🆕｜sd3

[lilac latch]

give me the prompt

[wet sparrow]

can confirm lol

[vestal cypress]

this is beautiful.

[patent acorn]

this feels very sd1.5 base

[jolly swan]

A better question - does SAI have any PR people left.

[lilac latch]

true

[vestal cypress]

when its not humans, the model does great

[coarse wedge]

Anyone getting good images without people in them at least? Since it seems to be useless at that.

[lilac latch]

SAI is dead.

[tropic aspen]

Also, we wouldn't be in this situation if you guys weren't acting like you have a superiority complex against anything that isn't the base model, and acting like it's other people's fault

[vestal cypress]

every image i have is a masterpiece.

[fleet meteor]

Cute 🤗

[vestal cypress]

that isn't a human lol

[subtle otter]

yep way better, thankjs

[raw wren]

that's out of date info. The current situation is people wanting to pay for the license but not getting a response after contacting stability ai

[vestal cypress]

thanks 😄

[hardy surge]

nice yoga

[fleet meteor]

almost

[mental bison]

Pony isn't a base model, pony is a finetune of the current SDXL, and they were capable of monetizing it. they won't be capable of doing that with SD3 if they can't get a deal. And this isn't just exclusive to Pony but many other finetunes.

[faint breach]

The person suggesting to them not to have public business dealings directly with a smut model creator might be part of their PR department

[silver sluice]

what's the prompt you're using for those?

[coral sable]

awkward moment when pony steps on pwsy

[finite hollow]

it's not really a prompt, it's a complex old workflow from sdxl

[abstract nymph]

he's grimacing in pain because he yoga'd his body into an amorphous pretzel heh

[keen falcon]

hands still got much issue

[dim geyser]

So they should get a deal, SAI needs to be able to monetize as well.

[compact forge]

just dont monetize it?

[subtle otter]

have you tried different cfg ?

[rain current]

noooooooo

[hardy surge]

remember, its not the model, its the prompt 😄

[mental bison]

That's what Astral is trying to do.

[coarse wedge]

Lmao

[tropic aspen]

You are quite literally running in circles through the same 5 points lol

[bitter hearth]

Weird it does fine on the api

[lament glacier]

Hi all

[bitter hearth]

That one might be too much

[compact forge]

thats hopeful atleast

[lament glacier]

How can I download and install SD 3

[dim geyser]

Believe me, I'm aware.

[dim geyser]

So best of luck to astral?

[mental bison]

💀 bro do you think it is cheap to finetune? Do you think the good finetunes of sdxl like juggernaut and pony are just charity?

[sick cedar]

[lavish osprey]

I can see tes fighting from miles away lol

[faint breach]

The API isn't using a basic default workflow with bad settings either

[patent acorn]

so what WENT wrong with 2B? dataset censorship ig

[ornate oar]

Do you use the full version of SD Medium or only with Clip?

[modern drift]

I think something is not right with the current comfy workflow or they uploaded the worng files. SDXL actually did good out of the box

[terse folio]

hey folks I wasnt around for the sd 1.5 or XL release so im curious with your impressions of this one vs those previous ones on release

[icy coral]

So it's a matter of waiting, even more hilarious

[ornate oar]

Full version > Clip

[storm saffron]

Maybe it should have been trained on 1 instead of 3 then?

[bitter hearth]

API, I just said, don't know what they are using there

[fleet meteor]

Bro think if its expensive just print money

[hardy surge]

too early to tell for me

[silver sluice]

i agree man SDXL medium is a bit underwhelming at the moment, doesn't feel next gen

[minor spindle]

WHy is SD3 so horrible what went wrong and how can they release something like this

[mental bison]

boyistrippin

[hardy surge]

its not exactly fast

[raw wren]

do we know that this is the reason why? We don't know the status of their licensing team right now

[silver sluice]

i think maybe we're so used to SDXL + loras that we forgot how just a bare base model is compared?

[static cairn]

dont worry this one just needs a few more months of training before its ready for local release! the girl isnt safe-shaped enough. hand me a scalpel...

[patent acorn]

finetuners will save sd3 in a month ig

[compact forge]

idk but i think realism engine is better than both of those and that one isnt payed 😅

[tropic aspen]

I'm pretty sure we all expected the base model to blow after all the "safety" bs that SAI talked about

[mental bison]

no way it's that bad

[bitter hearth]

[wild coral]

this one includes t5 right? Can Stable Swarm use it natively?

[lilac latch]

sdxl medium?

[rain beacon]

just like how it saved 2.0? nah

[mental bison]

Right

[tropic aspen]

If there's even motivation to fine tune because of the licensing

[limber stream]

So far SD3 is very nice, but @lavish osprey you guys are gonna need to remove the moronic 6000image limit from the creators license. Most likely the dumbest thing ive ever seen in my life. 😂

SD3 will fade away as fast as SC with that in place, it makes no sense, isnt enforceable.
Telling people how much compute they can use of thier own compute, come on now 😂

[lilac latch]

Censorship.

[hardy surge]

any upscale tests from anyone?

[raw wren]

at least leaving it open that its another issue gives them an off ramp for the issue

[faint breach]

they're only able to because of the relase in the first place. thanks obama.

[ornate oar]

Dudes and dudettes, you need to read the README and use the right model and text encoder.

If you're only using "withclip", the model's performance will degrade a lot.

Also, you need to prompt very differently. The model is very good!

[barren spindle]

its not working when I tried sd3_medium.safetensors with the tripleCLIPloader. I put the 3 text encoders in my clip directory

[silver sluice]

sorry i meant to say sd3 medium, getting mixed up lol

[fleet meteor]

That´s happen when you censor a model thanks to MFs that say they´ll sue you

[dim geyser]

Most finetunes aren't paid, although I do support finetuners getting paid.

[faint breach]

People are newbs they just want to be mad about sucking. and blame someone else for it

[wild coral]

People don't know how to prompt for it yet, so they think it's bad, it happens everytime there's a new model

[hardy surge]

any upscale workflows already working?

[patent acorn]

only if...

[mortal mesa]

get em Lykon do the for the 101st time thing

[finite hollow]

if you use the sd3_medium_incl_clip model @pesky then you can ignore the triple loader and take the clip directly from the loaded model

[real terrace]

I wanted to do the "It's alive!" scene but well need to think the prompt

[tropic aspen]

Only if what? SAI had better licensing?

[silver sluice]

to be fair I am trying to use some of the pony prompts with the score_ tags so that's probably messing it up too

[faint breach]

Waddup Axel. I hope your new movie isn't as bad as 3 was

[bitter hearth]

Show us the true girl lying on grass prompt

[viral plaza]

To be clear if you missed earlier conversation: it's a limit of 6k commercial images. So eg a generation-as-a-service company will hit the limit, but if you're generally for personal usage there's no limit, and if you're generating images while testing your commercial finetune or whatever, those images aren't commercial so the limit doesn't apply

[finite hollow]

@barren spindle if you use the sd3_medium_incl_clip model @pesky then you can ignore the triple loader and take the clip directly from the loaded model

[wheat schooner]

I really want to believe this

[viral plaza]

There's an internal discussion atm about making this more clear

[ornate oar]

Yes but then the model will perform much worse! Just FYI.

[barren spindle]

I know, but wasnt someone saying you get better results with the triple loader?

[patent acorn]

im not at expert at this but id say yes

[minor spindle]

Show me a woman lying in grass do it!

[marsh violet]

OMG guys look!
||cascade result||

[limber stream]

@viral plaza you gonna need to make that a lot more clear for sure, i had a feeling it was something like that

[errant dust]

Tell me, why am I getting different results when I use the separate CLIP files as opposed to the embedded ones? Is that normal? Just some weird RNG effect (yes, the seed is fixed and same)?

[hexed dirge]

at least it knows Batman

[finite hollow]

why is it worse @ornate oar and where do i get those other files from then ?

[patent acorn]

mutilated

[raw wren]

why is the 6k limit only for SD3? it seems like an afterthought

[abstract nymph]

at least it can do banks!

[barren spindle]

it seems the model that includes t5xxl (without triple loader) as well as trying to load t5xxl with the tripleCLIPloader breaks it for me

[wheat schooner]

Look at that lettering, thats impressive

[mental bison]

So if someone that makes a finetune and shares it with people across the world, will that commercial license apply to all of them or just the finetunner? Example if one of the users gets past 6k commercial used images, what happens?

[faint breach]

better than any base model we've had before

[viral plaza]

because it's new and generally they don't want to add new rules for old models

[icy coral]

SD 1.5 bros are out on suicide watch

[jolly swan]

I have about 120k$ of equipment around. Pony is expensive. I would like to burn only most of my money and not all.

