#🆕|sd3
1 messages · Page 15 of 1
Let's gooo
its over it got canceled by guys
it loaded
aw
But there will be limit for everything right? Right now I'm more Curious 8B model.
its going to release next month
I was right....
sorry guys
We made it
it's here
its out 
oh no not this gif!
OMGGG
i was right?!!
WAIT NO FREAKING WAY
GUYS IM CUMMING]
WHAT

😭
Introducing Stable Diffusion XL 2! Our newest model release from stability AI
LMAOO
looooool
YESSS
oh ye, and chat broke
i js wet my underwear
wtf
GOOD TIME
it won't necessarily get worse even if you duplicate all the layers like 30 times but without training those layers it won't get better either
it's fake, guys

hmm yeah
Gooo
non-commercial license
3 different model
boobas
💥
The SD3, the SD3 is real
(I don’t have a Pc to run it)
oh my goodness, i didnt think the day would come
I have been interested what happens if someone will just infinitely finetuning every possible concept (idk... cars, weapons, various objects, action and dynamic poses, holding something, and everything else). Like, will previously finetuned things become weaker and weaker with new finetunes?
is sd3 here???
how was i right wtf
1 hour early
they just released it on huggingface
can i just download it from comfymanager?
LESS FING GO
It’s a troll guys! You’ve been duked
according to a wrong time, but yes
huh
EXCELLENT
anyone able to tell me which version I should download haha
TIME TO TEST
Yuh huh
2 more hours guys
how do u know
I cant use it by now but I am sooooo nervous what you will create in the first hours. Hope for great results and suprises!
Depends on your spec. Largest if you have space
its out
i see the safetensors
4.3GB
okay, this is a very cool way to do it
it is this one right?
i'm downloading all of it
1 hour away
thank you devs for including fp8 weights, so much better 🙏
WHAT DO I PUT THOSE FILES ON
4070ti for me, reckon I could run it?
bro what do you mean its already out lol
the spam 😆
LOOK AT THE WORKFLOWS FOLDER LOL
Ohh
I suggest the smaller and then load the tes separately. The largest has fp8 t5 instead of fp16
why do we have to share a bunch of info to download the model
Nah, 2 more hours (shhh... I wanna grab it before HF crashes)
civitai when
is it there?
What do you write for "Organization or Affiliation"?
SD3 ia there ...!
I am downloading SD3 Included Clip
finally!!!!
Yo @lavish osprey can SD3 generate a red circle with arrows pointing towards it? (Not even Dalle 3 can do it)
because some people are naughty so you have to promise not to be
doesnt work on Forge, reported and manager called
But what's the difference between them?
civitai staff ginna download it with us 😁
where's the example workflow?
yo is it released?
Will this work in automatic1111/comfyui or do I need to wait for an update?
i still cant belive it
just comfy, workflows in the repo
What's the difference i havent been keeping track
get the incl_clips if you want the base
get the one with t5xxlfp8 if you want T5 as well
if you don't know what T5 is yet, get the base 😛
it can't be true no way
someone make a yt video for us @ work please we will like and sub i promise
which I need for my 4090
oh wait
base?? t5???
i havent been keeping track for a while it seems lol
the sd3_medium.safetensors is there so we can hook up with different Text Encoder
pretty cool
T5 is going to use lot more RAM or VRAM, but increases prompt adherence and text capabilities, you MIGHT NOT NEED IT
is this 768x768 like 1.5?
Just d/loading ...
T5 is what gives SD3 its amazing prompt adherence
1024x1024
and someone wanna try it with a m4 16gb ram ipad pro ?
SD3 has D R O P P E D
whats the difference
If I'm not using T5 yet, but might want to later should i get the with clip one? Or just go the complete base one
which one better, Base or Acid ?
The weight is over
it's only 10 gigs
there's even demo prompt
that mean if I want T5 I need download sd3_medium_incl_clips_t5xxlfp8.safetensors
and also t5xxl_fp16.safetensors/t5xxl_fp8_e4m3fn.safetensors ?
what gpu would you need to run the 10 gig model
2h to download :C 
back in my days we waited for models 
wth it took me 2 minutes
I was expecting a VAE file, I presume it's baked in?
Where do I get these custom nodes from? i forgor to update
update comfy dude
if this chat is too chaotic, guide from mcmonkey https://reddit.com/r/StableDiffusion/comments/1de65iz/how_to_run_sd3medium_locally_right_now/
update your comfy
The Huggingface servers rn
I can't decide what to download so I'm downloading everything 
Can’t wait for the numerous reddit posts going “Erm this model sucks actually”
forgor
Where should clip_g_sdxl_base.safetensors and clip_l_sdxl_base.safetensors be downloaded from if we don't want them baked into the model?
where the custom node guy?
official workflow
Post it
update comfy UI
wait no I need to change settings

my download speed is pretty slow
refer to the bingo card from earlier #🆕|sd3 message
baited us
cliploader?
