💬｜general-chat

[thin vessel]

u dont wana flex on the comfy crew

[crude notch]

just like sdxl, sd3 has different style of prompting

iirc lykon uses sd1.x / sdxl prompting

[buoyant moss]

idk I think SDXL and SD1.x have pretty similar prompt styles

[thin vessel]

he writes 8 different iterations of "ultra-detailed" all with different weights throughout his prompt?

[charred mesa]

its interesting cause he replied to someone who asked for an image with a SD1.X style prompt: this build doesn't enjoy this style of prompting, but here it is

[opal hedge]

Why try something new when you can generate portraits of generic women in generic settings

[crude notch]

most sd1.x style prompts are based on tags, sdxl is based on mostly natural language; sd3 is weird and uses both

[buoyant moss]

hmm

[charred mesa]

sdxl is based on mostly natural language???

i need to read the paper then

[crude notch]

natural language prompting*

[opal hedge]

Does sdxl work with natural language?

[crude notch]

at least in my limited experience

[crude notch]

yea

[buoyant moss]

@crude notch very possible! But I scraped 250 thousand prompts from this discord that people used for SDXL, and eyeballing them, they look like mostly tag based

[opal hedge]

I meant sdxl lol

[lone hawk]

yea, sd3 likes natural language

[buoyant moss]

The ones that made it into showdown and pantheon are mostly tag based anyways

[opal hedge]

I don't think sdxl works well at all with natural language

[charred mesa]

^same

[buoyant moss]

My impression is people tend to use tags for SDXL

[charred mesa]

exactly

[crude notch]

it could also just be my bias to using so few sdxl models

[honest spear]

it has 2 text encoders, one is probably better at natural language, but they work pretty much together

[opal hedge]

I think SDXL is better at language than sd1.5, but it's not a big difference. SD3 with T5, however, now that might be promising

[charred mesa]

interesting

[charred mesa]

honestly we should be excited about how it performs without T5

[crude notch]

t5 is the "oh we need to make text" tenc

[charred mesa]

unless people find good ways of quantizing it at like 4-bit

[crude notch]

good news!

[charred mesa]

other LLMs are great for it but idk about T5

especially at 4.7B

[crude notch]

t5 with 4bit? actually works well!

[charred mesa]

oh thank god

we have hope

[crude notch]

t5xxl? even better

[opal hedge]

Why not go for Bitnet 1.58

[charred mesa]

well doesn't it use T5XXL?

[crude notch]

yea

[charred mesa]

you need to TRAIN at 1.58 bit

its a common misconsception, you can only use that to TRAIN, not Quantize

at least that's what everybody keeps correcting us about

[thin vessel]

i put bigsby back on my s23 so i could have all these voice commands for my smartTVs, ended up training my own voice model for bigsby now I AM my very own personal assistant it is werd AF

[lone hawk]

news: https://twitter.com/EMostaque/status/1766890231687450933

[opal hedge]

Really? When I read the research paper it says it converts fp16 weights into ternaries

[charred mesa]

well idk anymore then 💀

[charred mesa]

thanks

Inference code is going to partners shortly and discord tests next few days I think that will ramp steadily. API access soon as well.

[opal hedge]

So it looks like we still have a week at least or so to go before release

[crude notch]

i wish

sd3 isnt even done yet

we might get a sd2.9

[charred mesa]

yeah a discord test

[opal hedge]

They're still fine-tuning it, but imo is that even necessary when it's custom weights that are what everyone will be using

[charred mesa]

its gonna be like SDXL 0.9

except this time it won't get leaked 💀

[thin vessel]

what is the size of the checkpoints

[charred mesa]

but yeah its like 3/4th the training from my guess

[crude notch]

60gb for fp32

30gb for fp16 with all included

[charred mesa]

okay 30GB for FP16

jesus

[crude notch]

there have been arguments to drop the tencs and make it 15gb fp16

[reef shard]

with or without t5?

[marble lintel]

jesuuuus

[charred mesa]

then again, 8B guys

8B is huge

[crude notch]

the smaller models were a last minute thing

[opal hedge]

8B w/ T5 I though was going to take up around 16GB Vram

[charred mesa]

💀

nahhh this doesn't sound fun

I guess those will have a delayed release

[crude notch]

oh yea 8b will run on 12gb barely

[marble lintel]

thats ok

[charred mesa]

unless they train that much faster

[charred mesa]

that sounds bad if you mean like 2 minutes/image or worse

[crude notch]

dont tell people but the smaller models are just the big models chopped a bit

[charred mesa]

what about xformers + T5 int4

[charred mesa]

[crude notch]

this is wrong, i forgot that it didnt work

lol

all models trained seperatly

same arch though

[charred mesa]

ah okay

[crude notch]

i expect something like 5 img/min on a 3060

[charred mesa]

I wonder if this will take off or people are just gonna use the 2B model lmao

5 images per minute wtf?

that sounds fast

[crude notch]

with max optim

[charred mesa]

OH but I use highresfix

😬

[crude notch]

tome, torch.compile(tensorrt), 4bit t5

[crude notch]

2.5 img/min

[charred mesa]

tensorRT 🤔

[crude notch]

or worse

[charred mesa]

that still sounds good

if I get 1 image per minute then its all fine tbh

I have a 3080ti though

(12GB pleb)

[crude notch]

oh something like 3img/min is a good guess for that

with hires

[charred mesa]

GOD that sounds good

[limpid tapir]

Or the community is going to split 🪓

[charred mesa]

this is what I'm worryin about

[limpid tapir]

Same here

[charred mesa]

unless Loras are compatible

[marble lintel]

switch to LCMs

[charred mesa]

same architecture, same clip, different DiT weights?

[charred mesa]

worried about prompt adherence, which is why I would use it in the first place

[limpid tapir]

Who knows, we never had different size models have we?

[charred mesa]

yeah, only different architectures

[limpid tapir]

Are LLMs interchangeable with their loras?

[crude notch]

this is what i expect

cant wait for nai4 to leak and give different clip n vae

[charred mesa]

512 -> SD1.5, SD2.X | 768 -> SD2.X | 1024 -> SDXL, Stable Cascade

[charred mesa]

lmfaoooo

[crude notch]

no-one uses sd2.x

[opal hedge]

I'd be okay with 2 images/min if they had good adherance to prompt and lacked artifacts and poor anatomy

[charred mesa]

yeahhh

I used to use it to make highresfix photos

and I like the paintings

[crude notch]

the best model for 2.x in my opinion is wd1.5

[charred mesa]

I never tried it

[crude notch]

and that model isnt that good

[charred mesa]

I saw WD1.5 images and thought they were mid

[crude notch]

yeah wd1.5 is mid

[charred mesa]

I used illuminati diffusion or base model for photos and oil paintings

[opal hedge]

Imagine if they censor SD3 the same way they tried to censor SD2

[crude notch]

but the entire point of wd was "hey, we're doing the same as nai but opensource lol"

[buoyant moss]

My 2 cents

[charred mesa]

#1045349359044280360 message

I wanna make funny images like this but with SD3

[thin vessel]

what is the low end as far as GPUs needed to push it?

[honest spear]

It's meant for further finetuning or merging, but now that we have Animagine xl wd is kinda disappearing as a project from what I see

[buoyant moss]

I'd rather they do not use T5, and use something like BGE-M3 instead

[thin vessel]

doesnt sound even possible for me

[crude notch]

the goal is to run it on 6b iirc

[crude notch]

why?

[charred mesa]

also I was one of the few who tried UniPC at 1 step and got somewhat coherent images (before LCM, Lightning and Turbo models)

[crude notch]

t5 has proven to be great for image models

[buoyant moss]

M3 is a dedicated text embedding model

T5 is not

https://huggingface.co/BAAI/bge-m3

[charred mesa]

Can't wait to try SD3 in comfyui in like 1-2 months (if at all)

[crude notch]

the issue is that sd3 training started a LOT before that released

[buoyant moss]

And third party benchmarks show it beating the OpenAI text embedding model

[crude notch]

should act as sd1.x models

[buoyant moss]

Ah good point, M3 came out like 6 weeks ago

[charred mesa]

????