[icy coral]

Talk about being salty

[slow cliff]

6k commercial images/month is a handful of power users. None of the other limits on the Creative License are relevant in light of 6k limit. If the goal of the Creators license is really "less than 1 million active users" https://stability.ai/license, you probably want an inference limit of about 1000 images/user a month (there'll be both left and right tails of the curve) so something like 1,000 images * 999,999 users = 999,999,000 images/month seems much more reasonable. Round it to a billion images a month.

[bitter hearth]

Lazy cat

[wild coral]

First thing I noticed is lower CFG

[raw wren]

ask cogvlm to describe a similar image and then prompt with what it says

[wild coral]

2 CFG seems to be best

[bitter hearth]

Small noodle soup

[faint breach]

/Be rich dad. Give daughter a corvette for her birthday. she wanted a pink corvette and he gave it to her withoout her bf watching and now she's on the ground throwing a tantrum.

vibes

[runic tusk]

I just moved them around a bit to condense my view

[noble coyote]

Once I get past HuggingFace Authentication (I can't!) I'll join all y'all 🙂

[viral plaza]

Each individual or business has a separate count for commercial usage.

If a model is commercial, and you use it, that's not commercial usage necessarily - that's just personal usage of a commercial model.
You're only commercially using images if you're selling the generation as a service, or including them in a professional marketing campaign, or things like that

[compact forge]

oh, it's 6k images created A MONTH, and i think those only count for commercially used images. so that excludes like 99.99% of people

[wheat schooner]

Have any devs addressed the criticism yet? Are people just trolling or is the model actually like that?

[opaque cargo]

almost with 2 cfg

[fiery dirge]

Having the same issue

[viral plaza]

ye

[opaque cargo]

nvm

[wild coral]

Second thing I noticed is more steps are needed, about 50 steps seems good

[atomic geyser]

https://cdn.discordapp.com/attachments/1021140136731349034/1250477122271707137/xnviewmp_2024-06-12_17-48-49.jpg?ex=666b1512&is=6669c392&hm=ae3d641c4431cb18b63d2dd2bf26bc2a24da554902686de9721dd365a43d5540&

[barren spindle]

no one has answers here

[bitter hearth]

6000 images a month per user?

[atomic geyser]

truly next gen

[hexed dirge]

There must be something wrong

[viral plaza]

people figured out that "woman lying in the grass" makes a mess and are generating 100s of copies of it

[abstract nymph]

I make 6k images in a day

smh

like, that's easily in a day

[viral plaza]

the model's not perfect, esp. at complex human poses. If you want perfect humans with complex poses wait for finetunes

[rugged nova]

woman in any pose not standing makes a mess

[gusty trail]

Did anyone reproduce this prompts?

[bitter hearth]

not just lying in the grass bro

[compact forge]

as long as you dont distribute them commercially you should be good

[fleet garnet]

So I just installed swarm, and put in the SD3 model. Any tips on what settings to use? (steps/cfg/sampler...)

[abstract nymph]

but ye, it's not a very useful commercial license imo

[wide pagoda]

if you let stableswarm set up your workflow it just works https://old.reddit.com/r/StableDiffusion/comments/1de65iz/how_to_run_sd3medium_locally_right_now/

[faint breach]

Devs don't need to address most of the criticism with the way it's being delivered. Once someboyd writes a mature and honest report about their findings, then tha'ts worth responding too. Knee jerk comments like "The model is shit!" don't need dev responses

[raw wren]

they have no control over the images themselves, so I though it was the act of using the model commercially to make those images

[bitter hearth]

[abstract nymph]

I guess it would exclusively be for personal use like doing commissions or something

[marsh quail]

xDD

[viral plaza]

steps 28 cfg 5 sampler euler scheduler normal, as per https://new.reddit.com/r/StableDiffusion/comments/1de65iz/how_to_run_sd3medium_locally_right_now/

[patent acorn]

butt naked real?

[mental bison]

can someone use this prompt?

A man with a mustache and a monocle, upside down view

[ornate oar]

Why is it worse?

Because SD3 has a new text encoder called T5, but also the old XL encoder called Clip or Clip Vision.

If you don't use T5, you'll get worse results.

All the files are in the repo.

[bitter hearth]

you can identify something like a butt in there ?

So... Cute...

[compact forge]

makes sense. sdxl was the same way

[frozen lynx]

guys u using negative prompt?

[viral plaza]

people complain when we hold the model back to train more, people complain when we release the first version of it, there's no winning people will always complain

[marsh quail]

Dude everyone is using T5 here...

[wet sparrow]

sit pose works decent too ngl

[patent acorn]

but it doesnt have the curves does it count

[bitter hearth]

You killed Shrek

i cant

[ornate oar]

I just reacted to comments saying "just use withclips, then you don't need triplecliploader".

[patent acorn]

who invited man

[abstract nymph]

SD3 can't actually do cats.,

a cat laying in the grass on its back laughing

[subtle otter]

you cdan see it was train on terrible stock photo, it has that studio lighting even on exterior pics

[lilac latch]

Nobody will even utilize SD3 8B. It'll take more than 24GB of VRAM easily.

[faint breach]

People also don't understand that a base model isn't like a new version of dreamshaper or protovision or whatever. It's an entirely new base architecture with entirely new base latents

[bitter hearth]

performing has the association

[runic tusk]

The model is very capable, but just like every other model released, people want it to be perfect every time for every prompt right out of the gate.

That's not reality.

[subtle otter]

this is gonna take some much time to retrain and unbias the model

i'm desperate

[bitter hearth]

want me to compare the prompt to SD 1.5

[marsh quail]

Bro, try to train a person on shit like this.
Thats a problem for finetuning too

[young blade]

there's always MJ they can move to, complaining for something free that cost the creators millions to make is kinda crazy

[ornate oar]

And it's much better than XL. SD3 finetunes will be banging. 🙂

[dusk silo]

should we be using normal scheduler? if i use exponential or like euler_ancestral it starts looking like silicon

[mental bison]

worst case scenario it will take months for a good finetune

[faint breach]

I think it'll be fine. SDXL base is the same with complex poses, but people train fine on it

[runic tusk]

For what point? I could easily find seeds with 1.5 that would have similar issues rendering the prompt.

[subtle otter]

it seems so...

[sterile oasis]

Is it free to use here?

[faint breach]

yup. but that's beside the point.

[subtle otter]

any way to know the dataset ?

[lavish osprey]

without tes it's 16gb

[barren spindle]

anything wrong with this? its throwing error

[hexed dirge]

why not Dreamshaper?

[faint breach]

The finetuners will be fine. Astralite is being weird about not wanting to pay anything ever

[raw wren]

Or Dalle 3 via bing chat or coze

[lilac latch]

I get horrible results without T5.

[wheat schooner]

You don't need tes?

[vestal cypress]

apple silicon people pulling up like

[faint breach]

dreamshaper author already finetuned sd3 2b

[silver sluice]

i'm not sure but if i were to guess id say the file didn't download properly, got corrupted, interrupted, etc

[icy coral]

Morons did the same thing when sdxl came out, only to advocate in favour of sdxl when sd3 came out. Rinse and repeat. They never learn.

[hexed dirge]

Yes I know

[rugged nova]

Is it true that stability is not allowing finetuners commerical licenses if they have nsfw in thier finetunes?

[mental bison]

no idea, all i know that big finetunes are in a delema with the comercial license

[atomic geyser]

oh wow its horrible

[fleet garnet]

Gueess I'll need to wait a bit for custom models, not quite impressed with it so far. But I understand it is a base now

[subtle otter]

fuck licenses

[tropic aspen]

Because the fine tunes saved SDXL

[velvet ruin]

It's new, people just like to complain.

[barren spindle]

already redownloaded them a second time already

[hexed dirge]

mmmmmm I don't think so

[faint breach]

why shouldn't stability earn money?

[dull star]

could you please show comparisons or pics, I actually forgot to test this lmao

[bitter hearth]

Dangerous hunter

[subtle otter]

peopel taking merit for things they have nothign to do with

[mental bison]

isn't it making profit off of it? i saw that astralite would gladly buy the enterprise sub

[raw wren]

we have no idea, it could be due to the layoffs making it difficult for the licensing team to do their work or other internal company issues

[silver sluice]

try running the update script and see if you get any errors

[faint breach]

base model is the core of all finetunes. Stability is responsible for all of them being made.

[lilac latch]

Sure. Give me a second.

