🆕｜sd3

[dull star]

wow

[dry wave]

Might be subjective. In my opinion Cascade makes worse images than SDXL, in particular when compared against custom models. Sure, people will now say that you cannot compare base models against custom models and cascade could be so much better when trained. BUT there is no good custom model for cascade. Maybe because nobody is interested, or maybe because it just doesn't work

like my feeling is that Cascade is already heavily overfitted on midjourney images and won't improve much by fine-tuning

also I found that training cascade doesn't work as well as training sdxl. But I know other people have different experiences in this regard 🤷‍♂️

[hallow lion]

i just hope i wont have to make another update in 2 weeks (eheh) when this thing drops and it cant run on anything less than 24 gigs

i got the tail end of the sd15 era and sdxl was too much for my old comp

but sinc eit was many years old i said ok lets go

but 6 months late ri really dont feel like upgrading again

[runic tusk]

Did you add --api in COMMANDLINE_ARGS?

🤷

I don't use it.

[silver bridge]

Tho u think I can use it eith my 8gb rx 6600 xt on Windows and u think it's worth learning?

Great I'm in the wrong channel

For this question 💀

[low stone]

You can use the sd3 api in comfy and then pipe it into any other comfy node to use detailers or hand fixers or just sdxl refine it with some denoising.

[severe phoenix]

if after all this waiting and they release a model that still spits out images that look more like these api images without any distinct improvement especially with human anatomy. i'm going to blow a gasket.

[kindred hemlock]

giraffe confident expression, pixar style, expression

#🌠｜show-and-tell giraffe confident expression, pixar style, expression

[charred holly]

/txt2img A white man , dressed in an OpenAI logo black T-shirt , writing on the blackboard with white chalk ，the blackboard have two pieces and can be moved up and down. The content he wrote is "Transfer between Modalities. Suppose we directly model P (text, pixels, sound) with one big autoregressive transfoer. What are the pros and cons?" Shot from the back.

[real terrace]

wow pretty impressive, I though cascade wasn't use anymore.

What workflow do you use?

[lucid swift]

nothing speachial. just good prompts

you have to talk to it in a natural lagugae

[real terrace]

I tried it but it took a lot to generate and I couldn't test prompts enough, results were of all kinds

[lucid swift]

cascade is just next lvl. the base is good. but it has much more potential. it lerns 16x faster / 16x more then other diffuion models, it can output very high ress, it has a very big clip model alsmot 2b.

[lucid swift]

i think ther is still a free demo on hugging face spaces

i like that it understands more complex stuff like "a person wherring clothing made out of trash bags" but it does not like mispellings lol

[real terrace]

I used in Comfy

I got some nice generations

But there was something, like too "perfect", cartoony, it made it unrealistic, beetween other kind of results

[lucid swift]

yes i agree. you have to include stuff in your prompts to make it less perfect.

like nosy phone image a older date and stuff like this

but the model is fine tuned on "ascetic pleasing " images. so a finetune on worse stuff would help. but the alternative is trying some promt stuff

[lucid swift]

try to use stuff like this "grainy iphone photo of a black backpack with the word LUMA embroidered, new yor city metro, posted on Reddit in 2013
"

[real terrace]

ok, I'm trying to set it up running again, ty

[lucid swift]

then you can get stuff like this

it looks more real and less cartoony. but the model sometimes still wants to make it lok perfect. thats why a lora is probably needed

[real terrace]

what resolutions do you use? and that's why asked about the workflow, I use this, only that worked. But took like 10 min to generate in its default settings https://civitai.com/models/119257/gtm-comfyui-workflows-including-cascade-sdxl-and-sd15

[lucid swift]

just use the basic workflow for now https://comfyanonymous.github.io/ComfyUI_examples/stable_cascade/

i just made this one with this prompt "grainy iphone photo of a bag , new york city metro, posted on Reddit in 2013"

it still looks a bit like a studio photo but its getting there

[lucid swift]

what gpu do you have?

jummy xD (cascade)

[real terrace]

RX 6700, 12 GB VRAM

[real terrace]

I'll try to run it when I set it up, ty

[lucid swift]

12gb shuld be enogh. so it shuld not take 10min for you. i think you can add stuff like a lwe vram option to comfy ui

[real terrace]

No idea, I never found "flags" for Comfy or any optimization, I just run "HSA_OVERRIDE_GFX_VERSION=10.3.0 python main.py"

[lucid swift]

i found this "python main.py --lowvram --preview-method auto --use-split-cross-attention" i think you only need the first one

(cascade)

now this one looks good/real (cascade)

[real terrace]

It looks really good

[hallow lion]

You're an expert at promting by now.

[visual jewel]

$prompt full body,

[remote holly]

learning "soon" word until sd3 release #2
French 🇨🇵 : Bientôt 👨‍🎨🖌️🥖 🥐 🗼

[storm gorge]

boy

[noble coyote]

My prompt = peruvian arpillera streamline moderne molly-mae hague natural nature majesty victo ngai henri rousseau vladimir kush tamara lempicka andrea kowch

Llama 2B prompt = Create a vibrant and eclectic pop art collage inspired by the fusion of Peruvian arpillera with Streamline Moderne, featuring Molly-Mae Hague as the central figure. Incorporate elements of natural nature, majesty, and fantasy to create a dynamic and imaginative piece. Collaborate with artists such as Victo Ngai, Henri Rousseau, Vladimir Kush, Tamara Lempicka, and Andrea Kowch to bring their unique styles and perspectives to the piece. The resulting work should be a bold and colorful celebration of artistic collaboration and creativity

[real terrace]

Oh I got it running, it generated really fast

[lucid swift]

nice!

and it looks good

[nimble pumice]

can anyone help me with this?

idk how to genrete imagess using this discord

[lucid swift]

you can ether buy artisan and can genrate it in that channel or here for free https://glif.app/glifs/clv4ca3aq0004xklr050jusec

we just post generated images here

[real terrace]

I'm trying the Galaxy Time Machine Workflow again, still takes long

[real terrace]

yes indeed, pretty neat and realistic

[lucid swift]

i think it uses also sdxl and stuff

you dont need to use it

[real terrace]

only if you use the face/hand detailer I think

It expend most of the time with model b and c

It generate images of 1920 x 1280, I guess that's why it takes more

After model c it shows a preview

Also what was messing with the generations was this SD XL Styles, for some reason they are broken, so I bypassed them. I guess it doesn't need it, and they add much nonsense.

well this is the same prompt with GMT workflow

I guess I'll try prompts with the basic one and maybe then try with this

And for some reason output is store in Temp folder

[lucid swift]

cool

[muted dove]

G T M 🤷🏻‍♂️ 🤣

If you're using v4 of the Cascade workflow, it does a double pass over the C stage, so will take longer. You can either bypass that, or use v3, but that uses the unet models. Should be easy enough to change around anyway.

[real terrace]

oh ok sorry

[muted dove]

No need to be sorry 🙂

What GPU are you using?

[real terrace]

RX 6700 12 GB VRAM

[muted dove]

That will work, but I wouldnt expect it to be fast. Cascade was a struggle for people with smaller GPUs

[real terrace]

it takes like 7 minutes I think, it is ok for the large scale almost finished image

[muted dove]

I have a 4090

[severe phoenix]

dude is this cascade too???

[muted dove]

This image Prompt executed in 123.69 seconds, including model load times.

The workflow may be embedded, but it's a different one to the ones I've posted on Civit.

[severe phoenix]

is this cascade? gaddammn its good but why did it take soo long? is cascade that much slower than sdxl?

[muted dove]

It is Cascade, but that is doing 2 passes in Cascade, 1 in SDXL and then Ultimate SD Upscale.

...and a little post-processing in the flow too 😄

Also Cascade

[muted dove]

To save memory, it clears the model after it has run, so has to load it each time as well.

I changed the SDXL model to "Boltning Hyper"... Prompt executed in 66.48 seconds
Not a bad time for a 2K image

[remote holly]

Stable cascade ?

