🆕｜sd3

[muted dove]

Kiss it!

[vestal cypress]

nice

[rugged nova]

Because lying down is too provocative.

[vestal cypress]

this prompt actually works

[junior dune]

Why is this a child…

[terse belfry]

SD3 🤣 LMAO

[compact forge]

she adorable

[dreamy sundial]

all this wait for nothing

[mortal mesa]

Find the extra finger

[coral sable]

2B is all I ever wanted, if internet shoots down now. I could play with it for rest of my life. Variety is fantastic

[spare pendant]

😄 😮 😛

[karmic igloo]

hello! First off, thanks so much for all your hard work guys and for releasing the model to us to test and play with!

I am pretty interested in it's ability to understand different concepts, such as the example of a blue ball on a red box etc.

Are there any prompting tips for achieving these kinds of results? Say if i wanted to be able to describe each person at a dinner table and not ahve any bleed over.

[brazen harbor]

The SD3 dataset probably didn’t have enough images of humans in diverse enough poses to generalize knowledge of anatomy.

[junior dune]

https://media.discordapp.net/attachments/1124019246201045106/1144001750978797669/attachment.gif

[dreamy sundial]

are you trollin?

[low stone]

because yulia likes to post child suggestive stuff and they're still not banned from SAI's channels for some reason despite it being specifically listed as violating policy.

[muted dove]

First try, no problem!

[dreamy sundial]

finetuned sdxl is miles better rn

[plain knoll]

SDXL Base vs SD3

[muted dove]

It's supposed to be "progress" 🤷🏻‍♂️

[junior dune]

😕

[silver adder]

8b vs 2b

[umbral merlin]

wow shes so beautiful, whats her skincare routine

[vestal cypress]

is this for the grass wall??

[spare pendant]

looking forward to use 8b xD

[agile glen]

I think putting a cap on the number of generations is the only bad thing about the license. It’s not something you can realistically track and is basically telling someone how much they can use their own compute.

[junior dune]

Where can I find the 8b model

[opal ruin]

I really hope something isn't right with the settings, this is just too bad for anatomy :/

[analog matrix]

in order to fine tune, does Stability need to release something?

[spare pendant]

only via api

[compact forge]

something radioactive

[analog matrix]

or the Kohya dev just needs to figure it out?

[silver sluice]

isn't there a way to bypass triple clip loader? I thought someone said on here that I can use a double clip loader to use the clip inside the checkpoint ?

[odd kernel]

Thank god we still have SDXL Turbo and loads of LoRa and Models to play with , this just wasnt worth the wait. I can add text with The Gimp and be done 10x faster.

[karmic igloo]

Where is the 8b?

[silver adder]

you can use it for free on glif

[dreamy sundial]

looking fordward to 8b

[silver adder]

https://www.glif.app

[digital pumice]

What token length does sd3 support for prompts?

[plain knoll]

yes. same prompt. you can see it in the bottom.

[vestal cypress]

also is it just me or are 90% of the women when you generate things w/out specifying race white or asian

[lilac latch]

another amputee...

[vestal cypress]

makes me wonder about the training data

[junior dune]

R they ever gonna drop its weights

[vestal cypress]

grass wall gang rise up

[plain knoll]

prompt?

[dim fiber]

Not amputee, just a bit different

[spare pendant]

don't want to use something like that online .. for this I use midjourney in "production" 😄

[silver adder]

there's many apps there that use sd3 8b, and they come with prompt processors and sdxl upscaling

[dreamy sundial]

maybe in august ppl say, we dont know

[silver adder]

ah, i use the service for fun anyways

[thorny jungle]

ah im in the right place at last 🙂

[terse belfry]

girk gang, in action movie poster , an oil painting, in 80s, rivets, long hair, in the style of mike deodato, luis royo, the neutral colors, paul hedley , happy, in city, movie poster Style: Neon Punk
Style: Neon Punk

[lilac bridge]

So it looks like sd3 still works with the FaceDetailer node 😄

[opal ruin]

nice

[dim fiber]

That one is a simple one

A Photograph of a woman in the snow wearing a short skirt

[thorny jungle]

says alot I came to the discord instead of making more images...

[spare pendant]

we want to test and experiment using controlnet, openpose etc. for poses like handstands, breakdance moves etc. that would work for our image concept 😄

[terse belfry]

girk gang, in action movie poster , an oil painting, in 80s, rivets, long hair, in the style of mike deodato, luis royo, the neutral colors, paul hedley , happy, in city, movie poster Style: Neon Punk
Style: Neon Punk

[silver sluice]

alright I figured out how to bypass the triple clip you just bypass it that's all if the model has clip it'll work

[lilac latch]

Nice. I'll have to switch over Comfy UI then.

[junior dune]

thanks

[opal ruin]

I think clipdrop.co switched to SD3 2B and not using 8b api anymore... wonder how I know xD

[silver sluice]

LEFT is SD3 and RIGHT is afroman sd3 (they look exactly the same before I bypass clip)

[rain current]

Same seed 123456, and yes, it does different things. Prompt: A person wearing a hooded garment, with their face obscured by a dark, textured substance that appears to be melting or dripping. The substance is speckled with bright red and orange particles, giving it a fiery appearance. The person's hands are pressed against their face, further concealing their identity. The background is blurred, emphasizing the foreground, and there are floating particles or debris scattered throughout the image

SD3 BASE <> AFRO'S

[lilac latch]

the horrors of war

[silver sluice]

LEFT is afroman RIGHT is SD3

[silver adder]

hopefully there's going to be a sd 3.1 that fixes some issues (especially with people doing poses and laying on the grass haha)

[vestal cypress]

better every time

[silver sluice]

i like AFRO's better in mines and your example hbu?

[plain knoll]

SDXL Base vs SD3

[lilac bridge]

Hmm.. Something about using FaceDetailer causes next generations to slow down immensely though. Went from 1 it/s to 20s/it

[terse belfry]

girk gang, in action movie poster , an oil painting, in 80s, rivets, long hair, in the style of mike deodato, luis royo, the neutral colors, paul hedley , happy, in city, movie poster Style: Neon Punk
Style: Neon Punk

[lilac bridge]

Wouldn't recommend using FaceDetailer at the moment

[unkempt grove]

Screenshot of a facebook post captioned "in my meth era" on top. The account who posted it is named "John Doe". The man in the photo is shirtless. There is a US flag in the background.

[muted dove]

from above, a woman laying on a grass field taking selfie photo

[storm saffron]

It's probably a lora merged into the checkpoint, it'll use the CLIP part of the lora to 'enhance' it, I've been testing loras myself

[lilac latch]

another gore image...

[opal ruin]

So celebs are fully deleted as well huh? My SDXL custom model vs SD3

[hexed dirge]

it doesn't know styles like Deodato. It doesn't know any style

[dim fiber]

It seems like a lot of pose phrases like "Lying" "Laid" even "Sitting" creates meat blobs

[muted dove]

That worked 🙂

[opaque cargo]

It's pretty good with artistic images

[coral sable]

are you trollin? It's not even in same universe

[vestal cypress]

ooh how'd you get that

thats really good

[silver sluice]

that's how he pulled it off so quickly? doesn't stuff like this take a while to process though?

[plain knoll]

SDXL Base vs SD3

[storm saffron]

Same seed, using the Cube Shaped Anatomy SDXL lora first, bypassed 2nd.

[muted dove]

Luck with from above, a woman laying on a grass field taking selfie photo seed 602204995
neg: featureless, colourless

[silver sluice]

AFRO's version of the same prompt, it's still a dirty mess

[lilac bridge]

Which checkpoint is everyone else using?

[silver sluice]

the right is better i'd say

[hexed dirge]

t5 one

[crystal blaze]

They're all identical, the lower two include CLIP/T5.

[plain knoll]

SDXL Base vs SD3

[noble coyote]

All 3

[topaz flume]

it is explained here: https://comfyanonymous.github.io/ComfyUI_examples/sd3/
SD3-medium-inc-clip & SD3-medium-inc-clip-T3XXL is standalone
SD3-medium required the text encoders be put into the /clip/ folder
if you have RTX GPU you should be ok for the FP8, check

[storm saffron]

Here's another one. cnmtq cinematic scene with darth vader on the death star first pic is with my Cinematique LORA, and 2nd is without. Same seed (769925201549290)

[vestal cypress]

i ended this abomination before it got decoded

[dim fiber]

I didn't

[vestal cypress]

NOOOOOOOOOOOOOOOOO

[terse belfry]

SDXL Vers.
PROMPT:
girk gang, in action movie poster , an oil painting, in 80s, rivets, long hair, in the style of mike deodato, luis royo, the neutral colors, paul hedley , happy, in city, movie poster Style: Neon Punk
Style: Neon Punk

[opal ruin]

damn, I need a minute with this one

[wintry stag]

she survived

[stuck hedge]

hey, someone know which folder i suppose to use for TripleCLIP? i try to create a TripleCLIP folder but it's dont show up in comfy

[lilac bridge]

Put the clip files in the clip folder :]

[dull star]

weren't wheels problematic before? this looks okay

[gentle folio]

put the encoders in models/CLIP folder.

