🆕｜sd3

[viral plaza]

0.o?

[dusky thistle]

the effect is greatest with the magic RES sampler i have, but yeah

we've been poking around over on the L2 server trying to pinpoint what's going on halfassedly

[viral plaza]

Wait can you confirm that fp32/fp16 are near identical with the same setup? ie it's specifically fp64 where something weird happens?

[dusky thistle]

fp32 and fp16 are also different but it's not as dramatic

[jolly swan]

Is there any chance you can get someone with the right paycheck to confirm this?

[dusky thistle]

i believe deliberately setting fp32 made no change

so it's not something that's accidentally fp8 somewhere

and a lot of us have replicated it

so it's not just my crazy setup

at fp32 or fp16 the images tend to be more creative - more combined objects, more transforms, and also more artifacts, broken fences, that kinda stuff

[viral plaza]

not at 10pm sunday night where the person in charge of that is chilling in taipei with lisa su, but i'll poke to try to make sure it gets answered clearly before launch

[dusky thistle]

fp64 things really come together

[noble coyote]

As somebody is talking about sigmas (noise scehdule?) - what exactly do they do?

[dusky thistle]

more detail, more cohesive images, and the it/s is unaffected

[candid holly]

One of these days after I build my new PC, I'm gonna have to sit down over a weekend and learn comfy...

[dusky thistle]

so it's something small somewhere

it's def not a bottleneck step

[jolly swan]

Thank you!

[dusky thistle]

comfy is a lifestyle, not a weekend 🤣

[cobalt moon]

this right?

[noble coyote]

ComfyUI is a doddle to learn! 😄

[viral plaza]

sigmas are like weirder magicer timesteps - the timesteps you see as the user are 0,1,2,3,...,20 but the model sees weirder numbers that go on along a curve

[dusky thistle]

sigmas are so much easier to understand imo

[viral plaza]

if you scroll up to clownshark's comfy screenshot you can see some graphs of his chaotic scheduling

[patent acorn]

so sigma

[hallow talon]

Will it be possible to train a lora for SD3 with the 2B model if I have a 4080?

[dusky thistle]

scheduling the amount of ancestral noise, the color of the noise, crude momentum implementation, c2...

[noble coyote]

OK, so it is a curve which influences the noise-schedule and there is obviously a sigma-sweetspot to maximise image quality?

[dusky thistle]

and some other shit getting injected into the sampling process and mixed in with hard light

that thing is a monstrosity

[viral plaza]

yes that there is christian, our CEO, and relevant final decision maker on the licensing topic, standing next to lisa su

[viral plaza]

yeah

[static cairn]

do you mind sharing what's being worked on for 8B? does it need more training before it feels worthy enough for a release? my team wants to train SD3 but we're debating whether we should wait for 4B/8B. 4B seems ideal for us

[hallow talon]

thanks! Looking forward to getting started with finetuning when it releases.

[viral plaza]

it needs more training first yeah. Right now our best 2B looks better than our best 8B on some metrics, so we need to improve 8B enough that the scale boost is worth it before 8B is relevant

[candid holly]

with my new PC being able to iterate a new image in under 20 seconds should make learning by doing much easier. having to wait 40-80 seconds per image means that you spend your time waiting playing around with multiple things only to find in your next image, one of the tweaks you just did borked the whole thing

[noble coyote]

So as I eagerly anticipate June the 12th ... will having an 8Gb VRAM RTX 2070 ... will that disappoint?

[torpid forge]

alright, definitely looking better now that I'm not forcing the prompt to stay under 75 tokens

[viral plaza]

all the recent training work was on 2B

[tiny lintel]

i only have 12gb, you’re still good with that.

[patent acorn]

wait what about the minimum vram to train a 2b lora

example a T4 gpu

[static cairn]

thank you for the answer. i look forward to the release of the models and hope the stuff applied to 2B proves promising for the larger models!

[viral plaza]

can't say for sure exact numbers, but a T4 has 16GiB VRAM which is well more than enough

[patent acorn]

ohh

[dusky thistle]

absolutely. my abilities with comfyui didn't take off till i got my 4090

[viral plaza]

2070 should handle SD3-Medium fine

generally SD3-Medium is a bit faster and easier to run than SDXL is

[noble coyote]

Thank you for the reassurance

[patent acorn]

its prob well lower resources than sdxl to me

yeah im righ

[viral plaza]

so whatever your SDXL perf is - expect a tad bit better than that

[torpid forge]

nice

[candid holly]

How well would a 3060 12gb do?

[viral plaza]

that's more than good enough yeah

[dusky thistle]

is 4b something we can expect this summer or should we buckle down and commit to building stuff for 2b?

[patent acorn]

what does the devs think about this though?

[noble coyote]

Not to spoil the party, but features like eyes, fingers and limbs are still poor - even for all the hype and chutzpah surrounding SD3?!

[viral plaza]

ngl i have no idea what the schedule is gonna be for the rest of the models. 2B is june 12th, rest are whenever they're ready

[dusky thistle]

great, thanks, appreciate it

[viral plaza]

Building against 2B is worthwhile imo regardless because most of the initial work is knowledge&tools, and those will transfer to the other scales

[dusky thistle]

okay great

good to hear!

the loras won't but if the tools will, perfect

[viral plaza]

2B is better than 8B on humans

not perfect

[dusky thistle]

get the training down with the fast train 2b

[viral plaza]

but like... RealVis-SD3 or whatever is gonna come out and you'll have awesome humans

[dusky thistle]

now where does 8b shine? is it in quality? diversity of concepts? transforms/creativity/combining stuff?

[patent acorn]

who makes the SD3 i wonder if they can use the method here

[dusky thistle]

my biggest desire is to have something that can combine concepts that sdxl tends to separate into discrete ones

[viral plaza]

we were looking into quantization of SD3 before, that project got deprioritized sadly but there's definitely potential there, and it'll be awesome once somebody gets it working

[dusky thistle]

clown + shark = clownsshark, not clown + shark

[noble coyote]

There a guy @Glif named FABLAN - he is running SD3 generation for free?!

[patent acorn]

ah so ya gonna release the unfinished quants code or?

[noble coyote]

How can that be so?

[cobalt moon]

wasnt doing it able to oof the quality of SD3

before ever Q6 or Q4

[viral plaza]

right now 8B doesn't shine much other than maybe sheer breadth of knowledge. Once it's trained to catch up it'll probably win out on everything

[candid holly]

How does SD3 handle prose? Right now I'm finding that Ideogram handles prose the best by a long shot

[opaque zodiac]

will it be free?

[dusky thistle]

gotcha, yeah, even if the quality isn't outstanding, i'd love for something that can do a lot of new concepts

i've got my dora -> crazy noodle fiesta sampling psycho ward workflow pipeline down

i'm training on crappy images and getting razor sharp ones out now

[noble coyote]

I find that text and prose in SD3 works about 85% of the time ...

[viral plaza]

SD3 is generally better about this, not always perfect. We did experiments with if you focus on T5 training (ie deprioritize CLIP) it gets even better at this (at the cost of some of CLIP's stylistic understanding)

[viral plaza]

dooooubt it

[dusky thistle]

i've already seen it from the api

[noble coyote]

About on a par with Harrlogos2 LoRA in SDXL

[viral plaza]

chat moving too fast, idk what this is referring to

[opaque zodiac]

hey dear

will it be free?

[dusky thistle]

the API version has been able to do transforms that dalle3, mj, sdxl, etc all failed miserably at

is the api the 2b?

[noble coyote]

#🆕｜sd3 message

[viral plaza]

uhhh no confident the answer offhand. Generally all text understanding is smarter, for more specifics, uh, Try It And See ™️ ? :D

either now on API the 8B or june 12th with the 2B model itself

[cobalt moon]

oh wait

what's the thing with Core SDXL

[viral plaza]

on june 12th the SD3-Medium weights will be free for noncommercial use, yes

[viral plaza]

oh, idk, presumably just using SD3 API and funding it himself or something? idk you'd have to ask him

[cobalt moon]

there have been a lot of misinformation of Core SDXL and SD3 API

[patent acorn]

for real someone should post all those information here to reddit

[cobalt moon]

I am doing right now

[trim arrow]

weights are the uh... model thing, right?