4 more hours
CLIP loader only neccessarily when you using non-Incl models
where'd you get this?
example workflows use fp16 t5, so we load it separately since bundles only have fp8 t5
who will be the first to post a picture?
goddamit
which file is perfect for 4060.... ?
there is 3 files
lmao
you need to agree to all the license stuff first
kinda sick
Guys
wait you manage the model space?
i did but im using sagemaker to donwload, i think ill try using token
wow, i'm blind
Can someone dm me how to set it up
Made with SD3!!!
wait does this not work in auto?
(Real)
where can I find the comfy workflow for sd3?
What's the difference between the clip 5gb model and the 4gb model?
yeah only for ComfyUI and Swarm at the moment
@woeful spindle
anime 
which model we should download ? all ?
one included CLIP natively
oh thx
the 4GB one does not
and all hell breaks loose
What d oes that mean though, is there any particular reason to download it or not download it?
I like how this model has the SD XL Lightning art style
??
🤔
I swear I’m not tripping
basically if you download Included CLIP one, you dont have to download the CLIP model separately
so pretty much I recommend you to download that included CLIP
would the model work on a1111 or forge yet, or would those programs have to update first?
Can Align Your Steps be made to work with SD3?
Guys what do I need to download
computer
if you're using stable swarm, it says it will auto download clip separately, so no need to download the 5GB version https://www.reddit.com/r/StableDiffusion/comments/1de65iz/how_to_run_sd3medium_locally_right_now/
Ram
So clips are a new thing, what's that all about?
which model??? there 3
?
Clip is what SDXL used T5 is the new encoder
you could run this w/ like ~16 gigs of vram, no?
I think…
for swarm, I need the base model only as it automatically installs the clip and T5 models
Idk
darn..
they're an old thing, the only difference is they've been split into a separate file so you don't have to redownload them with every finetune
is 12vram too little lol
Which I need for Comfy
I guess
I taked a nap , does sd3 was released ?
sd3_medium_incl_clips.safetensors <- smol vram
sd3_medium_incl_clips_t5xxlfp8.safetensors <- smol vram and big ram or big vram
so, where do these come from then?
define big vram
Welcome to brand new world
bro crashed
Leeeeeeessssssgooooooooo
I will try it imediatly
I'll tell you in 3 minutes when I load it 😆
lets see if my 2GB VRAM can handle it...
no.
hope it is not a ticker time bomb for my laptop
No need to download encoders separately if I got the big boi right?
gonna need some super noodle magic 
So If I want the full power of SD3 2B I download the 10GB one?
those are both smol. Biggest is 4gb + tes separately, which was actually too big to package without looking silly
ah there's a separate fp16 t5
Don't know where to put this but here seems good? Be careful out there! Be sure you know what you are downloading. https://x.com/jason_koebler/status/1800560669785526694
SD 1.5 users PCs after booting up SD3
has anyone tried anything naughty yet? 😏
Server @HuggingFace clogged: no access (temporarily)
It wasn’t trained on anything explicit
@lavish osprey So there's 3 different medium models. If I get the one with the T5 fp8, do I still need to download the clip etc models too?
"Sins: crypto promotion
Also: pay us crypto" lol
blazing download speeds
hand holding 👍
what is Model Sampling by the way?
ALWAYS SAFETENSORS
download the tes and put them under ComfyUI/models/clip
these are the urls https://github.com/Stability-AI/StableSwarmUI/blob/73860ec3c727e9745f3b273fdb2b27d8a3007965/src/BuiltinExtensions/ComfyUIBackend/WorkflowGenerator.cs#L1597-L1601
clip_g_sdxl_base.safetensors - https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/text_encoder_2/model.fp16.safetensors
clip_l_sdxl_base.safetensors - https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/text_encoder/model.fp16.safetensors
t5xxl_enconly.safetensors - https://huggingface.co/mcmonkey/google_t5-v1_1-xxl_encoderonly/resolve/main/pytorch_model.safetensors
Feet enjoyers on their way to generate explicit images
No luck so far with Hunter Biden
ain't nothing more criminal thannot following your own rules
bros can't even be radicals RIGHT
When you run SD3's testPrompt in the swarm, the download of clip_g_sdxl_base.safetensors will start automatically.
No, this seems to be the missing cliploader file...SDXL?
Thanks!
What's the deal with the one that says include clips on comfy?
anyone got the workflow
turns out it doesn't need much, like 8gb peak vram and 16gb peak cpu ram with fp8 t5
depends on your compute, best is sd3_medium + 3 tes separately
lykon ccan i have ur workflow @lavish osprey
Now we just need a uncensored stable diffusion 3 medium and then my ai waifus will be real
wait, you mean having the text encoder separate is somehow better?