[neon wagon]

quality and weird glitches are even more worse

in sd3

[viscid pivot]

Lmao, how can you call a model "revolutionary" when it is incapable of even creating a normal image of a person? How can you slaughter you training dataset to a point, where people become eldritch horrors, go online and say "Oh we created a revolutionary model, its so goooooood" and then release a complete mess?
its even worse than sd2

[wide pagoda]

any particular reason we've gone back to euler as the standard sampler, after DPM++ 2M was preferred for SDXL?

[normal raptor]

so how did Lykon get such good images?

[subtle otter]

they should have private partenarias to make it more rentable for them and release some model to the public retaher than disapearing liek theyr did

[faint breach]

A base takes millions of dollars to train and they give it away. License conditions for commercial use isn't a bad thing. The entitlement in this room is palpable

[mental bison]

my dude... all fine tunes are made from the base models that stability makes

[lavish osprey]

you do.

[raw wren]

He does want to pay though? I though that we've already been over this. The ball is in Stability's court right now

[barren spindle]

trying redownload again, then I'll try that

[dusk silo]

i get either:

[marsh quail]

Lykon, whats going on here xDDD

[subtle otter]

yeah i know and getting sued doesn't sound like luch fun

[limber stream]

@viral plaza what is the stance regarding, lets say, an artist using sd3. Would an image they post on a service like patreon or wat not, count to the 6000. Or are we literally talking about if i guess, its an automated process pumping out images as service to users ?

[silver sluice]

i mean it's a decent model

[bitter hearth]

SD3 is arguably better than 1.5

[icy coral]

And fine tunes will save sd3 as well, as they did 1.5, which was hilariously bad at release

[faint breach]

well maybe you can start a company with your own capital and do business that way.

They're charging for licenses is their business model, and that's fine

[mental bison]

i agree

[compact forge]

how much does that pony finetuner even make to not count in the creator license? 🤣

[silver sluice]

you could try updating while it's redownloading

[craggy ridge]

Haters gonna hate

[neon wagon]

the qulatiy is lower as sdxl and 2.1 vanilla

[tropic aspen]

That's if people will be motivated to fine tune off of it considering the licensing

[bitter hearth]

Kek

[faint breach]

He likely is making far more than the license would cost him

[viscid pivot]

Hate? Brother, just go and generate an image of a woman sitting on grass

[modern drift]

I can't even recreate the images from the paper, can someone try these " translucent pig, inside is a smaller pig."
"A crab made of cheese on a plate"
"A swampogre with a pearl
earring by Johannes Vermeer"

[raw wren]

the issue is with the 6k image a month limit, not the profits

[dull star]

bro that image LMAO

[verbal epoch]

If I want to upscale do I need to change the tile width and height also?

[neon wagon]

yea

[dusk silo]

my settings are messed up

[viscid pivot]

SD3 is decent at generating stylistic or anything that does not contain women

[lavish osprey]

paper was 8b

[viscid pivot]

at this point, you could even say that SD3 is freaking sexist as hell

[subtle otter]

it a shame we can't have stake in it or something else, now we indeed to have to launch a new company

[lavish osprey]

8b has far more knowledge, trained for more time on more data

[mental bison]

I wonder if sd3 will kill sd1.5, if sd3 gets further optimized and made to run easily on 4Vram i think it could kill sd1.5 because most users only have 4vram (sd1.5 users)

[bitter hearth]

gotta decide which one is funnier tho

[short arch]

if I downloaded the t5 model, how do I use the t5 stuff, or is it automatically done?

[minor spindle]

how did you get this

[compact forge]

ALL the images generated with his model does not count towards his own 6k

[barren spindle]

"SD3 CAME OUT!! WHAT ARE YOU GUYS MAKING?!"

".....hundreds of images of women in grass"

[icy coral]

It's been out for 2 hours lol. Communication will be sorted, they want to make money too

[faint breach]

I see no problem with the peopel spending a ton of money to make these base models, setting reasonable terms for their commercial use

[neon wagon]

its extreamly good in follow the promt right but fucks over legs,arms,hands,hight,eyes,distant faces,light and patterns

[sterile oasis]

How to create?

[lavish osprey]

you can probably run quantized versions of sd3 on much lower vram, and use only 1 te

[raw wren]

Everyone is just going around in circles here. Something gets debunked and then 10 minutes later a new batch of people are back to the same arguments

[tropic aspen]

We were told the licensing would be dealt with before the launch.

[sterile pendant]

You still use it with sd3, just make sure to pair it with sgm_uniform

[silver sluice]

that kind of looks like a crab made of cheese and crab

[dusk silo]

medium with text is running 30gb on my machine

[modern drift]

Damn, so 2B was not all you need after all...

[mental bison]

wait is the current sd3 fp32?

[opaque cargo]

"image of a girl sitting on the grass laughing, the image is comfy and there is a subtle light coming from the trees around her" 2cfg 50 steps

[icy coral]

Bummer

[faint breach]

soon ™️

[coral sable]

can't complain, it's 5 fingers

[rugged nova]

All these issues can be fixed by finetuners. The real issue in on whenever or not stability is denying fine tuner's enterprise licenses due to nsfw or not.

[barren spindle]

bout 2 hours

[craggy ridge]

https://tenor.com/view/bruno-trolling-encanto-trolling-has-officially-begun-gif-24829029

[viscid pivot]

Its a gain the argument "SD3 8b is far better than 2b", yea so WHY do you release the garbage 2b version xD

[finite hollow]

is this the wrong file to use ?

for the triple Clip loader ?

[golden cliff]

is the 2b model the same on being used in the artisan? I couldnt even recreate those

[viral plaza]

I think an image posted directly to patreon would count yes (but, yknow, if you individually post 6000 images to patreon you're insane)

[raw wren]

I don't think the license specifies its 6k per individual. Its 6k for the entire operation based on the wording

[barren spindle]

looks like a torch version issue??

[onyx sphinx]

update python dependences with the bat file

[fleet meteor]

Yep we need community models, the potato and the french fries supposed to be "fighting"

[subtle otter]

i meann not all compannies are willing to share their model, that's my main concern, i am fortunate enought to have ressource to somewhat train a model, but i know fort sure i won'ty publish the latest version open source

[finite hollow]

how do i do that @onyx sphinx ?

[bitter hearth]

[short arch]

Does T5 automatically happen in any prompt I write, or do I need to use a specific node?

[static cairn]

8b needs more time to safe-ify the anatomy. please understand

[sterile oasis]

"a big chair"

[silver sluice]

lol i played with that one yesterday

[desert garnet]

nice hands bb

[viral plaza]

if he's selling image gen as a service, yes it's 6k for the total service.

[onyx sphinx]

file update_comfyui_and_python_dependencies on the update folder

[silver sluice]

a crab made of cheese?

[dry wave]

if you use the trippleClip node then it's in

[viral plaza]

if a finetune is released, anyone who uses it / buys it / whatever does not affect the 6k

[compact forge]

that interpretation makes absolutely no sense whatsoever in the context of how these loras and checkpoints are distributed.

[mild bramble]

I am ready to pay anyone for a custom checkpoint on sd3 that is solely based on realism, hit my dms

[short arch]

I just downloaded the 10gb model

[mental bison]

I need someone to get their ass on quantizing sd3 (and to make a1111 webui support sd3)

[viral plaza]

the images from the paper were on SD3-8B, not the current 2B release

[patent acorn]

5 legs

[viral plaza]

the 8B is better at complex understanding

[limber stream]

@viral plaza for sure. Im just more concerned for the future in some sense. What would happen if lets say AD trained some stuff for SD3, and we ended up having people creating single animations with like 1000'frames' ? etc

[fleet meteor]

Are you using the t5 model? I couldn´t recreate it withouth it

[mental bison]

will 8B come out in august or?

[mortal mica]

Looking at these images sd3 seems to be. Ummm. Lacking in areas. Lol. Hopefully fine tunes will fix it? If people decide to finetune it.

[barren spindle]

still same error after updating and redownload

[silver sluice]

ther'es no smaller pig inside but that's pretty decent

[onyx sphinx]

yes

[viscid pivot]

yes, it hasn't gotten the h.p lovecraft treatment yet

[onyx sphinx]

yep

[viral plaza]

I don't know exactly how video would play into it. I would imagine the appropriate intent-interpretation is it counts once per video rather than once per frame

[subtle otter]

is stehre a realease date for the 8b ?

[viral plaza]

don't have a planned date yet

[subtle otter]

how much more ram is it gonna take ?

[lilac latch]

First one is T5. Second image uses CLIP only.

[ornate oar]

Make sure the files actually exist.

[modern river]

so when is sd3 horse molester edition?

[silver sluice]

im sure with better prompting?