[cobalt moon]

still writing

[noble coyote]

Will there be an affordable Commercial-Use release?

[opaque zodiac]

is it full sd3 or part?

[patent acorn]

i thought screenshotting haha

[viral plaza]

ImageCore is a workflow/finetune of SDXL, "ImageCore" is a placeholder to indicate "whatever the current best we have for general image generation" not including beta models like sd3

[candid holly]

If I were to give you a prompt I use to test the capabilities of models, would you be willing and able to show me it's output, or do i just have to wait 9 days?

[cobalt moon]

oh ok

[viral plaza]

"weights" is techspeak for "the model file itself"

[trim arrow]

I figured, thank you! I just wanted to make sure.

[noble coyote]

"How long do we have to 'weight'?!" 😄

[viral plaza]

presumably yes, not entirely clear yet, going to try to get it clarified with the execs in charge soon

[patent acorn]

but if cha screenshotting, you should censor the ordinary member names here

[trim arrow]

God it must suck to be you rn with how fast chat is going with all these questions

[viral plaza]

yeah they made that joke in the announcements: "The 'weight' is over: SD3 released on june 12th"

[trim arrow]

OHHHHHHH i missed the damn joke in there lmao

[patent acorn]

wait i didnt see it

[viral plaza]

It is the full SD3-Medium, and not yet the Small/Large/Huge variant models

[patent acorn]

the joke

[viral plaza]

sure why not

[viral plaza]

I have trained so many years for this

[patent acorn]

https://cdn.discordapp.com/emojis/910912676224434227.gif?size=48&quality=lossless&name=mw

[viral plaza]

I am putting my discord fingey muscles to the test

[trim arrow]

https://tenor.com/view/jimy-carry-gif-3864991880073036063

[patent acorn]

now test your type speed

[cobalt moon]

still writing...

i actually have my midyear exam tomorrow

so

good luck for me

[patent acorn]

good luck

[opaque zodiac]

what are the other versions?

[patent acorn]

ur faster it is

[viral plaza]

Different sizes of the model, with different resource requirements and theoretically once trained different quality. So the small model for example uses half as much VRAM/time to generate, but theoretically gets about half the quality, vs large is double/double.
In practice the quality different is less precise, and mucked up by the fact that only Medium is trained well so Medium is best quality atm

[gusty gale]

By 'Stable Diffusion 3 Medium' they are referring to a version smaller than 8B?

[viral plaza]

Medium == 2B

Small is 1B, Medium 2B, large 4B, huge 8B

[vernal ocean]

From XL to medium? Big downgrade?

[candid holly]

Okay, rewrote it in prose rather than tags:

In the Hall of the Thousands Doors, an awe-inspiring representation of the afterlife, the atmosphere is heavy with the scent of incense and damp stone. Quiet souls wander amidst dim torchlight, casting long shadows across the cool gray stone walls adorned with intricate carvings and symbols. Endless hallways stretch out from a central hub, each lined with doors featuring unique, ornate designs and precious materials. Whispers hint at the lives contained behind each door. In the center stands Aponyx, the Raven Goddess of Death, her majestic yet terrifying presence watching over her domain. Above her colossal throne, a mesmerizing portal shines with an ethereal light, offering a glimpse of the City of Heaven.

[opaque zodiac]

ok,sd3 is it launched in site or in gpu?

[gusty gale]

Ah, I see. Will there be a comparison of the different sizes?

[patent acorn]

there is a comparison

i mpretty sure thats the ice cube thing on twitter

lykon did post one too but its a guess game

[candid holly]

Yeah, when they're all finally ready, would love to see a side-by-side comparison of different versions using the same prompt and seed

[viral plaza]

SD3-Medium is much better quality than XL, with a slightly lower resource cost

[viral plaza]

SD3 Medium will be available to run locally on June 12th.
SD3-Huge (8B) has a beta version available on API

[patent acorn]

nobody talked about the gens sd3 b2 will be 512x512..

[viral plaza]

right now 2B looks better than 8B as it's been trained the most, we can't really do a fair overall comparison until the others are trained more

we have some comparisons laying around somewhere to show the current state of thing

[vernal ocean]

Really? Back down to 512??

[noble coyote]

SD3@Glif (FABLAN) ... free-to-use ... 9 seconds/generation!

[opaque zodiac]

then i don't need gpu?

[viral plaza]

that is a giant prompt and i'm not sure what you were aiming for but it looks cool!

That's on a 2B release candidate model

[patent acorn]

@viral plaza is it really 512x512?

[viral plaza]

SD3-Medium is 1024x1024

[viral plaza]

no

[patent acorn]

OH then ive been misinformed LOL

[opaque zodiac]

maybe edit this ar?

[viral plaza]

You can either use a GPU, or the API, at your own preference

[noble coyote]

SD3 into i2i SDXL is a very powerful look!

[patent acorn]

iirc ppl on civitai discord, somewhere twitter said it being 512

[gusty gale]

Yes, aesthetic wise it is given to be better and probably know more things due to being trained more. But I'm wondering if coherency is also better currently?

[viral plaza]

There's a bit of confusion because a few months ago (?) I posted a 2B Alpha model that was 512x512 and some people don't realize that in the months since work was being done lol

[opaque zodiac]

cpu?

[patent acorn]

got it then

[viral plaza]

running an AI model on a CPU would be very very slow i would not recommend that

[viral plaza]

argh okay ill go find the civit discord in my list lol

[patent acorn]

fuck yea

!

[viral plaza]

here's same prompt on the old 8B Beta

[patent acorn]

theres so much noises

[viral plaza]

ye, overall better

[candid holly]

Hmm, not terrible, but I think it lacks the ability to really handle the prose. Here is ChatGPT (left) and Ideogram (right) as comparisons

[viral plaza]

you can compare the images above^ as a reference between 8B and 2B

[cobalt moon]

oh my god

[vernal ocean]

Is the prompt style changing again?

[cobalt moon]

that astraliteheart situation is long to write

[noble coyote]

I think this iteration of SD3 Medium should be nicknamed 'Hamlet'!!! 😄

[patent acorn]

he was typing out the yapping for 15 mins

wait what gender is astral

[noble coyote]

2B or not 2B ... that is the ?

[patent acorn]

nah

keep it a sub version of SD3

[opaque zodiac]

artisan is sd3?

[patent acorn]

yeah?

[candid holly]

Yeah, it's pretty clear that the new 2B is better than the old 8B, but both fall pretty short of ChatGPT and Ideogram in terms of detail imho

[viral plaza]

Artisan is a discord bot with a variety of features - SD3 is the default result of /dream but you can also use Core or whatever

[noble coyote]

Hamlet x 4 - 8B or not 8B

[patent acorn]

nooo just keep it as sd3!!

[viral plaza]

bear in mind 'ChatGPT' (DALL-E 3) is estimated to be something in the range of 20B-30B or somewhere huge like that, so... unfortunately hard to beat. It also rewrites your prompts for you with the LLM and probably reduces the confusion from the complex prose

[patent acorn]

it would be boring for an iteration of sub sd3 be called that

[viral plaza]

torcello's just making jokes and does not have the power to rename the model lol

[ocean lance]

are you using clip vit-g and clip vit-l again? and are they the original models, or did you fine tune them?

[noble coyote]

DallE-3 is so horribly overfinished ...

[patent acorn]

i thought he was trying to call a name of the model for himself

[muted dove]

What's the story with Cascade? Why was it released and then abandoned, even on here?