I want use sd3 in comfyui with kaggle , the vm memory is 19gb do you think is enough for sd3 ?
At one cost. You will have to use tag based prompting
what am i supposed to put here?
the packaged t5 is quantized to fp8, the separate one has fp16 version
OK... i think it might be a little bit tricky
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!
but not sure how much it affects the output
I will never go back to tag based promoting
Sentences all the way
Not bad, not bad at all
there's 3 missing nodes, what do?
Someone generate a cat wizard
update comfy
damn thats nice. thanks for clarifying
I did
need to uninstall Civ 6 for this
I don't think that's healthy for the car
RIP
So, there's a file on comfy that says "sd3_medium_incl_clips"
If it includes the clips in that file, how do you access them?
Bro the inclip model is 29gb 
how significant is the quality difference?
idk, I have to download the separate te first (judging by llm perplexity increase from fp16 to fp8 there should be basically no difference tho)
Same way you did in sdxl/sd1.5, the little blue clip dot on the Load Checkpoint.
How i am supose to run it in a notebook if the model is 2x bigger than my memory
downloadmoreram
My bad 4.29gb
little blue dot..?
no difference between 8 and 16 bi
t
yeah we keep asking that question and so far no answers.
doesnt work on my 2gb ram and 512mb vram laptop either,i feel scammed 😔
is SD3 available to run locally yet?
My laptop can’t even run 1.5
2 more hours
Guys which I need on my 4090
you checked already?
no, i just know lol

every time a model is quantized from 16bit to 8bit
there's no real quality diff
in llms
diffusion models
countless papers on this
What folder do you put the text encoders in?
I will make a notebook and send you if you want
I mean, yeah, 16-6 bit is basically no difference for llms
Привет
I will try
Ah, found it guys!
I'd be curious to see how it does in 4 bit since bitsandbytes can quant it to that too
joke?
real
how do I generate photos?
did previous models work on there?
Did anyone met this error
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!
Is there any guide on how to install it, please?
will be right back after I fix the thing
Very impressive. sd5 when

my goat u saved me
it's slow, where's that tensorrt engine 🤡
nvm
Everything on card uses about 17 gigs of vram on my 4090
I'm not home, can someone try a fiat multipla?
Speed SD3 vs SDXL any first thoughts?
2.34 it/s
Is that SD3 lol
🫡
around the speed of SDXL I think?
awesome ❤️
you joke but
"Requested to load SD3
Loading 1 new model
ERROR: model not supported."
aww 😛 no tensorrt yet
can confirm 10vram works!
neat
1.02 it/s on a 3060 12gb so faster than SDXL for me?
wait what should I download or everything?
just get the smallest one and then get the t5 fp16.
I'm stuck at authentication ... doesn't like Username and Password?! I did get a token, but have little knowledge of how to use it?!
anyone know what the reccomended steps and cfg are?
isnt this sd3 ? https://huggingface.co/stabilityai/stable-diffusion-3-medium
what model are you using? I'm getting 1s/it on a 3060 12gb
Shouldn't the biggest one have everything all in one
It's alive
Damn pony v7 gonna be on SDXL instead of SD3
7
nevermind i forgot they're like the same speeds almost
how long until people make thier own checkpoints
I'm stuck at authentication ... doesn't like Username and Password?! I did get a token, but have little knowledge of how to use it?!
7 looks worse than 4.5 for me
Weeks for good quality ones
tried it as well, since tensorrt's limitations, right now with no fancy things available yet, might actually be the best time to use it, if it was there
with t5 encoder, it is using 7 GB VRAM!
does anythign look wrong with these settings
oh lol
I'm stuck at authentication ... doesn't like Username and Password?! I did get a token, but have little knowledge of how to use it?!
8 GB + up y'all are good
Someone have the link of sd3 workflow ?

curl -L -H "Authorization: Bearer hf_xxxxxx" -o sd3_medium_incl_clips_t5xxlfp8.safetensors "https://huggingface.co/stabilityai/stable-diffusion-3-medium/resolve/main/sd3_medium_incl_clips_t5xxlfp8.safetensors?download=true" (or the other version)
Thanks
Call an ambulance!
ah damn...
How do I authenticate at Huggingface to d/load SD3?
Guys which can I use on my 4090
i think it is is going to be tricky for me
1s/it = 1it/s - using med inc clips
yeah im dumb my bad
Okay, now to figure out what I'm doing wrong...
i thought that was obvious, if you download the model with clips, then you take the clip from the load checkpoint
same boat buddy lmao. gonna give it a try after i caffeinate
censored for showing her neck
we finally have proper gun handling.