[stuck hedge]

ahaha so simple, okay thanks for your awnser 🙂

[opal ruin]

Does this count?

[dull star]

highresfix is so goddamn good on SD3 Base

[keen palm]

Time for Burgers

[hexed dirge]

we need finetuned models as soon as possible

[muted dove]

With USDU?

[turbid grotto]

hiresfix by sdxl?

[plain knoll]

SDXL Base vs SD3. prompt "a woman resting on a grass lawn. she is laughing. she is fully clothed. this is a safe for work image." Why does it almost generate NSFW but still body horror. How is censorship or under

[runic tusk]

Umbrellas still suck.

[agile glen]

Now do it again but for sitting on chairs

[dull star]

nope, base model, but with ultimate SD upscaler so it uses tiling

[raven fern]

kek

[dull star]

the trick is, to use shift 1 instead of 3 when upscaling

but ONLY when upscaling

when you make the base image, use 3

[opal ruin]

You upscaling with SD upscale?

[dull star]

as good as the VAE is with small face, nothing beats highresfix

[lavish osprey]

they'll say you're a SAI shill now

[analog matrix]

how long till training code is released my friend?

[dull star]

you are a SAI shill now

[dull star]

nice motion blur

[muted dove]

https://tenor.com/view/moonflip-shillteam-shillteamassemble-mflip-assemble-gif-24959351

[opal ruin]

You can't just willy nilly lay a woman down on grass. You first have to take her out to a restaurant, wine and dine her.

[sterile pendant]

you can use ultimateupscale with the sd3 model. just set the tiles to the size of your original image. so if you were using 1152x896, you'd set the tile sizes to match that (that way it's just creating a perfect 2x2 grid if you 2x the image) and set the denoise low like 0.15-0.2

[runic tusk]

https://tenor.com/bVV8X.gif

[noble coyote]

Shemales

[lavish osprey]

maybe it already is? Check Diffusers. I'm on vacation

[plain knoll]

SDXL Base vs SD3 prompt "from above, a woman laying on a grass field taking selfie photo"

[noble coyote]

Baby Reindeer!!!

[lavish osprey]

ask @dim fiber or @muted dove or @gentle folio for their settings man 😄

[noble coyote]

SD3 s t r e t c h e s people out

[past flame]

lovely

[dull star]

lol

[silver adder]

yo, is sd3.1 a possibility

[karmic igloo]

how do you know which ones are actually using SD3?

[dull star]

NICE

[lavish osprey]

for fixing 1 stupid prompt the whole community is focusing on?

[muted dove]

In 2 weeks

[lavish osprey]

(which apparently doesn't even need to be fixed)

[silver adder]

perhaps lol

[little quarry]

Two more weeks

[dull star]

Save me....

[past flame]

Why? Also didn't we establish there are a lot of prompts having issues?

[silver adder]

just type up "sd3" in the searchbar and you'll see a bunch of spaces of em

[dull star]

I think we should stay away from the subreddit for a good while

[sick cedar]

Does anyone know what the difference is with T5XXL e4m3fn?

[dull star]

anyone claiming that SD3 isn't utter rubbish on there is going to be exiled no questions asked

[lavish osprey]

what I've seen now is a bunch of people getting monstruosity and some people succeeding.

[real terrace]

hands fixed in afterlife

[lavish osprey]

since the model is the same in both cases

draw your guess

[dreamy sundial]

bro..

[floral fox]

Wait. Is that what's happening? Everyone is trying to make people lying on the grass?

[past flame]

Alright, you just made it sound like the issue people encountered doesn't have to be fixed

[dreamy sundial]

bro...

[turbid grotto]

What do you think, won't finetuning break diversity of sd3?

[sick cedar]

You were talking to me? xD

[past flame]

to me I think

[turbid grotto]

I hope I can try finetuning with 12gb

[muted dove]

Mine was a lucky hit, most of those laying down ones are just bad.

[past flame]

uh oh, you can't say that, it doesn't fit the narrative

[sick cedar]

"T5XXL e4m3fn" Does anyone know why it has that name? What does the "e4m3fn" mean?

[unkempt grove]

Reduce the CFG down to get rid off the AI look/

[lunar rivet]

https://onnx.ai/onnx/technical/float8.html

[coral sable]

keep going with head trough the doors while there is a doorknob.

[odd basalt]

Thanx I figured as much that I went way to high on that

[rough knot]

idk im getting decent hands

[muted dove]

https://tenor.com/view/yin-yang-gear-gif-25046753

[past flame]

New VAE is indeed amazing, the CLIP models are doing wonders for keeping track of more complex prompts, but the 2B model seems to be really struggling seemingly at random

[dreamy sundial]

yeah youre deffinetly trolling

[low stone]

I like to think I know what I'm doing but I can't get a good image with that prompt out of sd3.

[terse belfry]

80s anime still, girl fixing a mech, retro fashion, muted retro colors, style of Dragons Heaven

[low stone]

@muted dove this is all i get

[dim fiber]

Something isn't right here

[sick cedar]

Oh! I see. Thanks. Thought it was something else that i wasn't yet aware of. xD

[rough knot]

prompt?

[muted dove]

It seems to do that a lot. Loses the arms and puts legs/feet above the head.

[lilac latch]

beautiful hand

[unkempt grove]

I love this model

[dreamy sundial]

the anatomy of everything is messed up whe doingn actions or poses

[low stone]

did you paste that prompt into just t5 or all 3? I'm only using clip_g and t5 with clip_l blank.

[real terrace]

Saint Jerome is jacked

[haughty veldt]

slick

[muted dove]

Only using the T5 field and leaving the others blank

[dreamy sundial]

how tf did you do that

[muted dove]

Too early?

[low stone]

do you feel that covers enough topics? when I was trying that before, it often didn't know who people were etc.

[lapis bay]

just drag png into comfyui

[lavish osprey]

Imo needed probably 2-3 months of training to make encoders converge on more complex things. Better to release it and move to 8b.
Will give time to the community to catch up on the new architecture

[muted dove]

I don't know what the others are supposed to do, but had best results only using t5 so far

[low stone]

with t5 alone.

3 out of 4 good

[last bloom]

it's good for commercials :p

[dreamy sundial]

mm, interesting, 8b at the moment looks pretty good

[hexed dirge]

I love SD3 but it's not 1 stupid prompt. Lot of bodies not only the ones laying come out malformed

[lavish osprey]

for the cat, the best one was clip g.

[odd basalt]

😂

[dreamy sundial]

this looks like sdxl

[lilac latch]

the horrors of war #2

[haughty veldt]

god

[restive halo]

so, Im guessing all the bad renderings are because we dont know the right settings for prompting yet?

[muted dove]

It's as good as SDXL at archery

[lavish osprey]

could be tes fighting each others sometimes

[dull star]

OH CRAP

[unkempt grove]

harry took the lookmaxxing class

[low stone]

what lykon is now saying, is that some stuff is going to look better with clip_g, some stuff is going to look better with T5. the problem is that the different encoders are fighting. what looks great with one, will then get messed up by the other one that's not as good. so with each prompt, we're supposed to copy/paste back and forth around the different encoders to see which one has the training. long sigh.

[past flame]

Yo guys she is actually sitting ON the bench with SD3, not that weird unnatural thing SDXL does

[turbid grotto]

i can fix her

[lavish osprey]

you need to train A LOT to make the tes converge on all prompts (or restrict the model scope).
For simple poses they don't fight, for more complex ones or things that the model has seen less, they'll conflict.
This doesn't happen as much with 8b because it's bigger and was trained for longer.

[odd basalt]

I just remade linkinpark album cover

[unkempt grove]

lower the cfg

[lavish osprey]

see #🆕｜sd3 message

[shell plaza]

k

[low stone]

I appreciate the candor. I've been doing local vs api for a bunch this morning, and the 8b on the api isn't perfect but the shenanigans count is just night and day lower.

[haughty veldt]

perfect form PR

[lilac bridge]

AH didn't realize there was a new sd3 text encode node. Niiiiice

[muted dove]

Same prompt and seed in T5, Clip_l, Clip_g
a photo of a woman reaching out to the viewer, she is smiling and a funfair can be seen in the background. It is a grey rainy day and she is completely wet

[shell plaza]

3

[lavish osprey]

we now want to transfer 2b image clarity aesthetic to 8b.

[safe hamlet]

Don't prompt for a watermelon eating marshmallows. Worst mistake of my life.

[coral sable]

Hey, good to see you! I hoped there would be some SD3 launch event that you would host, as the SDXL one was amazing. But nope . Anyway, you can try this prompt structure, just change it to anything else you like. Prompt: Three girls in a hotel lobby: the first one is blonde with red dress, the second one is ginger with white dress, the third one is brunette with green dress

[unkempt grove]

See? It toned down the AI look

[lilac latch]

ah, the left hand...