[keen palm]

Hopefully the threadmills are included when 8B is released.

[mental bison]

well the current one is a bit easier to run than sdxl

[radiant ledge]

how large are your text encoder files?

[lavish osprey]

you can already test it on apis

[marsh violet]

Now make him lay down in the grass

[ornate oar]

That ? looks suspicious. Are they symlinked to the wrong target?

[limber stream]

@viral plaza Ok makes sense, sweet. But ya, it would prob be good idea to expand a bit on that 6000images part of the agreement. Just so it doesnt like, scare off lets say AD creators from giving SD3 a go in the first place and what not

[wild coral]

it looks like SD3 can gen higher that 1024x1024 without weird duplicates though

[crude yarrow]

Album art material right there.

[subtle otter]

with 2 pass it still take 12gb, it's alright but a tiny bit more than SDXL for me

[coral sable]

What's number of new SD model?

[ornate oar]

It is less restrictive in terms of resolution!

[barren spindle]

copied and pasted them from downloads into the directory

[lavish osprey]

not without weird noises around the edge

[modern drift]

Understood, I was expecting even the 2B to outperform sdxl though with new captioning and facny new architecture, even a little bit of prompt adherence would have been good.

[marsh quail]

Ehm yes, cause it creates weird duplicates even on 512x512 ???

[sterile oasis]

#1237459938901491852 a person

[lunar rivet]

cascade could do different resolutions better tbh

[dusky thistle]

heunpp there

[dull star]

you mean it's dumber? it doesn't look so worse

[wild coral]

at extreme ranges sure, but 1536x1536 or 1152x1536 work great

[dusky thistle]

using custom fractal initial starting latent

[bitter hearth]

some prompts it just refuses to follow

[lavish osprey]

it definitely does lol.

[lilac latch]

I actually cherrypicked the second one (2 tries). The first picture is more detailed.

[subtle otter]

damm ok cool thenat least one good news ^^

[craggy ridge]

https://tenor.com/view/whatisgoingon-ohgod-crazy-gif-7867392

[subtle otter]

do you happen to know the base training resolution ?

[modern drift]

use the text encoders separately and use comfy ui. I am on 6GB vram and can generate an image in 32 seconds on a 3060

[dreamy sundial]

trash model

[worthy trout]

i guess it ignored the white part because samurais aren't usually white

[viral plaza]

it does outperform SDXL on most things (not all). Especially on general image quality and complex scene understanding SD3-Medium is way better than XL. When you get into specifics like human anatomy... yeah wait for SD3 finetunes

[subtle otter]

saying thios cause some ratios seems to have weird artefacts

[desert garnet]

model respects proper race

[dreamy sundial]

im so dissappointed it looks so bad

[viscid pivot]

Im really trying to compose and not be too rough, im glad we are getting new A.I. But why are your mods and employees always saying contradicting things. Like for example why do they say that 8b is way better than 2b, and yet you release 2b as revolutionary? That does not make sense at all

[viral plaza]

(you can get awesome looking humans out of SD3-Medium if you don't ask it to perfectly pose the entire body, eg a portrait shout usually comes out great)

[bitter hearth]

thats uh... racist

[lavish osprey]

I think there is also a "learn to use it" factor.

[wild coral]

here's a native 1536x1536 gen

[ornate oar]

From what I noticed SD3 has very precise prompt following, this is good and bad, because you need to specify what you want in lots of details.

Example: "abstract background" in SDXL would result in pretty backgrounds.

With SD3, it's literally random noise until you specify more.

[abstract nymph]

I'd; gladly take score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4 , which was obviously an expensive mistake since they don't have a hundred million dollars of investments they aren't planning on paying back to re-do it, over this:

[mental bison]

for 2b? 1024x

[wide pagoda]

anyone manage to run SD3 with only clip-l or clip-g yet?

[mental bison]

i have no comfyui knowledge so i won't bother

[subtle otter]

it is different for the 8b ?

[faint breach]

Seems more like people intentionally misunderstanding things then contradicting messages

[bitter hearth]

Does (stuff:69999) work?

[tacit kraken]

"a blonde woman getting a massage, oiled skin, fuji 400h, shallow depth of field, chiaroscuro, moody, gorgeous"

[dreamy sundial]

no way they really thought this was prepared for release

[desert fjord]

If someone wanna use SD3 Medium in hf spaces, https://huggingface.co/spaces/Nick088/Stable-Diffusion-3-Medium

[craggy ridge]

Anybody scored bingo yet #🆕｜sd3 message ?

[bitter hearth]

The weight

[lavish osprey]

we didn't spend "a hundred million dollars of investments" on 2b by the way lol

[mental bison]

pretty sure it's the same but just more pictures

[desert garnet]

nice cat

[abstract nymph]

pretty much anything other than standing or sitting tbh 😔 some synthetic data may have helped here

[bitter hearth]

no

[modern drift]

I am not able to get such results, even compared to sdxl 1.0. Can you double check if you uploaded the right checkpoints lol or if the comfy workflow on github is correct?

[bitter hearth]

apparently not

[barren spindle]

this clip loader error is going to drive me crazy

[viral plaza]

The 2B we're releasing is better quality than the 8B we have, and the existing SDv1/SDv2/SDXL models.
8B once it's fully trained will be the awesome SOTA lead model, but it's still in training.
SD3 in general is a revolutionary awesome new architecture

[worthy trout]

here's the thing, i think you might have it configured wrong because i've seen other people's generations with sd3 where it looks just fine

[subtle otter]

ok thanks fopr the innfo

[sterile oasis]

how can I create here?

[pseudo isle]

https://www.youtube.com/watch?v=xag3NU0Ckuk&ab_channel=MarkuryAI

https://www.youtube.com/watch?v=xag3NU0Ckuk&ab_channel=MarkuryAI

[compact forge]

but wasnt that the advertised improvements? "realism in hands and faces, is achieved through innovations such as the 16-channel VAE" or is that only foreshadowing the finetunes

[pseudo isle]

https://www.youtube.com/watch?v=xag3NU0Ckuk&ab_channel=MarkuryAI

https://www.youtube.com/watch?v=xag3NU0Ckuk&ab_channel=MarkuryAI

[abstract nymph]

fair enough, it was a gross exaggeration (8

[barren spindle]

ban for spam?

[eager wraith]

Things are popping off!

[faint breach]

whats the point of score_9 on every image prompt though? that's good for a refine for people that want one specific macro aesthetic, not a base model

[viral plaza]

just G or just L technically works but is silly to do vs G+L

[bitter hearth]

cant make unsafe content if you cant make content

[dreamy sundial]

why does it look so bad

[lavish osprey]

git gud?

[faint breach]

skill issue

[dreamy sundial]

stop lying

[wide pagoda]

In the long run, I expect "LLM rewrites the prompt by default, but also provide the option to send a raw prompt" to end up the standard for AI prompts
it's what udio has

[civic quail]

@pseudo isle square

[dull star]

Nice, welcome back

[abstract nymph]

nah it's to essentially filter out the low quality art but there was a bug in how it was trained iirc

but it also lets you still use it to generate a huge range of styles without censoring out that low quality art

[dreamy sundial]

?

[faint breach]

layer 8

[finite hollow]

any help on this error ?

[wide pagoda]

but is this possible with the released nodes?

[remote holly]

Sd3 hugging spaces are avaible , chek the sd3 model card

[pseudo isle]

im

[viscid pivot]

I agree, that prompt following is a lot better, but anatomy is bad. Saying "learn to use factor" in a model that supposedly should be easy to use and generate great images using real language prompts, should do that without having to prompt engineer like crazy. That is training issue and most likely cenorship beyond anything ever done before

[dreamy sundial]

for them tbh

[wild coral]

wait a minute... why in stable swarm when I press import from generate tab are the clip models labeled SDXL?

[ornate oar]

I think so, too. Many SD wrappers have "prompt magic"!

[faint breach]

DK rap

[subtle otter]

quality apart and let's suppoose we have a good finetune model inn the next month or so ' best case scenario) what about the controlnnet, thta is still lackinng for SDXL

[viral plaza]

the 16-channel VAE thing is referring to small details, eg distant faces in groups in an image, or the eyes, or etc. where XL and v1 previously kinda just utterly failed until you upscale the image a bit to make those details large enough to process well

[viral plaza]

wot

[onyx sphinx]

Thanks for SD3 StabilityAI devs, awesome work again ❤️

[ornate oar]

Also, SD3 has much less AI artifacts.

[neon wagon]

can someone test the same seed and promt but with different shedulers,steps and so on ????