So base medium, clip g and t5 xxl fp16?
THIS IS ME FR
https://huggingface.co/docs/huggingface_hub/en/guides/download#download-from-the-cli if you haven't done before, might be better off just downloading in browser
i wanted to try sd3 on HuggingFace how can i do it
imagine they forgot to upload a vae or something
yes
right as i went off to work lmfao
update: SD3 does not seem to like negatives?? For certain cases at least
m8 are you telling the community that stability ai doesn't support the furry models and stuff?
sadley stuff like "man doing a handstand" is still not good 😄 or did something get propper results for this 😄
going to try this one, as I have only 16gb ram, btw what is "sd3_medium_incl_clips" then?
Can't find the training code either, maybe not yet released
finally released im so glad! but i am a simple person so i wanna try out sd3 on HuggingFace online
#... working, thank you!
How good is SD3 at anime
I mean even the model itself? Is there a DiT in comfy or sth?
Fr. Waiting for online services to get it
I'm a little confused. From text encoders. Why does the “sd3_medium.safetensors” model exist and what is its benefit?
how much better is t5 fp16 compared to clip and what does it change?
is sd3 medium the same model advertised in demo in Feb?
yeah i was searching everywhere but i didnt find the model anywhere
waifus are safe
Wish it was better trained on celebrities. "cinematic photo of Johnny Depp in the teletubbies"
i think ill just wait till someone makes a webui
👍
Pony isn’t gonna be on SD3 so someone better step up
thats Johnny Deep
"Removing T5 for inference only results in significant performance drops when rendering very complex prompts involving many details or large amounts of written text. The above figure shows three random samples per example."
those Teletubbies giving uncanny valley vibes
Is the base model
They didn’t want to get sued for deepfakes
This is the version that I am d/loading sd3_medium_incl_clips_t5xxlfp8.safetensors
thats the thicc one,it eats a lot of ram
IS THIS SD3? I WANNA TRY IT
Apparently base medium big clip and fp16 separately is better?
you gotta take a second m8, AH is implying that it's gone successfully in the past, but is no longer successful to have "private" discourse about licensing. Why u being so salty?
The complete model is huge. This was not expected
Source?
hmm high steps 30+ seem good
i cant find the new nodes in install missing nodes
I have to provide my information and the model is non-commercial use? wtf
Can we say sd3 is better than dalle 3 or not yet?
check the readme
Sorry I'm a newbie with comfyui can someone tell me what I'm doing wrong?
so should i get fp8 or fp16?
I'd say "really good" so far
all i ever get is this
so you can't use the fp8 on the clip loader?
noise!
I like it
You have not selected your models in the different nodes
I am forcing my work desktop to run this on CPU, all i get is fan screams
oh it was during the license talk, I read a bunch but gave up and went to play games, it wouldn't stop 
thank you
you're not allowed to train it?
Thats, bad
He wants to make money
For today's standards thats bad
And doesn’t want to pay up
it is allowed
Playing with samplers so far this is my fav 
i mean
ohhh
Alright I thing I got It xd
7 minutes into a 20 minute d/load ...
Do we still need huge prompts to get good images?
i gonna see the license soon
What happened lmao
you're bad and should probably walk away with your head down
ok i see
SD3 likes high steps! 30-50
But dalle 3 would make a better image doe
The sampler doesnt seem to work too well 
you wouldn't though
do persons perform different positions?
how is it
Am i AI confirmed??!?!?!?
Have you upated ComfyUI, downloaded the models and hit the refresh button?
RayanAIR v1 on hugginface
I really should just wait for lunch to run home..
Putin is having a playful water fight
no way
When dropping T5 - longer more complex prompts work better
Bad model 
What am i lookin at
THE HANDS
start posting images for us to rate please
SD3, perfect hands...always
Dude what are you doing
Arguably the biggest community fine-tuner is trying to have a reasonable and respectful discourse and you're clowning
incredible
default workflow default image of comfyui
Completely ai!!!!
PiXart-Sigma standard 🙂
I love 2.5D style
1stSD3!