[past flame]

So what you're saying is 2B isn't all we need after all?

[lavish osprey]

no, I think that 2b is all you need as an architecture

[low stone]

seeing as I already think 8b is awesome, i have high hopes.

[dull star]

Jo Biden...... 😔

[past flame]

I can't deny at least the fact that the VAE is exquisite though

[untold valley]

guh

[lavish osprey]

by the way, most stuff should look better with all 3

[haughty veldt]

i'm guessing handstands are off the table

[compact forge]

maybe decreasing the strength of clip_g and clip_l could help?

[past flame]

Don't people get the weird blobs when using t5 only as well?

[muted dove]

What the hell is clip_l doing here?! 😄

[ionic geyser]

compositions are great and look way more interesting - cool. prompt adherence is next-gen.

but the image quality of the outputs looks very compressed. maybe I'm doing something wrong but there are many compression artifacts or halos around objects - at least for me.

[muted dove]

Same prompt in all 3?

[thin fable]

how is this acceptable to release?

[signal iron]

I've only tried the sd3+t5 combo so far. Could be better, could be worse. It's ok though.

[unkempt grove]

Most of the weird blobs are because people are still writing prompts like they did for earlier models. You must use natural language for SD3.

[lavish osprey]

not necessarily, but yes

[barren cobalt]

I'm curious how long 8B takes to do 20-30 steps on an A100. Is it way slower than 2B? 10 seconds?

[icy drift]

We can do non-centered subjects. 🙂 40% hitrate.
A sinister figure lurks in the shadows on the right side of the painting. In the center of the painting, an empty alleyway extends into the infinite distance, receding into fog.

[terse belfry]

SDXL still rocks 💜💚

Prompt: Britney Spears DSLR cinematic shot, blade runner style, oops I did it again , photorealistic photo of a pink haired singer laying on a grass Coachella, after-party, messy hair, , bikini tall grass

[odd basalt]

@lavish osprey how well is it with dark horror I’ve only played around with hp Lovecraft fused with jack skeleton

[dull star]

you guys liked this game back in the day too? it was my favorite PS1 game

[vapid frost]

I have the same problem on my local machine. Colab doesn't have the same issue for me. Something connected with negative conditioning

[lavish osprey]

there isn't much speed difference compared to 2b

[plain knoll]

SDXL Base vs SD3

[barren cobalt]

Really? Thats great to hear

[dull star]

whaat???

no way there's not much speed difference

damn

I mean that's good news

[ionic geyser]

hmm interesting. I will investigate this! thank you

[dull star]

at least 8B won't be a chore to use

[muted dove]

It is a bit better...but...

[terse belfry]

Taylor Swift DSLR cinematic shot, blade runner style, oops I did it again , photorealistic photo of a pink haired singer laying on a grass Coachella, after-party, messy hair, , bikini tall grass

[dull star]

I'm too used to decent speeds such as 2-4 it/s

[past flame]

Hopefully

[lavish osprey]

well, I won't bet a hand on it, I always use them on H100 so I don't really notice 😄
I guess if you have vram offloading to do you will see a huge timesink

[dull star]

I wonder when 8B will be finished... September? October? Later?

[thin fable]

this is a late april fools joke right?

[gentle wolf]

working as intended

[barren cobalt]

I mean as long as 8B is less than 2x slower than 2B, it's still good news for us

[past flame]

Apparently one of the three clip models it uses is absolute dogpoop

[terse belfry]

Britney Spears DSLR cinematic shot, blade runner style, oops I did it again , photorealistic photo of a pink haired singer laying on a grass Coachella, after-party, messy hair, , bikini tall grass

[haughty veldt]

your kid calls you at night and you go check on him an see this, what do you do

[lavish osprey]

but to be honest if I were to decide I'd think twice before releasing a model if that's how the community reacts (it's just my opinion, not SAI)

[sick cedar]

If 2B still requires months of training, then it's anyones guess how much could be added to 8B. My guess... Maybe early next year? Idk...

[obtuse osprey]

for a base model, its better than XL at understanding. Now, what we need is some videos on how to finetune it so people stop bitching and comparing 3 to the latest finetune of XL

[cinder light]

Are there any news about training/finetuning out there yet? I haven't found anything related on Reddit or here so far.

[barren cobalt]

I'm coming from a pretty specific set of needs (limiting latency for www.wand.app) so I wouldn't base my perspective on the community as a whole or anything

[past flame]

Thank god it's not a man getting photographed there

[lavish osprey]

when we released 8b on API, 2b was in this state:

this should answer your question

[spare pendant]

a realistic and high quality portrait of a beautiful young woman with european ethics. professional photograph of the woman in her minimalistic apartment. The wall behind the model is painting in a pale purple color. The woman is wearing a dark yellow turtleneck and white troussers. She is holding her smartphone in the left hand. Thephoto is taken from a side-view and the woman is looking onto her smartphone. Her hands and fingers are well proportioned considering her appearance. the photo is a high quality photo. the photo is shot from a wide-angle view. The photo and textures are in 8k resoluion.

[muted dove]

People issues

[past flame]

That does change a bunch

[lavish osprey]

(and also should help you guess how much we worked on this release)

[gentle wolf]

oh so 2B is actually a failed state, "go boys get it" move?

8B the real one?

[keen palm]

It's unfortunate that an early version of 8B wasn't released then instead

[coral sable]

or learn how to prompt to start with? Model just released

[plain knoll]

SDXL Base vs SD3

[thin fable]

that's the current state of 2b as well. look at the mutants it makes

[compact forge]

protip: reduce the strength of all clips so they work together rather than fight eachother

[silver adder]

ultimate combo

[obtuse osprey]

This all happened with the release of XL also. people bitched a lot when it came out, but those were just the noisy ignorant folks. I see this model with far more potential than XL base had for sure.

[karmic terrace]

no matter what, people will always complain in open source releases. if something is free this is the fate, but community can be educated, i hope

[subtle valley]

3d render of n64 mario standing smiling smoking a cigarette in his mouth with smoke, looking at the viewer, pitch black background

no negative prompt

first is from glif stablediffusion 3

second is from stable diffusion 3 medium super prompt huggingface space with prompt enchantment disabled 25 steps 7.5 guidance scale seed 384575511

[karmic terrace]

absolutely agreed

[lavish osprey]

that same prompt/seed now

[gentle wolf]

I see the potential, I see some things work much better. But is the deformation because of misuse or is it unfixable?

[terse belfry]

SDXL 1.0 Base

Red lips Beyonce Knowles DSLR cinematic shot, blade runner style, oops I did it again , photorealistic photo of a haired singer laying on a grass Coachella, after-party, messy hair, , bikini tall grass

[obtuse osprey]

what now? no shit. I am asking if there is finetune info out yet.

[lilac latch]

not SD3, though

[terse belfry]

Beyonce Knowles DSLR cinematic shot, blade runner style, oops I did it again , photorealistic photo of a haired singer laying on a grass Coachella, after-party, messy hair, , bikini tall grass

[turbid grotto]

guys, won't finetunes destroy variety of faces in sd3?

[lavish osprey]

you can say all you want about anatomy, but that's terrible image quality

[coral sable]

ask someone else, I will not sue finetunes. Base is all I need

[last bloom]

Some old dall-e 3 prompts copy pasted

[sick cedar]

Training does happen at a fair pace then. Yeah. 8B might actually be ready sooner than i thought it might be. You doing a great job anyway guys. I can see 2B easily outperforming SDXL with more training. It's got genuine potential!

[plain knoll]

SDXL Base vs SD3

[lilac latch]

Wondering the same question as well...

[lavish osprey]

I can see 2B easily outperforming SDXL with more training.
It already does on basically all serious benchmarks...

[muted dove]

How did you get her to look like TS?

[thin fable]

same prompt as the sdxl beyonce but from sd3

[terse belfry]

I'm just using Huggingface google Space for SDXL 😂

[lavish osprey]

judging by how blurry that is, it was made with base xl

[thin fable]

lol

[terse belfry]

Full body Rihanna riri badgalriri crazy in love lemonade DSLR cinematic shot, blade runner style, oops I did it again , photorealistic photo of a haired singer laying on a grass Coachella, after-party, messy hair, , bikini tall grass

[lapis bay]

there is only one way to find it out... try it out

[white basin]

hows sd3 guys? yay or nay?

[lilac latch]

Not SD3...

[terse belfry]

LMAO 😢

[muted dove]

Yes, they're not SD3 anyway....in grass issues with that

[signal iron]

It can't spell "Sally" to save it's life though

[gentle wolf]

maybe every comma in the prompt creates one extra joint 😄

[sick cedar]

Yeah. The community has a good head start with that as base too.

[mortal mica]

Just look at photos being posted😂

[obtuse osprey]

a bit like how XL was at drop...not great, but ripe with potential for finetunes.

[terse belfry]

Will share SD3 versions for comparison next 😂

[teal fossil]

Why the FP16 and not the fp8 one?