[severe phoenix]

so how long does normal cascade setup take to generate?

[remote holly]

Can stable cascade run on rtx 3060 12go ?

[muted dove]

Yes

[muted dove]

Try it. I don't have a basic setup, and it'll vary between GPUs

[severe phoenix]

ok but is it usually longer than sdxl or just about the same?

[real terrace]

yes

[remote holly]

Ha thanks , do you need some ajustments like -low-vram ?

[muted dove]

I'd guess it's slightly longer, because it's a larger combined model size and has several stages to run. The result is more important to me than the time to run it.

@severe phoenix

[real terrace]

cascade

[remote holly]

Low vram = more inference time ?

It's impact the quality of output ?

[muted dove]

Cascade
Prompt: manic firestarter

[real terrace]

@remote holly I'm running this workflow with the Comfy UI cascade models, 12 GB VRAM on Linux, no flags, 68.40 seconds 1024x1024aa

[muted dove]

Windows 11 with 24GB VRAM

With a complex workflow

That is also using SUPIR

[severe phoenix]

ok thanks

[real terrace]

the GTM workflow it takes oh wow 12 minutes! but generates in 1920 x 1280

[muted dove]

If you don't need the size, you can disable the upscaling to speed it up.

....or is that without it already? I suppose it must be.

[remote holly]

That's cool , i will try it when i recive my pc

Can you blend images with cascade

[muted dove]

Yes

[real terrace]

blend?

cascade

[muted dove]

Also Cascade

[real terrace]

where is that in the workflow? And I wanted to ask you, why it doesn't put generation in "Output", I found them in "temp"

[muted dove]

Probably because of config issue in the save image node. Have you checked the path?

[muted dove]

Very easy with Cascade

[muted dove]

Does the save image node have a coloured box around it after you run the flow?

[low stone]

I put in manic firestarter and just get this.

[real terrace]

does it need the full path?

[real terrace]

I need to check that

[muted dove]

Change "Quality" to 100

Should never have NaN in there

Actually, right click and rebuild node. Reconnect it if the noodle vanishes.

You probably don't want overwrite enabled either.

[muted dove]

I've actually replaced that save node with

[real terrace]

SDXL

Cascade

It needs some more cooking time

[muted dove]

I find Cascade does nice composition/images, but nearly always needs refining in some way.

[real terrace]

yes in general lighting is really good and things are in place

[muted dove]

Nice img2img too

[real terrace]

like this, everything is in place, the isometry (!?) is perfect, but it lacks stuff

[muted dove]

What's the prompt for that?

[real terrace]

Isometric Cutaway - An Image illustrating a diorama of an Alchemist in a simple Alchemy Lab, Bauhaus, Scott Uminga, elegant, abrupt, (in the style of Codex SEraphianus:1.4)

Clownshark attack (cascade) aaaaaah

[muted dove]

Not what I was expecting 😄

I had cinematic prompt style selected

[muted dove]

What's it lacking?! 🤷🏻‍♂️

[real terrace]

it is a good drawing, I don't know

Some details here and there, but no much more

[muted dove]

Did you expect something more like this?

He's trying to invent a fire extinguisher to put out that fire 🤔

[real terrace]

that has sdxl upscalling or something?

[muted dove]

Yes, they all do

[real terrace]

yes maybe, the drawing style I don't know, it needs some finetune or something, but I don't know, like it's never a clear style, it that makes sense

[muted dove]

Perhaps because you've prompted a style it doesn't know?

I changed it to
Isometric Cutaway - A photo realistic diorama of an Alchemist in a simple Alchemy Lab, Bauhaus, Scott Uminga, elegant, abrupt, (in the style of Codex SEraphianus:1.4)

[real terrace]

i prompted many drawings, animation and illustration styles and it always was a little bit meh. But maybe it is like you say.

[muted dove]

Isometric Cutaway - A black and white diorama of an Alchemist in a simple Alchemy Lab, Bauhaus, Scott Uminga, elegant, abrupt, (in the style of pen and ink line drawing:1.2)

[real terrace]

but also made some style that were really neat aa

[real terrace]

yes that looks more like it, a clear style of drawing

[muted dove]

I think you just need to be more specific with how you want it to look.

3D Model style

[real terrace]

this is a really good clown shart BTW

[muted dove]

@dusky thistle

Isometric Cutaway - A photo realistic cinematic diorama of an Alchemist in a simple Alchemy Lab, elegant, abrupt, (in the style of Game of Thrones:1.2),

[real terrace]

yes I'm finding that some "crazy" prompts that work in SD XL, they kind of glitch in cascade

SD XL - Cascade

[muted dove]

Ewww! Maybe my prompts are just not crazy enough 😄

[real terrace]

Oh I'm loving cascade now

[muted dove]

Cascade

[muted dove]

A dragon spaceship! 😮

[tardy tide]

A crocodile submarine

[muted dove]

😄

[real terrace]

compact dragon

[muted dove]

Cascade

[real terrace]

cascade

[hallow lion]

Cascade will never forgive us.

[muted dove]

Cascade ❤️‍🔥

[hallow lion]

refine that s$it!

make loras and controlnets

actually u can refine it with sdxl in comfy

lol

yes

cascade is better

[dull star]

non commercial 😔

[muted dove]

The majority of us aren't interested in the commercial requirement.

[dull star]

me neither

but its still technically a downside compared to sdxl

[real terrace]

what does it imply?

[dull star]

nothing that affects us personally, if you make images for yourself, it doesn't matter

[muted dove]

No large companies/groups will be interested in improving it, because there's no financial gain.

[dull star]

and that

[dull star]

Same with Stable Diffusion 3

[real terrace]

more cascade

[muted dove]

Cascade

[muted dove]

Nothing suspicious here, move along.
(Cascade)

Makes you wonder why there's no dedicated Cascade channel here, doesn't it? 🤷🏻‍♂️

[dull star]

yeah

lmao its archived

why

[muted dove]

Exactly...WHY?!?!

[dull star]

#1207078178510872636 this will show as locked or unknown or whatever if you dont have access to archived channels

😔

what I am sad about is that bokeh is getting out of hand

[muted dove]

Is it?

@real terrace Didn't you want unclean images of this?

Did you want it so the bag wasn't so much the central focus in the image?

[real terrace]

yes I don't why they replaced for sd3 that isn't even out yet

[real terrace]

impressive

[hallow lion]

There was one!

[real terrace]

I recreate the node

I think it is messed up

[muted dove]

Looks fine, you just need to select what you want.

[real terrace]

output_path and filename_prefix seems to be mixed

[muted dove]

That's the default, so it creates a dated directory each day and puts images in it.

Change them around if you prefer.

[muted dove]

Cascade makes a good job of JW

[remote holly]

Amazing, can you try with these images ?

I want to see what's done with liminal spaces

[muted dove]

They're way too small for Cascade

[muted dove]

I upscaled them 3x, but the pinhole camera image quality is too bad

You need decent images to do a merge, Cascade resolution is higher than SDXL.

I did another 2x upscale on the images, but the merge results are all similar...

[barren spindle]

when SD3 downloadable?

[remote holly]

Ha i see , do you think 1024*1024 image are good for merge ?

[remote holly]

Soon

[muted dove]

Should be

[lucid swift]

Cascde can nativly generate at that ress

[low stone]

John checking out those Wicks.

[dusky thistle]

that's terrible lol

[low stone]

yes it is. that's what happens when i have to sit in the car waiting for someone.

[dusky thistle]

Hahahah

Or... Waiting for sd3......

[rain palm]

seats are attached to the outter walls of the train, that's an interesting design, would work well in india xD

also, why are they teasing us with updates in #📣｜announcements when they know everyone's waiting for the weights xD

[dull star]

ngl this is starting to make more sense

in no way we are getting SD3 in May

and considering that they will release all models at once, 8B has to still cook

for a very long time

[severe phoenix]

why dont they just effin say that though instead of doing the annoying 2 more weeks hype thingy

[crude yarrow]

I was going to disagree, but 1.4 did release around October 2022 and staff were teasing it in August. Might have a point?