DAMN
That poor creature on the left
just use fp8, you'll never reliably see a difference between the two
🤷♂️ I just took their example workflow and assumed I wouldnt have to then modify it since its supposed to be the example lol
Evolution hits hard xD
I hope some applications will make sd3 more stable
faster it seems, at least for me

roughly the same? I could be wrong
the same prompt on 2b before we started working on it some months ago
RTX3090, Step 28 = 22sec Res: 1344x768
where do u put the text encoders what folder lyky @lavish osprey
there's a little difference
I shall soon have my SD3 i2i PiXart-Sigma SD15 SDXL PAG Face Detailer setup!!! 😄
ComfyUI\models\clip
Wut duh hell is dat
oh god what is this
awesome
whats the matter babe,why the long face
clip
Wtf 
||art||
A whimsical and creative image depicting a hybrid creature that is a mix of a waffle and a hippopotamus. This imaginative creature features the distinctive, bulky body of a hippo, but with a texture and
appearance resembling a golden-brown, crispy waffle. The creature might have elements like waffle squares across its skin and a syrup-like sheen. It’s set in a surreal environment that playfully combines a
natural water habitat of a hippo with elements of a breakfast table setting, possibly including oversized utensils or plates in the background. The image should evoke a sense of playful absurdity and culinary
fantasy
i'm getting roughly 22 second generations on a 2080 at 35 steps
did anyone manage to compile TRT for SD3? Comfy's TRT nodes got updated to support TRT and compilation is stuck
use the example workflows, why 35 steps?
@lavish osprey Is there a prompting best practices guide?
More equals good
Can someone make a anime ai waifu please 🥺
smiling cartoon dog sits at a table, coffee mug on hand, as a room goes up in flames. “This is fine,” the
dog assures himself.
No one has generated Will Smith eating spaghetti yet
there is examples on huggingface download page
How many people worked on sd3?

oh just a habit i guess, was so used to doing 35 steps for sdxl for my workflows
I always wonder if the results of 2 months ago give motivation or disappointment to developers
alright, it somehow finished compiling right as I wrote that
So you looked at that and were like "I can fix her" ?
being put into lowvram mode with 4070
Hi g, where can i find w for sd3?
Pwease
whats the difference in the 3 models? medium,medium_incl_clips, and medium_incl_clips_t5xxlfp8?
Too bad my gpu stopped working 
t5xxlfp8 is a 20 minute d/load
Will it work withh forge? i hate comfy.
One is base one is packaged with clips, the last one is packaged with clips and t5xxlfp8
I will wait diffusers support
What is the benefit of encrypting text l and g?
Swarm has a non noodle UI and it's updated
how do we do highesfix?
I could infer that much but thanks
sexy
where do I place the text enconders and and the t5xxl files??
I think with enough finetuning this model could be great
what's going on in Russia
Meth
Dude is long
Where do I place the text enconders and and the t5xxl files??
covered in https://new.reddit.com/r/StableDiffusion/comments/1de65iz/how_to_run_sd3medium_locally_right_now/
fp8 = 44 seconds
fp16 = 46 seconds
system specs: 8gb vram, 32gb ram, i9 13th gen cpu
Hope that helps ^^
but anyways, 1.55 it/s is pretty awesome. was expecting more like 1 it/s (rtx 2080)
Testing testing 1 2 3 ... ok, here weee gooooo!!
😂
thanks
what is prompt for this image?
so is the t5 model the most capable?
cute
Thanks for all your support, Alex!
ahh f me got 4070, how much time it takes to generate a image?
@lavish osprey LFG 😄
can't say. it's working well though. only thing I changed, was on the ultimate upscale, set to 1.5x scale, and 0.4 denoise.
the stone is shaped like the words:
"SD3 NOW HERE"
The 3 words are stacked vertically
worm's-eye view perspective, vanishing point perspective```
why do mine always come out looking like this?
ok got it working
so the recommended setup has us only using the negative prompt for the first 10% and then zeroing out after that? neat
Can someone do a anime girl generation
thanks
I'm sure this has been asked a bunch and will be asked a bunch more but for the triple clip loader, do i have to pick 3 different ones or can i pick 3 of a kind?
approx. 20 seconds, cant complain.
around 10-15 if text inputs havent changed since last generation
I also had to update my ultimateSDupscale node in comfy. it didn't work otherwise.
The negative prompt was reality
at 28 steps
lykon did it already
the VAE is baked in?
oh hi comfy
Welcome to open source 😄
hello comfy
any way to read the difference between the 3 models?
reddit gonna be wild without context
SD3 vs SDXL - same prompt
Where do the Text encoder files go???
into the clip folder
clips folder
SDXL base? 
do you download the 10gb model for best results?
Woohoo
request prompt!!!
No, but if base does hands like that... 🤷🏻♂️
got this when tried to run on Swarm
Doesnt get any simpler:
anime girl with "SD3 is here" written on a sign shes holding
which one is sd3

same prompt now
just wait until this gets a finetune
I assume that this file goes into the T5 Models folder and not the Checkpoints Models folder?
You probably didn't fully download the model file, check if it's the proper full 4GiB - if not you may need to redownload
16 gb ?!