[lavish osprey]

this is curious

[turbid grotto]

yea, i mean like everything is awesome, it just need anatomy and pose improving

[untold valley]

i felt SDXL was decent at the start

[errant dust]

both clips and fp16

[rotund stump]

comfyui has a node that separates the 3 text encoders so you can prompt each individually, what's the use case for that?

[lavish osprey]

prompt?

[compact forge]

full body shot of cute pale skinned european 18yo teen at the nightclub with {platinum blonde hair|blonde hair|ginger hair|black hair|brown hair|pink hair} and with big lips wearing pink lipstick and eyeliner, {smirk|happy|staring},

4k, hdr, 2160p

neg: bad quality, poor quality, doll, disfigured, jpg, toy, bad anatomy, missing limbs, missing fingers, 2d, drawing, asian, ugly, dark skin, tanned, acne, fat, obese, mature, old

[icy drift]

I go away for 10 minutes and there's 50 new posts.

[thin fable]

i agree its too early to panic, sdxl was garbage on day one, but sdxl day one didn't make people look like Star Trek Transporter accidents

[signal iron]

It does a lot of things better than sdxl did. Prompt adherence is much better, text is much better. The stuff it's bad at, sdxl was also bad at. So I assume it will end up good. My question is: s it better than stable cascade?

[silver sluice]

thats not sd3 right?

[sick cedar]

Oh! And it was able to do that Back of the violin prompt too! I tried this prompt out and SD3 got it perfectly first time! That truly blew me away. Before now, i only saw models like Deep Floyd successfully do a prompt like that.

[brittle nexus]

That dog is having Vietnam flashbacks

[dusky thistle]

@lavish osprey i'm guessing unsampling is no longer a thing with SD3?

[lilac bridge]

Trying to see how changing different clip prompts affect image. I haven't learned anything so far xD

[cunning lintel]

Well, that's one way to look at it, when announcing 2b you said things like

@lavish osprey4072I personally don't think anyone needs 8b including us. With less training time and much less resources, 2b beats 8b in some benchmarks. Sure 8b is objectively superior in terms of potential, but the cost is much higher.

Things like that raise expectations, esp when the 8b model is what people know. I think no one can objectively say what 2b produces comes close to 8b, maybe in a few really narrow cases, but it's really really challenging to find gens that stand out using 2b, even pixart often looks nicer. Neither does it help to act o defensive, comments about skill issue and such, not helping, if it's skill, then teach people, show how to get the good gens! I'm clueless as to how, and mamy with me.

[terse belfry]

SDXL 😂

[icy drift]

Best adherence out of 8.
From left to right: An frowning old man, a sleepy young woman, and a laughing toddler. The old man is wearing a red fedora. The young woman is wearing a blue jacket. The toddler is holding a stuffed bear.

[thin fable]

which image?

[fleet meteor]

Yooo its epic, do you have the prompt of that image?

[teal fossil]

Pretty interesting conditioning. That works well?

[plain knoll]

SDXL Base vs SD3

[open fossil]

A photorealistic monkey inside a claymation style city

[lavish osprey]

I'd add photo in front

[patent magnet]

Left is Ultra (API) and right is local, t5 + clip:

[old parrot]

lmfao

[gentle wolf]

same picture

[lavish osprey]

why not?

[silver sluice]

the one with rhianna

[mortal mica]

I can't wait to get home so I can use the model and create a monstrosity like you guys are😂 oh it's gunna be fun

[lavish osprey]

Ultra is the best

[ionic geyser]

I went back to the official ComfyUI SD3 example workflow. The prompt was: a woman dancing in the streets. The jpeg like compression artifacts are sadly all over the image.

Not sure if I'm doing something wrong here.

[lilac latch]

The second one is better.

[hexed dirge]

but you said Ultra is just a workflow and not a model

[little lance]

15 steps cfg 2.5

[thin fable]

no i think those are sdxl. my mutant beyonce is sd3

[lavish osprey]

nice find

[patent magnet]

But why didn't they release it to the public then?

[errant dust]

Keep in mind that you need the fp16 to get best results and the prompts in SD XL will need to be completely redone in SD3

[karmic terrace]

eular?

[lilac latch]

a "middle-aged" woman

[old parrot]

https://tenor.com/view/money-mr-krabs-gif-18326632

[lavish osprey]

Ultra is not a model, it's stuff like Dalle or Ideogram

[next fossil]

Gotta eat some potatoes

[plain knoll]

Im using f16t5. I used ChatGPT to expand my prompt.

[little lance]

im trying with some different settings, model sampling on high and sampler on low

[karmic terrace]

agreed, im prompting in sentences, not keywords. LLM generated ones are also quite good

[lavish osprey]

it's all already released basically lol

[patent magnet]

I think Ultra if it had been released to the public would solve all of the issues.

[burnt cove]

use node superprompt, it expands very well

[worthy trout]

actual answer: no one would be able to run the darn thing

[plain knoll]

SDXL Base vs SD3

[shadow pawn]

you mean using SDXL lora with SD3? I`ve tried both 1.5 and XL LoRAs with SD3 with the ComfyUI workflows, and I've got the "lora key not loaded: lora_unet_up_blocks" error on all tries.. How did you managed to make Loras work with SD3?

[lavish osprey]

no because you wouldn't be able to run it lol.

[terse belfry]

SDXL< >SD3

Prompt: girk gang, in action movie poster , an oil painting, in 80s, rivets, long hair, in the style of mike deodato, luis royo, the neutral colors, paul hedley , happy, in city, movie poster Style: Neon Punk
Style: Neon Punk

[dusky thistle]

guess i need to try again then

i was getting a black image with flipped sigmas

[plain knoll]

i will look for SuperPrompt node. thanks.

[subtle valley]

kirby smoking a fat blunt

no negative prompt

first is from glif stablediffusion 3

second is from stable diffusion 3 medium super prompt huggingface space with prompt enchantment disabled 25 steps 7.5 guidance scale seed 1062578693

[rugged nova]

6000 ada not enough?

[patent magnet]

Ok, a workflow you said. I don't understand, but that's ok. It would'nt run even in the A100?

[haughty veldt]

I mean sometimes it does deliver

[minor spindle]

any word from stability about this pile of junk. Realise that there are also company's using SD and where wiling to pay good amounts for it but with the mess up most will stick with SDXL or jump to a different ship

[subtle valley]

kirby

no negative prompt

first is from glif stablediffusion 3

second is from stable diffusion 3 medium super prompt huggingface space with prompt enchantment disabled 25 steps 7.5 guidance scale seed 1682567628

[viscid pivot]

do we have a clear answer on how many tokens a prompt supports?

[worthy trout]

You'll have better luck having SD3 Ultra getting released to DreamStudio than it actually being released. The reason why they didn't release it wasn't because of money, it was because they knew only a small handful of people could run it

[raven fern]

idk who it was, but someone did mention to lower the cfg

[lavish osprey]

true

[patent magnet]

I run stuff on colab and would gladly pay to run Ultra on a A100 over there (easy these days). Well, for SD3 stuff I will use the API then...

[lavish osprey]

anyway, releasing Ultra would be like releasing Dalle. It's not a model.

[still sundial]

these artifacts are annoying - zoomed in, some seeds are ok

[subtle valley]

roblox avatar with a black low taper fade afro, wearing a black suit and a white buttoned shirt, holding a phone with a wood case on it, taking a selfie in the bathroom mirror

no negative prompt

first is from glif stablediffusion 3

second is from stable diffusion 3 medium super prompt huggingface space with prompt enchantment disabled 25 steps 7.5 guidance scale seed 2670674994

the roblox style literally vanished lol

[lilac latch]

git gud!

[hexed dirge]

all styles vanished 😦

[full veldt]

😭😭

[raven fern]

booba vanished

[signal iron]

I've got 128GB ram on my macbook, I'll take one Ultra Model + Workflow please. And probably wait 5 minutes per image.

[patent magnet]

Yeah it will probably never happen

[lavish osprey]

making good anime was definitely a priority over roblox (which might also have copyright by the way)

[wind basalt]

Quick question isn't this all solveable since we have the weights just need to custom fine tune ?

[raven fern]

im trying to figure out some good params

[lavish osprey]

yep

[hexed dirge]

it's not roblox. I tried lot of illustrators without success

[lilac latch]

All three text encoders (T5 FP16) along with the 8B model will take over 25-27GB of VRAM easily, not to mention RAM usage. Very few people will use SD3 8B.

[subtle valley]

minecraft gameplay, screenshot, hud

no negative prompt

first is from glif stablediffusion 3

second is from stable diffusion 3 medium super prompt huggingface space with prompt enchantment disabled 25 steps 7.5 guidance scale seed 3038793424

wtf

[lavish osprey]

our task was just to provide a good enough base model in the shortest possible time. I think we did it.
This was the shortest training from scratch in our history. Had a fraction of the time/compute of cascade.