[buoyant moss]

Probably wayyy too late for SD3

[charred mesa]

you mean like no refiner, I know about that

[crude notch]

in terms of nodes used

welll

no more samplers to pick from

[jade mason]

SD3 when?

[charred mesa]

oh yeahhh

[crude notch]

not today, check back tomorrow

[charred mesa]

different sampling techinque

[jade mason]

Ty

[charred mesa]

I forgot to dive into the sampling

[crude notch]

also t5 worked for the closed-source deepfloyd

[crude notch]

tldr, LINE!

[charred mesa]

LINE?

[buoyant moss]

https://huggingface.co/BAAI/bge-m3/resolve/main/imgs/others.webp This was the third party benchmarks comparing OpenAI embeddings vs BAAI M3

[crude notch]

the sampler is a line

[opal hedge]

I wonder what architecture SAI will adopt for their new video model

[charred mesa]

lmao

if the single sampler is fine then sure I guess

[jade mason]

Starting daily “SD3 when?” messages

[charred mesa]

nothing to argue about, no placebo with "this sampler is better than that sampler"

[crude notch]

probably a variation of sd3

[buoyant moss]

I get that T5 may have worked for deep floyd, but T5 is also huge

But yeah, I agree. M3 came out too late for them to experiment with it

[charred mesa]

bro I'm using Mixtral 8x7B at int4

[buoyant moss]

Only came out in 2024 Jan

[crude notch]

there was this sd3 image i wanted to post but it got deleted (e.g. no sharing!),

[charred mesa]

#1045349359044280360 message

[jade mason]

I can’t see, sad 😞

[charred mesa]

it was fun before LCM came out

[charred mesa]

#👥｜roles message react to these

[crude notch]

also get ready for yet another vae

[charred mesa]

yessir

16 channel vae lmao

[jade mason]

Wait, 16 instead of 4?

[crude notch]

also, all i've said is public info (if you know where to look)

[astral goblet]

which one?

[astral goblet]

deep floyd is open source. it was the model that had a restrictive license, but the weights are still released. the training and inference code is permissive. https://github.com/deep-floyd/IF/blob/develop/LICENSE

[crude notch]

sure

[astral goblet]

helps to look for the info right? 😉

[lone hawk]

oh boy here we go, six fingers, bad sword, no character interaction at all: https://twitter.com/Lykon4072/status/1766914556146696427/photo/1

[pine fiber]

that’s what I was concerned about, I wonder why that’s the case

[charred mesa]

four fingers on the left hand

idk what way to look at the glove

[pine fiber]

actually six

look at the knuckle

[trim nymph]

model still not finished

[charred mesa]

idk how much it can improve in the last quarter of training

it has improved a lot since the half point

[lone hawk]

yep, it improved

[charred mesa]

idk if hands specifically will improve

I personally don't care about perfect hands, but it should be expected from an architecture upgrade (or just a next installment in general)

[worthy bone]

Let's hope so https://twitter.com/EMostaque/status/1766890231687450933

[lone hawk]

oh god, the hands... 😄 https://twitter.com/Lykon4072/status/1766924878223921162/photo/1

helicopter is a disaster: https://twitter.com/Lykon4072/status/1766924765925662741/photo/1

this is bad.

[fervent thunder]

i hope sd3 is bad so we never switch and stay on sd1.5

[buoyant moss]

I don't think hands are a deal breaker for me

But I recall before the SDXL release, we saw lots of pictures with text as well

So hard to judge how good SD3 is based on these (understandably) cherry picked pictures

I don't fault them for cherry picking

[charred mesa]

https://twitter.com/Lykon4072/status/1766925737104159187

[charred mesa]

hold onto your papers fellow scholars!

[crude notch]

bad news

[fervent thunder]

i pray for sd3's downfall everyday

[fervent thunder]

wym? all the comments i see are saying it's bad

[lone hawk]

text is good in this one, pixel art cyberpunk style

[crude notch]

there are so many examples i cant post that are actually good

[charred mesa]

3/4 of this community is: no corn, no care

[crude notch]

damn fair enough

[charred mesa]

and also hardware requirements

[crude notch]

sd3 wont be able to do that unless you ask nicely

[charred mesa]

but if T5 at int4 is viable I'm happy

I don't generate corn myself

[fervent thunder]

we should just stay on 1.5 then

[charred mesa]

I want funny images

[crude notch]

lewd....

[charred mesa]

see, my suspicion was correct lmfaoo

[crude notch]

sd3 can do weeb though

[charred mesa]

thats not enough

people want 1girl big booba

[fervent thunder]

why even make sdxl or sd3? 1.5 is already enough for booba

[charred mesa]

honestly people will train nsfw in anyway, this doesn't look like SD2.X

[crude notch]

"hey windows xp still works well enough"

[fervent thunder]

SAI should shift and do crypto/web3

[charred mesa]

if people train SD3 with corn images with detailed examples, you could get impressive images instead of just big boob girl portrait

that's my guess though

I want to know if it's as censored as SDXL or not

if its just SDXL level of censorship we'll be fine

[fervent thunder]

doubt it will be better than sd1.5 finetunes

[crude notch]

probably sdxl censorship

[crude notch]

remember when people said "yeah sdxl wont be that good"?

[charred mesa]

Stability: "Mark.. This is Good news...."
SD community:

[fervent thunder]

still kinda mid ngl

[charred mesa]

shit my bad

[crude notch]

sure whatever you say

[dry trellis]

I read comment from Emad that it would be better at fine tunes than 1.5.

[fervent thunder]

SD Cummunity

[charred mesa]

lmao

[charred mesa]

damn fr?

I thought finetunes will only improve it as much as it improved previous models

[dry trellis]

Yeah...I'll try to find it.

[charred mesa]

I do expect finetunes of SD3 to handle new concepts better though

[fervent thunder]

the day when sd insert version number can make can make perfect hands everytime while running on 12gb vram is they day i switch from sd1.5

[crude notch]

dont take everything emad says as truth

[dry trellis]

I guess it was this that I saw

[crude notch]

emad does emad things

[charred mesa]

yeah...

[dry trellis]

Nevermind it won't let me add a screenshot

[charred mesa]

it don't matter tbh

emad promised us the early phase test since feb 29th

twice

[dry trellis]

https://twitter.com/EMostaque/status/1766575722724749633?t=nfzs_9wtmHsv49DLXUw12w&s=19

[charred mesa]

and still nothing

[dry trellis]

Lol yeah that's true

[charred mesa]

so yeah

[potent spire]

i dunno, SD3 seems to get the overhype as MJ V6 did get

[fervent thunder]

the problem with fintunes is when people learn to train the models well it will be already time for the next sd version

[dry trellis]

Supposedly this is last version....we'll see I guess lol

[fervent thunder]

100% sd3 is gonna be sdxl clone that runs slower

[fervent thunder]

x to doubt on last version

[charred mesa]

yeah emad said that it wont get better, but uhhh 💀

[potent spire]

why do you stick with 1.5? Because of spec requirements and amount of plugins?

[charred mesa]

it will become more efficient, even more intelligent, it will get even better anatomy

there's no final model

[potent spire]

unless the company goes down

[crude notch]

the arch is a lot different

[fervent thunder]

by image quality i mean

[potent spire]

or unless Adobe acquires Stability AI

(wont happen)

[charred mesa]

god 🤢

[fervent thunder]

imagine this is gonna happen before we get sd3

[charred mesa]

mistral got acquired by Microsoft

[potent spire]

hard laugh but impossible

[charred mesa]

only 2 open models released then they made like 4 closed ones that outperform the open ones

[potent spire]

it makes zero sense for Adobe to acquire Stability AI

[crude notch]

eh, someone will leak the model if it does happen

[charred mesa]

SD2.9

SDXL 0.9 leak moment

[fervent thunder]

not likely the testing is happening through discord and only small group of people have sd3 weights

[charred mesa]

yeah the devs

and lykon

[crude notch]

if the company ends, i am sure a dev will leak it though

[charred mesa]

100%

[crude notch]

lykon is not a dev confirmed

[charred mesa]

jkjk

[fervent thunder]

dev will leak a unfinished model that only beats sdxl base

[crude notch]

hahahah

sd2.9 already beats some finetuned sdxl models

some

not all

[charred mesa]

which do you think it doesn't

DreamShaper?

Juggernaut?