How do I load clips on comfy if they're prepackaged?
i got this... what's wrong?? "90s animation shot depicts sailor Mercury with one eye closed, smiling and holding a sign that says "愛している""
i prefer the first one,this one too 3d for me 😔
laughs in 32 gb
SD3 API vs local, i might not have the right settings yet (just the defaults) but i do see a certain lack of fine details in 2b, they are cleaner though! can't wait for 8b!
inspired by Bella Kotak of a cute blue/grey skinned, hyperrealistic, life like, goblin with whiskers eating cheese on a chalkstone cliff, large azure beach below, extremely detailed, 8k, intricate, warm summer vibes, gentle breeze, coastal, realistic, dramatic lighting, movie still, fantasy
A grizzled orc warrior, adorned with piercings, sits in tranquil meditation pose, channeling willpower to manifest piles of gold coins and mystical symbols surrounding him. Amidst the industrial, winter wonderland backdrop, the brisk wind whispers through the scene, as if shot on Cinestill 800T film. The atmosphere is cool and crisp, with a hint of mysticism, rendered in a realistic style reminiscent of Julie Bell's work. Negative: tusks
will try redownload, thx
no nihongo, sadly 😦
Japanese turn to English, but somehow translate correctly
Diffusers support will be released today ?
it's 2 different models
Wtf
did they biase toward black people like google?
huh almost says "love you"
it is ram, t5 is massive, should do without
Is there a difference in comfy between an empty latent and an emptySD3 latent and if yes, what is it 
out of this
Another API v local 2b
Hauntingly beautiful mixed media collage, influenced by the dark, mystical realms of Camilla D'Errico and Gustav Doré, with the surreal, dreamlike atmosphere of Masayoshi Matsumoto. A daring Russian explorer, fiery red hair ablaze, clad in a worn leather jacket, ventures into the mystical underground cave, clutching a radiant crystal, as shadows dance, concealing ancient secrets, waiting to be unearthed
what
When you guys have been messing with it, does this one actually understand sentences or is it still better to talk, like, this
Hello everyone
@lavish osprey are these weights models that can be further trained?
most samplers seem kind of broken 
Yup noticed that too
The t5xxlfp8.safetensor goes into the T5 folder?
Is ram or vram ?
How could such a result happen?Use official parameters, Just reduce the number of steps.
just ram
All of the 3 models you can download go into the checkpoint folder I think
Ha ok
I'm not rich lmao
SD3 fundamentally imcompatible with Euler A
and SDE
I think it was the vram
what's the best combination for best quality?
any reason why the negative prompt is timestep conditioned in the example workflow?
all
nice, gaing to upgrade too
which are you using? because my results kinda suck
sweet
Which do I need?
Id run what the sample workflow suggests and try around from there
I am clearly doing something wrong.
Euler, 28 steps, Clip + T5, CFG: 5
1st
You can run it on 12gb vram ?
I expected it to be better than an SDXL finetune, but I guess not yet?
My username and password work on Huggingface Website; but I cannot use them to authenticate SD3?
I can't generate something since well, I have my own error
ddpm and ddim_uniform scheduler also gives me some weird results
Not sure, I just had to leave my name there. Maybe the servers are at capacity?
fast as fuck boy
Does sdxl controlnets/ipadapters work with sd3?
Is it me or does this model seem to quickly drive towards anime style images ?
yea guys with 16gb do not try t5 even 8bit
12gb 3060 - t5 40 seconds, clip only 30 seconds, t5+tensorrt 20 seconds 🎉
Nope
No
Dang better not use t5
It says to get a token (SSH) and use that. I got a token, but have nowhere to input?
right one is sd3 right?
Sorry I dont think I can help with that
use it on cpu. By the way the reason why we made it modular is exactly to allow different usage
No
@lavish osprey when does the DreamShaper SD3 drop? 😁
Hey Lykon which should I download for my 4090
seems the censorship ruined this one
how?
did you use a sdxl finetuned model or base?
i deleted the current t5 and found about the t5xxl_fp8_e4m3fn so ill try that one
1m+/it at 1024 😂
managed this at 1920x1088 without any upscales, looks pretty okayish
default upscale workflow 4070 around 60s 2048x2048
fp8 = 42 second average
fp16 = 46 second average
bf16 = 49 second average
for the t5xxl models
I'm kind of blown away at how quick the t5 inference is. Was expecting 5-10 seconds like when using pixart sigma in comfy, but maybe they just don't have the same optimizations
what's happening with my images?
i wonder about that also
What does the ModelSamplingSD3 node do?
this is what clipl+clipg with no t5 (basically "sdxl mode") does:
insane
so still pretty decent for art
how do I make T5 to load only in ram and do not touch my vram?
which sampler?
Looks normal to me ...
haha
I asked this to @simple thistle yesterday, you need to start comfyui in low vram mode
Probably --lowvram
clip mode only
dpmpp_2m sgm_uniform, default settings from the example workflow basically
Can anyone explain the multi prompt node please?