The API argument does fall apart though.

[dull star]

idk why they keep saying soon

they want to stretch the hype

🤷‍♂️

[crude yarrow]

Soon is just the term they use to placate people asking for a date. Always has been.

[dull star]

"guys it'll be here just wait a few more days" so that we'll be kept interested

[crude yarrow]

Unless staff has been giving actual timelines like "we are planning on a release near end of May", then assume soon tm doesn't really mean much.

[dull star]

civitai blog had a "we've heard that it will come at the end of may"

but I dont remember the exact wording

[crude yarrow]

More often than not recently, they just drop models with no warning.

[raven fern]

why not come and just say something like "look guys, it's actually not going great, it might take more time to release the weights" instead of nothing 😦

[dull star]

alex mcmonkey already addressed that 8B is very undertrained

[cunning lintel]

It's all conjecture, the utterly bizarre part is SAI just refuses to make any announcement. I fondly remember the SDXL trajectory, never was there this promise of "soon" the worst was a week?? delay when release we announced. And we got the 0.9 leak. And the model in bot. and more talking devs.

[dull star]

but nothing similar has been announced collectivelly from stability

[raven fern]

sdxl 0.9 leak was kinda cool :3

[dull star]

I love how stability made a post about it

https://www.reddit.com/r/StableDiffusion/comments/14s04t1/happy_sdxl_leak_day/

like bruh

[raven fern]

i mean im still patient and technically have lots of stuff to play with already, but just wish they gave us some form of information, even if it's bad news

[cunning lintel]

yeah, this silence combined with occasional hype posts on twitter is more annoying than that it creates interst

[crude yarrow]

Leaks to force a model release do seem to happen way too often...

[dull star]

I hope a leak wouldn't force them

8B has a long way to go still

I dont want them to stop training prematurely

[cunning lintel]

they don't even announce "oopsie that testing where you sign up for a bot, not gonna happen" while it obviously isn't gonna happen anymore... be open and transparent, but all these empty/false promises only lead to more skepticism

[raven fern]

but do they have anything ready, like for example is the 2B model trained but they are waiting for all the versions to release all at once?

[cunning lintel]

They might/might not, we can speculate 🤡

[dull star]

I was hoping they'd release 2B first, therefore I was confident in the may release, but they probably want to release it all at once

and yes, its all just speculation

😬

[raven fern]

yea i wish they released the smaller one so we can play with something while waiting

hopefully it doesnt go past June

[dull star]

does he know?

[raven fern]

as long as you know more than Jon Snow it's good 🙂

worst case scenario we just pay alex mcmonkey to leak something :3

[crude yarrow]

Fairly confident SD3 is not the next model that SAI is going to release.

[cunning lintel]

skipping straight to SD3.1 😀

[sterile pendant]

We aren't entitled to anything from them. Updates and two-way communication are nice, sure, but don't expect them.

[dull star]

?

[sterile pendant]

Yeah I sit on the fence about this as well, or if they do release sd3, it will be far different from what they originally planned or intended. I know they brought up issues about overtraining recently, so that's what kind of made me start to think they're going to redo a bunch of it but maybe with a different spin

[dull star]

overtraining?

[sterile pendant]

Yeah they overfit the current models

[dull star]

oh I know about the overfitting

lykon said that

[sterile pendant]

Which is what I just said...

[dull star]

do you mean they lack diverse data?

oh

I thought they meant different things for all this time

[raven fern]

an option is to just release whatever they have currently (bad checkpoints) and let the community fix it with finetunes anyway :3

[dull star]

this was 27 days ago

so I suppose a lot has changed since then

I really hope that a bunch of well captioned datasets come for finetuning

like juggernaut X or whatever

(even if that specifically ended up as a disaster)

[cunning lintel]

I disagree. If they were a company that just drops models, sure. But SAI announces SD3, lets people sign up for bot testing, mention model soon, mention model in 2 weeks.... And NONE of all they things they announced come to pass... Meanwhile they do post teasers on twitter, ,meaningless images... If a company on one hand hand says things will happen, these things don't come to pass, then responds with nothing but silence, but does manage to keep putting teasers on twitter, i strongly feel that company should also keep its users which they made announcements to up to date when these announced things don't some to pass

[dull star]

Valve:

[raven fern]

half life 3 confirmed

[dull star]

valve does next to no communication with the fanbase lmao

[raven fern]

😦

[cunning lintel]

But honestly, actions say more than words, it's clear SAI (not the employees, the company as such, it's clear people like mcmonkey try) doesn't take its user base serious in any way except as hype-cattle, the closing of sdxl bot that would be back soon with no word about definitve clusre whatsoever, closing cascade channel while still in good use, opening up a waiting list for something not coming to pass all show nothing but disdain.

[raven fern]

if sd3 doesnt happen, i guess people will now focus on cascade perhaps? hmm, but i still want sd3 😢

[cunning lintel]

I think something like pixart, the thing most lacking in SDXL is prompt-following. And maybe a surprise ella release will go a long way for that as well

But SD3 will happen 🙂

[crude yarrow]

I'm speculating that the freesound stable audio model is going to beat SD3 out the door.

[dull star]

oh crap yes if they release the free version

I'm waiting for that

hopefully finetuning will be easy

[raven fern]

wait what happened to the reverse engineered ella sdxl thingy? they said they will release very soon last time i checked and i still dont see it...

[dull star]

I wonder how hard it will turn people away from the non-commercial license lol

SD3 too, but for now, stable audio

[cunning lintel]

Was by the same guy that released morbius, and het got a very very cold shower 🙁

[raven fern]

i would love to try stable audio or some version locally :3

[dull star]

apparently its very vram efficient

like 4-6GB iirc

Finetuning will be the deciding factor if I'll care about it

[raven fern]

oh.... mobius... so those guys just say stuff but never release? sounds like the Alibaba group LOL

[dull star]

cuase if I could fine tune it on songs I like and it sounds decent, then I'd be amazed

its all instrumental though

[raven fern]

i mean cant you fine tune it to include vocals? :3

[dull star]

ehh

we'll need a massive community finetune then

even instrumentals would be cool tbh

[raven fern]

yea

[dull star]

especially if its offline, free, forever

[cunning lintel]

i think there will be weights, just not "free to do whatever with" and lots of hype. They haven't said they aren't releasing ELLA, just silent about it

[raven fern]

meh

[dull star]

people gonna just use suno for serious use though, cause its only $10 with commercial use, and stability is $20

and it has vocals

[raven fern]

but can suno do instrumentals only?

[cunning lintel]

suno and udio are sooooo goood

[dull star]

suno can do instrumentals yes

[raven fern]

kk

[dull star]

both instrumentals and instrumental+vocals too

[cunning lintel]

would be amazing to have an open model that's competitive, but somehow i'm doubting that

[dull star]

yeah

even an instrumental model that has decent coherency would be amazing

[raven fern]

i wish we had more projects out there to try and mimic some of the audio tech, i mean we have a lot for text generation and image generation, but not a lot for audio stuff

well im not counting tts stuff, we have a lot of those

[cunning lintel]

meta's audiobox (just fancy text to voice) seemed soo amazing.

[dull star]

I can convert an album into instrumentals using UVR5 with decent quality and just make a dataset using that

[dull star]

hmm

Like speech?

[raven fern]

idk :3 that was 2 months ago eh... 😦

but ok maybe they will release that

[teal fossil]

In 2 weeks.

[crude yarrow]

The team was aiming for end of May. As far as I know that hasn't changed.

[severe phoenix]

so emad knew all this stuff and yet he said sd3 was coming "soon". Bruhh wth, why hype this stuff up and do the lets pick ppl for access to showcase images when they knew it had all these problems and that it wouldnt be actually ready anytime soon?? this whole thing is just annoying. Cascade is a pretty good model why didnt they drop it like cascade? whats the need for all this nonsense?

