow]

i don't look in your shit

[wispy ether]

but you can

[wispy ether]

i keep thinking masterchief is the doom slayer

[wispy ether]

samus?!

[bold mist]

Wasn't expecting it to turn this prompt into a human, but turned out neat

[wispy ether]

ru using a vae

[bold mist]

I don't think so, unless it's prebaked

[wispy ether]

can u make doom slayer

[cyan snow]

no =]

[wispy ether]

🗿

what model is that

the 2.1 one?

from civitai

[cyan snow]

fuck no, i made my own model

[wispy ether]

bruh

LOL

[cyan snow]

ever heard of finetuning, well i suppose not, you didn't even know what is CLiP a day ago -_-

[wispy ether]

i still dont

cuz nobody explained it

and i never intepretated it as important

more like third party tool

[cyan snow]

BRO, with out CLiP diffusion isn't possible, especially SDXL

that's a key component

[wispy ether]

give me 1 reason to be able to know about it for no reason

[spring sail]

I miss vqgan+clip

but yeah clip is real important

[wispy ether]

then why did i never need it before

[cyan snow]

it's encoded into the model checkpoint

[wispy ether]

so why do i need to know about it then

[cyan snow]

it's the component of the model that makes connections between text and images

[wispy ether]

what does this do

[cyan snow]

uh, maybe because you installed the AI locally

[wispy ether]

and

all i do is install the folder tho
after i click webui-user.bat it installs everything automatically

[cyan snow]

and? i'm losing my brain cells by each passing moment

[cyan snow]

do you really not care about the coding and the technology behind the AI? stable diffusion is ment for people that know what they're doing

[smoky oak]

I just love when SD decides to nuke your whole PC and you have to deal with 5fps until you save everything on your PC to restart it

[spring sail]

no reason to gatekeep, even if someone doesn't know all the specifics or exactly what everything is doesn't mean they can't use the program

[smoky oak]

1 little missclick with a VRAM OOM, and your whole PC is toast until you restart

[wispy ether]

i just like making images that are unrealistic but looking realistic

[spring sail]

sometimes shit takes time to learn and people'll take a while to learn it, or they won't, but they'll still have fun with what they have. and that's fine

[smoky oak]

Nep is kinda a community troll in a lot of ways, to be fair. They ask for help, then disregard that help, and then blame you when things go wrong. It happens nearly daily, and it stems from them having a lack of interest/understanding of the tech being used

[wispy ether]

no

[spring sail]

ohh

[wispy ether]

i keep learning

its just like

[spring sail]

that makes a lot more sense

[smoky oak]

Most people who have been here for a while can attest

[wispy ether]

bro im not a 20 years experienced dude

[spring sail]

how old are you?

[wispy ether]

16

[spring sail]

oh, same age as me

[smoky oak]

It took over 5 hours and 9 people to help them install Kohya SS, and they still never listened and never got it working right

(I was one of the people)

[spring sail]

I was expecting you to be like 13

[wispy ether]

i got it working after time

i think fp16 was the problem

had to use fb16

[spring sail]

nod

[smoky oak]

They give off that vibe.

Really though, I do think that sort of thing is warranted with Nep, when it comes to not caring about how things work, and just "gimmie thiss, gimmie that" and "tell me how to do it, I don't care about why" and stuff

Can't even google anything for themselves :/

[spring sail]

out of curiosity, what are your like? computer specs. sometimes really old gpus cause some weird ass problems if you've been dealing w/ that a lot

[smoky oak]

They have a 3060

[spring sail]

huh

[wispy ether]

i used to use ai with a igpu on my old laptop

[cyan snow]

you mean bf16

[wispy ether]

yes

[spring sail]

on laptop or desktop? big difference

[wispy ether]

desktop

[smoky oak]

Desktop IIRC

yeah

[spring sail]

CHRIST

[wispy ether]

custom pc

selfbuilt

self put together

[spring sail]

that's a good amount of vram for sd

[wispy ether]

self selected parts

i learned over the years

[spring sail]

how did you build a pc yourself and you're. struggling with sd

[smoky oak]

@spring sailSeriously, they always have problems with everything and never learn, it really is just a losy cause at this point

I help people all the time when I can, but I have never once ever been able to successfully help them, cause they never listen

[wispy ether]

cuz sd isnt physical bruh

[smoky oak]

Nep is an enigma lol

[wispy ether]

an what

[cyan snow]

before i learnt python i was making custom watercooling loops, that's not surprising

[wispy ether]

an sigma

[spring sail]

maybe I'm just a little over experienced for a sixteen year old in comparison to you or smth but I'm baffled (I started messing w/ like, vqgan+clip type stuff when I was 12 on really slow google colabs)

[wispy ether]

https://tenor.com/view/daeth-funi-periflight-dies-from-gif-22212023

[spring sail]

idk I feel sd is a bit simpler than python

[wispy ether]

bruh

[spring sail]

not harrassing you though nep

[wispy ether]

sd is python

[spring sail]

💀

[wispy ether]

made with*

[spring sail]

yes, that's more correct

[smoky oak]

Same here. Nep is just a very immature person, and I don't mean that in an insultiling way, just a direct observation

[spring sail]

you don't need to know python to use it though

n yeah, very fair

[wispy ether]

mostly the direct observation hits the most

[spring sail]

I do wish nep luck in learning SD better

[smoky oak]

Same here

[cyan snow]

eh. not necessarily, NONE of SD is possible without the diffusers python module.