[barren spindle]

because the damn text encoders in the sd3 release arent working for lots of people

[hallow talon]

Having the same issue

[desert fjord]

Hello dark

[abstract nymph]

they likely filtered out most images of humans until they couldn't get sued for being able to generate illegal content from the base model

[dull star]

hello

[barren spindle]

no u

[viscid pivot]

nudity isn't illegal

[craggy ridge]

[lilac latch]

Very horrible anatomy.

[viral plaza]

try with SingleClipLoader. I don't know offhand if comfy added support for that because it's very silly to do

[faint breach]

well that's made up

[frozen lynx]

using different Scheduler gives very different results

[neon wagon]

i dont have access @barren spindle

[finite hollow]

i still get this error and the update_comfyui_and_python_dependencies.bat didnt go through with the above mentioned error

[turbid grotto]

pose problems due to lack of training data?

[neon wagon]

site is down

[viral plaza]

i think you replied to the wrong message

[faint breach]

THE WEBSITE IS DOWN https://www.youtube.com/watch?v=uRGljemfwUE

[viscid pivot]

maybe, there are so many messages here lol

[viral plaza]

SD3's textencoders are literally copy/paste the SDXL one and add an optional T5-XXL

[neon wagon]

i am at work

[barren spindle]

and im still getting after downloading redownloading them 4 times to ensure its not a corruption issue, and running update all in comfy manager, and checked for update of comfyui

[viral plaza]

if you have SDXL models with finetuned tencs you can try them on SD3

[modern drift]

wait im confused, better in quality how? like lighting etc or prompt adherance? If so why are api and paper results of 8b so much better ?

[fleet meteor]

😭

[dreamy sundial]

sdxl better tbh

[silver sluice]

i mean you might as well just trash your whole setup, just create a new folder, donwload a fresh copy of portable comfy and just use that, then bring whatever it is you had over to the new folder to ensure a fresh working install

[viscid pivot]

But again, if you claim to be on paar with midjourney or dall-e that also includes usability and they are just a lot better at creating poses and anatomy with simple prompts

[rain current]

My SD3 is broken....

[faint breach]

will the sd15 clip layer be like quake 2 code in future AI models? 20 years down the road and it'll still be in the newest generative models

[dreamy sundial]

yeah its not looking good for stability.ai

[desert garnet]

SD 1.4 was better and hotter

[exotic schooner]

You can use both you know

[dreamy sundial]

its just trash

[faint breach]

skill issue

[finite hollow]

@silver sluice could you give me a link of where to download the fresh comfyui best ?

[dreamy sundial]

no

[lavish osprey]

https://x.com/kentelTICE/status/1800921581247709391

🅼🅰🆁🅲 🆀🆄🅴🅽🆃🅴🅻 🍓 (@kentelTICE) on X

@Lykon4072 And it works with all ratios !!!

Twitter · Jun 12, 2024, 4:02 PM

[silver sluice]

meanwhile dalle3 gets it on it's first try

[bitter hearth]

Big noodle soup

[wild coral]

so by default it's not using the T5 in stable swarm?

[viscid pivot]

nah, thats like 18/20 images trying to generate this prompt

[faint breach]

Guy has MJ as a profile pic but nobody wanna be like him

[dreamy sundial]

skill issue for stability 😂

[barren spindle]

bump, anyone know what to do for this? I've redownloaded the files many times to rule out corruption, and checked for updates for comfyui and all nodes

[silver sluice]

SD3 is still confused about what a cheese-shaped crab looks like

[viral plaza]

quality refers to, like, the visual details n stuff generally.

SD3 in general has much better prompt following than XL does.
8B is better at prompt following than 2B.
2B is better at visual details (like the stuff you see when you zoom in and look close) than 8B is currently (until 8B training gets further along)

[lavish osprey]

my day is ruined, let's discontinue the model.

[silver sluice]

https://github.com/YanWenKun/ComfyUI-Windows-Portable/releases

[viral plaza]

seemingly yes they keep shoving L into everything aaa

[silver sluice]

lol just help me out, what am i doing wrong? does SD3 need elaborate prompting to be good?

[dreamy sundial]

its just not well made

[modern drift]

Ah alright

[barren spindle]

what is the min torch version required for sd3?

[silver sluice]

join me brother, show me a good SD3 prompt that will make me a cheese-shaped crab im obsessed about it now

[lavish osprey]

try "cheese sculpture shaped like a crab", you keep asking the model to make a crab, it will make a crab lol.

[rain current]

No, I'm creating different prompts.... "A woman with brown hair is lying amidst a lush green field of tall grass. She is wearing an orange dress with white polka dots and is looking upwards, possibly lost in thought. The field stretches out as far as the eye can see, with the woman being the central point of interest"

[viral plaza]

yes, edit this param to turn T5 on or off:

[desert garnet]

shes literally me

[wild coral]

thank you!

[viral plaza]

i'll probably swap it to default CLIP+T5 if you have enough RAM

[fleet meteor]

Nah sd3 is good, but we HAVE to create and use a community finetuns, its clear that the censorship of the model affect the generated persons (the same thing happend in 2.0) and the dataset used was also limited, it cant generate images of famous games or styles, but with the images that it can generate it looks good and it have good lighting and textures

[viscid pivot]

yea i mean "lying down"

[abstract nymph]

obviously, but CSAM and deepfakes are :)

[viral plaza]

uhhh what is that

[silver sluice]

yeah good point I kept said "a crab made out of cheese" but then I changed it to a "cheese shaped like a crab" but its still doing the same thing, ill try the sculpture bit and see if that helps SD3 understand intent

[viscid pivot]

So, thats nothing you can sue Stability for

[viral plaza]

that is not comfy official download

[abstract nymph]

🤣

[barren spindle]

bump, anyone know the fix? what is the min torch version required?

[wild coral]

welp :/ I'll try restarting

[faint breach]

you can sue people for anything and a judge determines if it has merit or not

[short arch]

If I downloaded the 10gb model what do I need to do to activate the text encoder?

[silver sluice]

yeah it's not it's like a preloaded version it's what i use and it works well, sorry if i should only link to the official release one

[viscid pivot]

Yea but it does not have merit, because you can't create deepfakes out of the box. Better just stop training on celebrities instead of nsfw

[modern drift]

better take it down, we had the recent malware distribution because of custom nodes, better be safe

[viral plaza]

Server -> Update and Restart

Your copy of swarm is a day or two outdated

[wind talon]

what are your guy's opinions on sd3 so far?

[fiery dirge]

Thats crazy detailed

[viscid pivot]

and once you use photoshop to create the deepfake, it has 0 merit in any form or way

[dreamy sundial]

it is NOT good

[craggy ridge]

but it is

[hexed dirge]

and you can say after 3 hours.......

[barren spindle]

base models generally arent great compared to the community finetunes

[viscid pivot]

SD3 architecture is good, training is bad

[fleet meteor]

For certain images it produces good results, but the trained data IS COMPLETE GARBAGE

[dreamy sundial]

how is it good, people are all deformed it looks unusable

[dusk silo]

orange flavor

[desert garnet]

its dead we lost time to pack it up and play with Elon model

[fleet meteor]

The problem is the training data

[silver sluice]

apples to apples dalle3 converted my prompt:

generate for me an image of a cheese shaped like a crab on a plate

to:

A piece of cheese sculpted to look like a crab, placed on a plate. The cheese has intricate details mimicking the crab's legs, claws, and shell. The plate is white and simple, with the cheese-crab placed in the center. The background is plain to keep the focus on the cheese-crab sculpture.

SD3 on the left, Dalle3 on the right

[dreamy sundial]

yes?

[errant dust]

You can also specify to use it alone

[wind talon]

i think the saving grace of sd3 will be how easy it is to finetune so far

[barren spindle]

what is the min torch version required by SD3?!

[faint breach]

reddit c*** box vibes

[dreamy sundial]

is it 2b?

[silver sluice]

@lavish osprey indeed sculpted was the key to getting it to look right

[weary snow]

Comparison between different clip models ? Is clip only bad ? How much ram should one need for clip+T5 fp8 ?

[silver sluice]

yeah running locally

[hexed dirge]

no. Did you see SD 1.5 standard model?

[fallen whale]

Are we sure that the model is implemented properly on local repos? The results from the api vs local are so different.

[dreamy sundial]

yes its trash by todays standards

[solar jasper]

If anyone gets a license for it..

[wind talon]

pixart will save us

[unkempt grove]

damn, drop the settings+prompt

[desert garnet]

shes showing her arms,ban

[bitter hearth]

Pix who

[faint breach]

i think the saving grace is that it's free and it's weights. good enough for me. The new VAE is pretty saved too.