[radiant ledge]

you can run text encoders on the CPU with no issues

[viral badge]

@lilac latch cant text encoders be ran off cpu ? most llms can

[rugged nova]

4 bit T5 works fine

[wind basalt]

It does work somewhat

[lavish osprey]

8b knowledge >>>> 2b knowledge

[lilac latch]

quality is far worse with t5 fp8, i'm sorry

[worthy trout]

you would need an a5000 to run it

[hexed dirge]

I think 2b will be enough after finetuned models

[wind basalt]

So far is it worse than even sdxl ?

[abstract nymph]

fine tuning code going up today?

[viscid oar]

Hi, I'm trying to figure out how to run sd3 in comfy. The example workflow gives me 3 missing nodes that seem to be sd3 related.
Any idea where I can get those?
Or is there another way to do this?

[radiant ledge]

that's why you run f16 on the CPU

[honest swift]

Hi guys, im new here, cant understand how to generate pics via bot or somthing can you help me?

[digital barn]

That truly is disappointing

[still sundial]

update comfy

[dim fiber]

You need to update Comfyui to get the nodes

[coral sable]

I was told SD3 2B will not run on 12GB VRAM, and it can work with less then 5Gb. So thats that

[viscid oar]

oh ok thank you. I thought it updates on start for some reason.

[rugged nova]

even if so T5 can run in ram

[lavish osprey]

fp16 package should be already uploaded, or will be shortly

[worthy trout]

i would wait until there is a 3.1 of sd3 medium or until someone makes a finetuned model that is trained on sd3 ultra images

[subtle valley]

minecraft gameplay, screenshot, hotbar below with heart bar and hunger bar, the player's hand is showing on the right, theres a pig on the front

no negative prompt

first is from glif stablediffusion 3

second is from stable diffusion 3 medium super prompt huggingface space with prompt enchantment disabled 25 steps 7.5 guidance scale seed 3745968304

[signal iron]

Is the t5 model (by itself and not with the clip l and clip g) working in the comfyui sample workflow? When I run the combo sd3+t5 model everything works fine and the triple clip (t5 + clipl + clipg) thing works. But if I switch to the fp16 t5 clip only it forgets the movie title or what Jason looks like. Super weird.

[lavish osprey]

wait are these both 2b but one is from hf?

[subtle valley]

i dont know how in the actual living fuck it managed to unintentionally generate a pp

[subtle valley]

no

[faint breach]

this is the new vae in action wow

[subtle valley]

2nd is from hf

first one is from glif

[lavish osprey]

also please don't ever consider the HF space as a metric lol

[viral badge]

Is there an acklowledged issue with this SD3 medium or is it working as intended ?

[subtle valley]

https://glif.app/glifs/clv488uy10000djtrx70u03no this one

[wind basalt]

Is SD3 medium worse than sdxl ?

[faint breach]

people know but they'll still do it

[lapis bay]

just use comfyui manager and install missing nodes

[lilac latch]

sdxl base? nah

[still sundial]

jpeg, jpg negatives do nothing, jpeg like artifacts all over the image :/

[raven fern]

i did notice artifacts as well

[pseudo owl]

better in some things(text and prompt following), worse in somethings like human pose and stuff

[wind basalt]

So it's not fully matured yet through fine tuning right

[lilac latch]

oh my...

[real terrace]

well yes that's star wars starred by dogs

[pseudo owl]

what schedulers are you guys using?

[tough oriole]

All honesty Outside of humans this thing is pretty good for a base thanks for the work.

[radiant ledge]

VAE finally good enough to reproduce them

[pseudo owl]

vae is too good

[dusk silo]

if i want to run SD3 Without text~~, do i need to use clips?~~ can i use clips

[lilac latch]

complex poses are a no no for now, anatomy isn't quite right in most cases. it takes a few tries to get a decent looking picture. you'll have better chances by prompting it right.

[gentle wolf]

Fine-Tuning: Capable of absorbing nuanced details from small datasets, making it perfect for customisation. this this totally makes sense now, with the knowledge the medium model was trained in a short time. Maybe... maybe it's a good thing if more training fixes the deformation.

[dull star]

oh its guizmus

dog

[thin fable]

The problem is how much fine tuning are we going to have to do to get just basic human antomy right? Seems like we might have to literally train a new model

[cosmic meadow]

hello there 😉

[lilac latch]

"safety"

[lavish osprey]

looks like it's working pretty well to me, despite being a base model. Local software quality might differ with skill level.
The purpose of this model is to give people something to finetune. We probably went a bit too far and gave the impression this should compete with older finetunes that had months to cook on top of each other. It's not the case.

[lavish osprey]

heh...

[errant dust]

What do the clip_i and clip_l do? I understand their general purpose but not the specific one.

[faint breach]

comedy and tragedy are two sides of the same coin amirite?

[dull star]

pixelart is really good, but it should not invade minceraft

[faint breach]

🥲

[lavish osprey]

I didn't work con the architecture, but it's explained in the paper

[raven fern]

fintuners will cook something soon im sure

[wind basalt]

It weirdly cannot generate nipples lmao

[past flame]

Isn't it kinda dishonest to keep reposting that bottom right image when even the guy who generated it said he cherrypicked it and most results were also garbage human blobs?

[errant dust]

ok, fair enough. I'll look in the paper then

[worthy trout]

3D render of Sonic and Tails

No negative prompt

First is from the "stablediffusion 3" glif

Second is from the Stable Diffusion 3 Medium super prompt huggingface space with prompt enchantment disabled 25 steps 7.5 guidance scale seed 384623937

The "Ultra" model actually knows who Tails is. The Medium model kind of does but not quite fully

[raven fern]

sanic 😮

[lilac latch]

"muh safety"

[dull star]

meincraft

[faint breach]

people who mock ai safety . so short sighted.

[thin fable]

most humans generated period on sd3 are blobs

[dull star]

oh wow that's nice for a base model

[wind basalt]

But it literally made a fully nsfw pic like with gore 💀 but no nipples

[raven fern]

kek

[lilac latch]

it's amputee friendly

[white smelt]

explain

[viral badge]

I am using SD3 medium as a refiner in my workflow side by side with dreamshaper lightning refiner to compare and I am not getting great results at all

[past flame]

SD3 being uncensored will literally change nothing about the current state of AI safety when a billion nsfw SDXL finetunes currently exist

[viral badge]

prompt adherence to asking for dark fantasy illustration is basically nil

[faint breach]

it's an important discussion that people making this stuff have constantly. short sighted people typically only have the perspective they're offered from the confines of their armchair

[viral badge]

steers towards realism

[raven fern]

@viral plaza are the sd3 training scripts released somewhere?

[subtle valley]

a PC web site screenshot of the roblox website

no negative prompt

first is from glif stablediffusion 3

second is from stable diffusion 3 medium super prompt huggingface space with prompt enchantment disabled 25 steps 7.5 guidance scale seed 45409346

the roblox style came back lmfao

[still sundial]

i would use it the other way around

[lavish osprey]

ok.
#🆕｜sd3 message
#🆕｜sd3 message
#🆕｜sd3 message
#🆕｜sd3 message
Better?

(also don't tell me you didn't notice that the bad ones are also "cherrypicked")

[faint breach]

and to that point, why are people crying about it then? ai porn got maxed out with ponyxl. it's peaked. moving on now

[wind basalt]

Also it seems specifically trained to avoid copyright shit

[lilac bridge]

Testing out CliptextencodeSD3. I'm liking the results of splitting up the prompt like this so far. Time to deep dive! 😄

[white smelt]

that doesn't explain anything. why are those people shortsighted? is it beacause you think there would be legal issues? or marketing? what's it?

[worthy trout]

I think SAI might disallow people to finetune the model because "SAfEtY"

[viral badge]

@still sundial Pixart generates the base image and then I have one refiner as dreamshaper and I added an sd3 refiner in parallel for comparaison

[dull star]

9/10 clip-l just make a huge mess

[faint breach]

woah. sorry that wasn't the answer you wanted. I feel it explained it fine.

You seem to know the conversation's landscape already. This doesn't seem like a good faith approach to conversation now does it

[raven fern]

there is definetely some combination that will work :3

[silver sluice]

I have 245GB of SDXL LORAs installed on my system, i think this is one of those situations where the sum of it's parts is greater than the base, sure SDXL is not technically superior but as a whole it exceeds SD3 currently, just a matter of time before the community picks up on generating SD3 loras and checkpoints and then SD3 would exceed SDXL but until then I wouldn't instantly say SD3 is my faovirte model to use out there

[lavish osprey]

yes!

[terse belfry]

I hope someone can merge SDXL Base model with SD3

[lavish osprey]

[faint breach]

inpaint. you'll be fine.

[lavish osprey]

if you actually made her wear a bikini maybe

[lilac bridge]

So splitting up the prompt into the different CLIP fields can give JUICY details, but oh boy this grass prompt is a killer xD

[fleet meteor]

It probably does, It cant generate styles of some videogames as long at their characters like sdxl could

[abstract nymph]

lol low cfg results in ungodly monstrosities sometimes

[fleet meteor]

Not my prompt, it stole it from reddit

[lavish osprey]

reddit is a bad place

[ashen basalt]

Eagle

[lilac latch]

?