[tropic aspen]

No one is going to use Cascade because of it's licensing

[hallow lion]

did SD3 leak? 😄

[dull star]

nope

[hallow lion]

😦

its still two weeks away

[dull star]

no commercial use but otherwise idk

it can make very clean images

I dont make images for commercial use

[hallow lion]

how is it with realism and eyes and hands?

is it as gooey as sdxl was on release?

[hallow lion]

Money.

[dull star]

yeah cascade looked a little too smooth on skin in my testing

[hallow lion]

but with some sdxl refining in comfyu its epic

but again no contorlnets nothing

so its even mor eof a crapshoot

u never know what u gonna get

[severe phoenix]

really? like we give a damn about commercial license, its pretty much unenforceable, that license stuff is for big companies with high revenues wh can afford to be sued to be sued millions in damages. stability dnt care about sueing our raggedy asses lool

[dull star]

yeah its mainly for big companies, its just that many people still get turned off

[severe phoenix]

honestly i used to think this but some examples i've seen on here have really blown me away

[dull star]

people are afraid no matter how non-enforcable this all is

https://civitai.com/models/334891/cctv-cascade?modelVersionId=375105

also this CCTV lora looks really good

rivals SD3 CCTV images

[hallow lion]

lmao

thats like the sdxl boring ppl or VHS loras

those are whaaaaacky

[neat wigeon]

imagine sd3 loras

[dull star]

poopmaster

[hallow lion]

oh yeah sd3 will be amazing

[severe phoenix]

people who? nobody gives a shit. if we ddid everyone in here would be anti-ai

[hallow lion]

but will it blend?

will you need 64 GB VRAM?

and 20 TB harddrive space

hmmm?

[dull star]

and idk about model trainers

if they are just in limbo because "sd3 could come any minute, why would I waste training on Cascade" or whatever

[hallow lion]

yes

cascade wa sdealt a bad hand

i dotn get it

perfectly usable better than sdxl

no one cared

[neat wigeon]

the wait better be worth it lmao

[hallow lion]

haha

[dull star]

epic

[tropic aspen]

Casecade was made by a team entirely separate from Stability that got funding from them IIRC

[hallow lion]

thats another thing

the hype is bad

the longer people wait the higher the expectations

eventually it simpossible to satisfy the expectattions

[dull star]

honestly I would have a lot of fun even with the base model

so that's why I'm still eagerly waiting

[neat wigeon]

happens with anything big

[dull star]

otherwise I'd be waiting for finetunes and not get excited at all basically

Lykon is probably getting a finetune out very soon after the launch probably

[hallow lion]

im basically waiting for the finetunes and mor eoptimized stuff

[dull star]

since he's a stability dev

[hallow lion]

turbo version whatever

[tropic aspen]

Let's be real here. Most ppl are waiting for the Pony SD3 finetune lol

[hallow lion]

i hope i can at leats run those locally

[dull star]

yes

[hallow lion]

if not the bas

[dull star]

ponySD3 wont even need T5

they can train on the clip models just like with SDXL

[hallow lion]

i dpont get pony

[dull star]

lol

[hallow lion]

it doesnt even do perfect anatomy

and it cant do nothing bot bodies

[dull star]

exact same clip models as SDXL

[hallow lion]

like seriously all it makes is a room with a woman in it

every single time

regardless of promt

[dull star]

lol

[tropic aspen]

You can say that about most SD finetunes tbh

[hallow lion]

lol

nah

they do good variety

[dull star]

yeah this is why I'm scared of SD3 finetunes a little

[neat wigeon]

i think the 3b will be more interesting than people thing

[dull star]

yes it can have good vareity

3B?

[hallow lion]

some amazing sdxl checkpoint sout there

[dull star]

we will have 800M, 2B, 4B, 8B

[neat wigeon]

fuck

2b

wait

[dull star]

minor spelling mistake

!!!!

[neat wigeon]

i swear it was 3b

😥

maybe im thinking of something else

[hallow lion]

im not technical so i dont really know what those Bs mean

[neat wigeon]

the 4b then

billion

[hallow lion]

i assume the more Bs the better

parameters?

[neat wigeon]

more params

[dull star]

2B will replace SDXL

[hallow lion]

so

basically the more params the biger the model

[dull star]

smaller parameter size, better quality, T5 running on CPU

[dull star]

essentially yes

[hallow lion]

T5 better be faster than Sigma

its sloooooooooooooow

[dull star]

uhhhhhhhh

well ummm

yeahhhh 😅

[hallow lion]

cascade is 5 times faster than sigma

and gives same quality

[dull star]

but that is only using clip models

[neat wigeon]

800m might have similar results but wayyy faster

[dull star]

eh

no

well similar I guess maybe

[neat wigeon]

boom~1

[dull star]

800M will replace SD1.5 (or not idk)

2B or 4B will replace SDXL

8B will only be used by 16-24GB users

[neat wigeon]

when it comes out ill train a finetune on cpu for the memes

🔥

[hallow lion]

i hope they will optimise even 8B for cats with less than 12 gigs

comfyu

fooocus

[dull star]

12GB with T5 with only CPU

[hallow lion]

whats the base size of sd3? still 1024x1024?

[dull star]

who knows

1024x1024 only for 8B

smaller models MIGHT get 1024px versions

but honestly

the 16 channel VAE probably carries it

[hallow lion]

u know with the excitement and readiness to pounce there better be every single sd3 version of every plugin available 30 minutes after the weights drop

controlnets supir comfyu

everything

[sterile pendant]

The one about overtraining was a discord post from one of the devs or team members.

[neat wigeon]

my idea exactly 🔥

[dull star]

a discord post

[neat wigeon]

sd3 comes out im speedrun making something

[dull star]

I need to find it then

[tropic aspen]

You know it won't. SDXL doesn't even have full CNet support

[neat wigeon]

like a colab, application anything

speedrun

[dull star]

SDXL got a new openpose controlnet, and its finally good

or rather, it actually works kek

[tropic aspen]

Does that work with Pony?

[dull star]

no idea, its SDXL based so possibly 🤷‍♂️

[sterile pendant]

Well the neat thing is that objectivity doesn't care if you disagree. They are under zero legal obligation to even release the model. At least there's the API to play around with.

[hallow lion]

whats it called?

[dull star]

https://huggingface.co/xinsir

canny and scribble too

[hallow lion]

this looks new

fresh out of the oven

[dull star]

in the openpose one, there's a "twins" version, and the creator said this:
It is a model with similar performance and different style. The pose will be more precise but aesthetic score will be lower.

yeah I tried this openpose one and its good

[hallow lion]

lol pokemon model

[dull star]

https://i.redd.it/efs0gqs9673d1.gif

see, it ACTUALLY works this time

[hallow lion]

yeha looks better

god i wish video would catch up

haiper does some ok things

but again

crapshoot

[dull star]

I want boring reality for SD3

as good as "low quality" works sometimes with SD3, something a bit more consistent would be good

[hallow lion]

connsistency is what wer eall after

Selecting different areas in comfyu is very good

i want to be able ti select a character or object or backdrop and somehow tell the ai to not change it anymore

only change angles and perspective

but leave textures and shape alone

[cunning lintel]

wtfbbq, what does legal have anything to do with it?!

[dull star]

https://www.reddit.com/r/StableDiffusion/comments/1bv83qt/update_on_the_boring_reality_approach_for/

[hallow lion]

:))

[dull star]

this is SDXL with boring reality

[hallow lion]

yeah these cna look crazy deceptive

[hallow lion]

not AI looking

[dull star]

exactly

[hallow lion]

u reeeeaaallly have to look

[dull star]

oh nice, an old photo of you and your frie-

🤨

[hallow lion]

hands, numbers, words, etes

but if somen know nothing abt AI

[dull star]

you really have to look to figure it out 😉

[hallow lion]

at firts glance they go yeah thats a pic

wow

boring jabba

[dull star]

#✨｜sdxl message

closest I have ever gotten without boring reality using sdxl

goddamn this aggressive depth of field

[hallow lion]

whats ur settings

sampler

[dull star]

haven't used this for a very long time

[neat wigeon]

joe biden lean

[dull star]

uhh

I think it used to be just like dpmpp SDE or whatever

[neat wigeon]

vidalcat

[dull star]

and highresfix

that's basically it

[stuck haven]

wth lmao

[dull star]

#1045349359044280360 message

back in my day we used to have SD2.1!!!! 👴

[dawn jolt]

did you train it or just use good promts?