[viral plaza]

yes the same clips as used in SDXL, and no sadly not trained. There's also a T5-XXL Encoder added in

[viral plaza]

researchers joined, made model, left, and SD3 outprioritized it.
The real value with Cascade was in the research concepts they shared, rather than the model itself. Unfortunately I don't think much of that made it into SD3 due to timing overlap, but hopefully future image models will incorporate the concepts (eg the complex latent compression or the two-stage setup)

[noble coyote]

Oh yeah - if it was down to me - "Torcello-2B! ™ "

[ocean lance]

thanks, I have been training the clip l model. would it be possible to train a lora for the t5-xxl? or is it even necessary, like for adding new concepts

[dusky thistle]

Idk about that

[patent acorn]

cant wait for another opensource model that will be a 10b+ to overtake sd3 (and opt)

[dusky thistle]

And I'm getting better quality images now with some sampling tricks than I've seen from any closed source model except sora

From sdxl

We've got the flexibility of custom tools, trainability... It's way more powerful than any closed source bs

[patent acorn]

i remember trying out Open-Sora-Plan and it generated worse than expected, wanna post it here

[dusky thistle]

I even have SDXL writing text

[noble coyote]

SD3@Glif

[dusky thistle]

It's far more capable than we think, just need tweaks and training on custom concepts

Doras are incredible

[patent acorn]

woman: there are so many pomnegrates im so hungry
horse: how hungr

[candid holly]

Gonna have to go bug the guy I know with his $300,000 8 H100 setup to see if he'll let me use a bit of it's power when his team is taking a break from trying to make AI generated games...

[hallow lion]

lol i got the email too so apparently i wa son the list but i never got to access sd3 before lool

What's wrong with me Emad? XD

[patent acorn]

eh i got the email after losing a match on mobile legends anyways

[hallow lion]

loool

[patent acorn]

that was like 3 hours ago

i think

[hallow lion]

Photorealism: Overcomes common artifacts in hands and faces, delivering high-quality images without the need for complex workflows.

Bold claims.

[patent acorn]

[viral plaza]

Possible to train T5 / as a LoRA: Yes

Do you need to: probably not tbh, just train CLIPs and use that. Can even transfer SDXL CLIP training directly onto SD3.

[cobalt moon]

oh man

there's like

10 more question to go

I already write 6

[viral plaza]

^ this is the real facts right there. No matter how good your closed source model is, open source always wins cause you can customize and control

[patent acorn]

if someone trained lora with danbooru tags on sd3 ill be dead out laughter

[cobalt moon]

I will

[patent acorn]

NOOOOOOOOO

[cobalt moon]

I mean it is inefficient to write CLIP tags for anime

where you could have thousands of detail to write

[dusky thistle]

yep. once you take the time to dig into custom workflows, custom nodes, start training concepts and really learning how it works under the hood... you look back at dalle3 and just laugh

[patent acorn]

nah i think the t5 xxl has solved that

[viral plaza]

if you think there won't be an anime finetune of SD3 out instantly you're crazy

[cobalt moon]

ask Lykon

[patent acorn]

there will be but the concern is about the prompt adherence (it might be the same as the sdxl anime finetunes)

[dusky thistle]

on that note - are we looking at a day 1 release of training tools?

[ocean lance]

have they got a pull request ready to add the sd3 pipeline to diffusers?

[viral plaza]

HuggingFace has had the source code and reference weights for months now so i sure hope diffusers training will be ready day 1

beyond that idk

[rain current]

[dusky thistle]

the one big issue with cascade on my end seems to be some kind of undertraining maybe? design issue? with stage B... it's odd, more than 7 or so steps and it actually starts looking noisier

[dusky thistle]

excellent

[viral plaza]

@jolly swan I got a reply from one on the team managing the licensey stuff: they're still getting it sorted but are expecting to have a clear answer for commercial users before launch

[patent acorn]

magnificient

[patent acorn]

hopin a great news

[robust junco]

Always those Doras 😄

[dusky thistle]

Dora the Mad Sampler

[open plover]

SD3 model will be dropped soon? Is that true?

[viral plaza]

Yep, June 12th

[sterile pendant]

yeah i was talking about doras the other day. it's crazy that an 8 dim dora can compete with 32+ dim loras, even 128 dim. really cool stuff

[dusky thistle]

and it's not just size

i've done some 1:1 testing recently

[robust junco]

Sampleress 😄

[dusky thistle]

even though the charts show doras being marginally better at rank 64...

the results are distinctly better

there's a sharpness and a cohesiveness to the images that i've never gotten out of a lora/locon

[muted dove]

Are your changes any good for realistic images, or just the stuff you create?

[sterile pendant]

this is for LLMs, but it still is relevant

[ocean lance]

how will the model be packaged? will the T5 files be separate?

[sterile pendant]

probably separate, like other models that use T5

like when you use pixart sigma

[dusky thistle]

they've helped with everything

[dusky thistle]

yeah that's def where the big gap is

what i found surprising was how significant the gap remains beyond that

[sterile pendant]

i'd have to look into it more

[dusky thistle]

it's not something you'd notice standing 10ft from the screen, but it's very much there

i'm getting images far better than i trained on in terms of quality

[sterile pendant]

here's a phenomenal breakdown of doras for anyone else interested in the topic:
https://sebastianraschka.com/blog/2024/lora-dora.html

[dusky thistle]

i had lots of blurry, kinda pixelated muddy stuff from cascade... came out great with the dora, whereas with locon and lora especially that crap carried over

[viral plaza]

most likely we'll have the SD3 MMDiT backbone and VAE in the model file, and then textencs separately

[muted dove]

Eyes and skin aren't great, but I presume it's unrefined. Have you shared your nodes on GitHub, or in the manager?

[viral plaza]

there's some disagreement internally about optimal format atm tho

[viral plaza]

"anime sucks"? I mean you can have your own preferences and whatever, but you can't deny millions of people love it

[patent acorn]

GAZILLION!

[dusky thistle]

yeah that's model to sampler to vae decode

any defects you see there are a result of the crappy lora

cuts down the incidence of misshapen eyes quite a bit

mutations, etc

it can do text

nodes need some cleanup then il'l throw em on github

[muted dove]

Great, thanks!

[patent acorn]

walter white elected -> meth is legal

[ocean lance]

Will the workflow in comfy be similar to SDXL where you can send a different prompt to each text encoder? Does it still take the resolution, target, and crop parameters?

[viral plaza]

Yes you can do different prompts to different tencs if you want.
It doesn't have all the secondary conditionings anymore no

[limber stream]

and no point refiner right 😂

[cobalt moon]

https://www.reddit.com/r/StableDiffusion/comments/1d6ya9w/collection_of_questions_and_answers_about_sd3_and/

ouch...

[robust junco]

If you look at the image there is more to it 😄 Then, only because millions like it doesn´t necessarily say anything 😄 Btw I like anime, though I´m more of a manga guy. It´s more the endless repetitive often mediocre anime girl portrait AI images, especially the lewd-kind that´s sort of spoiling it at least here. Even more so if the lewd stuff looks like 6-year olds.

[ocean lance]

Just checking, but does it support img2img and upscaling?

[viral plaza]

ye

[viral plaza]

i'm mcmonkey4eva on reddit not mcmonkey

[viral plaza]

whoa also on question 13 refuse to collaborate further that's not part of that

[trail palm]

Hey Alex! hope all is good on your side. Will you guys release controlnets as with sdxl?

[cobalt moon]

there's going to be quite a lot of edits lol

grammar stuff

[viral plaza]

uhhh not sure but i don't think so. The way to make controlnets is really clean and clear tho, mmdit arch has a great way to do that (add a stream), vs in unets it was kinda hacked in.

[dusky thistle]

what kind of vram requirements are we talking to train one?

probably >24gb?

[cobalt moon]

someone also asked whether if you have any opinions on the rumour that SAI is running out of money and considering a sale

( that someone from CivitAI )

[viral plaza]

not sure exact numbers. Slightly less than equivalent training applied to XL

[viral plaza]

i'm just a dev i can't talk about anything to do with executivey businessy stuff

[dusky thistle]

i presume like the loras, you'll need to train them separately for 2/4/8b etc

[cobalt moon]

I expected this lol since probably NDA policy stuff

[sterile pendant]

95% chance, yeah

[viral plaza]

yes the different size variants are effectively whole separate foundations until something like x-adapter comes along for it

[bold socket]

Does anyone have a link to any sort of information where i can inform myself about the status on the training of lora's on SD3

[teal fossil]

I'm so used to vaguely organizing my crazy noodle empires now...

[patent acorn]

ah damn, still wondering whats next after SD3

[dusky thistle]

my record is 541 nodes

over 400,000 characterss of metadata

[viral plaza]

SD4?