[fervent thunder]

in which chat can i ask for model recommendations?

[crude notch]

im a weeb, so most weebtunes are better than sd2.9

exceptions: wdxl0.9 (we undercooked it SO much)

[fervent thunder]

then i am asking here, i need a fantasy model that is not sexy focused X).

[dry trellis]

Maybe Dreamshaper?

[lone hawk]

you can control sexy with "sfw" and "nsfw" in positive or negative prompts, and many other tricks

[fervent thunder]

i tried, but still was often skimpy or something was only on the nipples...

[frank kernel]

are prompts like gatekept?

[sterile raven]

Sure but prompt adhesion is alot better. So thats a huge advancement. Even then i think its also a quality improvement

Thats without fine tuning

Imagine that juggernaut and dreamshaper sd3 would look like

MJ will finally go out of business. Hate them, bunch of liberal mods

[fervent thunder]

pfffft didn't expect that sentence to end with "liberal mods" and MJ is not going out of business anytime soon, they have a large paying user base and have more mind share. They can get investments easily and they can always just train a better model with no regard to if user base will be able to run it

[sterile raven]

Maybe.. it gets hard to make large progress after a certain point

[pale latch]

lol what? censorship and preserving a sense of decency, keeping things unchanging, is conservative.

[sterile raven]

Bro just look at their logo on discord. Nuff said

[pale latch]

personal liberties like not being attacked for being lgbt is also conservative. resisting disent, change, revolution, making sure populations stay cohesive, is what conservative politics is about

hating people because who they are, those are the people trying to change and disrupt

[pale latch]

it has a rainbow flag. case in point

[sterile raven]

Ye, liberal

[pale latch]

why hate people? its so weird

[potent spire]

they wont

its mass appealing

[pale latch]

it'll peter out like leap motion

[sterile raven]

Bro nobody is hating. Its an ai company no need to get political its weird. Either celebrate everyone or no one.

[potent spire]

the one thing i dont like is that MJ allows mockery of Jesus but not mockery of prophet Mohammad

[pale latch]

you're pretty mad about their rainbow flag. thats basically hate.

they're a private company and can do whatever they want. I think that flying the pride colors smokes people like you out and that's a good immune system for a community to have

[sterile raven]

K bud.

[pale latch]

Cause Jesus's skin so thick after those NiN. what did mohammed ever go through? 8 child wives? boohooo

[potent spire]

i dont care about the rainbow flag, i support the same rights for for example homosexuals (but for sure im not going to carry that rainbow flag ofc)

[pale latch]

/offending everyone

[potent spire]

[potent spire]

yeah they are a private company

their rules basically

[pale latch]

whose house? MJ's house

i dont see them crashing and burning though. they'll have a soft landing and just become irrelevant as the tech progresses

[potent spire]

dunno if they will become irrelevant

its hard to move people away from it even if a competitor offers better features for example

[pale latch]

once cellphones can run models at the same quality as mj , why woudln't they be irrelevant? they'll pivot maybe, or just not. the founder will likely move on to a new company. like he did before

[potent spire]

i mean depends on scenario, but realistically right now there is no end in sight

we can ofc mention speculations

i know that MJ was also supposed to be killed off by DALL-E but also SD according to a bunch of people on the internet which ofc didnt happen

[pale latch]

i dont think another saas will make mj irrelevant

ubiquity will

i could be wrong though. i also thought twitter wouldn't last because the sms protocol wouldn't last.

but here we are

[potent spire]

yeah we will see

tbh for a short brief of time for some reason i had the thought of MJ being killed off by D3 as well

[pale latch]

it probably ate a bunch of their subs over night

this just in. teamspeak is the discord killer. https://twitter.com/teamspeak/status/1766547230494748875

[potent spire]

i almost forgot TS still exists XD

[pale latch]

they're like winamp. still functional and pretty good by modern standards, but too old for the new kids

https://forums.winamp.com/forum/winamp/winamp-discussion/316135-winamp-5-9-final-released

oh no they added more since then

[toxic relic]

Hi everyone 🙂
I am interested in training my own Lora, I wanted to know what is the minimum amount of images or information for my lora to be well trained.

[pseudo lark]

does anyone know how I can do Aesthetic score finetuning?

[pearl ocean]

what about Ventrillo?

[hexed chasm]

Where's a cheap place I can run python code online, on a good GPU?

I want to run any new experimental stuff I find on Github

[potent spire]

what the hell is that :kekk

[rough sail]

You know, they should show SD3's ability to do DALL-E 3's claim to fame, I.E. it's ability to do multiple subjects in complicated/complex situations and from multiple angles.

[hexed chasm]

Can anyone help me?

[hexed chasm]

When is SD3 coming out?

[rough sail]

I have no idea.

[hexed chasm]

I'm registered on the waiting list

I saw the nice samples on their page - love the idea of incorporating text into image output

[pearl ocean]

😮

[potent spire]

gotta find out when its out

but i put already a bet that people expect too much for what they want

[pearl ocean]

SD3 is taking ages!

[shell tendon]

during WW3

[fervent thunder]

https://twitter.com/Lykon4072/status/1766934948752113735/photo/1 prompt: "girl holding three pencils in right hand and sunglasses in left hand"

[pearl ocean]

If you guys could create any image you want, what image would you make?!

[shell tendon]

whatever would convince someone to immediately give me all their money

[fervent thunder]

gay alien corn

[pine fiber]

[hexed chasm]

I like comicbook & anime style images -- good quality ones

[pseudo coyote]

hello

how do I uninstall stable diffusion auto11111

[shell tendon]

select the folder and hit the del key

[pseudo coyote]

thats it?

wow thanks

it was a 60GB folder

[pearl ocean]

small

[shell tendon]

mine's only 1.95 TB... just 135,774 folders and a mere 1,020,975 files

[iron depot]

how much ram should I have for SD

[fervent thunder]

Anyone looking for freelance work? Need a few things made, only experienced ppl pls.

[pearl ocean]

😮

[shell tendon]

just checked, only generated 55,790 images since last november

[pseudo coyote]

😮

How many GB is yours ?

[rigid tusk]

Does anyone have access to SD 3

How would I get access?

[pearl ocean]

idk.
but video games I got are way more

[shell tendon]

just give it time

SD bulldozed my gaming hobby

[pearl ocean]

you'll never game again!

[shell tendon]

certainly possible with the current trajectory lol

[pearl ocean]

I play totk every morning before work 🙂

[shell tendon]

i binged the absolute f outta that for a month or two last summer

built every crazy thing imaginable and then that was that

[reef geode]

/create

[pearl ocean]

/delete

[small bronze]

https://thequantumbyte.substack.com
This is a very cool weekly newsletter on Tech, Science, Math, AI and Personal development.
Subscribe it for free.

[static schooner]

Nothing more true

[pearl ocean]

when Sora comes out, it might be more true

[static schooner]

I merged Action Figures with Pictures so i can have unlimited poses and randomly any character i desire

[shell tendon]

only if i can run it locally

legit zero interest in any cloud crap

[pale latch]

age and full time work bulldozed mine

[shell tendon]

got the full time work part down

that killed my interest in fetch quests thats for sure

lol

[pale latch]

remember the game blastcorps? you had to bulldoze cities before an out of control truck carrying an armed nuclear missle ran into any buildings. now that's a game i can get into

[shell tendon]

nope never played that one ha

[pale latch]

rareware

[shell tendon]

i like the concept lol

ah yes

[pearl ocean]

now a Full-time image generator 🙂

[shell tendon]

haha

[static schooner]

50% Image Gen, 50% ParaVM

[jovial wraith]

good morning everyone!

i'm getting pretty darn close to pixel art animations with acceptable consistency, so i'm in a great mood! :D

[placid wasp]

Question. is there a particular model that is better at creating art as opposed to humand. I'm using Juggernaught XL currently and though I am happy with tghe output for the scene i described I just wondered if people have specific models they use for specific situations, You know the old adage "Use the right tool for the job"

[jovial wraith]

i'd recommend looking into art style loras and specific checkpoints. have a look around civitai and i'm sure you'll be able to find something you'll like!

[worthy bone]

Hi Yellow
Since I'm thinking of developing a game with RPG Maker and I need sprites, are you saying that stable diffusion is great for these things?