[dull star]

I did make a lora but it was a failure

[dawn jolt]

so the Biden images were made with promts?

[dull star]

and a special finetune of SDXL

https://civitai.com/models/241445?modelVersionId=272411 this

using an nswf image is okay, but it gets really annoying when it sometimes makes nipples or whatever for no goddamn reason

even if you spam it in the negative prompts

#1045349359044280360 message

cosxl before cosxl

even though this was jsut primitive offset noise

[raven fern]

i know about that, but there is still no sdxl segmentation model (i mean technically there is one i think, but the filesize is ridiculous for a controlnet)

[raven fern]

we dont talk about 2.1 :3

[real terrace]

so I'm "loosing" generation power (!?) if I generate in 1024x1024?

What's the square native resolution?

[raven fern]

not sure about native, but for me, cascade makes better squares i think around 1536x1536 from what i tested

[sterile pendant]

Basically, you're complaining out of entitlement, which I had already brought up. What I said didn't change your entitled stance, so I was letting you know that they have zero need to keep you in the loop or even release the models at all, from an objective standpoint because nowhere does it legally state they have to. You just want them to and are mad that they aren't giving you updates about what they are doing or going to do. If they do, great. If they don't, then oh well.

[raven fern]

https://tenor.com/view/gimme-what-about-bob-bill-murry-i-need-gif-19552065

[real terrace]

maybe I'm biased because I saw it here, but I instantly recognized the people as "SD 1.5 people" xD if that makes sense

like a common denominator of people that the models end up doing

all with the same smile and expression at once

[rain palm]

hmm, is the new CEO gonna do a Sam Altman ?

[fiery wharf]

no,hes gonna pull a Sam Bankman

[turbid grotto]

Maybe the real sd3 was the friends we made along the way

[muted dove]

It's just a cleverly designed wall, made to look like a train.

[lucid swift]

@muted dove 1024 but it can also do much higher ress because it was trained on multiple resolutions

[muted dove]

https://huggingface.co/stabilityai/stable-cascade

It handles 1536x1536 without needing to upscale, larger too.

[lucid swift]

Yes

[lucid swift]

Heh its alredy out. At least one version xD

[lucid swift]

Cascade is 3.5 and 1

[dull star]

Huh

Stable audio 2??

[dull star]

whaaaaaat

I'll honestly wait for the official open version

guys!!!!1! very real

[lucid swift]

the end of may interesting xD

[dull star]

tomorrow is left, I wonder if they won't release it, or they won't release it

if it's cause of 8B still being undertrained, then no problem 🙏

[muted dove]

They could still release the smaller one for now.

[dull star]

2B would be nice

that's the only one I would expect to come out tomorrow or today

[muted dove]

https://tenor.com/view/chris-farley-motivational-speaker-matt-foley-get-it-on-rock-and-roll-gif-13919525312128553605

[dull star]

not expecting it though

but its plausible

[lucid swift]

sme x.X

[cobalt moon]

I expected June 1

but eh

[fiery wharf]

just need to give the SAI CEO a few millions and he will release it,i swear its 4 real this time

[idle current]

*billions

[fiery wharf]

*trillions

[lucid swift]

galons

[left parrot]

Afaik the only mention of 'end of may' for the SD3 release date was in a civitai newsletter, nothing official.

[fiery wharf]

only official news we got was that they are broke and looking to sell

[dull star]

what do you mean? "we've heard" is a really credible source

[cinder junco]

The last two posts from Lykon that I can see (through Google) were on May 24th. The second to last post said “still cooking”. So I don’t think it is likely that we will get SD3 for some weeks yet.

[dull star]

yup

8B def June or July (July more likely, idk how long they want to train)

2B could come in June though

[cinder junco]

Wouldn’t it be more likely that the smaller models are distilled from the largest model, rather than having completely separate training?

[cobalt moon]

distilled hurt too much

on quality

it is much much better to use cut-off dataset against limited steps training.

sure things like Lightning or Turbo could help

( SD3 Turbo's on the API btw )

[cinder junco]

What do either of those things have to do with a smaller model? The number of training samples or epochs seems independent to me (and that more would always be better).

[cobalt moon]

the relationship between parameter and dataset-training is... uhhh a lot of math. so it is a little bit tricky to explain.

If I understand correctly, parameter is a calculation inside neural network to give a data using their learned perspective

or gonna ask GPT

[fiery wharf]

basically "i dont know"

[cinder junco]

Parameters are just the number of weights and biases in the network. Fewer means fewer nodes and connections.

[cobalt moon]

it also have to do with computer programming

[cinder junco]

A smaller network will train quicker (fewer steps and less time per step), but I’ve seen scaling laws that indicate things keep improving with more training and more data.

[sick cedar]

When was this posted?

[dull star]

idk early may or before

[sick cedar]

Hmmm.. We might not get it by tomorrow then. My guess was mid/late June.

Worst case scenario would be july, i thought.

We already know that the current version of SD3 is already looking very consistent now. Lykon showed us a anime image the other day that looked like it came straight out of a finetune. So i think it's almost ready. At least one of the models is.

[echo pumice]

https://tenor.com/bRd1k.gif

[sick cedar]

Imagine if it is tomorrow tho? Imagine. Lol! A day away.

[dull star]

only 2B or 800M or whatever

8B still needs cooking

[sick cedar]

@dull star Yeah. 2B would still be amazing tho. I think it will be way better than SDXL (If you use T5 with it.)
Don't you think so?

[dull star]

yes

but we'll see

how much worse its gonna be in prompt adherence compared to 8B

[fiery wharf]

now imagine if SAI goes bankrupt before july

[dull star]

leaks:

[fiery wharf]

only thing they been leakin is money

🤑

[sick cedar]

Idk.. I personally am comparing it to what we already have. And i believe it will be better than that.
I think 2B will know less "things" objects, names for things etc. Though i think 2B will still be competitive to a degree with 8B, because of the T5 text encoder will be able to merge a lot of concepts together to make new ones without the need for LoRas. (We will still need them tho.)

I just think we would need Loras less than before.

But yh. That 8B is gonna be a monster.

[dull star]

if its better than pixart then fine

[sick cedar]

It appears that SAI still have set a goal to reach with SD3. If it wasn't at all possible, i doubt they would have tried to continue with it. So in my personal opnion dispite everything going on i think the full training of SD3 is still relatively safe.

[sick cedar]

I believe it would be, yes. Not sure by how much tho.

The custom finetunes will pretty insane too, i think.

[dull star]

8B is the closest to dalle and ideogram

pixart is smart, but still behind 8B

[sick cedar]

Yeah. I've seen Ideogram too. Strange it came out at a simular time to SD3's announcement. I've also noticed that Ideogram makes similar images to SD3, in terms of composition? It almost looks as if it's a finetune of SD3's architecture. I don't think it is, because it was available way too early for that. But it seems to work in a suspiciously similar way.

If Dall-E 3 & SD3 are compared, for example. They don't look nothing alike with the same prompt.

[sullen moss]

Be calm and wait two weeks...

[fiery wharf]

im sure they will recover like FTX 🙏

[cunning lintel]

i don't see similarity at all (3x sd3, 3x ideogram, same prompt, no magic prompt)

[dull star]

its partially made from the deepfloyd devs

its a pixel based diffusion model or whatever

ideogram is better than SD3, but finetunes will probably gets us closer

[mortal mesa]

lacks diversity funny that

[dull star]

its weird how if you give it a prompt and you don't use magic prompt, you get very very similar images

[low stone]

Lykons Twitter post on left, my own raw sd3 on right.