[patent acorn]

possible

[sterile pendant]

Tack on like 300mb to every png lol

[viral plaza]

400,000 characters is 0.4 MB

[sterile pendant]

I know I'm just joking around lol

[dusky thistle]

things kinda went downhill with this one

and thank god i finally made a node to do this

[cobalt moon]

holy mother

..

what are you actually doing

[patent acorn]

NERD SHIT

[elfin plank]

rofl, that looks like my students doing unreal engine visual scripting. Visual scripting is not a reason clean code should not be maintained! 😄

[dusky thistle]

the modulation continent^^

i call it greenland

[patent acorn]

NERDUI

[elfin plank]

no, just no

[dusky thistle]

yeah this was my worst ever

this shit is what happens when i'm experimenting

[torpid forge]

https://tenor.com/view/-gif-8162509

[dusky thistle]

the contintents of sigma nodes now look like this

[sterile pendant]

I work with UE4/5 every day and have been doing so for like eight years. I've lived in noodles for so long that I'm a pastafarian now

so much so that i actually prefer noodles over code

[dusky thistle]

it's a balancing act

if your organizing is costing you more time than it saves digging through messes, you should stop organizing

if it's the other way around, it's time to organize

espec if you're going to be using it a lot

but experimenting? trying new ideas? make a mess.

all just about getting shit done

[sterile pendant]

yep, noodles are amazing for experimenting. granted, in UE4/5, you can just make a million functions to clean up the graphs, but still, early on, you deal with a lot of noodles while shits getting hammered out

[torpid forge]

I organize when my brain wants to rage quit

also, someimes what looks organized to me surely looks like a monstrosity to otthers

[elfin plank]

At least group and comment stuff :/ Don't teach yourself crap routines.

[dusky thistle]

current WF... organized as hell to me, but that's cuz i got the components memorized

[elfin plank]

Future you will appreciate the 5 mins you take to do that.

[torpid forge]

building those crap neural pathways up, bud

[viral plaza]

hey google can i ban someone from a discord just for inducing brain damage visually

[dusky thistle]

i'll be honest... that one became a practical joke

took every single node i had in all my open workflows

dumped them into one

and chained like... 12 samplers in a giant ring randomly combining everything

[torpid forge]

but why you do that?

[sterile pendant]

shit clown, i forgot about the plugin node i was talking about a while back where i was wanting to make some kind of bspline editor for doing the sigmas

when we were experimenting with stuff back then

[dusky thistle]

oh, yeah, that would be AWESOME

[patent acorn]

https://tenor.com/view/xqc-spotify-mood-want-a-break-bruh-gif-24599734

[torpid forge]

man, I have had some monstrosities myself. but never started off with that intentions

they organically turn into monsters

[sterile pendant]

i just dont know the framework for comfy all that well, but in theory, it's pretty simple. i could probably whip one up in UE5 in maybe 30 minutes, 20 of which would be relearning equations and stuff lol

[dusky thistle]

here's the full workflow i made where i combined everything i could

in the most complex way possible

[patent acorn]

still nono

[dusky thistle]

it gave me this.

[sterile pendant]

ive been digging all these ultra high contrast things you've been making. really neat stuff man

[dusky thistle]

thanks

it took modifying the samplers to control it

or it just got blown out or muddy

[sterile pendant]

just havent been active that much lately to talk as much on here. busy with some projects

[dusky thistle]

the dora really helped push things over the edge too

work finally died down for me after an 1812 overture style bombastic end in early may

so my form of relaxing has been binging on making nodes i've wanted for a while

[sterile pendant]

i feel you there

[sand jay]

@viral plaza do you have more demo pics of the current SD 3 medium? 🙂

[viral plaza]

sir you work here lol check on slack

[cobalt moon]

ask Lykon for it

[dull star]

Yay

A set date

This is nice

[teal fossil]

Well, with weight loading we could prune those... but without Unet most LoRA's won't be worthwhile at all, so we should just ReTrain everything on the new tech with better caption understanding.

[viral plaza]

textenc based models are super powerful, i wish people trained them more often

[dull star]

Is it possible on 24GB

[hallow lion]

what do these even do??

[patent acorn]

Make will smith eating spaghetti

[dusky thistle]

here's an example of just how much of a diff the sampling improvements has made...

before, after, and 1/4 the run time

that was trained on a number of images that had blurry edges

blown out contrast, etc

[ocean lance]

Highly recommend use this for training the clip-l model. Ran it twice with 200k images, then averaged the result. I think it helped improve the output from SDXL. https://github.com/zer0int/CLIP-fine-tune

[torpid forge]

I think I know stuff until I come here, lol

[hallow lion]

There's always someone with a more complicated workflow...

[dusky thistle]

i'd like to see the one more complicated than mine... i actually would, not being a lil shit lol

[patent acorn]

before sd3 comes out, is there a magic prompt node for comfy ui like ideogram?

[dusky thistle]

the 541 node one or whatever

[torpid forge]

no no no

[viral plaza]

what? running sd3-medium? Yes

[dusky thistle]

aimingfail made one called the clownstyler. it takes prompts, and makes them insane

[torpid forge]

oh nice

[dull star]

Oh no, like training textual inversions for example

[viral plaza]

oh, yes that too yes

[dull star]

Lessgoooo

Thanks

[patent acorn]

sheez 😭

[torpid forge]

he made a model with some of my images, lol. haven't chattted with him in a while. he said he'd gotten into the llms

[hallow lion]

my worst

:)))

[cobalt moon]

at least you organize it

[hallow lion]

ah yes

otherwise i go nuts

[dull star]

Oh wait this is 2B, even DiT training might be possible

[torpid forge]

I just put the boxes underneath, foret what those are called

the magnetic boards

[viral plaza]

yes

[dusky thistle]

you don't even have any nodes overlapping? live on the edge... try sticking a save image node underneath a ksampler

[dull star]

It's for 8B where training on 24GB might be finicky

[dusky thistle]

fingers crossed the 5090 is 32gb

if it's 28gb i will die

[patent acorn]

it says 28gb right

for 5090

[cobalt moon]

LLM peeps even wanted 46GB

lol

[dusky thistle]

i mean really, it should be 80gb

nvidia is fucking everyone

[cobalt moon]

I feel like there is not need to keep adding up VRAM

like hell

[dusky thistle]

the fact AMD failed so horribly to offer real competition is killing our field

[cobalt moon]

at this point 5090 is going to be all-round GPU rather than gamer-orientated.

[torpid forge]

why has it been 24gb for like 5 years?

[torpid forge]

should they not meet demand with supply?

[hallow lion]

its not the complexity that kills it for me its the slowdown, a few ipadapters and maskign and it slowls to a crawl

😦

[dusky thistle]

by not offering more than 24gb, they can charge another $10k for the next 56gb

$180/gb

which currently costs like 3 bucks

[torpid forge]

I saw that picture of the h200 I believe?

[dusky thistle]

there's no equal to their high end cards, and there's no way around vram requirements for training, so they can charge pretty much anything they want

[torpid forge]

openai bro and nvidia bro and then some other guy

[dusky thistle]

there's really no reason we shouldn't be able to buy consumer equivalents of a100s

[torpid forge]

it's like a small fridge

or medium maybe

[cobalt moon]

that's SXM version of H200

[viral plaza]

AMD is finally competing in the datacenter, they just need to get that improvement into consumer cards too

[cobalt moon]

H100 also have SXM version

[dusky thistle]

that's great to hear... hope that helps you guys out on the bottom line side

[cobalt moon]

SXM is like CPU socket with no cooling whatsoever, and also NVLink each other more efficiently compared to PCIe version

[sterile pendant]

because contrary to popular opinion, gaming GPUs don't really need more than 16gb of vram right now, unless the devs are bad and don't know how to manage texture pools correctly. also, allocated vram != used vram. you can make a pixel game like minecraft allocate 85% of the VRAM if you want, even if it's only going to use 100mb of it, but if you check the gpu usage, it will show that a ton is "in use."