[bitter hearth]

[wind talon]

pixart sigma

[hexed dirge]

SD 1.5 weights were one thing. Finetuned models another one. This will be the same with SD3

[errant dust]

@viral plaza What is the difference between the fp8 text encoder and the fp16? I get quite differrent results, which is why I ask

[dreamy sundial]

but it cant even generate people doing actions, its so censored

[fleet meteor]

That is the thing that i dont understand, why people have to contact SAI by email instead of buying the license like its another random software?

[upper snow]

Less precision, the text encoder takes up less space in memory but is more inaccurate

[solar jasper]

We need chinese models to become better kekw. SAI cant keep destroying our hopes.

[craggy ridge]

https://tenor.com/view/stop-fighting-hose-parenting-hose-off-gif-12003125

[dusk silo]

i love it such fun

[odd kernel]

Awfully slow I stick to Turbo for now.

[lilac latch]

the hand on top of the head looks like ps2 low poly, lol

[faint breach]

Important to note that Pixart has commercial restrictions too. Theres a limit on how many end users can use the model through your service. So a place like Civit has to license it to tencent

[dusk silo]

i love this

i love life

[faint breach]

if Civit doesn't want to pay tencent for licensing, Pixart may never take off

[errant dust]

That is actually matching my results, notably with text.

[dusk silo]

people complaining about sd3 just havent gotten it to work right

[desert garnet]

im confused rn,where is the girl????

[wind talon]

pony guy has shown interest in experiementing with a pixart finetune, so perhaps one day..

[craggy ridge]

mmmm corn

[dreamy sundial]

really baf

[lavish osprey]

looks like it worked 😉

[compact forge]

idk i was supposed to make a gigachad

[lavish osprey]

git gud

[fleet meteor]

True

[eager wraith]

Can SD3 make me a pizza with horrific toppings?

[desert garnet]

dead model SAD!

[dry wave]

because this is how it'salways done in business 🤦‍♂️

[viscid pivot]

I created a dataset with different angles, left and right hand poses and so on, around 1500 images captioned with gpt4 vision and curated. Once training is added to kohya or onetrainer, i will try out how good poses can be improved

[faint breach]

its a controlnet hidden image you have to squint

[dreamy sundial]

different model 😂

[shrewd cairn]

So, whatcha guys think so far?

[fleet meteor]

Then why some people CAN´T get a license?

it makes no sense

[solar jasper]

Pony guy, the goat, or shall i say, the horse kek

[rugged nova]

Sitting is too risque huh?

[lavish osprey]

no

[viscid pivot]

git gud? its still deformed. Look at the leg

[lavish osprey]

here another

[weary snow]

Difference between clip only and clip+T5 fp8

[odd kernel]

18:14:26.383 [Info] Generated an image in 398.86 (prep) and 1197.11 (gen) seconds

[hallow lion]

so its out?

[viscid pivot]

also bad

[dreamy sundial]

it looks so different

[shrewd cairn]

I'm reasonably enjoying myself. Oversized apples...

[bitter hearth]

No 2 hours

[faint breach]

this is what you call gaslighting. somebody accomplishes something other people can't, so they tear it down instead.

[hallow lion]

pls dont do a cascade on this and wait for 8B and not train this and make tools for it

[silver sluice]

dalle3 does the same thing as SD3 where prompting it this way makes a crab shell made of cheese flesh

[bitter hearth]

[solar jasper]

Is this supposed to be impressive?

[lavish osprey]

I mean, it's definitely better than the shit some of you are posting 😄

[lilac latch]

First image is T5 only. The second one has only CLIP loaded.

[desert garnet]

yes,my soul is out of my body as i keep getting error on comfy

[eager wraith]

banana for scale?

[hallow lion]

ouch

[verbal epoch]

I found a pretty sick upscaler

[faint breach]

Can't. My roots go too deep.

[dry wave]

again, this is how it works in business. You contact the company and make a deal. I assume it's not you who try to buy a licence but someone else. I wouldn't care so much about information from third sources

[dreamy sundial]

looks the same as sdxl

[lavish osprey]

no, but looks nowhere as horrific as your gend. Git gud.

[shrewd cairn]

How many of us actually slept last night? lol

[viscid pivot]

Yea, we are posting with a model from you guys.. using simple prompts that you guys said could create great results

lmao

[opaque sedge]

She has three legs and one of the arms is really weirdly bent.

[subtle otter]

better prompt adherenhce my ass, it was : cinematic, interior design of a modern living room, designer furniture, polished granite stone, marble annd golden accent details, maculine design, trendinng on AD

[dreamy sundial]

cant even do words properly lmaoooo

[faint breach]

skill issue

[shrewd cairn]

What.....?

[dreamy sundial]

stfu

[shrewd cairn]

lol

[lavish osprey]

unskilled feedback is not really useful. Learn to use the model first.

[desert garnet]

thicc

[solar jasper]

Brother i can feel your hands shaking behind the screen...

put the computer down..

[subtle otter]

stop coping with the skill issue it's a terribhle model and you know it

[lavish osprey]

I keep getting decent ones all the time

[fiery dirge]

how long until finetune scripts get made

[viscid pivot]

Sure, why is midjourney and Dall-e capable with the same prompt? 😄

[faint breach]

maculine design!

[wind talon]

pixart seems to have no issues with this prompt

[glossy quiver]

some people better get ready to apply for unemployment benefits lmao

[odd kernel]

Lots of work for the community I guess.

[hallow lion]

these dont look to bad

refine a bit with sdxl for now and have a ncie day

😄

[dreamy sundial]

they do, and a lot

[weary snow]

Which clip model are you using ?

[subtle otter]

yeah ever heard of stark ?

[finite hollow]

i installed the new comfyui portable, but it doesnt have all those nodes, and it doesnt want to install the with the install missing nodes in comfy manager either, why is tha t?

[lavish osprey]

dalle is not a model, mj is a finetune (of xl? maybe).

Why are you comparing a base model to a service? Compare Dalle to Ultra lol

[faint breach]

imagine being so gross and toxic that you wished people lost their jobs because you're bad at prompting

[subtle otter]

don't try me on design mate

[dreamy sundial]

is this good?? really?? can't do "very deep learning"

[viscid pivot]

@lavish osprey also you keep saying learn to use the model first while posting garbage results all the time?

[shrewd cairn]

In this economy? Don't have to be replaced by AI to get unemployed. lol

[opaque sedge]

Her legs are completely wrong.

[tropic aspen]

Holy shit what did I come back to lol

[desert garnet]

i see a third leg here

[silver sluice]

but it's not made of cheese, it just looks like a crab made out of a golden baked bread shell or something lol

[solar jasper]

Stop sitting on SAI's lap, the model is trash.

[faint breach]

surely part of the dall-e tool chain comprises a model

[hallow lion]

hands baby

[shrewd cairn]

It got a bit crazy while you were gone.

[weary snow]

How much ram is needed for clip+T5 fp8 ?

[craggy ridge]

Don't lol

[rugged nova]

I can not get a single decent gen of a woman sitting. Just a woman sitting.

[solar jasper]

the face of every person using SD3

[dreamy sundial]

LMAOOOOOOOOOOOOO

[silver sluice]

try running the update script after downloading the latest portable version, that's what I did, i also verified the update script points to the real github source for updates. also if you have tortoise git installed you can just right click and run git pull on the ComfyUI directory

[tropic aspen]

Don't what?

[dreamy sundial]

yeah this is unusable

[wind talon]

it's a very special cheese!

[faint breach]

i'm over here using a free model. it's fine.

if it's trash then you got seriously skill issues. generative ai been out for years and you still haven't figured it out? Figer it out

[unkempt grove]

It fails miserably at generating human bodies in any context

[wet sparrow]

skill issue (c) Stability staff

[wild bear]

Im glad that someone from stability staff, on the day of release, is just telling people to "git gud" and then post ok-ish images as a gotcha. SD3 might be better down the line with some finetunes but this launch is ROUGH.

[faint breach]

skill issue

[wind talon]

how is that a skill issue lol

[opaque sedge]

They are not even ok-ish, look at the legs, this is body horror.

[subtle otter]

bro is paid by SAI

[static cairn]

skill issue (on behalf of the bakers)

[viscid pivot]

Then dont say its our fault that your model doesn't behave as you advertised. Just say go finetune it, we didnt bother to train right

[faint breach]

i mean, lots of people making images that are fine is why

[finite hollow]

the update script doesnt seem to work in the new comfyui

[dreamy sundial]

it is trash, and the skill isuue is from stability.ai, i didn0t make the model

[celest hill]

thats wild

[faint breach]

what? since when!? I guess they're paying me with free models.