Anime-style dramatic scene featuring a determined white-haired, red-horned woman in dark, glossy armor with red lines, brandishing a fiery red blade, intense battle unfolding with robotic soldiers in matching armor, background ablaze with orange flames and flying rubble, towering gray cliffs frame the action, sunlight piercing through clouds, crisp details, and contrasting fiery and cool color palette.

[dull star]

the right one doesn't even look anime

still a good image though

[cunning lintel]

For that prompt, it's a lot of improvement 🙂

but questions questions questions 😆

[dull star]

no cherrypicking, just put photography, cinematic photo in the negative prompt

[cunning lintel]

depends what you look t, one is holding a sword, the other has a sword fused in the arms, one has robot soldiers running around, the other robot soldiers fused broken in the ground, in one the girls has arms, torse legs, the other she lost her arm i think 😢

[cunning lintel]

Still same diff, fused sword, two left arms 🤷

[dull star]

cherrypicking will still be required

[cunning lintel]

New one is so much better, but.... questions questions questions :p

[craggy ridge]

https://x.com/thibaudz/status/1796217532153581637?t=rbtZLj_5Wyz7-orS4mPKQQ&s=19

Thibaud Zamora ∞ (@thibaudz) on X

i think it's the most important tweet of the week.

Twitter · May 30, 2024, 4:30 PM

👀

[dull star]

are these beta sd3 people?

who got access to do finetuning or what?

2B finetuned will be fricking sick if true

[craggy ridge]

🤔

[dull star]

so it is

But yeah I suppose we'll see 2B in June then

[craggy ridge]

All the hopes up yes

[dull star]

8B 😬 ehhhh

[sullen moss]

But SDXL has 3.5 B...

[dull star]

yes

https://media.discordapp.net/attachments/1089974139927920741/1227267272058474607/whats-the-latest-update-about-sd3-v0-qlg8jt5z9xsc1.webp?ex=6659e1b7&is=66589037&hm=0c26dedae4dd9a732fbbdaaa3dde29d50368b57410d879a2b12376dd99c052a5&=&format=webp

afterbeforewhatever

[sullen moss]

I was disappointed by the news about the 5090, only 28 vram...

[dull star]

LMAO

[dull star]

yes lol

[cunning lintel]

such a simple prompt/scene, those img say nothing sadly :/

[dull star]

yeah its just Text

I want to see how good 2B is in prompt adherence

[wide pagoda]

That's why it's called SDXL and not called SD3

[sullen moss]

I want SDXXXL 😂

[low stone]

I just went through 18 generations on sd3 with the following prompt and it could never do all of the letters. "8b is ice too" in letters made out of ice. Golden hour, in the arctic.

The closest it got

[runic tusk]

Okay. Here you go: ||https://civitai.com/models/167764/sdxxxl|| (NSFW model)

[dull star]

8B is ice too 🔥

[sullen moss]

[cunning lintel]

that's why the tweet has 4 images :p

[sullen moss]

It will also be interesting to see what the new model from Ideogram will be like

[cunning lintel]

ideogram seemed to have stepped on the brakes. First i noticed a message "25 free prompts become 20" and now when i try to generate an img, it says "our slow queue is currently full, please upgrade or try later"

[low stone]

Totally and I realized that after I saw how hard it was having it. Kind of weird when it has done longer stuff before

[low stone]

Yeah I saw that. In their favor, the image quality has been slowly increasing since they launched. I just wish they had an upscale option. I've been a paid user since the beginning since I wanted the private generations.

[cunning lintel]

i understand fully, it's just so noticeable at the same time, i never got that slow queue, and now reduced free gens and slow queue.

ideogram is really curious, composition of the gen is always the same, something odd must be going on there. seems like it does your regional prompting workflow 😉

[low stone]

That's a good call out. I generated a lot of images to use for training and even though the details of the subjects changed, they were always the same pose. Some people have referred to that as overtraining, where there's little change from seed to seed

[cunning lintel]

maybe, another explanation could be ideogram is stacked models, a lowres one for composition, which then works much like a controlnet to guide the high res one. Either way, I'd be surprised if it turns out to be just a single pass

It's so crazy literal with prompts, i had some really weird ones, and it just bends the image till all objects are in it. Unlike dalle-3 (and sd3, much more so) that puts aesthetics first and seemingly ignores things it can't fit it.

All of which makes it a great additional model, it behaves so different

[low stone]

They really do have the secret sauce on actions. They have actions that no one else has , or will do because of censorship. Concerning stacked models, that sounds like most of what I'm doing these days. Pixart/ella/hunyuan refined with sdxl

[woeful spindle]

maybe the model behind the sd3 api was never the 8b model 🤔

and they generated all these pics with 8b

[low stone]

Yeah, based on what's been said before, it's definitely not as good as the current 8b

Plus they have upscaling etc internally. The api is raw 1300 res

[dull star]

1300?

[dull star]

yeah its 8B, just not the current version

it cannot be

https://tenor.com/view/nooooo-it-cant-be-ninjago-zane-gif-25290061

[lucid swift]

oh i did not know that who of them worked on it?

[dull star]

no idea

[lucid swift]

did it get anounced i thoght 32?

[lucid swift]

they cant burn infinite investor money lol

[lucid swift]

they plan on making a new one? 😮

[lucid swift]

thats how cascade works lol

[lucid swift]

only 4gb more that is nothing 😭

[sullen moss]

https://videocardz.com/newz/nvidia-rtx-5090-new-rumored-specs-28gb-gddr7-and-448-bit-bus

[lucid swift]

😭 man they didnt upgade vram for 4years and then they only upgrade it for 4 gb.

wtf man

[velvet ruin]

I'll be amazed if they truly are only bringing vram up that much.
Leaves dangerous room for competitors to outshine them.

[tropic aspen]

I mean tbf, what competitors does Nvidia have?

[neat wigeon]

Not really anybody

[velvet ruin]

Currently not a whole big push to outcompete, but that always only lasts so long with anything

[left parrot]

28GB is more than enough for gaming, while not enough to compete with nvidia's server GPUs, that have a much higher profit margin

Nvidia's main competitor is itself...

[dull star]

cannot let plebs get high vram easily

[neat wigeon]

Just get a random ai gpu that was used for cypto before

[left parrot]

Hopefully AMD or Intel or some chinese outfit will eventually release a reasonably priced GPU aimed at AI hobbyists, but it looks like it will be years before anyone catches up with nvidia

[neat wigeon]

Amd seems to be leaning against that and focusing on mid range gaming cards

For intel i dont know much but i think they have been making a few

[left parrot]

I might still get a 5090 if the performance increase is good. Even if it can't run bigger models than the 4090, if it can generate images substantially faster it would be a nice. AI art is all about iterations

[low stone]

A closeup of hyper detailed Cookie Monster face, with fiery yellow eyes and an angry expression. The background is a dark gray with sparks flying around him. illustrations, comic art, and cinematic light effects. Dark fantasy setting with smoke. a dark tone with focus stacking

[dull star]

https://vxtwitter.com/Lykon4072/status/1796251820630634965

this is 2B with highresfix I assume

and if this is really just 2B with highresfix I'm excited

this is good enough for a base model AS LONG AS WE GET VARIETY

@low stone

if it weren't highresfix this would be a distorted blurry face (I zoomed into this image)

[sterile pendant]

Yep, like I was talking about the other day when people were thinking they'd magically be able to train their waifu generators with the 8b model lol

[dull star]

I wonder if lykon is teasing 2B because it'll come tomorrow (May 31st), or if they are just trying to say that "look how good 2B looks despite being smaller"

probably the latter lol

but yeah 2B will be trainable offline, so my interesting will rise if it's absolutely the case

[dull star]

You can't train 8B.
I know...
But they can..

8B won't make lora-type models viable even for 24GB sadly

I'd love to train concepts and stuff so badly for SD3

I hope the prompt adherence is at least better than pixart sigma 🙏

wait wtf?

so these were just upscaled with like esrgan or swin-ir or whataver?