[cobalt moon]

basically cooling was settled by air conditioner inside server room who often have air circulation

[teal fossil]

Does 2B have a 75 or a 512 (T5 iirc) Token limit?

[viral plaza]

bothn't

CLIP is 77, T5 is 512 or whatever, but you can also just stack multiple CLIPs as-needed

[dull star]

I wonder what will happen if we plug longclip into it

In like comfyui

[frail tulip]

hi guys

[ocean lance]

I didn't like the results with longclip

[frail tulip]

https://tenor.com/view/hi-hi-there-gif-16107842475380108536

https://tenor.com/view/munsonated-gif-26549638

https://tenor.com/view/anyone-there-goofy-anyone-there-gif-11686233

[patent acorn]

long clip?

[frail tulip]

oh hi there

[cobalt moon]

https://tenor.com/view/electro-boom-gif-18170714

[patent acorn]

ohh this guy

hes resistant to ELECTRIC!

[frail tulip]

https://tenor.com/view/yuji-itadori-suku-romen-sukuna-ryomen-sukuna-yuji-changing-to-sukuna-gif-10017632583639619047

[cobalt moon]

oh yeah hey

[patent acorn]

no this channel is sd3 discussion not your jjk general

[frail tulip]

https://tenor.com/view/annoyed-fuck-frustrated-gif-15331740

[patent acorn]

so hi

[frail tulip]

btw hi

[ocean lance]

less gifs, more sd3

[frail tulip]

sd3?

https://tenor.com/view/how-confused-but-how-puzzled-calculate-gif-16247428

[patent acorn]

hmm i think this would be fine

[torpid forge]

fair enough. well perhaps that may have something to do with them making sure to stay within the hardware contraints most people have?

[cobalt moon]

yo

[patent acorn]

NO

[frail tulip]

yo

[cobalt moon]

this is not your usual interaction server

[frail tulip]

then?

[patent acorn]

welcome to the server where the world does not revolves around you!!!!

[frail tulip]

what do we do?

[patent acorn]

this is FUCKING SD3 DISCUSSION AND SOME AI SHIII!!!!!

sorry for being aggresive

[frail tulip]

ayo chill

[cobalt moon]

if you keep posting this you will get oof.

anyway

[frail tulip]

i am new here so i dont no anything

[dull star]

Oh will we get controlnets or not?

[patent acorn]

you only should discuse sd3 in this channel (either ai stuff because why not), not the usual interaction

[frail tulip]

waht is sd3

[patent acorn]

wait

how did you

even join

??

[cobalt moon]

even you want it just go to #💬｜general-chat

[patent acorn]

who invited bro

??

[cobalt moon]

SD3 is a new generation text2image diffusion model

[frail tulip]

i want a channel where u discuss abot bots

[cobalt moon]

pretty sure you have heard of SD1.5 or SDXL

[frail tulip]

no

https://tenor.com/view/how-confused-but-how-puzzled-calculate-gif-16247428

[cobalt moon]

it is a new member in the family

[patent acorn]

theres this talked about somewhere in this channel

[patent acorn]

so then ur a rando that randomly join servers

[frail tulip]

wait

[cobalt moon]

probably not at launch date but soon after launch

... actually depends how successful is the 2B

[frail tulip]

is this a family

[patent acorn]

No

[frail tulip]

then

[patent acorn]

that was a sarcasm

[cobalt moon]

what?

[patent acorn]

kinda like a new member to the server

[cobalt moon]

bro

[frail tulip]

何

[patent acorn]

@frail tulip who invited you anyway

[sterile pendant]

the average GPU has 8gb of vram and ~5% of pcs have a gpu with more than 12gb of vram. https://store.steampowered.com/hwsurvey/ if you want to see what people are working with. as a game dev, i can assure you that pretty much all pc devs use these kinds of stats for performance targets (unless you're some big AAA company trying to make the next cyberpunk2077)

[patent acorn]

its like ur so clueless

about stable diffsuion

[frail tulip]

私は新しいです

no one invited me

[patent acorn]

ah my lil bro dont read rules

then

how did you came across this server?

[frail tulip]

i explore discoverable server

[patent acorn]

ah then ur so clueless

[cobalt moon]

calm down bro

[patent acorn]

go to anime hangout

not here

[frail tulip]

no

[patent acorn]

alright

[frail tulip]

i wanna learn smt abt chatbots

[patent acorn]

but

[cobalt moon]

oh chatbots?

[frail tulip]

https://tenor.com/view/ryan-reynolds-wtf-what-the-gif-23856906

[cobalt moon]

that's Large Language Models' job

[patent acorn]

the most chatbots are on Llama chatbot discord servers, stable diffusion is image

[cobalt moon]

or LLM for short

[patent acorn]

yea

[cobalt moon]

yeah Stable Diffusion is text2image

[frail tulip]

https://tenor.com/view/what-do-you-mean-james-engvid-what-are-you-saying-what-did-you-say-gif-24308110

[torpid forge]

well I will not dispute that. I'm not really up on gaming these days

[cobalt moon]

( although we have #1098025024541167646 channel )

[patent acorn]

its bad i suppose

im loving llama 3

[frail tulip]

ok

what do u even do here?

wtf

[patent acorn]

discuss ai image genereation, which is Stable Diffusion

this one is opensource

the nowadays u see locally are Dalle 3

[frail tulip]

that is only what i am talking abt

[torpid forge]

you need to chill there, bud

[frail tulip]

i joined the right server then

[cobalt moon]

someone is damn new alright

DALL E 3?

[patent acorn]

https://cdn.discordapp.com/emojis/1092035766034640946.webp?size=48&quality=lossless&name=mrphmmm

[cobalt moon]

well DALL E 3 is intergrated into ChatGPT

[patent acorn]

n oi mena i see that everywhere in my place

if ur confused

[frail tulip]

i ma leave

[patent acorn]

goodbye

[frail tulip]

cya

[torpid forge]

whata is happening?

[cobalt moon]

before that DALL E 2 is a completely independent website on its own

[frail tulip]

dayum

[sterile pendant]

fair enough, just saying, these are the kinds of things that companies like Nvidia factor into how much vram to slap on a card. there have always been higher vram cards though if you go the quadro workstation route.

[cobalt moon]

no ChatGPT stuff

[patent acorn]

yup i do not like dalle

[torpid forge]

why are people tribal in regards to AI models?

[cobalt moon]

well, psychology

[lavish pier]

is the sd3 model which is releasing bad compared to the api one?

[cobalt moon]

you know when one product with revolutionary feature get introduced

[torpid forge]

yeah, tthat is the technical answer

[obtuse fractal]

Same reason people are tribal about gaming platforms - the one I pay for is the best licks boot

[cobalt moon]

how will the people react

it's not just AI, but like, everything.

[patent acorn]

exactly just dont post anything AI stuff to artist discord server lol

[cobalt moon]

you show them boring stuff they dont even bother you whether you release the thing or not

[cobalt moon]

it's a hot topic at the end of the day

[patent acorn]

though twitter has misinformed them and leading them thinking its "theft"

[sterile pendant]

why do people get tribal about their favorite sports team, their favorite music genre, their favorite foods, etc etc. it's just people doing people things and it's mostly just the vestiges of hundreds of millions of years of evolution that got us to this point.

[torpid forge]

people are only doing themselves a disservicee if they deprive themselves of learaning about things simply to feel like they're on a team

[torpid forge]

yeah, that's something I've actually been reading about a lot recently

[patent acorn]

humanity moment

[robust junco]

thank you for the effort

[torpid forge]

so many of these things that were once necessity

[obtuse fractal]

100%! It’s not team midjourney, team sd, team dalle… It should be team best value for money and team best safety features to ensure AI doesn’t murder us all

[torpid forge]

it was advantageous back when our ancestors were cavemen. I guess maybe it still is in some ways

I just like learning about different things

[sterile pendant]

just observe our ape cousins, you'll see the same stuff. even in packs of other animals as well, dogs, cats, mice, etc etc. it's pretty common in the animal kingdom to not want to be the outcast, because that usually signifies something wrong with the individual and the urge to reject them is to keep the "broken genes" out of the pool. it's observed in pretty much every creature on the planet with a brain

tribalism is just an extension on all that

since we have more cognitive capacity and all

[patent acorn]

what about people calling anyone that uses ai "tech bro" (not ai bro in this case)

techphobic much!