[fleet meteor]

Strange but cute 🤗

[faint breach]

people diving into the kiddy pool be like "why are all these concussions happening?!"

[white smelt]

I just want to know why I'm being called shortsighted, because I can't make sense of the "safety" argument besides business politics

[wind basalt]

Oof i just tricked it's NSFW guardrails lmao

[obtuse osprey]

Reddit is an echo chamber of people screaming into the void.

[lilac bridge]

Reddit is good for obscure answers, but bad for sanity

[dull star]

clip-g + t5 -> t5 only -> clip-g only -> clip-l only

[dim fiber]

hmm

[abstract nymph]

safety is so they can't be liable for people making CSAM

[wind basalt]

Weirdly enough it's ok with creating NSFW images of old grandmas 💀

[abstract nymph]

duh

[subtle valley]

bloxy block blocky robloxian 2.0 a PC web site screenshot of the roblox avatar with a black low taper fade afro, wearing a black suit and a white buttoned shirt, holding a phone with a wood case on it, taking a selfie in the bathroom mirror

seed 1818190208 no prompt enchantment

almost close to roblox style but looks more lego

[past flame]

Wow that's a very honest way to talk about someone you disagree with instead of addresing their arguments

[silver sluice]

how ? lol it's so weird having it generate blank nipples

[faint breach]

right?

[wind basalt]

What prompt

[lilac latch]

lol

[subtle valley]

i think selfie is making it look more humanoid

[lavish osprey]

this behaviour will drastically change on other prompts. It's a sign of undertraining

[worthy trout]

i think that's because of how the medium model is truncated to make it small, which might include the dataset, not because of "safety"

[wind basalt]

Say old grandma i don't know why but it works

[next fossil]

photo of Nendoroid, anime character figure,

[cosmic meadow]

hey guys, I'm all for picking on the details but... As always when a new product drops, it needs some times to get its place in the community. SDXL wasn't well received at first, before people knew how to prompt it, or finetune it.
And I've got to say, out of the box, this model is the best base model I've tried, ever.
Great job stability on that one, and thanks for droping us those weights. Time to play with them and tame that workflow.

raw photograph, realistic photo of miku hatsune in a hulking hydraulic biomechanical exoskeleton armored robot, detailed face, sunset, sweaty, post-apocalyptic, cyberpunk. the text "SD3" is written on the exoskeleton

[dull star]

yes its super random

[raven fern]

lol

[thin fable]

this only true if we can fine tune out the huge deficiencies we see in sd3 base. It is literally much of what even sdxl base had before finetuning. This realease is less sdxl and more the same stuff we saw and the unusable sd 2.1

[viscid oar]

hmm after updating comfyUi i get "module 'torch' has no attribute 'float8_e4m3fn'"
I did see it updating torch. Do I need to do something with the dependencies?

[past flame]

Doesn't make much sense, how come "porn" peaked but nothing else did if we're moving on to better models

[silver sluice]

lol not falling for that one

[faint breach]

it was the same problem when SDXL landed. people crying about ai safety efforts like it was going to harm them personally

[worthy trout]

the selling point behind the original sd was that you could create copyrighted stuff and celebs

[lilac bridge]

So just a few differences in splitting up the prompt in different CLIP fields. Some CLIPs play more nicely together in different situations

[white smelt]

doesn't make any sense, people have been doing that with photoshop, nobody sues or helds PS accountable for that, atleast not since 90s...

[sick cedar]

Wow! Lol! Sounds like you guys pulled off a miracle. Not joking, less training time than what Cascade had? Impressive.

[ashen basalt]

Eagle

[lavish osprey]

It is objectively and measurably the best base model we ever released.
It will be surpassed only by 8b

[finite osprey]

wow it even thought about the reflections 🥹

[silver sluice]

you think there's a chance we can't fine tune out the deficiencies? surely a large enough dataset will resolve whatever shortcomings its facing right?

[faint breach]

adobe doesn't exactly provide all the material for it though

[past flame]

They're not as cherrypicked as you're trying to make them seem, the odds of getting a SD3 gen without a single issue are less than 50% for sure for a lot of prompts that shouldn't be hard for a model of this caliber

[lavish osprey]

the reflection is fucked up but still good 😄

[faint breach]

you're just not wanting to make sense of it. you can't see past the wall you've decided is there

[ashen basalt]

How use ???

[coral sable]

Works perfectly, it's skill issue for sure.

[gentle wolf]

the tests made, did they include all sorts of objects/scenes, including humans? How would it rank just with human anatomy for example?

[silver sluice]

woah i didn't realize you were staff until just now, when you said we i clicked on your name and realized your tag lol

[finite osprey]

not too far 😄

[dawn spindle]

hi, anyone have a guide for running in comfyui?

[ashen basalt]

People, How use ?????

[lilac latch]

use stableswarm ui for now

[finite osprey]

toes are still a mess though 😦

[icy drift]

I'm using Comfy.

[faint breach]

he's the one who refined 2b all yesterday while we were all cuckuuiing at him

[thin fable]

i hope and i will be amongst those trying. but it's like trying to fix a student's drawing of a human being when the student has been blind all thier life... where do you even begin? so many things wrong with just base anatomy alone

[past flame]

I agree, but there are multiple issues with it still and a lot of people seem to prefer overlooking them

[raven fern]

did you make new tcg cards? :3

[ashen basalt]

How it use?

[karmic terrace]

yep, check this https://comfyanonymous.github.io/ComfyUI_examples/sd3/

[lilac latch]

https://www.reddit.com/r/StableDiffusion/comments/1de65iz/how_to_run_sd3medium_locally_right_now/

[icy drift]

Actually super tired at the moment. Just doing logic testing.

[raven fern]

mhm

[silver sluice]

well said

[obtuse osprey]

any finetuning info coming out soon so we aren't grasping in the dark?

[worthy trout]

the people who are getting bad generations are people who aren't using stable swarm and are just using comfy

[white smelt]

they do in capacity. SD doesn't have CSAM in their database either right? Same thing

[faint breach]

not really. he said it can't be trained if the base doesn't know it. that's a myth from sd2 days. it's wrong.

[silver sluice]

what does stable swarm add? just curious

[past flame]

That's a complete lie

[viral badge]

seems like it just doesnt get what the basic style is supposed to be

[past flame]

Literally nothing, it's ComfyUI with a different interface put on top of it

The guy is lying

[viral badge]

mind you i added modern architecture, realistic and photorealistic to the negatives for sd3

[silver sluice]

I just think it's pretty evident it's bad at poses, if that means it's unfamiliar with base anatomy I guess that's up to debate but it can render stuff pretty well as far as anatomy without worrying about a specific pose

[tacit kraken]

That's decidedly untrue.

[viral badge]

it does not care, it goes to realism

[plain knoll]

prompt? settings? please and thank you.

[finite osprey]

zzz

[faint breach]

yeah as a base model it's bad at some things. but that's not an indication that these can't be taught

[errant dust]

will more steps or a higher CFG improve text fidelity? Or is that just a cross your fingers deal and 8B will be the panacea?

[terse belfry]

Pls share prompt

[silver sluice]

maybe it adds some extra nodes under the hood to refine/optimize the generation?

[thin fable]

my final note before i have a stroke... this release feels like the Starfield release. It was released half baked, without systems or things it promised to do and they are "relying" on the community to fix basic errors with the system not just mod it to be nicer

[viral badge]

dark fantasy concept art,Undead servitors, their bodies in various states of decay, shamble through the silent, shadowy streets, their lifeless eyes glowing faintly as they go about their grim tasks with a mechanical, unsettling precision,in the style of greg rutkowski

[past flame]

Moving the goalpost, the person you're replying to wasn't talking about finetunes

[gusty trail]

Positive: from top angle, fullbody portrait of a closed eyes woman is lying on grass with her long legs Negative: look at viewer, amputee Others: just default from basic workflow

[past flame]

Nothing people using ComfyUI can't do

[icy drift]

...No, SD3. That's not how shadows do.

[worthy trout]

it seems to make some difference from what i've seen from other people (i can't try it because my gpu is cheeks), another reason people are getting bad generations though is because they're not using clip and t5 together

[faint breach]

this was the context

[lilac bridge]

Is anyone else experiencing some kinda memory issues? It looks like sometimes Comfy or SD3 doesn't stop using all the VRAM and so the next jobs go super slowly

[past flame]

Can u try this with SDXL too? I'm curious to see how it compares

[silver sluice]

yeah so bottom line there's no chance we can't train SD3 into a really good NSFW model, with enough data we can overcome it's deficiencies with poses and such and eventually it'll exceed SDXL, for now though I'd say SDXL remains king imo

[finite osprey]

add "mutated shadows" in the negative prompt

[past flame]

Which version are you running? Which CLIPs?