[jovial wraith]

as of now, not really. however, as my workflow improves, it'll become better and better at it.

RPG maker uses chibi sprites, right? i haven't really looked into those for animation just yet. that may be a fun side project

[worthy bone]

yeah exactly

[jovial wraith]

i'll go look into it. i got a free copy of RPG maker XP, so i may try and use it as well

[worthy bone]

thanks Yellow 🙂

You would be doing me a great favor

[nova zodiac]

Try Dreamshaper XL or Blank Canvas XL

[placid wasp]

Thanks

[placid wasp]

Thanks

and both of them work with SDXL? or are checkpoints imune from versions?

[nova zodiac]

The checkpoint will have a base model - which is effectively the shape of the “brain” (model architecture).

Those two are both SDXL models so will work if your system can run sdxl 1.0

Those two also dont use the refiner model btw

[placid wasp]

Thanks. its still all double dutch to me but I am getting there. I am not using refiners yet so thats not an issue. I have an SDXL install on my local machine and it is running well so hopefully it will go according to plan. Thanks for your help.

Ahjh well just another 13GB to download. I'm glad we don't use 1200 baud modems anymore. we would all be in the dark ages.

[nova zodiac]

Yeah 300mb down 100mb up fibre to the door so damn good for stable diffusion!!!

[still glacier]

baud, I haven't read that in a long time

insert ''Í was there Gandalf" jpeg quote

It still applies to embedded stuff like arduino projects.

[placid wasp]

I need to sort out my conectivity. i have 500Mbps cable, but then pipe it round the mains in the house where I am getting 190Mbs at my sedktop.

Its good to know the old tech as well as the new. gives you perspective. I remembe running a BBS system back in teh early 90's think it cost about £400 to get content onto it with all te telephone bils for connecting around

[still glacier]

Don't know how it works where you live but most ISP advertise xxx mbps/gbps and under deliver. Reading the fine prints it's always "yadi yada best effort networks yadi yada advertised speed may not be reprensentative of real life speed yadi yada give us money you don't have choice anyways"

[placid wasp]

same here but if i connect in the front room i get upwards of 400Mbs. Just did a check on wifi from the back office and im hgetting 94 dowen and 68 up on my phone

rediculous, just did data on phone and peaked at 1.2Gbps, susyained 980Mbpos down with an up of 78Mbps. maybe a tether 🙂

[still glacier]

what a strange idea to live in a house with walls made of lead

:p

good luck figuring out your lan issues

[placid wasp]

bricks are quite good at dulling waves especialy when you are going through 3 or 4 walls

Thanks, but no real issues. just niggles. it works faster than I can

[still glacier]

I lived in a farm house with 1m wide interior walls, I feel your pain.

[placid wasp]

Nice house by the sounds of it. bugger to heat in winter

[still glacier]

bedrooms are next to the fireplace. It's quite easy to heat. Just have to get used to chopping and storing wood.

fiber internet is not a thing however 😄 not even ADSL can reach it.

it s wimax or satellite only

[placid wasp]

Memories of my childhoos in big farm houses in Scotland. one of the chores was splitting and stacking logs. a dirty job when it had rained. Chatracter building stuff though.

There is a village about 5 miles ffrom where i am in south east england and their internet is still delivered via copper cables. population of under 500, so little likelihood of them improving their megre 5Mbps. mates come here to download stuff

[still glacier]

similar vibes http://news.bbc.co.uk/1/hi/world/africa/8248056.stm

Anyways I ll stop rambling and give back the mic to stable diffusion content.

[placid wasp]

love that RFC 1149 IP over Avian Carrier. April fools joke back in the 90's

Thats a great story about winston though

[icy quest]

hey in dreambooth, do u train it, and then generate the checkpoint?

[placid wasp]

I asked perplexity your question as i like to see what comes back on a subject I know nothing about. https://www.perplexity.ai/search/in-stable-diffusion-W1bxi9MiQXCZbv57155TeQ

I cant say for its full completness but there should be enough there for you to answer your own question

[loud quest]

Anyone got any tips for fixing hands on inpainting? It seems like if I put it on Only Masked mode it adds a bunch of random stuff, and if I put it on Whole Picture it works better but I just get different kinds of malformed hands 😅

Faces were easy enough but I'm having a hard time figuring out hands

[icy quest]

are u using negative promps to avoid mangled hands/

and also idk, make a batch of 24 till u get a non mangled one

[foggy halo]

when sd3 release

[timid valve]

i have a question, if i make images with Stable diffution can i sell them ? like comics e.t.c ?

[still glacier]

Depends of your country, your business and the models you re using. Can't get get you proper legal details about that through discord, that's not really the place to have those discussions.

[loud quest]

I'm using "bad hands, deformed hands, extra fingers" as a negative

Anythin else I could add?

[jovial wraith]

there's also a bad hand embedding iirc

and you can use ADetailer

[icy quest]

bad anatomy missing fingers

how do u do something higher weight again

[jovial wraith]

select what you want to give a higher weight, then ctrl-uparrow

[icy quest]

hm doesnt work for me lol

[quaint scarab]

Howdy everyone! I recently found out how to get consistent faces using LoRAs using SD in A1111, but now it's time to use my own face for my project. Does anyone have a link to a good tutorial to making their own LoRA for this purpose? I assume that would probably be the best way since it's working so well, and I can take unlimited pictures of myself.

[icy quest]

uh im pretty sure its exactly the same as the stuff you did before...

just take like 20 slightly different pictures

[odd kiln]

Do anybody know about a model able to generate a Booru type prompt from a human natural language prompt ?

[icy quest]

no clue what that means

[odd kiln]

Like you input "A girl standing" and you get "1girl,brown hair,long hair,standing,..."

[icy quest]

cant u just use the booru interigator on irl images?

[odd kiln]

Yes but it necessitate looking for an image matching my idea which is kinda limiting me

[still glacier]

generate an image with natural language and use that with the interogator then ?

[quaint scarab]

no I used a pre made LoRA lol'

but anyone with the same LoRA can reproduce similar images, so thats no fun

[still glacier]

otherwise no, I don't know about such a booru translator

[icy quest]

bruh

you have to learn to train a lora

[quaint scarab]

I went really deep into deforum and image masking to make crazy half ai/real videos of strippers at the strip club I used to work at

yes, I do. Is there a good resource for that or just watch tons of random youtube videos?

[icy quest]

idk i havent done it yet, i went through hyper networks, to dreambooth, and ill try lora after

but there should be a train menu

[quaint scarab]

cool I have never heard of either

[icy quest]

i hear its even faster than dreambooth

[fervent thunder]

there no sd3 room?

[still glacier]

sd3 isn't released yet. So no.

[charred mesa]

https://twitter.com/EMostaque/status/1766890231687450933

yeah this is the last news I have seen so far

Inference code is going to partners shortly and discord tests next few days I think that will ramp steadily. API access soon as well.

Discord testing will hopefully be this week 🤷‍♂️

(just a wish, not a fact)

[dry trellis]

Yeah we've heard "next few days" for a few weeks now lol

[charred mesa]

^^^this too

I just hope it will be this week

it's been almost 2 weeks since feb 29th, which was when they promised the test apparently

[dry trellis]

Me too...hyped to test it

[trim magnet]

the more u talk about it the sooner it will be released

[trim nymph]

[shell hazel]

SD sora model when?

[trail lion]

Use controlnet open pose, and if possible use the base model, as derived checkpoints tend to introduce more such deformities as a side effect of training/merging

I can't remember the last time I had to fix a hand, people must be using crappy models, or 1.5

[static cape]

I just hope SD3 will properly work with Controlnet again... there were so few well working CN's for XL and they never worked well (at least that's been my experience).

[trail lion]

I imagine if it's an architecture change again, it would be a delay again before we had those models, as with XL ... But hopefully that hurdle will be less since presumably many of those prior efforts can be leveraged

[placid wasp]

noob question. If i install SD3 into a seperate path I take it I can keep runnungf SDXL and SD3 seperately?