[spring sail]

if they aren't a troll I'll try n offer advice when I can

[wispy ether]

i remember trying to use something without a gui

almost got braindead

[smoky oak]

But yeah, anyways, hope that little bit of information helps explain why Tdg may have responded that way. Generally speaking, they are a really nice person (Tgd)

[spring sail]

working w/o guis is always funny, I fuck up on that a lot myself

n yeah! that makes sense

[smoky oak]

@spring sailDo you mind if I DM you?

[spring sail]

I thought nep was getting ruthlessly bullied for struggling a little LOL

uhh, sure, what about?

[smoky oak]

A small continuation of this convo, little more details in private

[spring sail]

sure sure, I typically keep my dms closed but luckily for you I actually have them open on this server so I can access the bot dming me

[wispy ether]

sd isnt my only interest tho

[wispy ether]

my perspective

[oak osprey]

@smoky oak did you see my comment about my inference code bug

my discord bot was overriding my cfg to 10 all the time

sooooo

thats why i had incredibly crispy crap for so long

[oak osprey]

you can uae tensorflow instead but it sucks

use

[cyan snow]

that wasn't my point, i meant that SD is impossible without python modules

[oak osprey]

Oh, you can do it in c or c++ but itd also suck

[cyan snow]

huh, but as far as i know, stable diffusion is meant to be used with python, even more so, pytorch

[oak osprey]

theyre just weights, a vae, and text embedder

[cyan snow]

yes, but none of them can function to create images without the modules. at least as far as i know.

[oak osprey]

you can use it with jax or onnx or tensorflow etc

[cyan snow]

those are all libraries, stable diffusion isn't able to run on it's own

[oak osprey]

i certainly didnt get far with them when i was just starting out but thats because you have to convert the ckpt first too

it works fine lol

pytorch has more developers using it

its easier by far

[cyan snow]

you can't just put a diffusion model on a PC out of the box and expect it to create images

[oak osprey]

microsoft and google keep trying to compete and fall flat

i dont get why that is even a point to make

no one is dumb enough to think that

are they?

[cyan snow]

you'd be surprised, but yeah, i understand what you're saying

[oak osprey]

first tried to do inference by asking gpt4 how to load a pt file

what a disaster

[cyan snow]

that won't end well

[oak osprey]

it had me write a bunch of code

none of it worked

[cyan snow]

makes sense

[oak osprey]

it assumed a ckpt is a tensorflow file

[cyan snow]

GPT-4 assumed that? or GPT-4 LLaMa?

[oak osprey]

when it wouldnt load, i asked and it was like try pytorch. but i didnt know how good that was, i only knew of tf

gpt4

[cyan snow]

weird, i thought it already had way more than enough parameters to be as capable as us at coding.

[oak osprey]

stable diffusion is newer

pytorch wasnt as big

a lot has changed

the huggingface modules used to have totally different parameter names and loading models through transformers was hard to figure out too

[cyan snow]

still, isn't GPT-4 trained on more recent data than chatGPT?

[oak osprey]

nay

[cyan snow]

by chatGPT i mean GPT-3.5

[oak osprey]

2021

[cyan snow]

damn

i ran once LLaMa locally, and it knew about stable diffusion =/

[oak osprey]

that thing is garbage tho

gpt4 knows about stable diffusion

[cyan snow]

it's good for generating prompts, and it's local

probably the best local LLM model

[oak osprey]

it doesnt know how to do anything correctly with it in code

[cyan snow]

i use it for prompting, so that doesn't really matter for me. but i bet if you finetune it specifically for coding it might be able to code atleast decently.

[oak osprey]

chances are slim of that

fine tuning a llm needs very high quality data it hasnt seen before or you risk mode collapse

this is in general a problem for humanity now because llm are flooding the internet with text

hoovering up data for training like that is over

[sterile temple]

amazing how text llms can require so much ram, when people use to say stuff like you can fit all of a massive library of books onto a dvd

and yet images take up a lot more disk space in comparison

[proud dagger]

Weirdly enough, while you can fit all of wikipedia's text in so many megabytes, the text used to train LLMs is huge - terabytes of plaintext

They're not trained on a library of books, they're trained on all text ever basically, cause they need to know about everything ever

a chatbot needs to be able to answer list every key event in japan in the year 1602 perfectly (or any other obscure overly-specific question like that)

whereas with SD, it's not really a problem rn that you can't say give me a dog that's not too fluffy because the bot's too dumb to know what not means and just gives an extra fluffy dog instead

[oak osprey]

SD just decided that it wasn't TOO fluffy. it's subjective, see

[oak osprey]

FANASY WAAAAARLD

[wispy nest]

I see the prompting as a sort of, "here is a card with the word dog on it, now, show me a picture of a dog." That's why I it feels like writing full sentences to the ai doesn't do anything because what would a "card with the word has" have a corresponding image to the ai's "library of image references?" There's like billions of different possibilities.

A dog can look very different from image to image, but if someone "has something" then it can either be worn, held, dropped, at home, etc and that's what I think is why the ai draws weird stuff hovering around, having it fused to other prompt words, and more. There's just not enough fine-tuning for normal sentences yet. :P

[feral pebble]

In a1111, When making non square images, I often I get the repeated patterns. I thought there used to be a way of generating at 512x512 for some cycles, then scaling it and continuing... ?

(actually it looks like the thing crashed.. but n/m that)

[wispy nest]

Highres. fix might be the thing you're thinking of?

[feral pebble]

@wispy nest that's it. thanks. (I think with my window size it didn't show the width/height scaling sliders)

danke!

:))

[wild sorrel]

You can start with square then outpaint to the sides to