[torpid forge]

well they're probably not comfortable with ai

makes them anxious

so they externalize it and call people tech bro

[cobalt moon]

remember those industrialist back in 1800s at England

[obtuse fractal]

Same as people calling the crypto shills crypto bros

[cobalt moon]

there always had peoples fighting against the idea of an industrialist and sought to protect their green land

it is the same

if you knew it

[obtuse fractal]

Early adopters of tech have always been mocked by the same people who later adopted it

[torpid forge]

People don't like the idea of the reality they grew comfortable in changing

It means the things they're good at might become less relevant

[cobalt moon]

this remind me one Tom Scott video

[patent acorn]

except crypto are kinda worse in my eyes

oh nvm

i misenterpreted that lol

anyone that mines ig

[obtuse fractal]

lol

Ive made good money with crypto

But crypto bros are annoying af

[torpid forge]

People also tend to dismiss things they don't understand

Crypto bros are like "influencers"

More sure of themselves than they should be

[patent acorn]

ive been hating em even the nft bros

but are ai bros worse? i dont think so lol

[obtuse fractal]

Ahh totally agree

[dull star]

I mean fron stability

[obtuse fractal]

I bought an NFT. Never regretted a decision in my life more.

$100 I’ll never get back 💀

[patent acorn]

you poor soul

YOU POOR SOUL.

[cobalt moon]

uhhhh

[torpid forge]

nfts seem useful as a concept. but not the spaceape crap

[sterile pendant]

I wouldn't say dismiss, reject would be a better word. A large portion of people are immature and can't stand when others know something they don't. They then go on to not bother trying to learn about the new thing, and then they just reject it.

[cobalt moon]

I remember correctly all ControlNet models was uploaded by Illyasviel

but is he a Stability dev/staff?

[patent acorn]

cant even mention the dark humor ones lol

[obtuse fractal]

Thank you, thank you. Oh btw wanna buy an NFT? Guaranteed to moon 🌙

[patent acorn]

no thank you

[torpid forge]

it's mostly based in a lack of honest self reflection I believe

[patent acorn]

id screenshot

;)

[torpid forge]

I don't even know about these dark humor nfts. I guess I haven't kept up

[obtuse fractal]

There’s like 50 other people with the same NFT as me and they’re selling for like $5 now lmao

[patent acorn]

in some cases harrass the ai user

[sterile pendant]

Mixed with childish narcissist tendencies that they never grew out of. Those types of people will always just nope the f away from people that don't reinforce their behaviours

[patent acorn]

there are, and i would get banned for talking about it lol

[sterile pendant]

And therefore, they never learn or grow up

[torpid forge]

narcissists lack a full sense of self so they literally can't self reflect. well they could. but it'd require acknowledging everything they've pushed down since they were a child. most will never do it

[stray cedar]

SD3 finally gets announced and this is what the convo turns into on the SD3 channel...

[sterile pendant]

yuuup... but anyways, i'm not going to dr phil rant about it any longer because i can't stand dr phil (he's a massive narcissist and a megalomaniac lol)

[torpid forge]

what do you want to talk about, buddy?

[stray cedar]

hmmm how about... SD3

[torpid forge]

oh, by all means

go on

[sterile pendant]

what is there to talk about other than it comes out in a couple weeks?

and maybe see the same waifus spammed over and over again lol

[patent acorn]

dont know how to defend this one of course
and the convo is over byee!!

[patent acorn]

the resources

there isa lready a reddit post for the infos

[torpid forge]

AI short circuits the idea to result path. it will facilitate an explosion of creativity

[dusky thistle]

Can't wait for the prompt zombie Armageddon

[patent acorn]

still wouldnt accept that because "theft"!!

[torpid forge]

oh well

they're on the wrong side of time

[patent acorn]

wait for 4 years and i will be able to reference AI stuff without getting harrassed

[torpid forge]

they're literally fighting against the idea of democratizing creativity

[patent acorn]

waiter waiter! failed gens please!

[torpid forge]

they won't be remembered in a positive light

[sterile pendant]

oh lord, and the complaining about them not looking good enough just because it isn't some hyperoverfit model they're used to using that spits out "adult women" with the facial proportions of a younger-than-adult-woman... oh and the bobs not being gud enuf

[patent acorn]

i will say they waste their time arguing on twitter, they will say i waste my time ai genning "slops"

[dusky thistle]

I can't wait to fulfill each other every request

[sterile pendant]

SAME lol

i'm going full troll with it using the typical old men in tiktok poses lol

[torpid forge]

well also people act like the models won't become more sophisticated. like they're these static things that will just be

[patent acorn]

whats a "sophisticated"

[torpid forge]

it's an adjective

[patent acorn]

whatever!

[torpid forge]

(of a machine, system, or technique) developed to a high degree of complexity.

[hallow lion]

We need AI to fix the real estate and the economy. 😄

[patent acorn]

AI Governments when

[torpid forge]

never

[sterile pendant]

not going to lie, i still use the base sdxl model for like 2/3 of the things i make, unless it involves people. but i rarely make people unless it's waifu trolling

[torpid forge]

who will train them?

[patent acorn]

hey man can you unban tax evasion
"im sorry, but"

[torpid forge]

well I'm more talking sd5 or 6, or whatever comes after. it's going to get exponentially better at what it does

[hallow lion]

I am going to Mars and throw your tea into space. I refuse to pay taxes to earth. Havea nice day.

Free Mars.

[desert garnet]

do not listen to those analog bros,heard they still clean their buttocks with wood based paper

[torpid forge]

what about the 3 sea shells?

when's that happening?

[sterile pendant]

hah they dont know how to use the three sea shells

[patent acorn]

ikr they make me braindead!!

[craggy ridge]

So. Many. Messages.

[dusky thistle]

i use base + my dora

base is flexible

i like that

[torpid forge]

I use whatever I want

SD3 image for you guys

[hallow lion]

Im glad Dear Leader has been trained in.

[patent acorn]

my hero

[hallow lion]

All hail Kim.

[sterile pendant]

yep, it's ultra flexible. all these other models really do is push out variety in the model for waifu closeups

[torpid forge]

all the models have everyone in them. I got dalle3 to give me Xi

turns out the key is just spelling atrociously

[dusky thistle]

there's a few decent ones but yeah 95% is that shit

[hallow lion]

u just got winnie the pooh

[dusky thistle]

big waste of all the training ability we have

there's already trillions of pr0n pics online, do we really need ppls 4090s cranking out another 90k a day

[torpid forge]

nah, I got copilot to flip out about stuff microsoft did, censoring things for china. and then started in on making images to illustrate

[hallow lion]

did google fix the black pope divesity?

and asian nazis

[noble coyote]

SD3@Glif

[desert garnet]

yes we do

[sterile pendant]

the only other models i use are realvis4, tuggernuts9 and pg2.5

[hallow lion]

yes

[restive halo]

didn't they originally say we are getting Controlnets with the release? It sounds like that's no longer the case?

[hallow lion]

BastardV1 is good for promt adehrence

it gets ANGLES right

top down, low angle etc

[noble coyote]

Knot Essdeethree

[hallow lion]

make one wiht XI and Kim kissing

[patent acorn]

can sd3 generate winnie the pooh?

[dusky thistle]

so long as we get the tools, who cares

we just need the weights and the tools

[restive halo]

sorry, who cares if we have the tools, as long as we have the tools? what are you even saying

[hallow lion]

i think so

[dusky thistle]

what are you even saying

[hallow lion]

only famous artists were left out i guess

and getty/shitterstock watermakrks

famous people are fair game

[restive halo]

I asked if we are still getting controlnets, which are some of the main tools. You said who cares if we are

[torpid forge]

which artists were left out? I'll test that out

[hallow lion]

theyre in the public domain and attention whores anywayz

[patent acorn]

can sd3 generate this?