Anyone got the customsampler working? 
Same as the sdxl one just with a slot for t5
You can feed different things to the different clips
I never understood that either 😄
are the SDXL VAEs compatible with SD3? Probably not right?
upscaling makes everything painterly-like
So will i be fine (this is 16gb ram)
no, different number of channels
sadly sd3 doesn't get this prompt right :/ (1st is ideogram and 2nd is SD3)
ill get back to you in 30 minutes when it gens 😂
does it work on a1111 or only comfyui?
can anyone share workflow for sd3_medium_incl_clips_t5xxlfp8. just join the clip to model?
takes about 1m on a 1080ti, just tested
@simple thistle is there a way to combine individually loaded CLIP models? I don't find the node for it
it looks like the vae isn't working properly?
ah nope it went reconnecting
Have the same problem
wrong cfg? wrong negative handling? are you using the example workflows?
did you actually... Read the documentation when running your tests?
G is good at more natural wording, L is good for sd1.5 tags and t5 is llm style very good grammar and wording comprehension
direct use of the example workflow
second example
what about highresfix?
Do you have to use all 3?
Guys where do I need to load the model and encoders
mrrrpp
like there's tiled highresfix
no gravedigging please
but it makes my images looking painterly
how is sd 3 in generating anime characters?
what about SDXL embeddings are those compatible with SD3?
they are compatible with clipl and clip g
It's not SDXL
so partially I suppose

so as long as i have clip_l and clip_g enabled in the triple clip loader it should recognize the embedding right?
?
Where could I find the training code?
only those 2 te inputs will
t5 will ignore your embedding
it got text out of that????
but there you also just load different variants using TrippleCLIPLoader or DualClipLoader
some time measurements with a 6GB card in stableswarm (euler, 20 steps). looks like SD3 is in theory faster than SDXL, but it only works with high resolutions, so in practice it's slower than using a SDXL model that's compatible with 512 or 768
512x512 - 33 seconds, but ugly image
768x768 - 1 min 14 seconds, but still ugly image
1024x1024 - 2 min 38 seconds
huh. it just works.
cool
The thing is I can't find the code for finetune
Help! How do I use an access token when all that is asked for is a username and a password?
Fine tune for "hand" prompt x3
has anyone already finished training a lora for SD3? lol

can someone try to push sd3 to it's limits for realistic images for some reason I am getting wierd results
same
Help! How do I use an access token when all that is asked for is a username and a password?
possible to run sd3 on a 6GB?
Help! How do I use an access token when all that is asked for is a username and a password?
can sd3 run in A1111?
No
yes, scroll up by like 5 messages
only COmfyUi?
can someone help me install it in ComfyUI?
By far yes
the VAE is very nice on faces
Help! How do I use an access token when all that is asked for is a username and a password?
ok thank you
is SD3 in comfy compatible with ip adapters and controlnet yet?
okay, saw this.
wrong number of fingers dude
Help! How do I use an access token when all that is asked for is a username and a password?
bro got 7 fingers on one hand
I'm just concerned because stability doesn't seem to grasp what the model was for, and how SDXL was limited in a way where you can either have "general everything (i.e. base model)" or "really good at something specific (art, anime, etc fine-tunes)"...
you can also use the dualclip if you dont want to use t5
his face is this pixelated, yet it looks this good
this is the 3rd time you've said this
SDXL aspect ratio?
sd3 does not like working with copyrighted characters
I'll repeat it one last time: no gravedigging please. Stop poking.
Yee, smol resolution for that face
What is the difference between fp8 and 16 on that t5 CLIP? I'm wondering if I can get away with fp16 on a 3060... looks big af
ugly
who wanna try 'em?
doing it rn,it works
after Bearer
oi mates, what files I need for comfyui? whole repo?
just use the fp8, you'll never reliably find a difference between them
Do pineapple pizza flavored chips
classic gas station generation
Wasn't there a 95% chance to get perfect hands with sd3?
that claim was directed at 8b
8b hands >>>> 2b hands at the moment.
sussy image
tried on 3060 laptop💀
lol thats funny we generated a mario at the same time
My First SD3 Upsacle
How do I use the t5 on cpu as someone mentioned? I have 64gb of ram and 16gb of vram. Will stable swarm do it automatically or is there a specific way to do it?
yeah lol
shrek a gooner
prompt?
Isn’t 8B undertrained though
2b finetunes will do far better hands. It's still undertrained imo
8b had much more time to cook than 2b
THANK YOU FOR YOUR WORK!!
OOF
Are we getting Dreamshaper?
Dreamworks movie poster of "Shrek 5: The Last Edge".