[broken cave]

use the openxl hands fix lora, aka the slider lora for fixing hands

[astral goblet]

its now 5 days after emad said "tommorrow or the next day"

[trail lion]

Meh, days, weeks, doesn't matter

[astral goblet]

it matters to me

[astral goblet]

theres tons of controlnets. they work great. even have tile models now. the open pose has been the big one that didn't work for a long while but i've just substituted depth maps instead

[charred mesa]

They will release controlnets alongside SD3

iirc

or at least within probably the 1-2 weeks

[meager sage]

Tomorrow means 5 months

[trail lion]

That would be fantastic, skeptical of it being that fast

[charred mesa]

I am trying to find that tweet but twitter f*cking sucks

I can't search from a user

[meager sage]

Avarage x moment

[charred mesa]

but I remember emad promising that they'll be launching SD3 with controlnets

they are training them

hopefully canny, depth, openpose are guaranteed

and possibly tile

[meager sage]

How much options u think the discord bot will have

[charred mesa]

none for controlnet

[meager sage]

Or just "image"

[charred mesa]

hmm

it will probably have aspect ratio option

don't know about CFG

probably a style option

[meager sage]

Hopefully

[charred mesa]

I want to try how different I'll have to prompt

[meager sage]

Would be good for logos, posters, backgrounds etc

[meager sage]

Does it do that enhanced prompting that I see some things do using an llm

Idk much about sd

[charred mesa]

if I get massively better results with CogVLM style prompts

not -> "cinematic film still, candid cinema, medival knight infront of castle, HD, intricate"
instead -> "A cinematic film still image of a Medival Knight infront of a castle that has ivy growing on it. The medival knight is holding his giant metal sword at his waist height. The sword has a big red ruby gem on it. The sunlight is casting a strong light on the medival knight's head which causes bloom."

[static cape]

Good for you then.

[astral goblet]

not really about me, or you. its about accurate information.

[meager sage]

Interesting

[static cape]

And use cases and resolution / ratio etc.

And Openpose was one of those I tried often that always failed. Depthmaps & other CN's also don't work well if the source image is drawn - those worked better with V1-5.

[charred mesa]

It's gonna get better for people who would prompt it more like explaining an image to a human instead of computer, please make X in Y with Z in the style of A

But it's great that the dataset only consist of Half CogVLM and the rest is regular Laion Clip-like tags, so it's still hard to mess up, but you are rewarded for exploring and thinking creatively

[astral goblet]

canny for sketches is better. depth-anything preprocessor can handle most anime, but it'll break when you get into doodles / kids drawing.

[meager sage]

As soon as sd3 is public ima speedrun make a colab for the memes

[trim nymph]

im curious how well it works regarding styles/textures.

[astral goblet]

focusing on what doesn't work, instead of using whhat does work, kind of limits you

the proverbial you. we're still not making it about you or me

[trim nymph]

ive already blocked u days before

[static cape]

Funny. I just told a friend that SD3 memes will become a thing.

[charred mesa]

I wanna know how much 8B matter. Does it know more characters in the base model for example?

[meager sage]

😈

[charred mesa]

Easier to make Disney & Pixar movie titles

lmao

[astral goblet]

more parameters is more knowledge yeah. potentially has better character knowledge. but they'll be built on the same dataset if i'm not mistaken

[charred mesa]

The dataset always contained the characters I wanted, but they never worked

not even in SDXL, which was the biggest model at 3.5B (Base of course)

[astral goblet]

captioning of the base model is a big deal too. i think 3 is going the ways of other modern base models and captioning so much better

hard to know the difference between anime characters if they're all captioned as anime girl

[charred mesa]

hmm with CogVLM and the premade tags idk how it will end up

[meager sage]

Where did sd get it's dataset from 🤔

[charred mesa]

CogVLM must know basics about popular anime

[static cape]

I appreciate your input about XL CN's, but I don't appreciate the condescending tone.

Take that however you like, without taking it personally of course, because this is not about you or me.

[charred mesa]

Laion

https://laion.ai/blog/laion-5b/

[astral goblet]

you mean this tone?

[charred mesa]

for example this, you can look at the others

[astral goblet]

i tend to reciprocate the energy i get.

[meager sage]

Interesting

[charred mesa]

god at this point #🏞｜general-with-images should be named 1.x channel and this #💬｜general-chat should get image perms, its so annoying not having to post examples and previews

that's what they use it for 9/10

[astral goblet]

i think it's built around it but it's not purely laion's set

[meager sage]

By built around were more images removed or added

[dusk badger]

guys which ones harder to make, lora of a artist' style or lora of a character?

[static cape]

Same. It's like a perpetual back and forth of "nuh-uh, you!"

Makes it hard if both parties can't leave the unwanted convo without having the last word.

[astral goblet]

a room without imageperms is good to have. thats why they're seperate now i ifigured. you give #genchat image perms and it'll look just like gen with images

[charred mesa]

hmm true

[astral goblet]

i'll just not offer advice anymore. somehow it turns into arguments.

[charred mesa]

maybe if embeds could work, like twitter updates then 🤷‍♂️

[static cape]

Ahhhh... That's better. ☺️

[meager sage]

And gifs 🗿

[astral goblet]

just use gen with images if you want that. any chatroom with embed perms is going to become that.

[charred mesa]

yeah that sucks

gifs count as embeds

[trail lion]

I would say they are roughly equal in difficulty. Your expectations will set the difficulty bar once you have a basic mastery of the process for each, subject vs style

[patent scroll]

I solved it 🔼 @pseudo jetty. It was all about the source of the seed RNG generator.
Both DirectML and ZLUDA installations had GPU seed enabled by default.

With DirectML the CPU seed and GPU seed are the same, and NV seed is different.
With ZLUDA the GPU seed and NV seed are the same, and CPU seed is different.

So the GPU seed sides with CPU on DirectML (torch+cpu) and with GPU on ZLUDA (torch+cuda)
So for any images generated on DirectML's CPU/GPU RNG, just need to change the ZLUDA seed RNG to CPU to regenerate the same images.

[maiden sluice]

Just now: "This week, @xAI will open source Grok"
~ Elon Musk

[astral goblet]

lol i love that he's flipping out about open ai this whoel month, but then grok

he won't release weights. calling it now

[dusk badger]

I tried making a style lora and it barely resembles the artist, it resembles more of the base model. Also would you recommend training them on colab environment? im planning to buy a pro version since my pc dont have very advanced gpu

[patent scroll]

That was 8 hours ago though.

[meager sage]

Tf is grok

[astral goblet]

Elon's LLM

[meager sage]

Hmm

[astral goblet]

also a classic science fiction term for empathetic understanding and communication

[charred mesa]

isn't grok like "based" or less aligned?

[astral goblet]

i hate that elon is shitting all over the word

[charred mesa]

I don't remember

[charred mesa]

lmao

[meager sage]

Shitting is based

[charred mesa]

hell yeah

[astral goblet]

based is a shit word. it's been ursurped by race supremacists

[meager sage]

Sigma

[charred mesa]

well that's a based statement

[turbid bay]

Grok being open sourced isn't a big deal, it's a weak model with few redeeming qualities

[astral goblet]

the original meaning yeah. the way it's used now a days is demented from that

[meager sage]

If it doesn't use much computional resources might be useful ig

[charred mesa]

yeah and also how big is grok?

would we be able to run it?

[astral goblet]

i saw a prediction from an ai researcher i follow. end of 2024 we'll have at least 5 gpt 4 level models. gpt5 won't happen. scaling gpt larger won't improve it's capabilities anymore.

[charred mesa]

epic

[astral goblet]

open source and weights are different things. he won't release weights

[charred mesa]

ok so train them on 3000 googol Tokens then

got it

and use 0.0007538 bits when training

[charred mesa]

ahhh

[astral goblet]

now theres a word i haven't seen in some time. the proper googol

[charred mesa]

should have looked deeper in the news

[astral goblet]

almost looks fake

[meager sage]

1 with 100 zeros

[charred mesa]

cant wait for llama3

[astral goblet]

we needed the definition thanks dictionary.com

[charred mesa]

oh wait we're getting off track

[meager sage]

Np

[charred mesa]

lets talk about stability's LLMs lmao

[meager sage]

[astral goblet]

🦗

[pseudo coyote]

Hello

What does instant ID do on SD?