[low stone]

From what I've seen on artisan, they're using sdxl turbo to do a lot of the upscaling heavy lifting. One hopes that switches to sd3 when it's ready. I have faith that a 2b sd3 would still be great because I've been using Ella/pixart/hunyuan with tiny models and the upscaler output is fantastic. You add vastly more training even onto those little models which sd3 has, and it'll be great.

One of the biggest issues of upscaling those other sd3 competitors with sdxl, is that sdxl doesn't understand multisubject so you often lose character specific traits as the upscale stages progress . When you can upscale with sd3 that does, great things are gonna happen.

[sullen moss]

I'm still concerned about 2B, it's not enough for creating complex concepts

[dull star]

really?

[sullen moss]

Everything Lycon shows is beautiful, but I don't see any complex scenes in his examples. Detailed faces aren't impressive anymore.

[sterile pendant]

Right, but someone was trying to argue it the other day. I don't think it was you.

[low stone]

So all of these were created with Ella which is a tiny model which uses another tiny model. So for me, size isn't as important as concept training.

[low stone]

I don't think we're gonna get dall-e level concepts. Even with fine tunes. After doing some Lora training and seeing how easy it is, I realized that stability doesn't have the money to train all the concepts, and the community coesnt have the interest in training anything other than mostly porn.

I was kinda put off recently when I saw lately that the majority of Reddit model and civitai stuff has moved to pony derivatives. Something that goes back even further than sdxl, to make it into a better looking sd 1.5 with single word tags again.

[sullen moss]

I'd be very happy to be proven wrong

[low stone]

Dalle can do insane concepts.. ideogram beats it for actions, but dall-e can do transformations better than anyone

And I don't see sd3 ever catching up to even current day dall-e.

[cunning lintel]

And throwing away all the fine details the new VAE allows 👀

[sullen moss]

I have to agree... And it saddens me greatly 😔

[cunning lintel]

The ELLA paper said the limitation was the captions that just weren't good enough. SD3 used the same vlm for their captions, so i'm not expecting much improvement either

But hopefully it will atleast understand longer prompts compared to the api

[static cairn]

Pony (the most popular SDXL finetune) already stated they are intending to train on 8B so already he's wrong. So yeah, another way to try and dodge releasing the 8B because "heh its not like you could use it anyway!!!"

Are they ever just going to admit that they're keeping it API only because it's the good one?

[low stone]

Absolutely. It saddens me that my cosxl raw or cosxl merges always look way worse than just regular sdxl models. It means I can't upscale with the better color range.

And hyper sdxl doesn't work with cosxl models.

That said, I just ran a bunch of complicated prompts against sd3 and it beat pixart,hunyuan, and Ella for prompt adherence every time. So even with a 2b sd3, it'll be better than what we have now.

[dull star]

hopefully yes

[low stone]

Cookie Monster with robot arms, puppet hands, a stone head with ruby eyes, dancing on the moon over a massive pit of garbage. Sd3 pic / pixart / Ella

Ideogram ^^

[cunning lintel]

A detailed anime scene featuring a young woman with cropped white hair and luminous yellow eyes; she has prominent black horns extending from her head. Dressed in a tight, advanced black suit with glowing orange patterns, she stands amidst large-leafed tropical plants that shine under the bluish hue of bioluminescent bubbles and moonlight.

[low stone]

I don't know if you used turbo, but I feel like I was able to get sharper lines out of sd3

[cunning lintel]

Just SD3

Oh, and downscaled by discord 😬

[low stone]

boo hiss!

[sterile pendant]

Do they have at least a dozen or so a100s and probably three of four months worth of time? Since sd3 is a dit based model, you can easily look over in llm communities to see what kind of resources it takes to train a small llm in the 8b range.

[lucid swift]

Current dalle is shit

[lucid swift]

Youndont need a big model for complex concepts. The te desides if you can do that

[lucid swift]

Why is everyone so aggressive to them. That will not help

[low stone]

They're doing the best they can given the circumstances. 3/4 of his comments on twitter at this point are just fending off attacks, stemming from comments by people who aren't with the company anymore.

[low stone]

yeah, and gpt4o dalle is going to be nuts in comparison. i'm almost a little scared at what it'll be capable of compared to everything else we've seen so far.

meta has realtime diffusion and then 3d modeling of that scene and animation, on demand. there's no way the new dalle won't have all that and more.

[lucid swift]

It feels like open ai makes ther models look less real willingly

[lucid swift]

Real time is just turbo

[low stone]

their turbo is better than sd3 regular.

as far as prompt adherence.

[placid belfry]

has the model been released?

[neon wagon]

look how long it has taken in the past from api release to weight release........ there is an pattern.You will not see any "model" before July.

[desert garnet]

as you can see the copium has moved to July,after July it will move to August or September

[noble coyote]

"SD3 is like a greasy-piglet: it defies anyone to grab a hold of it!!!" 😄

[desert garnet]

schrodinger's SD3,its here but also it isnt

[left parrot]

I'm not among those worried that SD plans to pull a bait-and-switch and keep SD3 closed, everything indicates that they are actively working on the model and the weights will be released when ready. But I am concerned that the longer it takes, the higher the risk that SD runs out of money before we get our hands on SD3.

[blissful hamlet]

movie poster, Triangle-shaped food wrapped in green bamboo leaves, known as Zongzi, glutinous rice dumpling, exquisite plates, beautiful cutlery, Pop mart style, Pixar style, Depth of Field, Ray Tracing, Front View, Intricate Details, Unreal Engine, Octane Rendering, Best Quality

[dull star]

https://www.reddit.com/r/StableDiffusion/comments/1d4r3tn/stability_ai_is_hinting_releasing_only_a_small/

[cinder junco]

People casually ignoring Lykon confirming that everything will be released open source as has been confirmed many times. 🤦‍♂️

[dull star]

yeah lmaooo

OMG SD3 CEO COMING

[desert garnet]

my ceo is not a liar,he may be a cheater,tax evader,gambler,deceptive but never a communist

[neat wigeon]

😭

[craggy wave]

Was supposed to be released in May no? Why they don’t give a date instead of saying soon :/

[neat wigeon]

it will come out one day 🔥

[craggy wave]

Yep hope “ soon “ :/

[sullen moss]

What the best integorator for sdxl on your opinion?

[cinder junco]

They probably don’t know and don’t want to guess after their initial predictions didn’t pan out. I’m guessing that either the training loss has a different trajectory than they were expecting due to the new architecture or the training methods didn’t work well enough and they had to restart a few times to make tweaks. They could even be making tweaks to the architecture itself. I believe that all this means that, on release, the performance should be much better than it was for the version analyzed in the paper.

[craggy wave]

Ok :/ so we can wait again a bit of time, was thinking the 2b was almost ready

[cinder junco]

It could be as little as a few weeks, I don’t know. But I imagine that they will need to do things like safety and performance testing after the training is done, so it won’t be as simple as train and ship.

[desert garnet]

its coming,just give more money to sai

[craggy wave]

🐱

[hallow lion]

fk 2B i want 8B

Stop hinting!

Keep your hints in your pocket Emad.

I want 16B too and 32B

I want flying cars and a private resort on Mars.

[noble coyote]

8B will require huge VRAM... most of which this community will never own, let alone have access to!! 2B/Twobee doobee doo will suffice!!!!

[cinder junco]

And a diamond pony!

[desert garnet]

2B = 2Weeks it all makes sense now,the numbers dont lie

[craggy wave]

2 weeks from now or from in 1 month

[hallow lion]

are they making that much money on the API to keep postponing release...

come on...

[cinder junco]

Macs will be able to handle it with no problem. I just want to know what the resolution scaling will be like, since I like to generate at 4K sizes. I can do that with SDXL and hiresfix, but I’d like to be able to do it in SD3 8B.

[hallow lion]

yes what about the rest of us numerous windows peasants?

[desert garnet]

2 weeks in venus time

[craggy wave]

[cinder junco]

You can use 2B, buy a new 5090, use 8B quantized and accept the quality loss, use the API, or buy a Mac. What are you expecting, magic?