[torpid forge]

yes

[sterile pendant]

you'll get them eventually? or would you rather them wait another month or two to release the model

[patent acorn]

oh please

[desert garnet]

keep my wife's name out your mouth (Taylor Swift)

[patent acorn]

and this i hope

[torpid forge]

but can sd3 generate this?

[noble coyote]

SD3@Glif

[restive halo]

without the tools, and with tooling only getting made around a small model, it's about the same for me, or maybe worse this way tbf

[patent acorn]

pooh handshakes with kim when

[patent acorn]

i think it could generate sculptures like that

[torpid forge]

so can I. in fact I generated that image right there

[patent acorn]

oh ._.

oh shit fish gyatt 👀

[noble coyote]

SD3@Glif

[torpid forge]

can sd3 generate this?

yes. yes it can

[noble coyote]

A Coelacanth?!

[teal fossil]

Will it be like Cascade that instead of as a Part of the Model (like SDXL) we'll be able to use the Tences individually (and thus also Finetune them and use others like T5 variants)?

That would also mean that we can use Longclip (once both versions are available).

[sterile pendant]

i modeled a super basic model in zbrush based off a trout, but i added some scales and arapaima to the prompting

[noble coyote]

SD3@Glif

[patent acorn]

wholesome!

[sterile pendant]

i made it as a joke for my idiot fishing buddies that are obsessed with fish like they are attracted to them

[teal fossil]

I guess I'll try that, even if I'm pretty sure I'll ReTrain everything on 2B.

Does SD3 use Clipskip 1 or tricks like Clipskip 2?

[patent acorn]

i need to take a nap rn

[patent acorn]

men will see this and say hell yeeah

[noble coyote]

This is what all the hard work has produced!!! 😄

[sterile pendant]

ahahah yeah, god i hate it here lol

[sullen moss]

12 June, almost two weeks 😅

[lavish pier]

9 days

[desert garnet]

2b 2weeks,numbers dont lie

[teal fossil]

Can we train SDXL DoRA's with Kohya or OneTrainer already?
It's not the same as Lycoris, right?

[noble coyote]

Good prompt coherence

[sullen moss]

8B 8 weeks. LOL

[teal fossil]

I guess we'll also get the Vae separately?
For XL for example the Finetunes FP16 fix Vae is the go-to at the moment.

[lucid swift]

I thoght medium was 4b

[desert garnet]

it tried but it wasnt meant 2b

[muted dove]

This is the most organised I've ever seen any of your workflows! 🤣

[noble coyote]

2B or not 2B - hey, let's call it Hamlet instead of Medium?!

[teal fossil]

Did you create a node that let's you control the strength of a LoRA per step of Inference?

[dusky thistle]

It's actually very organized lol

[dusky thistle]

I haven't, I wish

[teal fossil]

I always train both. Tenc training was absolutely a game changer for Dreambooth etc. (basically enabled proper v1-5 FineTuning iirc), but in my experiments neither Unet nor Tenc alone make big enough changes to grasp new concepts properly (and separate them properly).

[noble coyote]

Knot Essdeethree

[hallow lion]

looool

[cerulean granite]

a gnome

[noble coyote]

SD3@Glif

[outer elm]

@suchamazewow gut

[rapid ivy]

I thought because I paid for stable membership I'd get SD3 as a part of it but you're asking for commercial use for us to fill out a form.

Do you need that if you have stable membership?

[hallow lion]

Hmmm, number of legs questionable...

[torpid forge]

indeed, that's happened multiple times

[hallow lion]

oh no

before it was hands

now its legs!? :0

[restive halo]

so how exactly does the fully non-commercial license for sd3 work? Can you not even use it in youtube videos that have ads on them?

[muted dove]

The license hasn't been published yet, so nobody knows for sure.

[restive halo]

I wasn't sure if one of the staff who posts here hasn't said more

[cobalt moon]

they still sort it up

the license things

[noble coyote]

I suspect that SAI will make some money by selling Commercial Licenses ... but the Community can knock-itself-out for free ...

... as long as they're not selling stuff!

[restive halo]

the previous license was reasonable, you can use it commercially if you have less than X users (which presumably is the same as viewers for videos), and over that you pay

[muted dove]

That's what they said, it should be ready before the release of SD3 anyway.

[restive halo]

but in the email they implied you can't use it commercially at all for sd3, and you can't just pay you have to do a private deal

[cobalt moon]

"at all"

dude

it doesnt mention it at all

[restive halo]

'If you would like to discuss a self-hosting license for commercial use of Stable Diffusion 3 please complete the form below and our team will be in touch shortly.'

[cobalt moon]

just the word non-commercial license

and form

read that again

[restive halo]

it implies you can only do commercial if you contact them and do a private deal

'SD3 Medium weights and code will be available for non-commercial use only.'

[cobalt moon]

man

[restive halo]

only = at all

[noble coyote]

It would be very hard for SAI to regulate sales of SD3 material ... they'd have to have a "finger in every pie" to make it work!

[cobalt moon]

and where's the non-commercial license looking at?

[restive halo]

what

[noble coyote]

In the old days, MJ said that you could sell material up to a certain level of remuneration. Anything above that, and MJ would want a cut, want a percentage

[restive halo]

that's similar to how the current SD license was, but in the email they say you can't do commercial 'at all' which is a change from the 100k users thing

[cobalt moon]

you realized both still have some differences

even though they did have some relation to each other.

[noble coyote]

Adobe do 'commercial' with Firefly: they allow so many free uses/month inside of Photoshop - then when that batch has finished - you must top-up, or wait until next month

[cobalt moon]

plus I am thinking "SD3 weight will be available for non-commercial use only" can imply those who paywalling their trained model.

[hallow lion]

wtf is that 😦

Emad! That's 6 fingers!

[torpid forge]

https://tenor.com/view/counting-fingers-123456-hands-gif-27529933

[lucid swift]

WhAt

[teal fossil]

Stack Clips? Like A1 / Comfy prompt truncating?
Please elaborate.

[teal fossil]

Did you feel an actual improvement with Longclip? I gotta say I didn't. It didn't get more concepts right than the usual prompt interpretation in Comfy.

[hallow lion]

On a scale of 1 to 10 how exhausted are you?

I am a solid 8.

C A F F E I N E is extinct.

[cunning lintel]

#🆕｜sd3 message #🆕｜sd3 message
I really don't like the 2B sample, way too contrasty with hard lines while nothing in the prompt calls for it, and while not so noisy, I'd call it smeary instead, fine detail got smoothed away. I kinda expected this effect was from overdone postproccessing when i saw that teaser tweets, but it seems not. I'd like the posted 8b one if only it weren't so unusable noisy.

Either way, tried the prompt in the API as well, it's different to both (follows the prompt less it seems, no goddess to be seen, but the overall aesthetic is way more similar to the 8b sample posted, not the weird 2b output), so that begs the question, what's running in the API? I'd suspect a first iteration with smaller text context and that that has been cranked up since.

[remote holly]

Sd3 for 12 june , the end of sd3 loop paradox lol

[left parrot]

Is it likely that we can run SD3 Medium with comfyUI on the 12th?

[sterile pendant]

Yes, comfyanon is a beast at getting new things up and running quickly. Even if he doesn't have it already ready, you can always just run the sample code from the hf page when it goes up, but you'll only really be able to do basic prompting

I'd imagine comfyui will have it ready before the models go live, or within 24h of the release.

[plucky flax]

Hey guys can you tell me which channel should I go to for help in using stable diffusion in local

[gusty gale]

I'm pretty sure he even already has it ready, as people with access to the model mentioned ComfyUI workflows with SD3 a while ago

[coral sable]

Fantastic news SD3

[dull star]

well definitely not with SDXL, it was already smarter than 1.5

[sterile pendant]

True that, yeah I'd assume he works with them on stuff since I've heard a lot of them use comfyui internally.

[woeful spindle]

I wasn't lying when I said it's 2 weeks until the SD3 weights drop

[dull star]

SD3 when???
SD3 then

[gusty trail]

SD3 when-> SD3 Large When

[dull star]

So SD3 4B?