[charred mesa]

if I recall correctly it recreates a face in Stable Diffusion

it tries to resemble an input face

[trail lion]

I use vast, but colab pro probably works fine

[astral goblet]

it's an sdxl thing. i use it with the controlnet extension on a1. catches identities and faces with one or more photos, then reconstructs them into new pics

[pseudo coyote]

Oh how cool. Will try it soon

[astral goblet]

theres ip adapter face swapping too which works good and on sd15 models

[pseudo coyote]

Is there any model which changes the expression of source face. For example: wide open smile to make it closed smile?

[astral goblet]

both instant id and ip-adapters will "attempt" if you prompt that way

[pseudo coyote]

I mean not to generate a new image. Just the same source image but different facial expressions

[trail lion]

Or try just prompt + inpaint, controlnet seems over kill

[pseudo coyote]

I will look into it soon. Is there any tutorial on how to do it?

[astral goblet]

potentially? maybe inpainting with those same tools? there's this thing but , it's on its way not here yet. eventually i believe in expression shifting. https://humanaigc.github.io/emote-portrait-alive/

[pseudo coyote]

Yeah I've heard of it but it's for lip syncing. I am not sure if they will release the code though

[astral goblet]

i tend to hate tutorials and just go for it and figure things out. automatic1111 makes such exploration fine but i'd never attempt that on comfyui. every custom node developer has their own ways of doing things and you've got to follow along with docs to hook things up always.

[pseudo coyote]

What I want is: to change only the expression of the face while retaining the rest of the details of the image

ComfyUI is too way advanced for me

[astral goblet]

says they will. the rumors say they wont' because they're chinese. i hate such xenophobic rumors.

the research is published and other people will implement it eventually if they don't. it's a good release regardless of rumor mills

https://humanaigc.github.io/emote-portrait-alive/content/video/16比9视频结果/song_god.mp4 best example

[pseudo coyote]

That's for videos.

[astral goblet]

its also more than lipsync. theres a few of those. this one emotes the lip syncing

thats why i brought it up. is very expresssive

[trail lion]

Inpaint, select the head, masked only, take out all tokens that refer to other body parts or scenery,change the prompt, use smirking, or play with other descriptive terms, use denoise of maybe .45

[pseudo coyote]

There is an app called FaceApp which also changes facial expressions, I was wondering I could do the same with SD

[astral goblet]

i'd also turn on controlnet to offer face guidance to the inpainting

[astral goblet]

that's a classic face swap. post process. just pastes the face on over top after the fact. works good for many cases. i think an extension called "face chain" might help tehre

[trail lion]

If your expression is very particular, then ip-adapter is how I would do it, find a portrait online with that expression

[pseudo coyote]

Not face swapping. Just to change smiles

[astral goblet]

hard to change the mouth without changing the whole face. if someone smiles without the rest of their face shifting it'd look alien

[trail lion]

I had one with a guy looking up with wide eyes like he saw a. UFO, used ip adapter

[pseudo coyote]

You have a point. But at least retain the same face without altering too much of it

[trail lion]

But it will clobber the look if you use a high cnet value

[pseudo coyote]

So now I use IP Adapter and Instant ID?

[astral goblet]

ideally. i won't say it'll all be automatic. you'll have to play around with it and practice your skill

[pseudo coyote]

I will do my best

Thanks

[astral goblet]

dialing knobbies is one of my favorite learning exercises. good way to get a feel for things. tutorials don't provide that true "groking"

[pseudo coyote]

And is there a model to change clothes without altering rest of the image?

[astral goblet]

not really. ip-adapter and fiddling with knobbies again

[trail lion]

Much harder, that one

[astral goblet]

i think there are models being worked on. salesforce would be interested in such things.

[trail lion]

Sometimes if it's a simple change, you can get away with inpaint sketch

[astral goblet]

segment anything works well for masking only clothes

[trail lion]

Like give the guy a brown jacket by literally coloring the area in brown

[true canopy]

is there something specific u gotta do to run sdxl? ive been using 1.5 for a few days, and i wanted to try sdxl but it cant even generate a 128x128 image, runs out of vram error?

[dry trellis]

Which UI are you using, Auto1111?

[true canopy]

no, comfyui

[dry trellis]

how much Vram do you have?

[true canopy]

8gig

[dry trellis]

it should work fine...maybe try a different model

[astral goblet]

comfyui should be able to fit sdxl into 8gb. make sure your vram isn't being eaten by anything else. check your task man

would hope you're using nvidia too

[pine fiber]

only do 1mp images

like 1024x

512x at the minimum

[true canopy]

why?

[quick girder]

hi guys is there any website like civitai that i can find custom trained stable diffusion models?

[pine fiber]

wasnt trained for anything lower

[true canopy]

i cant even do 128x128, something else is wrong

hell i cant evendo 64

[astral goblet]

nvidia hardware?

[true canopy]

no, amd

[astral goblet]

thats why

[true canopy]

so amd can do 1.5 but not sdxl?

[pine fiber]

not on 8gb

[true canopy]

i had no idea

[warm junco]

What's your GPU?

[true canopy]

its an old 5700 xt

[astral goblet]

amd lacks a lot of optimizations

[pine fiber]

also did u guys see the paper that can make sdxl beat dalle 3 in prompt understanding

[astral goblet]

you're better off loading it up on linux and generating images there

[pine fiber]

weights release in a week, should get it in comfyui

[astral goblet]

the efficient large language models paper fuck yeah

[warm junco]

With that your out of luck with using sdxl :/

[astral goblet]

won't run those on amd

[true canopy]

ill buy a nvidia card eventually, just wanted to atleast try sdxl once, but guess not lol

[warm junco]

You should stick to 1.5

[true canopy]

alright thanks guys

[astral goblet]

go linux. do it. you'll love it. after the pain subsides

[real kernel]

yes it seems amazing altrough I only have skimmed the paper

[warm junco]

But on 1.5 you should be easily able to do 512x512

[true canopy]

yeah i can do 1024x768

but thats it

[warm junco]

Okay

Thats good. For higher res you would need any form of tiled upscale

Like SD ultimate upscaler extension

[true canopy]

yeah ive only been at it a few days, but ive gotten great help here

ive tried tiled upscaling and inpainting and stuff

ive been thinking of getting a 4090, but yeah ....$2500 in my country

[fervent thunder]

Honestly hold off if you can

Rtx 50 series is not too far away

If you can wait for 2025

[true canopy]

yeah, if 4090 is 2500 what is 5090 going to be ....

[fervent thunder]

According to leaks and what nvidia said, it'll be worth buying over 40

maybe not 90

but maybe 80

u will have to see the memory too

but overall there has been a trend

much better performance for slightly higher price

logically that is a better choice

[true canopy]

slightly higher? i doubt that

[fervent thunder]

Compare rtx 30 to 40 prices

And then performance

They say the difference with rtx 50 will be even bigger

If you're ready to pay premium price for top of the line, you might as well wait for the latest one to release

And get more value

[astral goblet]

RTX 50 will launch higher and 4090 will stay relatively same price. Nvidia isn't about to leave money on the table. Nobody is catching up to them anytime soon

[fervent thunder]

I am guessing

if 5090 vs 4090 is going to be a bigger diff than 4090 vs 3090

The 5090 could cost 200-300$ more than 4090

MAX

or ppl wont buy it

and $200-300 for that type of increase in performance is a good deal

[static cape]

I heard a rumor that the 5090 will also merely have 24GB VRAM... which means that for AI purposes you might as well just buy a cheaper 4090 or 3090.

[fervent thunder]

Where?

Also, does bandwidth matter for ai generation?

It will have insane bandwidth according to leaks

[static cape]

Somewhere on Reddit. A so-called "leak", but we'll see if it turns out to be the truth.

[charred mesa]

memory bandwidth?

[astral goblet]

Its not a rumor. its an industry fact. TSMC won't be putting out larger chips for a coupel years. it's the density of each chip which defines the market

[fervent thunder]

vram bandwidth

[charred mesa]

For LLMs it matters iirc

[static cape]

So far the biggest limitation for working with AI Models & training / fine-tuning was VRAM.

[fervent thunder]

It could be 24

[charred mesa]

or the speed of the vram

[potent pecan]

Anybody have any recommendations on how to generate consistent AI influencer (with a consistent face) in different backgrounds/ content situations using stable diffusion, focus, or any other tools?