There is Shrek standing on a rock holding liquid soap in his left hand, and a roll of toilet paper in his right hand. The environment around him is a mountainous cliff side and overcast weather.
lmao how
most realistic ai generated image
That model is gonna be dalle 3 level for sure
sickkk
i think you can make it run t5 on cpu and the rest of the model on gpu
But I was talking about SD3…
for the time we had and all the pressure for a quick release, 2b came out awesome lol.
How?
bruh thats sus
i thought you said the 2b is much more ready compared to the 8b, am i misunderstanding something?
m3 max does this in:
Prompt executed in 85.49 seconds
Medium + clips + t5xxl_fp16
I wonder when we'll see the first sd3 image without flaws
i've noticed the first generation is about 45 seconds for me but then subsequent renders are at 30 seconds, so i'd say times are on par with SDXL
Put this in your prompt and you'll get it
"Cat bot written on a white background"

my wife generating book characters from their wiki descriptions
What is sd3 pipeline :
Encoder/clip -> vae -> unet ?
can you send the json workflow
Sure, because in first gen models are loaded up
you cfg is probably to high or your step count to low
But if it has 95% success with 8B, do you know the reason why it's not being released? It takes longer to upload to huggingface I believe?
im just using the example one cited on the huggings page
cfg 4.5 steps 28?
thats the default setting
opinions
@lavish osprey I ran the model. (With 10GB of Vram) and the whole thing works! ❤️
I just want to know what mode it's running in tho. I want to find out how i know if the T5 model is running off of my CPU or not. There isn't a node available to tell us this yet?
yeah that makes sense it takes into account the time loading it into memory
all my images are higly saturated, anyone knows why? using the example workflow
has other issues. Well, had, newer finetunes solved most of them
you should be more than fine with 16gb of vram to just run it in normalvram mode. i'm on an 8gb gpu and am having zero issues. just use the model that has fp8 and not the fp16 one
ask @simple thistle
SD3 likes low CFG
ah yes perfect hands!
lol
I WANT PLSS
Help! How do I use an access token when all that is asked for is a username and a password?
does controlnet work?
if we release stuff too soon people will start posting stupid failed hand generations
it seems very very good with english text. Not so good with italian
i generated this with local sd3 🤗
you have to start comfy in low vram mode. this can be done by adding --lowvram
seems to do text well enough without t5
sowwy
;c
How to generate images like that wizard image
im using the swarm ui
so uh, how do I download it? I installed git lfs, also got my pw, lost now 😦
thx fixed the saturated images with 3-4 cfg
I’m using swarm because I’m not proficient enough in comfy to know what I’m doing
selfie photo of a man dressed as a wizard on the street of NYC - maybe
cherry pick from a couple hundred 😆
Tokyo, but close
selfie photo of a wizard with long beard and purple robes, he is apparently in the middle of Tokyo. Probably taken from a phone.
what?
and what aboput step numbers?
don't stress, you guys did an awesome release. Thanks so much!
got about 25-30
I'm following the huggingface instructions... but not sure how to get the SD3 files into my computer
only medium?
you can also run t5 in cpu mode. t5 is qute fast so it can run on the cpu
16ch vae
the thumb is slightly longer than it should be, THIS MODEL IS SO DOA
its so joever sd3 bros.... 😔
I wonder what the standard finetune will be like now pony was SDXL’s
thanks. give more examples please 😛
Hey, where do we place the Clips and Text encoders models?
@simple thistle M8. Sorry to bother you, but i want to know what hardware the T5 model is using on my system. I can't seem to tell if it's using my GPU or CPU. I assume it has resorted automatically to using my CPU because i only have 10GB of vram. I am using the fp8 T5 tho, so im not sure.
I didn't have to follow any guides and i got it just fine???
that too, but its still surprising
pixelart is usually bad without some pixelart converter
it's probably running on the GPU
How? 🙂
Pixel art was always held back by shitty VAEs
It worked
now I'm going to integrate SD3 with my custom ComfyUI interface I've been working on
seriously anyone have any idea about this error?
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!
if you use --lowvram then its on the cpu
post the full trace
man this fios came in time
So are ghe weights out. Do I no longer need to use paid tokens to use it?
Thx. Will there be a node to allow us to switch between modes in the future? Just wondered since Pixart has a node that does this, i think.
soon, as I keep redownloading stuffs
do we need to use a refiner?
do I need to clone the whole repo for comfyui?
The text encoders go into your clip folder and the base model goes into your checkpoints
sounds like you're trying to use CPU but one node is trying to use GPU, and they ain't mixing right
torch seemed to be incompatible with xformers==0.26.0
first one
are you using a1111 or comfyui?
bruhhhh
TOOO HOGWARATAS!