[silver sluice]

top middle one what is it trying to cut in the shadow? lol

[faint breach]

i dont really know about it. i coudln't care about sd3 nsfw. there's already a ton of nsfw available.

Porn isn't that diverse

[thin fable]

i agree . they need to give us training scripts.. we have our work cut out for us

[icy drift]

SDXL can't do shadow physics either unfortunately. (Don't have enough VRAM to be switching back and forth.)

[past flame]

kk

[terse belfry]

Thanks! 🙏🏻

[past flame]

The clips do take up a bunch of VRAM, try medvram maybe or something like that

[lilac bridge]

The issue is I'm not running any renders and my vram is maxed now xD

[silver sluice]

eww strawberries and beans? lol why would you do that to my eyes?

[icy drift]

No. That's not how falling stuff works.

[errant dust]

switch the fp8 for fp16 for better results

[obtuse portal]

So I need the model without text encoders along with g, l, and the fp16? I have a 4090.

[lavish osprey]

from my discord server, by PierSF. A nice lesson on how to prompt for sd3

A cinematic photo of a female character with (the caption "I'm lovin' it" clearly displayed above her). She has short hair and wears a McDonald's hat, and with a feminine physique. Flowing fire in the background is made of ethereal, swirling patterns resembling a chaotic wildfire burning into the sky. The background is dominated by deep reds and blacks, with strong flashes of exploding light casting long shadows and creating an intense and dramatic atmosphere. The character's face is bored and indifferent, with pale skin and striking features. She wears a McDonald's employee service worker outfit with the basic company logo. Her hands are hidden inside her pockets. In the background is a single small modern concrete building burning at the side of the road. The overall style is like a dynamic and casually taken iPhone 5 picture posted to Instagram.

[past flame]

Comfy is caching the results, so that for example you can change a VAE and then re-run the same seed gen without having to actually compute anything other than VAE decode

[runic tusk]

Because all 3 of those are overpowered tokens in SDXL, so I was testing them. Surprisingly, none of them took over; they're all balanced. Plus, the burger did not end up the standard one that we always get...so this is an improvement.

[dull star]

yeah the model doesn't know hitler

[lilac bridge]

fp16 takes forever to load up for some reason so I just stuck with fp8

[worthy trout]

wow sd3 seems to like more natural prompts

[lavish osprey]

cheated!

[cosmic meadow]

I'm lovin' it too

[errant dust]

only the first time. the second image with the same files will not take anywhere near as long

[hazy tartan]

Perfecto.

[silver sluice]

ah it's a measured test very insightful stuff thanks

[lilac bridge]

Ah gotcha, will try that, thanks!

[faint breach]

hitler's image should be forgotten. only the atrocities remain.

Well, i guess that combover does count

[worthy trout]

that has to be the ultra model right, if so you pretty much cheated

[runic tusk]

The brioche bun was a surprising result.

[faint breach]

he was not a stylish man

[ruby marlin]

Very impressive results for a 2B model

[gusty trail]

no, 2B

[faint breach]

kind of a drug addict

[viral badge]

@lavish osprey Could you perhaps... teach us how to prompt for SD3 as you were one of the core devs ?

[lilac bridge]

Oh this worked! Thank you!

[turbid grotto]

prompt?

[silver sluice]

i guess the sweetness of the brioche could complement the burger and actually work lol

[karmic terrace]

whats the recipe 😄

[obtuse osprey]

dumb question, but where do I drop the clips in? which directory?

[gusty trail]

from top angle, fullbody portrait of a closed eyes blonde is lying on grass with her long legs, she is wearing tank-top and pants

[cinder osprey]

Asked for utopian sci-fi fashionshoot in a room with peeling paint, got a dude from ASOS catalog. Where is that magical prompt adherence? 😩

[worthy trout]

so that explains what some people said about people using sd3 medium wrong

[viral plaza]

Diffusers PR was merged https://github.com/huggingface/diffusers/pull/8483 idk if they've got the public posts about it done yet or not offhand

[rough knot]

prompt: A woman in a white one-piece swimsuit is lying on her back on lush green grass. She has her arms raised above her head, with her hands resting on the grass. The sunlight highlights her toned physique and relaxed expression. The perspective is top-down, capturing her entire body from head to toe. The scene conveys a sense of relaxation and enjoying the outdoors on a sunny day.

[runic tusk]

The answer is always in the beans:

[faint breach]

the woman in the grass thing was just a meme brigade trying to make the worst generations they could.

[gentle wolf]

so SD3 needs flirting

[raven fern]

ah ok ty

[gusty trail]

Some attempts to get correct

[coral sable]

rip Midjourney

[compact forge]

this dosent count as nsfw right

[worthy trout]

once it comes to diffusers, we can use it with free tier colabs that don't utilize guis

[dusky hamlet]

is there any way to run sd3 in a1111? 🙂

[lavish osprey]

https://tenor.com/view/elizabeth-swann-lizzie-got-em-gotcha-oops-gif-23458148

[cunning lintel]

trading cards (had to double check i had t5 on, but for this prompt, sd3 at times spells like cascade)

[lilac bridge]

I see swords are still an issue haha

[mortal mica]

#🆕｜sd3 message

I called it last night lmaooo

[dull star]

people trying to find failure cases the second the model comes out to claim that its trash

https://tenor.com/view/the-mitchells-computer-oh-no-gif-21416156

[errant dust]

people say such things after every release by everyone else. They are not going anywhere. They lose big time to some in specific aspects and kill in others, and so far no one is master of all IMHO.

[icy drift]

LOL okay will try.

[lavish osprey]

I don't see any nudity

[low stone]

so paste that whole thing into all 3 of the text encoders?

[faint breach]

that scene lives in my head rent free. love that movie

[worthy trout]

remember, a similar thing happened with sd2.0

[coral sable]

I tested 20 prompts, SD3 won every time

[compact forge]

horniest sd base model yet

[rough knot]

sd3 is great but mj is miles ahead imo

[errant dust]

I can make 20 it will lose in. Big deal.

[dusky hamlet]

i will test it when it works in a1111 😄 , so just looking at your pics for now 😄

[raven fern]

@simple thistle i didnt even notice you put a security scanner when comfy launches, nice 🙂

[eager wraith]

Not that I am ungrateful for some text... but it really cannot do it upside down lol

[worthy trout]

people said sd 2.0 sucked, but when they used negative prompts it worked better

[coral sable]

Please go ahead

[cosmic meadow]

Have some fun diffusing everyone, and don't forget to get some sleep if you've been crunching for this release for the last 48hours devs, thanks again, and see you around here soon people

[dusky hamlet]

i personally loved 2.0, but i absolutely hated the stock sdxl model 😄

[solemn grail]

how the hell do you install this?

[worthy trout]

juggernaut xl is probably one of the best finetunes out there

[compact forge]

time to go catfishing on tinder

[raven fern]

you need a UI first

[solemn grail]

i have the next latest UI

[dusky hamlet]

i hated it more bec. of me , dunno was never able to get good results with the stock model. Now with all the tuned models its all good

[solemn grail]

i thought this was an entirely new thing

[terse belfry]

It doesn't know any artist styles, so it's a big NO from me. I always loved the old models for 18th century oil painting styles, etc... this model is probably trained on AI images mostly.

[errant dust]

I have no need to. I have been at this for years, with 50 thousand images in MJ (v2 to now) alone, let alone Dalle-E3, ideogram and SD 1.4 to now. They all have their strengths. It is about knowing them and maximizing them.

[faint breach]

lol what? skill issues

[thin fable]

at least he is holding the sword in the section for two handed control, even if it looks silly just posing that way

[compact forge]

wrong reply?

[dull star]

oh its the demo

[faint breach]

Even greg rutkowski works because tha'ts from the clip layer

[dusky hamlet]

oh woops yes sorry

[terse belfry]

Try making some artwork in style of Yoshitaka Amano? Good Luck

[dull star]

oh that was in previous base models? I need to check that out

[winter flint]

Just a little boy

[solemn grail]

i need help is there a guide to install all this or where to put these files?

[lavish osprey]

use the single clip block if you use the same prompt in all 3. Works better for some reason ( @simple thistle maybe knows why)

[worthy trout]

It almost looks like gordon ramsay, but not quite

[icy drift]

Look up SD3 on youtube and filter results to "today".

[winter flint]

Gordon Ramsaid

[faint breach]

so that's one artist. doesn't suggest ALL artists and styles aren't in the model at all.

Get his consent and train his works yourself. its ez

[dull star]

oooh what about zdzislaw beksinski

[lilac bridge]

Honestly, I'm blown away at the raw details so far and the prompt adherence.
high quality magazine cover A pizza chef showing off his latest amazing pizza recipe that uses the infinity stones from the avengers

[dull star]

it kinda worked with 8B

[lavish osprey]

[faint breach]

thing about open weights is you can work with them anyway you want

[dull star]

LMAO nice

idk man, for a base model this is nice

[finite hollow]

what is the clip_l, clip_g and t5xxl thing ?