[craggy ridge]

I meant rather don't come back for while, it is wild here 😁

[unkempt grove]

They are cherry picked

[ornate oar]

🤣

[solar jasper]

Then please, Explain how to proompt.

[dreamy sundial]

u gotta be trolling

[faint breach]

Dude got MJ as a profile pic and can't even make a free throw shot. /sucks teeth

[subtle otter]

why you copeing then, what' in ti for you to congratulate such a disapointment ?

[eager wraith]

more plz

[celest hill]

sitting doesn't seem so bad tho?

[wind talon]

it seems to be a dice roll at this point really. quite disappointing after seeing how stability flaunted sd3's ability to generate good hands before it's release

[craggy ridge]

https://tenor.com/view/spider-man-we-one-gif-18212100

[wild bear]

Idk,I feel like this is a bethesda moment of "the modders will fix the game". "The finetunes will fix it", sure finetunes will always be better but out of the box the model should atleast be decent.

[faint breach]

i'm not disappointed. Actually quite entertianed today

[dreamy sundial]

like.... which ones

[modest flame]

Pretty sure SD v1 and SDXL launches were dozen times better than sd3?

[viscid pivot]

if you ignore the 3 legs

[wet sparrow]

SOTA in rock generation

[faint breach]

well not yours

[finite hollow]

the otehr script gives me this error:

[subtle otter]

i bet yeah

[modern drift]

this config with dpmpp_2m gives much better results thatn the default workflow comfy comes with,

[celest hill]

its the future whats wrong with 3 legs

[hexed dirge]

one more foot please

[bitter hearth]

[faint breach]

/spit take

[dreamy sundial]

yeah now im trolling.. everybody can see its unusable stop coping

[ornate oar]

sadly, have to do other stuff than hanging out here, but appreciate the sentiment!

[hallow lion]

as expected everyones complaining... ill try it later when the dust settled a bit

[viscid pivot]

or the cutout hand

xD

[subtle otter]

for real these guys are mental

[wind flume]

it's skill issues everytime until we get the finetunes and are better haha anyway are CN models released?

[faint breach]

just a minority of people with seriously skill issues

[wild bear]

Maybe ppls settings are wrong or something. Different CLIP needed or something hasnt updated. Why are the results so different?

[dreamy sundial]

yeah show me one good one lets all see

[desert garnet]

try it now,whats the matter getting tired eh sunshine? ill drop u like a bag of potatoes

[wild bear]

HUGE SHOES

[faint breach]

naw i don't jump when clowns tell me to

[ornate oar]

The same happened when XL released.

[hallow lion]

i has a bit of skill i pass it thru my aweosme noodletown setups later

[solar jasper]

the only thing that truly matters

[dusk silo]

landscape seems to take much longer now

[bitter hearth]

Average woman's shoe

[dreamy sundial]

you cant, bc its trash lol

[subtle otter]

oh yeah great result, what a fine image to show how good sd3 is

[compact forge]

platform shoes was difficult to do on sdxl so thats a win

[rugged nova]

40 gens so far, not a single not mutated woman sitting.

[upper snow]

Yeah DPM++ 2M should be better than Euler in general, setting it up with an ODE solver might work even better since that's the traditional method for sampling RF models

[unkempt grove]

I DID IT!!!!!

[faint breach]

QQ

[dusk silo]

1920x1080 is too much on my vram... can i load this without text? just use 2 clip?

[faint breach]

may need to look this meme up @dreamy sundial it might be before your time

[wild bear]

"The model is good!" "Can you show us?" "No clown". Glad we are all being constructive here.

[hallow lion]

maybe we need to use diff smaplers here unipc is a bit of a hidden gem

[icy coral]

Try a more detailed prompt

[rugged nova]

How is being more detailed going to make it fuck up less?

[wild bear]

See like some of the results are ok. Why does it vary so wildly?

[dreamy sundial]

fucking clown

[rugged nova]

Im trying for the most simple prompt

[fleet meteor]

x2, im not dissapointed on the model, it works and if the base model can achieve this things the community finetunes will be epic, im just dissapointed on the dataset they used to train the base model, there are a lot of things that it just "forgot" and was possible on older sd models, of course the community will be able to finetune it and fill those missing gaps, but idk why SAI keeps removing stuff related to copyright when the competency is better and don´t do it

[frozen lynx]

when I was using API .. it seemed like the upscale made images way better,

[solemn raven]

Are there any papers, guides or other documentation about the nuances of prompting SD3 you'd suggest?

[lavish osprey]

doesn't look too much different from sdxl base by the way. Maybe you're comparing a base model to 1 year of finetunes or enterprise apis.
If that's your goal I suggest you compare our Ultra or Core services with those.

[faint breach]

Whose being constructive where? i'm not going to work when scrubs demand it. Ever.

I got nothing to prove to y'all lol. You're just clowns with skill issues sulking

[hallow lion]

put in some epic promts. does it like natural language/>

[eager wraith]

We need a skull issue

[dreamy sundial]

bro were comparing it to base sdxl, and its worse with anatomy than sdxl

[viscid pivot]

THE PROBLEM IS YOUR ADVERTISMENT. You basically said its the next coming of christ

[solar jasper]

Man you all have such skill issues
Then post your examples
n-nuh uh.. yall have skill issues, clowns! Clowns!

SAI...

[craggy ridge]

https://tenor.com/view/tell-em-again-gif-20633643

[hallow lion]

its a ne wmodel folks it has its quirks we need ot figure it out

[modest flame]

I genuinely think that the skill issue lies in the negative prompting

[wild bear]

Ok bud, we hear ya. Everyone is clapping.

[lavish osprey]

by the way, the same happened at sdxl launch, where people where comparing it to DreamShaper, RealisticVision, etc, and even using those as "refiner"

[faint breach]

just me. sulk harder

[quartz mulch]

Is it just me, or is this model just bad? At the very least, it's absolutely nothing like SD3 through the api

[icy coral]

the model doesn't seem as good at generating from nothing, it might need more handholding

[civic egret]

This

[rugged nova]

Ill do base sdxl side by side if you want.

[dreamy sundial]

bro you cant even show us one image with good anatomy, get lost

[faint breach]

i remember people prompting with just the refiner and then decalring that stability only ever releases trash

[desert garnet]

sdxl sucks bud,1.5 better

[viscid pivot]

I wouldn't even hate SD3 if you just had given us realistic expectations lol

[dreamy sundial]

do it pls

[lone jacinth]

SDXL 0.9 didn't output images like these for me.

[wild bear]

Gotta be some strong tea. She don't know where to look lol

[barren spindle]

just updated to CUDA 12.1 and updated torch accordingly, still getting same error

[civic egret]

Everybody's all up in arms at the very start of the release of the model. Y'all need to chill, learn the model and its quirks, then you can feel free to complain.

[subtle otter]

this guy is a troll mate don't answer to him, he is coping so hard

[faint breach]

i'm not going to publish anything i make today. The low skill babys are primed to kick over every sand castle they see

[dreamy sundial]

yeah

[dreamy sundial]

whatever get lost

[craggy ridge]

Some people troll for living

[modest flame]

Try and use t5 fp16

[silent tree]

GOOD JOB STABILITY XD

[supple gulch]

By the way, can we nsfw with sd3?

[faint breach]

Make a free throw. Can't do it

[silent tree]

i was waiting for this?

😂

[lavish osprey]

some sd3 2b vs sdxl base comparisons.

[coarse wedge]

Don't respond to flowwolf he's just trying to irritate you.

[viscid pivot]

They censored the model like hell, yet they allow biden to go in bed with trump xD

[fleet meteor]

Oops typing error

[dreamy sundial]

not the same model we have

[faint breach]

If you can't make a free throw right now right this instant, you're a troll.

[opaque sedge]

Bent elongated arm and hand, three feet, broken trousers, if you think this is good I have no words.

[rugged nova]

Now try that in anything not standing perfectly straight

[keen falcon]

lol

[rugged nova]

make them sit

[sullen pond]

But can it do people lying down?

[dreamy sundial]

nope

[coarse wedge]

It can barely do people standing up

[lavish osprey]

I did and posted it and had no issues. Understand that with 3 text encoders you need to be clear in your prompt, or they will fight.

[solar jasper]

Bu-but skill issue

[weary snow]

How much ram needed for clip+T5 ?

[wet sparrow]

mind share your prompts or setup? your gens looks really decent

[faint breach]

for real though. i am having skill issues. my hugging face download keeps stalling and failing.