[hallow lion]

I expect consistency.

😄

If 2B cna do consistency I am happy.

to be fair all this shit is magic really

[cinder junco]

I think the recent posts were to reinforce that 2B still has impressive quality, even if it’s not as capable as 8B. I think that is good news. Not sure what you mean by consistency.

[hallow lion]

5 years ago would you think u can type in whatever and get an image

consistency means i can recreate any part of the image i desire with 100% consistency between generations

[cunning lintel]

To me it kinda did the opposite 😂 compare the wizard to the earlier wizard announcing sd3 and the api, and the iceblock words to what the API can do, it's a bit of a step backwards

[cinder junco]

Are you talking about inpainting? What “part of the image”?

[cinder junco]

If you’re expecting no quality difference, then…🤷‍♂️. But I haven’t seen a direct comparison between the models on the same images. I thought the ones lykon posted looked great.

[hallow lion]

i want to be able to select a certain par tof the image ( like facedetailer cna select faces) and there is ways to selct a character or object easily in comfyui

and then iw ant to be able to "lock" that part somehow in promt and recreate it in subsequent generations

maybe even select it with a select tool like photoshop

[cinder junco]

It sounds like you’re asking a lot. I don’t know how that would be possible.

[hallow lion]

and tell the ai to only change the angle position and lighting for that part

and ;eave the shape and texture alone

henc eu can composite any scene and story telling becomes possible

[cinder junco]

If you need part of one image inside another, it would be better to mask off and do inpainting or outpainting.

[hallow lion]

well relighting a scene without changing the shape and texture has been achieved

also AI is starting to understand 3D there is text to 3d so lets go

so i hope soon u cna select an orange and somehow lock it down and reuse it other generations.

until this is achieved all this is just fun one off images

no stories no comics no movies nothing

until this is done

[cinder junco]

Character/object/scene consistency I think would be better achieved by LoRAs and stuff like that. It’s too much to expect from t2i, because a picture is literally worth a thousand (or more) words.

[hallow lion]

sadly loras dont achoieve it too

[muted dove]

This was with Cascade....
Prompt was: blocks of ice forming the words "Put it on ice" in a frozen arctic region on a sunny day
🤣

[hallow lion]

😄

[cinder junco]

How would you describe the subtle differences in facial features that differentiate one blonde babe from another? If you don’t know how to put it into words, you can’t expect the model to read your mind. Unless you overfit during training so every “blonde girl” becomes the same blonde girl.

[hallow lion]

well reactor can give pretty similar faces

or you can do ipadapter to kinda give the same outfit

not a strecth to think that when you select an orange and ask the AI to "lock it" it woudl create some kind atemplate like that in it smind and recall it in the next generation

like now you can have realtime loras with a few images ina folder orreactor face with just an image to guide the ai

why cant the ai make a mental note and keep it in it smind instead of us making a fodler and putting the images there

[muted dove]

Still trying...

[hallow lion]

Cascade is trying

Putin on ice

[muted dove]

Closest yet!

Perhaps SAI are just finetuning Cascade and selling it as SD3 🤔

[low stone]

I think I'm at the point where I'm going to stop using the SD3 api until after public release. Maybe it's because I've just gotten my workflows so good or something, but now the local models pixart/ella/hunyuan are all putting out pics that are better looking and more prompt adhereing than the sd3 api. The stuff lykon is posting on twitter, even from the 2b, look massively better than what I've been getting out of the sd3 api lately.

[muted dove]

Not quite what I was trying for 🤷🏻‍♂️

Fixed 🤣

[low stone]

sd3 vs. hunyuan. on more and more stuff, sd3 is barely prompt adhering at all. sdxl base is giving me better images at this point. half of me wonders if they're testing the 0.5b on their api right now. It feels like it got worse recently.

[hallow lion]

We need refiners and loras ans dpo and perturbed attention guidance and turbo and controlnet and apiadapter and LCM support

after that SD# will be better than those other things

[cunning lintel]

I think it's just first sd3 is new, wow, much amazing, try all nice prompts at at

Now SD3 has been there, tried all the nice prompt, let's try something new

[low stone]

maybe i've just been spending too much time with it... but even just the concepts have recently fallen off a cliff.

[cunning lintel]

ooops, sd3 doesn't do so well with something new 😢

[low stone]

everything i'm getting from it is this muddy mess now

[cunning lintel]

Maybe prompts are getting too complex

[hallow lion]

it needs to be refined with sdxl

[low stone]

countless school buses for countless children in a world where only school buses and children exist

looks what I got for this prompt.

that's just embarassing

[hallow lion]

reminds me of early sdxl

[low stone]

yeah exactly

[cunning lintel]

My main issue is, i'm not sure what's teased on twitter is better, yes, it's more trained, less broken gens

but an image with that many small people isn't showcased

[low stone]

people literally call out lykon for only posting simple images, and he responds back with 3 people holding little signs. He literally created dreamshaper which can do amazing stuff. Makes me feel like he's purposely avoiding.

then again he's posting on twitter for a high profile company, so he's probably limited in what he wants to post.

[cunning lintel]

Yeah, in one reply he said marketing would post prompts, obviously that means he isn't allowed to

@low stone it might be more than just sd3 getting "old", have you tried old prompts, i noticed two i tried getting different results now #🆕｜sd3 message

#🆕｜sd3 message

could be just unlucky rerolls 🤷‍♂️

#🆕｜sd3 message

Or not

[low stone]

sigh

[raven fern]

kek

the magic school bus

[hallow lion]

Needs boring school bus lora.

[rain current]

ideogram+sdxl

[dull star]

@low stone https://github.com/lllyasviel/Omost

automatic regional prompting basically

https://github.com/lllyasviel/Omost?tab=readme-ov-file#parameter-location-offset-area

[low stone]

hah i just saw that. first couple of examples, ok, that's kind of neat. by the last example where they've got seriously complicated images going on... really impressive stuff.

this isn't as simple as ella-sdxl, but it probably gets to the same place in he end.

but it looks kind of complicated to get going.

https://huggingface.co/spaces/lllyasviel/Omost

ooooo they have a demo spcae

[hallow lion]

https://www.youtube.com/watch?v=ZjbNXnbhnJY

see this is what im talking about but for images

select somethign and lock it - keep it as is in subsequent promts

[noble coyote]

It starts to compartmentalise the photo, like a grid!!!

[low stone]

first result from it is meh

no book, no monks, no shark in a robe.

hunyuan / Omost / sd3

it shows promise.

the only downside, is that the llm piece where it's writing out the "code" for the renderer, takes a LONG time. it's a very large amount of tokens to generate.

I guess the upside is that it probably works with any sdxl model instead of something like pixart or hunyuan where you have to train something new.

[noble coyote]

How can it accommodate so many tokens ... way more than the usual 77?!

[low stone]

I'm assuming that it's working on the image in stages instead of trying to throw all that at once.

[noble coyote]

Most likely...

[gusty trail]

Very cool idea

[noble coyote]

(I'm d/loading fp16.safetensors)

Missing module TRITON!

[low stone]

https://huggingface.co/spaces/lllyasviel/Omost you can also just try it out here if you want to see how it does with your prompt.

[noble coyote]

I have tried the online space - it eventually "cannot find a free GPU!" 😄

i.e. you've had your turn, now over to somebody else! 😉

Here is what the online space generated #🆕｜sd3 message

[proven pecan]

"a young lion sleeping on top of an elephant with a giraffe in the background in thelush berlin zoo" is too much to ask for I guess

[sullen moss]

And which model does it use for image generation?

[dull star]

its so weird, omost has a DALLE3 feel to it

I wonder if DALLE3 is just a pipeline that uses an SD3-like model and does some weak regional prompting

wait wtf the local version only takes 8GB of VRAM

[proven pecan]

Change the lion into a tiger cub, add a giraffe 😆

Oops I didn't mean to ping you @sullen moss

[sullen moss]

No problem