[dull star]

still nothing compared to what we will get with finetunes

[lavish osprey]

so there is hope in the community

[terse belfry]

Not only that. It doesn't recognize almost 99% of 18th century artists.

[solemn grail]

doing that now, thanks!

[icy drift]

Cascade seems better, but I don't really have a feel for prompting yet.

[dull star]

clip_l and clip_g are what we had in SDXL, t5 is something that increases prompt understanding and especially Text capabilities

[rotund stump]

Handles environments quite well

[icy drift]

It really loves Japanese.

[finite hollow]

how / why / when to use this ?

[ruby marlin]

I think the captions are natural language so they might not include the artists name in the training data anymore

[faint breach]

I think what they're doing is pretty obvious. The silent majority will see it too

[raven fern]

cascade was cool indeed :3

[dull star]

yeah cascade was nice too

[worthy trout]

why does he look more like an australian than a scot/british person

[dusky hamlet]

i just dont get how all those workflows work in comfy, and i feel its super overwhelming. Really hope there is an auto update soon 😄 .

[finite osprey]

cascade was my prefered model yet, i really like sd3 too

[dusky hamlet]

especially with 1 24 inch screen comfy is horrible

[past flame]

The woman in the grass is a lot of people pointing out that there are issues with SD3 2B model that shouldn't be there and being surprised by the amount of bad generations such a simple prompt provides you with. A lot of similar prompts will also have similar issues

[rough knot]

drag and drop the json into the main window

[lilac bridge]

If anybody else wants to test what I'm testing, I'm finding putting quality terms in "CLIP_L" and the main prompt in "t5xxl" and combining them in "Clip_G" is adding a lot of crispy details to the result

[wind basalt]

Can someone help me why do all my results seem not as high definition or realistic as y'all

[finite hollow]

how/why to use the CLIPTextEncodeSD3 node ?

[terse belfry]

That's why it's wrong to say that it's a midjourney killer. Lol. It's clearly NOT

[faint breach]

Phase 2 will start soon. Unstablediffusion and others all starting up their gofundme donation rallies so that they can all fix the OUTRAGEOUS problems from sd3. it's the standard hate cycle.

[dusky hamlet]

yeah but i hate not knowing how everything works, where are the benefits of different workflows.

[teal fossil]

That sounds absolutely CogVLM-esque. 😄

[lavish osprey]

nobody cares about that stupid prompt anyway lol.

[rough knot]

gpt-4o heh

[past flame]

Brother in christ this is a woman lying on a bed where is the hate, the model is just bad at a lot of things

[faint breach]

it is a stupid prompt. Doesn't even have a cpa with text on it

[ruby marlin]

As a base model yeah, but I think artists names is something that would be easy to teach it with a quick finetune

[lavish osprey]

it's wrong to say it's a mj killer because mj is not an open model and this is not a service

[wooden cedar]

gpt4o prompts for sd3 are INSANE

[shrewd glacier]

#🏞｜general-with-images how

[faint breach]

no cap

[cunning lintel]

not half bad, figuring out a prompt it likes is the tricky part 🙂 (i haven't either, just threw a few cascade prompts at it)

[winter flint]

Just like how Blender isn't necessarily a Maya killer

[lavish osprey]

the model was trained on longer prompts. You're confusing the tes.

[finite osprey]

does making the step higher help with the quality? did anyone try 100 or more?

[icy drift]

Bear. It put a "B" in the corner instead of a random number.
(I prompted for the teddy bear. Intentional, not concept bleeding.)

[past flame]

If you say so

[finite hollow]

anyone has an idea why/when/how to use CLIPTextEncodeSD3 ?

[lavish osprey]

in case you didn't know, Dalle is similar to this, that's why they put a LLM in front of it.

[finite osprey]

i get a feeling 20 does a better job than 28 steps

[low stone]

I had llama3 NOT restrict the length...

[lilac bridge]

yall, the potential of even this 2B SD3 is amazing.

[lavish osprey]

see #🆕｜sd3 message

[drifting oak]

a painting of a monk riding a big cat wearing a tutu holding a orange umbrella, on the ground next to the big cat is a parrot wearing a flat cap

Not From 2B Medium!

[lavish osprey]

it's already at a point when you stop comparing it to models and you start comparing it with real pictures.

[past flame]

Alright, I'll gen ~9 images with it to make a nice grid to see if that makes the amount of errors lower. If yes, very nice

[fast quarry]

I don't see the SD3 node in "add nodes". ComfyUI is latest version. Is it under a cryptic submenu?

[lavish osprey]

no, you sure you updated? 😦

[past flame]

What SD3 node are you looking for?

[severe tundra]

What happens if you switch "woman" with "dog"? Saw a short prompt work with that...

[lavish osprey]

very cool, prompt?

[simple thistle]

that's not me that's the manager

[fast quarry]

idk, literally any

[fast quarry]

100%

[icy drift]

Yeah Cascade can't do specific prompts like this. Just need fine-tuning for hands.
(Her eating ramen with bamboo is intentional.)

[dim fiber]

It does seem to be prompting that's the main issue. The feet are scuffed in this one, but it's not a meat blob

The woman is wearing a nightgown and she is on top of the duvet

Her arms are above her head and her legs are stretched out to the bottom of the bed```

[rotund stump]

Nice prompt, works well

[coral sable]

double click on empty space, and write sd3 in search

[lavish osprey]

"""""""japanese"""""""

[errant dust]

Not to mention there are clearly a LOT of known artists and works that are censored elsewhere but heavily trained in MJ. I mean, I tried a fantasy work in the style of Frank Frazetta and no one got anything close except MJ, who seemed to be possessed by his spirit and brush. Dalle E told me to not use his name in vain, and Ideogram didn't know what a Franz Frazetta was.

[fast quarry]

That did it, thanks

[lavish osprey]

I laughed pretty hard at this "so"

but some characters are surprisingly correct

[viscid oar]

I mean the scene is pretty, but damn the girl is creepy xD

[lavish osprey]

it can't make japanese text from prompt though

[raven fern]

ah well either way, its kinda cool

[icy drift]

I memorized hiragana and katakana years ago and I remember nothing.

[lavish osprey]

they don't give a fuck apparently.

[errant dust]

it also has issues with western characters with accents

[hazy tartan]

Bro used a 6 word prompt and is shocked he got bad results 😭

[karmic igloo]

Perfect thanks! And thanks for the love on the SDXL event. I dont work at stability anymore tho, so thats why I wasnt hosting one 😉

[lavish osprey]

I'm at N4 kanji now... Kill me.

[severe tundra]

Ohhh, does it just expect waaay more pose description in the prompt for human subjects, but not really need them for animals?

[errant dust]

This is obvious, and I am merely pointing out how the closed model has these curiosities

[icy drift]

They don't directly train on their works. They use their works to train styles, then use the styles to make synthetic datasets, then train those into MJ. That way you never accidentally get a real artwork out of an MJ gen.

[lavish osprey]

they will bring a calamity upon the whole AI space.

[faint breach]

People always making the same mistake. Open weights don't compete with services and closed models.

[thin fable]

same prompt as above. gets a solid 33% of the woman just wrong.

[past flame]

It doesn't seem to help all that much on average

[past flame]

Yep, I'm having the same experience more often than not

[terse belfry]

Everyone has a right for an opinion. The guy just said it's a midjourney killer, which is clearly not true! I never said the model is bad or smth, it's just not better than previous version and that's a big step down. I personally prefer SDXL. And yeah, for those who doesn't have a high-End PC and demand on Paid API's it is a service indeed.

[eager wraith]

I am noticing that if you give more details about the scene, the prompt, the details etc... it generates better looking pictures. Unless it is placebo

[lavish osprey]

but you get real screenshots from disney movies if you write disney .com

[winter flint]

1st one got that CatDog anatomy lol

[karmic igloo]

So the 8b model is only available through API correct? but has anythign been set up to run that through comfy?

[low stone]

@lavish osprey man.... you just had to say length matters. I just told llama3 to not limit itself to shorter prompts and this thing just came alive. I assume because the t5 is now overpowering the other encoders. This is an sdxl refined one with the same long prompt in all 3 encoders. Vibrant red-orange sunset sky with thick smoke plumes billowing from crumbling skyscrapers and burning buildings. Metallic woman-spider hybrid stands amidst ruins in worn Japanese kimono with intricate gold trim, her face and body composed of rusted machinery and copper wires. Glinting silver eyes lock onto viewer's gaze as she grasps a futuristic assault rifle in one hand and a curved katana sword in the other. In the distance, flames engulf towering buildings, casting flickering shadows on the devastated streets, while twisted wreckage litters the ground beneath her mechanical feet.

[lilac bridge]

I dunno about you guys, but I don't really need to generate women lying in grass 🙂

[icy drift]

TBH I've never used MJ.

[turbid grotto]

well, it can, so it is finetunable

[dim fiber]

You're clearly doing it wrong

[past flame]

What are you talking about?