[fervent thunder]

LLMs?

[true canopy]

theres lots of rumors of a higher vram 5090 as well it seem

[fervent thunder]

I only read 32gb

[charred mesa]

large language models, ones that generate text basically (Llama, Mistral, GPT-3)

or reply to your queries (ChatGPT, Claude, etc)

[fervent thunder]

ah

[astral goblet]

if i understand it right, right now we have 3gb gddr6 chips. 8 of those fit on a board. more if the board is bigger (enterprise). Supply chains matter a lot too since you can only produce 3gb chips so fast. In 2027, they'll start making 4gb chips.

[charred mesa]

I don't know how diffusion works exactly and how it utilizes memory speeds

[astral goblet]

thats when we'll get 32gb cards

cheaper 24gb cards too

[charred mesa]

would cheap 128/192bit 24GB gpus be viable 🤔, would people like them?

[astral goblet]

not really about diffusion specifically, but it sort of visualizes how cpus parallelize computation vs how a cpu does it. I feel like this is applicable knowledge to the subject https://www.youtube.com/watch?v=-P28LKWTzrI

[charred mesa]

like the 4060 ti but with 24GB of vram I guess

weak bandwidth but more vram

[astral goblet]

sounds like a way AMD would cut corners

[trail lion]

Don't know about the influencer thing, is that like a cute tiktok girl? Lol... You probably need to train a lora for consistent look in different backgrounds, but to get the training data you will probably need to face swap

[charred mesa]

https://twitter.com/Lykon4072/status/1767232325513773414/

not a complex prompt unfortunately

but it still looks good for a base model

https://twitter.com/Lykon4072/status/1767229515523084319 this one is more interesting

[astral goblet]

influencers are people on social media that hype brands and opinions. they're walking talking billboards basically

[true canopy]

wont releasing a gpu with high vram, somewhat lower the value of their super expensive cards like a100?

[astral goblet]

they won't compete with themselves yes. but also, supply chain issues

[fervent thunder]

no

[astral goblet]

4090 not having nvlink is a big example of them not wanting to eat their enterprise sales. companies would buy 2 4090s before they buy an a100

[true canopy]

they seem to be 10x the price of 4090 so

[astral goblet]

exactly. 2x 4090 is a lot cheaper for 48gb

they dont leave money on the table

thats how a video game hardware company has become a trillion dollar company

[trail lion]

Cool, so what would be the point of creating a fake one in SD? Rhetorical

[astral goblet]

dont' have to pay a walking billboard

answered anyways

[true canopy]

wait, no gpu in 4000 series have nvlink?

[astral goblet]

correct

[true canopy]

wow, thats so sly

[astral goblet]

they did put the hopper transformer engine in every 40 series though

[true canopy]

i had no idea

[astral goblet]

no sli either (heh)

i did once back when crysis 1 was fresh, and it still didn't run good

riser cable yeah. still not worth it imo. software has to be specific for it and it often causes bullshit like stutters as the cards negotiate

[true canopy]

i wonder if u can sli som old gpu and get decent speeds for generating images

[astral goblet]

sli and nvlink are better suited to research and development

[jade mason]

SD3 when?

[astral goblet]

multigpu rendering has JUST come out, but i don't think any ui has implemented it yet. one image being made across multiple gpus that is.

[shell tendon]

stable swarm can do it

[astral goblet]

swarm just queues images over many gpus i think

like a spool

even if you share gpu memory on one image, i bet it'd still be a single processor doing the generation

[true canopy]

i guess u can have 2 gpu and render 2 images at once, thus making it 2x faster, in a way

[shell tendon]

there's something that can actually generate one image using the vram of two gpus that arent nvlinked without a significant slowdown?

multigpu rendering doesn't really matter tbh, but multigpu training really would

[astral goblet]

i dont think anything like taht is a thing

[shell tendon]

yeah i don't think it can be a thing

[astral goblet]

100% . but 1 image per card at a time. if you've got 100s of cards, thats 100s of batches at a time. could work for video potentially if you can somehow manage consistency across keyframes

[charred mesa]

Which do you think will released first? Inference code for Ella (https://ella-diffusion.github.io/) or SD3

[astral goblet]

ella wont come out next week

in a month people will notice ella isn't out yet

[charred mesa]

lmao

We propose a novel lightweight approach ELLA to equip existing CLIP-based diffusion models with powerful LLM. Without training of U-Net and LLM, ELLA improves prompt-following abilities and enables long dense text comprehension of text-to-image models.

This is impressive though

[astral goblet]

yup

[charred mesa]

Stability training everything so that it could work with T5 perfectly and these guys make an adapter that makes almost any Local LLM work with current Diffusion models

[shell tendon]

hate hearing these announcements with "code to come later" "weights to come later" bet it's not just a week lol

[charred mesa]

same

I remember having "code soon" stuff and it either doesn't come out or it comes out almost a year later when nobody notices

[shell tendon]

IMO what it is are people who are scared they'll get "scooped" by someone publishing something similar first

[charred mesa]

https://jeya-maria-jose.github.io/rebotnet-web/

"Code Soon"

[shell tendon]

so they put out the report with just enough to make it publishable

when they know the code, etc is a mess and needs months of work to be suitable for the public

[charred mesa]

makes sense

[astral goblet]

i cheer them on. these are people trying to get their phd i'm thinking

[shell tendon]

that's exactly what i'm thinking yeah

the race to publish

[astral goblet]

publish first defend later

[shell tendon]

the problem is, the big reward comes from the "communication", the first publication

throwing in the other stuff later doesn't carry the same impact and it's better to chase the next communication

so tons and tons of stuff in scientific journals in general is like this... the big hit, the thrilling new method/discovery/whatever, usually with "studies to ____ are currently in progress" in the conclusion

and rarely do the details get fleshed out

that's generally the end of the road lol

[astral goblet]

its all scientific process. academia is a whole different world

[shell tendon]

Seen that crap thousands of times and I've done it myself lol

Yep

[charred mesa]

*Publishes paper with no code*
*Here's an amazing advancement in AI research*
*Refuses to release code*
*Leaves with PHD*

[astral goblet]

when newton published the principia he didn't even tell people how to understand it! Leibniz had to invent calculus to figure it out

[true canopy]

anyone got a RTX 4060 Ti? how is it compared to 4090?

[astral goblet]

time honored tradition

[shell tendon]

yeah

it does hold us back

[astral goblet]

naw. thats just magical thinking

[naive thorn]

wheenn i get access... does anyone have it yet in here

[shell tendon]

it is what it is but it's unfortunate that getting things working and building a solid framework dosen't reward in the way that the first "OH SHIT LOOK AT THIS REAL QUICK" does

[true canopy]

its like... 1/4 the price of a 4090 lol

is it somewhat fast?

[shell tendon]

it really does because we end up reinventing the wheel all the time in research

someone publishes a communication with the exciting new result, graduates, 90%+ of what they tried in the lab isn't ever published, along with the failures

[naive thorn]

bruh hello?

when they release this

[charred mesa]

nobody has access

[shell tendon]

so then other labs try to follow it up, and end up having to work out all that stuff again for themselves

[astral goblet]

double performance at 40 series level is quite a leap. more than doubling a 3060's performance

[shell tendon]

lots of redundancy

[astral goblet]

that would still happen believe you me

[naive thorn]

i hope. because this so unfair if my country is going to be the reason i will have to wait more

[astral goblet]

if every researcher was forced to release complete inventions, they'd put out a lot of wheels

[charred mesa]

naahh

I don't think it'll be region blocked

[naive thorn]

ok

[charred mesa]

this ain't bard or gemini

[astral goblet]

and alos, reinventing the wheel is a thing. it happens all the time. think we started with goodyear?