[desert garnet]

8b,i like them big

[dull star]

I like em chunky

[sterile pendant]

Probably in a few months after they see what worked and what didn't work with the 2b model. Also, to give them time to pull in some revenue and to see if it's financially worth it to fully train the higher parameter models(or wait for some company to buy them out)

[sullen moss]

SD3 Dzen)

[patent acorn]

SD3 how

SD what

[compact forge]

So we got the announcement to announce the announcement on june 12th?

[desert garnet]

i talked with cto last night in shower and he 100% real no lies no scam no virus

[dull star]

I was responding to xiaozhi about SD3 Large being 4B, just in case he meant 8B

I am aware that it will take a lot of time

I bet we won't see 8B until July or August

when 2B comes out and its good, I'm gonna buy 10$ worth of credits as a donation to stability

[gusty trail]

I thought large is 8B. if large is 4B, what is the name of 8B?

[dull star]

Huge apparently, but I'll look it up

[desert garnet]

thicc

[dull star]

Small, Medium, Large, Huge

[gusty trail]

SD3XL

[dull star]

yeahh

extra large

[turbid grotto]

extra huge

[dull star]

We're on track to release the SD3 models* (note the 's', there's multiple - small/1b, medium/2b, large/4b, huge/8b) for free as they get finished.

so yeah, 8B will come out

for free, offline, etc

[wide pagoda]

800M / 1B - "Small", same size as SD1.5
2B - "Medium", the first one that will be released
4B - "Large", same size as SDXL
8B - "Huge"

EDIT: this is wrong, apparently XL is 2B

[dull star]

basically

but 2B is as good as SDXL basically

[sterile pendant]

I think it might be even later than that, but we'll see. Personally, I don't have a dire need for it since there will likely be strong diminishing returns compared to the 4b model.

[turbid grotto]

btw they could make something like 16b for API or own service to compete with dalle and mj

[wide pagoda]

All except 2B are very much still WIP and not the current focus

[dull star]

Exactly, and the 2B model will keep us busy anyway, so waiting for the rest isn't so bad

800M might have to be their next target, so lower end people can get to work

[tough oriole]

Even if they gave us the 8B one they said it wasnt finished. It would take a while to even tune it so waiting is fine.

[dull star]

imagine FastSD-like implementation/app for SD3 running 2B-Turbo or 0.8B-Turbo or whatever

smart image generation democratized

its only T5 that has to be figured out

we need some quantization or ggml implementation so we can do stuff like IMatrix quantization at like 4-bit for T5

[sterile pendant]

2b will be massively more powerful than sdxl. The 2b size would be more like 8b in unet format. Sdxl has like ~3b in the unet, not counting the other model parts.

[dull star]

I remember 3.5B on the website, but I suppose that's the total and not just the Unet

idk why they didn't write that in the paper or the huggingface page

only in the announcement of SDXL

[wide pagoda]

I wonder what the hardware requirements for 2B are going to end up looking like. Forge is able to run SDXL on some surprisingly weak cards, but it's been abandoned and none of the other UIs ever managed to accomplish that

[turbid grotto]

I checked base 1.5 and just shocked how much finetunes improved it, I think there won't be any quality problems with 2b version

[sterile pendant]

Kind of like how llama3 8b outperforms llama2 80b, it's kind of pointless to compare model sizes between different architectures

[turbid grotto]

What?! It has almost 3 times more data and similar requirements?

[wide pagoda]

I'm mostly interested in what cards it actually runs on...

[dull star]

I can't wait to make Lora-like models for 2B

this is going to be so much fun

[sterile pendant]

So sd3 2b will likely be as if sdxl had 2-4x as many parameters, plus, don't forget we will have 16 channel vae

[dull star]

and Textual inversions might come back to fashion if we're going for ""cross-platform"" compatibility between the 800M and 2B models or whatever

just probably for styles, maybe not subjects

[turbid grotto]

is this transformer's magic?

[sterile pendant]

Sd3 doesn't use a unet, it uses a dit format like llms use. Transformer based.

It was just a bitch to pull off for image generation until more recently

Without having dog poop quality or paint drying speed

And without needing 9001gb of vram

[turbid grotto]

now I am even more hyped, why didn't they use it as marketing... or I missed something...

[sterile pendant]

It's in there, wanna say they call them mmdit blocks, but I can't remember off the top of my head

[dull star]

I suppose stability wants to do a llama moment where they actually train the model for a good time instead since now its more transformer based

if the results are gonna look like lykon's, I hope it will still have variety

[sterile pendant]

You can train it all you want, but if the captioning on the dataset is weak, you're just going to get more of the same. I know they used cogvlm for a good chunk of the dataset in sd3, but I don't know how well it works. I've used a ton of vlms and I know how much of a crapshoot they can be sometimes

[dull star]

yeah 50%

so that the lack of knowledge from cogvlm (pop culture or video games or etc) aren't as detrimental

[turbid grotto]

so there is still room for improvements

[dull star]

I'll make loras for video games and other stuff

its gonna be fun

when the model actually somewhat knows what you actually want

[turbid grotto]

I am curious, was dalle3 captioned by gpt4v?

[dull star]

think so

or whatever their in-house tool was for captioning

it is super detailed and smart

probably better than cogvlm

but cogvlm is still very good

[turbid grotto]

and big

[sterile pendant]

The biggest room for improvement would be in the t5 portion of the model and its ability to map dumb prompting to magic in the network

[dull star]

man this looks photoreal

gave this image to some people and they thought it was real

[wide pagoda]

They did sometimes https://fxtwitter.com/EMostaque/status/1760660709308846135

[noble coyote]

SD3 into i2i Searge SDXL w/flow

[dull star]

its great that they have both clip and T5, so T5 doesn't become a huge bottleneck for low end people

people can just disable anything they want

[desert garnet]

needs more cream

[dull star]

like this looks good, and it doesn't have a dreamshaper or juggernaut feel to the face

[sterile pendant]

True, but the fact they are still using clip for prompting as the main source of the prompt is kind of lame and is what will hold the model back a lot. If they had gone pure t5, it would likely be a monster, but as you said, would shut out the bulk of users

[dull star]

also look! no bokeh! no depth of field!

https://tenor.com/view/thanos-impossible-marvel-shocked-gif-15104180

[desert garnet]

but the pocket is wrong 🤓

[wide pagoda]

It's still uncanny valley, just not the sameface of previous models (which you wouldn't expect it to be)

[dull star]

in my experience, 8B was still smarter than pixart or ella, so I still think its still a win

but yeah I wonder how much smarter it could have been if it was T5 only

[sterile pendant]

Pixart is absolutely tiny in comparison though. Isn't the sigma 1024 version only like 0.8b?

[dull star]

now this has a finetuned model feel to it though

[dull star]

0.6B

its quite small

its still impressive for its size

and its better than cascade for example

[sterile pendant]

And look how fing powerful it is... I use it all the time for the prompt cohesion. So imagine if they had a much larger model...

[turbid grotto]

btw I learned how to deal with blurred background with sdxl, just sharpen image and put into upscale, maybe even 3d time, but needs a bit of color corection after

[dull star]

exactly

pixart is still the best model offline in my experience

for prompt adherence

[sterile pendant]

That's what sd3 could be if they went pure t5

But with actual image quality

[dull star]

Cascade:

pixart:

for a 0.6B model, that like 1 stage and way smaller than cascade, this is really nice

[sterile pendant]

Yep

[turbid grotto]

yea pixart is crazy

are they cooking something else?

[sterile pendant]

Such a powerful little experimental model

I hope so

Competition drives innovation

There's also a couple of those Chinese dit models I think as well

[dull star]

HunyuanDiT

[sterile pendant]

Huendit or something

[dull star]

haven't tried Hunyuan but I think it wasn't that good

[sterile pendant]

And some other one that starts with an L I think

[dull star]

Lumina-T2I

I think

haven't tried that at all

[sterile pendant]

Haven't tried either either

Just know they're out there

[dull star]

the skin detail is nice on 2B

the eyes look nice too

I remember it looking a little buggy in the Goku images

it was undertrained back then

[turbid grotto]

I liked it more than any sd1.5\sdxl portraits. Idk it looks more unique

hope finetunes won't get that 1.5 look

[dull star]

I wonder how much this "no refining, no upscaling" thing is true