[silent tree]

lykon how its possible you were showing very decent previevs for last couple month and now we got this? XD

[viscid pivot]

Yea and that are like hyper trained images in your dataset? They are the same tho, so nothing special compared to SDXL

[faint breach]

can't use the t5 model yet

[ornate oar]

Uses roughly ~16GB

[finite hollow]

i use the new portable version now and i did update and it did a lot of things, but its still missing EmptySD3LatenImage for example, and not letting me install it

[silent tree]

im not hating or something, just curious

[bitter hearth]

Eep

[lilac latch]

T5 only. StableSwarmUI. No upscaling.

[errant dust]

I will make the following suggestion to others: use the external file for the fp16 encoder instad of the embedded smaller fp8. I am getting night and day differences notably in text reproduction. Here is fp8:

[rugged nova]

No you did not, you posted a mutant

[weary snow]

Thanks

[lavish osprey]

yep

[errant dust]

and this is fp16:

[keen palm]

Careful when you walk down the stairs folks

[desert garnet]

SD3 sucks balloney sugma model better

[ornate oar]

oh, sorry, a bit more in total, but 16GB VRAM.

[errant dust]

This is using them without even the other clip files

[icy coral]

Already much better

[lilac latch]

Now my results look like garbage... Very pixelated images despite using the refiner.

[viscid pivot]

It is good at understanding itself hahaha

[bitter hearth]

can it do Booba? Sign me in

[weary snow]

Too Much

[unkempt grove]

CFG: 2
dppm_2m, 30 steps,

[desert garnet]

you are now banned from SAI

[lavish osprey]

git gud

[modest flame]

Foot is a bit iffy

[wide pagoda]

getting the scheduler wrong on purpose actually seems pretty cool for glitch art

[rugged nova]

2/4 of how many, and it still has mutated fins for hands

[celest hill]

90% of subjects coming out asian is wild btw

[wet sparrow]

thanks

[icy coral]

So you're just looking for arguments. Alright

[viscid pivot]

Bro are you trolling? look at your stairs

[rugged nova]

Im looking for a single decent gen not standing

[bitter hearth]

I wouldn't mind myself 👍

[faint breach]

yup. if you show results the goal posts change. it's why i don't post examples for these people. nothing is good enough for them

[viscid pivot]

look at your images before you say "git gud"

[compact forge]

i know why the anatomy is so bad. it was trained on your mom

[topaz flume]

https://i.gyazo.com/9f092f34b2b39751d6fdb3028b5fd7e3.png
What is a Download Mirror and why is Load Balancing a good idea.

[desert garnet]

i can do the same when sneezing on napkings after eating a burger at wendys

[faint breach]

git gud

[finite hollow]

my comfyui is still saying this:

but update comfyui says its newest version

[lavish osprey]

definitely better than this #🆕｜sd3 message 😄

[weary snow]

Thanks man

[lilac latch]

NSFW is dangerous according to SAI and its investors, so no. Wait for a proper finetune (months).

[craggy ridge]

https://tenor.com/view/mad-gif-20369678

[fleet meteor]

This probably happened because of the removal/censorship of some images in the training dataset, the community will fix it fast

[young blade]

i wouldnt call velveta "cheese" tho 😉

[celest hill]

git gud

[lavish osprey]

wasn't supposed to be... ?

[rugged nova]

Thank you, first worthwhile gen ive seen of anything not standing. Ill try that prompt

[faint breach]

AI Safety is a real concern. Those at the bottom can't see the view from the top though. Understandable that you wouldn't get it

[viscid pivot]

no his stairs actually look kinda okay

[finite hollow]

how/where can i get a comfyui portable version that works with SD3?

[tropic aspen]

If people are motivated to finetune because of the licensing

[dreamy sundial]

not same model

[lavish osprey]

yeah perfect image

[subtle otter]

this will take month to have something usable dude

[viscid pivot]

yours look like they have been cutout from paper and glued back together

[topaz flume]

haha

[lavish osprey]

literally downloaded from HF lol

[solar jasper]

Nobody is getting hurt, Lemme gen some booba.

[faint breach]

Very short sited.

You've got sd15 and pony already. go nuts

[dreamy sundial]

were not getting nearly the same results

[finite hollow]

how/where can i get a comfyui portable version that works with SD3?

[lavish osprey]

git
||good||

[fleet meteor]

Yes but this is not something it can be fixed with scripts or controlnets, its a pure training mistake

[ornate oar]

@compact forge just quietly posting banger gens. Go you!

[topaz flume]

github, just update

[rugged nova]

Your using the encoders seperately. Ill try tht next

maybe that is a issue

[finite hollow]

can you give me a link @topaz flume ?

[coarse wedge]

Garbage anatomy aside, why does it only want to make Asian women even with it in the negative prompt?

[mellow latch]

Yeah like.

[dreamy sundial]

whatever

[solar jasper]

Lots of them in the training i guess?

[celest hill]

"british woman"

[bitter hearth]

Oh my god this wizard is powerful

[south geode]

This is so sad to hear all those people bitching about SD3, just because release model is bad at hands and anatomy you say is bad overall, look at other things it can generate, or go use whatever model you like and pay for it, without fine tunes, flexibilty of comfy and programers who chat with you here.

[solar jasper]

Google Gemini Image gen returns...

[finite hollow]

https://github.com/comfyanonymous/ComfyUI this one ?

[faint breach]

question i'm curious about. would using the fp8 clip be better on a Ada 40 series card, since it has the hopper transformer engine? Or would that do nothing for the efficacy of the math?

[topaz flume]

sure https://github.com/comfyanonymous/ComfyUI
just install the portable and make sure it's update using the comfui manager

[lilac latch]

Did you generate this with SD3?

[compact forge]

ill try adding some booba to the prompt

[celest hill]

yes

[lavish osprey]

maybe we should just upload the fp16 version lol. Too hard for people to use them separately

[rugged nova]

Im going to try using the encoders seperately instead of the all in one

maybe that is a issue

[bitter hearth]

👍

[low stone]

@lavish osprey so here's where I'm at. sdxl had a lot of the same issues with hands/limbs etc. Over the last month we've gotten the align your steps, and the res_momentumized sampler. Between those 2 things, all of those limbs/hands issues have largely vanished. res sampler is out because it's like sde/ancestral. Do you think there's options for new samplers to fix this like they did for sdxl?

[finite hollow]

@topaz flume i got it from there, but it says it cant do the sd3 stuff and its not updating itself further

[tropic aspen]

Ah yes, because clearly ppl don't use AI image gen for generating ppl

[desert garnet]

doesnt work

[topaz flume]

probably a good idea haha

[ornate oar]

Left: sd3_medium + TripleClipLoader (clip_g, clip_l, t5 fp16)
Right: sd3_medium_incl_clips_t5xxlfp8

[lilac latch]

weird... all my images look pixelated as f*ck

[odd kernel]

And StableSwarmUI is shit too

[viscid pivot]

Bro, you found that most nsfw image! No get banned!

[barren spindle]

ok

[dusk silo]

dpmm_2m and sgm_uniform are the only config options that work?

[compact forge]

skill issue

[faint breach]

seems that way. after he showed the fp16 version i dleeted the model i had and am downloading the version without the clip

[celest hill]

this is trained entirely on asian women or something

[topaz flume]

no for sure it is working with the correct update

[storm saffron]

dpm2 also works.

[odd kernel]

Hope Forge is updated fast

[modest flame]

DPM++ 2m + simple works too

[modern drift]

is it? works on the first go with the comfy ui workflow

[fleet meteor]

Amazing lighting bro

[pseudo stone]

do I use t5 fp16 or fp8?

[lavish osprey]

do you want more quality or less vram?

[lilac latch]

thsoe sandals...

[ornate oar]

Have the VRAM for 16? Then use 16. If not, then not. 16 > 8 in quality.

[faint breach]

signs like this should be addressed and potentially bannable. this is just beyond toxic now. @lavish osprey thoughts?

[upper snow]

must be from Hong Kong

[rotund ibex]

I have no clue, hope my answer helps in anyway

[dusk silo]

ty, should i use gpu?

[topaz flume]

you want ComfyUI c8b5e08 version or newer

[fleet meteor]

This is what im talking about, it can generate very good images, just not about all the subjects

[dusk silo]

ty

[storm saffron]

if you get that option yes

[shrewd cairn]

It's a meme dude. I'm not trashing the model.

[icy coral]

Shocking that they don't want to be associated with weirdos making furry and real people porn

[remote holly]

https://www.reddit.com/r/StableDiffusion/s/c9sGvzFvjG

[lavish osprey]

deleted, better safe than sorry

[shrewd cairn]

I actually like the model.