[shell tendon]

nah, not saying forced... just saying it sucks that it's disincentivized

same thing with negative results... those aren't generally publishable but are pretty fn important

[astral goblet]

if no one ever reinvented the wheel, we'd be full flintstones stone carved wheels in wagon ruts

[shell tendon]

(coming from the perspective of a chem phd so maybe this dosen't translate as well to other fields)

[astral goblet]

you're a doctor of chem? cool. wouldn't have guessed

[shell tendon]

yeah, just not much value in rediscovering the same mundane shit over and over. it's like learning software without the manual cuz it's locked in a closet, you'll get there, but not very efficiently

[astral goblet]

getting teh entirety of the human collosus onto the same page. nobel vision.

really depends on how much you value a dolalr

[shell tendon]

4090 is 1.6k if you're a hawk about getting the FE

i already had my gigabyte 4090 oc but a few weeks ago i had a FE in my cart on nvidias site and almost bought it to sell at cost to someone but then got lazy and let it go

4090 has been so so worth it

[true canopy]

https://www.tomshardware.com/pc-components/gpus/chinese-workshops-recondition-nvidias-old-flagship-gaming-gpu-for-ai-rtx-2080-ti-upgraded-to-22gb-for-dollar499

lol

[shell tendon]

yeah idk i haven't had any issues

that was apparently ppl not snapping it on properly

[astral goblet]

i've got a 4080 for $1100 canadian. Was the best deal for over 8gb at the time. Canadian retailer prices are very gougey and the 4080 was getting downhyped online so it wasn't boosted in price.

[shell tendon]

and then buying junk cables that were supposedly superior (but inferior) replacements for the originals

[astral goblet]

I would've got a 4090 but i didn't have the cash to spare

less than 1% of buyers had issue but youtubers hyped it up for clicks and views like they were mass producing fire hazards

when investigated too, it was usually people who bent the connector hard or damaged it when yanking it out to retry

1% was a generous over estimate. it was a non issue. there would've been recalls if it had any significance. the reason you think it was so huge is because fo youtuber clickbait.

[shell tendon]

really the biggie with the 4090 is just be sure you use the support brackets etc

but that goes for any of the big 3080 3090s too anyway

[pastel turtle]

Before summer, most definitely.

[reef shard]

checking in again, has anyone gotten sd3 access?

[pastel turtle]

Really?

[reef shard]

and if they have, is it full weights or just some online interface

[pastel turtle]

I didn't.

[lusty beacon]

Any good upscaler for A1111

[pastel turtle]

Try looking at https://openmodeldb.info/

[lusty beacon]

Can you recommend any

[fervent thunder]

just download the largest one

download all of them

try all of them

you think this is science?

it's fucking art

there is no best one

[astral goblet]

all of the gan upscalers suck imo. i wouldn't use any of them. theres supir by google which kicks ass though. heavy to run but awesome results.

wait no, that one's by tencent https://github.com/Fanghua-Yu/SUPIR

[opal hedge]

I can't say for sure but that's what I read on an SD3 thread on reddit, and as we all know, everything on reddit is true

[astral goblet]

as i understand it, the 8b unoptimized will fit into a 24gb card. with fp8 half precision, it should almost certainly load into 16gb

[keen rose]

So there's some clothes I want to turn white when generating some art. But when I generated 10 images specifically with the prompt mentioning white clothes (and more specific words like white corset, white stockings) only one returned with the intended results, how come even when the AI proves it can read my intentions (as with that one example) it generated 9 ignoring the conditions?

[astral goblet]

Yeah prompt bleeding is a thing. Stable diffusion isn't exactly great at prompt understanding. Colors especially will bleed all over.

[opal hedge]

Hopefully those days are over with sd3... for the most part

[shell tendon]

reminds me, the other reason it drives me nuts when people publish and say "weights/code will be dropped in x days/weeks" is then you're stuck wondering if you should bother working on improving a workflow or not

[astral goblet]

There are a number of tricks. Sometimes you can paint the colors where you want them, then controlnet guide the image against an img2img situation, where it has base pixels and guidance

[shell tendon]

i've found the best thing to do is shit like that^

paint the color over it and img2img without specifying the color for anything

let the image guide it

inpaint controlnet can be handy too, or tile

[astral goblet]

generate the image, put it into controlnet for pose or depth guidance. paint over the image in "paint" or whatever other app, then paste that into img2img. give it a denoise. regen. that's often my process

more often i don't have a plan for an image and if it doesn't stick to the prompt exactly, i'll squint an tell myself "eeeehhh good nuff"

[shell tendon]

can also be done in comfyui pretty smoothly with frequency separation or some of the layer blend nodes

generate the image, clipseg and mask blur out the shirt, color or hue blend, or (better) frequency separate and set the color on the low pass

[keen rose]

that's a lot of new concepts I haven't heard about yet but thanks, it'll take me a while to research all the options being mentioned

[shell tendon]

then do iterative adv ksamples with low denoise, finishing each with frequency separation and injecting the low pass from the previous iteration into it

[astral goblet]

inpainting and controlnet. fun stuff to learn. really makes you say "okay, wait a gosh darn minute here"

These intel gaudi chips. Can we buy them to install into our home PCs?

damnit no. tehy're only in the enterprise space

[clever vault]

What is a good recommended model for realistic generations?

For general

[lusty beacon]

Anyone know if the SD3 Preview will be online?

or offline

[lusty beacon]

JuggernautXL

[shell tendon]

also, realismengine, helloworld, realisticstockphoto, realvisXL, albedoXL, many others

thinkdiffusion

[charred mesa]

online

it's gonna be a discord server where you get to test it with bots

we'll get weights when it's 100% done

[lusty beacon]

Mhh okay but there is no ETA?

[stoic nexus]

Can anyone help me? I installed stable-diffusion-webui by AUTOMATIC 1111 and downloaded Stable Diffusion model v2.1. But everything I try to generate turns out to be an absolute mess. I can't even generate just a cat or a person. My parameters: Steps: 20, Sampler: Euler, CFG scale: 7, Seed: 1377736898, Size: 744x1032, Model hash: ad2a33c361, Model: v2-1_768-ema-pruned, Version: v1.8.0

[lusty beacon]

2.1 is bad

use sdxl

or 1.5

[pearl ocean]

I use this one!

[lusty beacon]

🥹

[stoic nexus]

tried it. it's the same for me

I mean I can get absolutely nothing that looks normal

[lusty beacon]

Did u use a fine tuned model

its not the same as normal sdxl

[stoic nexus]

Do you mean it's ok that i can't even get a cat with v1.5 or v2.1?

[astral goblet]

2.1 is really great if you're prompting for anything but porn

stable video is based on 2.1. clearly is a capable model

[lusty beacon]

mhh

[astral goblet]

but if you're obsessed with pornography, well, okay, you've got a point

[stoic nexus]

could it be an installation error?

[pearl ocean]

SDXL can change a life!

[stoic nexus]

i downloaded it right now

i'm gonna check it

[astral goblet]

744x1032 is pretty big for 2.1. it's a 768x768 model. not sure what you mean by "total mess"

[lusty beacon]

but sdxl has other settings than fine tuned right?

[stoic nexus]

a cat with 3 heads without a torso for example

[astral goblet]

i think 2.1's biggest failure is it's really hard to train anything on it. often has wicked prompt comprehension and detail though. The unstable crowd funding campaign really poisoned the well around it and stability just abandoned it. sdxl is basically 2.2

[lusty beacon]

wait for sd3

[astral goblet]

sounds like resolution attention issues

[stoic nexus]

I just wish I knew what does it mean

[lusty beacon]

try 1024x1024

[pearl ocean]

I notice a lot of ppl complain about quality of an image, but they make them at 512 lol, gotta set to 1024, and enable hires fix if required 😄

[stoic nexus]

my pc went crazy with SDXL, i'm gonna reboot

[charred mesa]

only emad's promises

[stoic nexus]

is this ok? https://imgur.com/a/vxaGKSh

[lusty beacon]

No

[stoic nexus]

this is not how it's supposed to work I think https://imgur.com/a/gv9OYJC

[lusty beacon]

Try with Fooocus first it‘s easy and then you can switch to A1111

[stoic nexus]

@lusty beacon what would you change here to get a normal looking picture?

[astral goblet]

2.1 is a 768x768 base resolution. you need to generate at least that

[stoic nexus]

thank you, it's much better now

but I think I still have to reinstall it https://imgur.com/a/4CsCQvd

[astral goblet]

i'd get rid of artstation. openclip is a lot different to prompt against than 1.5. don't use the 1.5 and sdxl prompt cliches, since 2.1 only uses openclip. no dual clip layer

[molten quest]

Does anyone have a comfyui worlflow that would allow me to inpaint a person to change their cloths while using controlnet?