🆕｜sd3

[dull star]

@low stone

so unlike pixart you can highresfix it seems

[tropic aspen]

I mean, keep in mind it's going to be one of the smaller models as per another developer

Hopefully the 2b one, but I feel some ppl are going to disappointed when it isn't the 8b one

[low stone]

I will be. After working with pixart though, I understand that people don't want near 1 minute render times for first ksampler on a 4090.

Rendering with an llm text encoder is slllooowww

That said I think that was the power of cascade, first stage really low res and then just upscale that

I'm doing that now with Pixar and Ella. 512 first stage, then upscale

I've got a 4090 and I'm still doing that

[dull star]

2B is 512 anyway

[hallow lion]

Just fix the hands please.

I'll wait another 3 months just fix the anatomy.

[dusky thistle]

yeah... all depends on how big the diff really is

if it's like pixart 512 vs 1024, no biggie

if it's like sd15 vs sdxl, or sdxl base vs a sdxl finetune for concepts, then well, shit

Not gonna lie, the phrasing of that tweet is a lil weird

"as far as I know"

[woven dock]

tbf the base images shared by lykon on his twitter all look like an sdxl finetune

or better

[dusky thistle]

i'm not talking about the quality of the image

[woven dock]

They don't look any noticably better than an sdxl finetune but I'm still keeping my hopium up

[dusky thistle]

if you're patient and have a good finetune, you can make some really good looking shit even in sd15

[woven dock]

true

[dusky thistle]

it's about the underlying architecture and understanding of concepts, especially ones it wasn't trained on

[woven dock]

Some 1.5 finetunes look better than xl finetunes

[dusky thistle]

idk about that, but point is, image quality is a poor assessment of the capabilities of a model

ability to follow a prompt is important, but yeah, even more important, concepts

can you use it to make a shark mouth so huge that a city has been constructed inside of it?

shit like that

can you use it to make a spider with bat wings holding onto a witch (like someone on a motorcycle) flying on her broom?

etc etc

[dull star]

it does look like dreamshaperxl

[woven dock]

At the very least they look better than the api

[dull star]

it definitely will

he is using a local version

which is more up to date, uses more steps and uses highresfix

its gonna look better

api is watered down

and just raw result

[woven dock]

unfortunately the 8b, 4b, 2b and 800m versions aren't cross compatable

So more fragmentation yippee

[tropic aspen]

I mean, the dev team seems hopeful someone will make an x-adapter thing so that they are cross compatible

[woven dock]

I'm gonna be so disappointed if 4b or 2b catches on and people finetune it to make it look like 8b while 8b is left to the dust

[cunning lintel]

@dull star so that made me look at that series of tweets. I've been pretty optimistic about this whole thing, but working on the open release version, implies there's a closed version. Yay shitty low end model for the people 😥 https://twitter.com/Lykon4072/status/1791839648987156525

Lykon (@Lykon4072) on X

@cibernicola_es Since my team is working on the open release version, it would be odd if we end up not doing it.

Twitter · May 18, 2024, 2:33 PM

[tropic aspen]

Well yeah. They said they were going to release one of the smaller models first, and then maybe one day release the bigger models

[dull star]

smaller models first

is facts

comfy said that iirc

[dusky thistle]

big model never was what we were told

[noble coyote]

SD3@ClipDrop

SD3@ClipDrop

SD3@ClipDrop

[cunning lintel]

But this makes it seem more like only smaller models,, not just first. Time will tell.

tbh, seeing what pixart does with .6b params, comparing that to what sd3 produces, i wonder why 8 are needed.

[severe phoenix]

Maybe someone can direct him to this channel if he wants "MORE sd3 images". wth is even this statement lool. seems like needless fanboying. please when sd3? emad said it was "coming shortly" 3 months ago. i aint mad they got him outta there tbh. the fact that stability doesnt have a leading website in ai gen like leonardo nd co is just insane. wth is clipdrop and that stabledream thingy. Everything runpod is offering, stability should've had a better version of that by now. its crazy how far behind they are in that space.

[dusky thistle]

yes, the plan comfy shared was that 8b would not be released to the public

[woven dock]

nah they said all models will drop

[dusky thistle]

not 8b

[dull star]

I do remember him saying that 8B will come later, but with a stupid ass reason

[tropic aspen]

Source?

[woven dock]

do you have a link?

Lykon said all models will drop at one point

[dusky thistle]

it was a long time ago in ai years somewhere on here

honestly

i don't think any of it tells us much of anything

[severe phoenix]

i'm not that worried about this. the difference between 8b nd 4-6b ones isnt that great from what i've heard. besides most ppl dnt have the resources to run the 8b one

[dusky thistle]

i'm inclined to believe even top SAI employees don't know where this will all end up

[dull star]

idk about the fact that it won't come out at all, but I definitely know that 8B comes later

[dusky thistle]

unless they're ready to release before any negotiations for a sale begin

[dusky thistle]

true, most ppl don't... i do though as do a lot of ppl on this discord

so we're going to care

[cunning lintel]

probably not.

[woven dock]

you'd need at least a 16gb card to run 8b anyways

[dusky thistle]

won't be able to train 8b on 24gb vram, sure, but it's not like ppl aren't renting time on a100s anyway

[woven dock]

Maybe 12 if you do some offloading

[dusky thistle]

yep, got a 4090

if the 5090 has 32+gb vram i'll buy that too

[severe phoenix]

fair enough, but i dont though i'm not a high roller like y'all lool

[woven dock]

I'm probably getting a 4090 whenever the 5000 series drops

[dusky thistle]

either way, sdxl is what... 3.5B?

[dull star]

yes

[dusky thistle]

not rich but don't spend on much of anything these days

[cunning lintel]

Tough that sale "news" seems noting but rumours from an over eager reporter as well, no mention of sources anywhere.. if you keep repeating that every month, it'll be true once 🙂

[tropic aspen]

SDXL can run in less than 4gb using fp8

[dusky thistle]

got bored with eating out, got bored with drinking, and i have a big oled in the basement so i don't go to theaters or anything like that anymore

all my sports teams suck and i lost interest years ago so i don't spend on that either

adds up to a crazy amount of money every year

say on a weekly basis, you bought a coffee for $5, ordered a single shitty pizza for $10, and had a beer and a burger on saturdays at a bar conservatively for $20

that's a 4090 every year with some change left over

[woven dock]

sdxl finetunes are still somewhat undertrained for its parameters

unlike the thousands of 2gb 1.5 anime merges

[woven dock]

FP8 scales linearly with n-b of parameters

SD3 8b would be about 8.5 or so gb in size

[dusky thistle]

did you see the emad tweet?

i'd think he'd know what was going on, being the primary stakeholder... but who knows

[cunning lintel]

the sarcastic one?

[woven dock]

You'd need at least 12gb for the 8b parameters

[cunning lintel]

That seemed more like Emad being Emad and tweeting without thinking. He's posted denials like that last year as well, andhere we are

[dusky thistle]

could be^

[woven dock]

I still hope for 8b releasing and people making LoRA's/finetunes compatible with 8b

[gusty trail]

From Lumina paper, it claims bigger model would converge faster.

[dusky thistle]

this is a good point too

[cunning lintel]

That's going above my head. Yes, i understand the meaning of "converge faster" but totally opposed to that pixart is in a similar league as SD3 for many prompts, and it is trained way less than SDXL. Either SAI skimped on training SD3 (why would they) or there's more to it

[dusky thistle]

interesting

[dusky thistle]

it's also entirely T5

don't know how much of a diff that makes

[cunning lintel]

Not that pixart is the be all end all, but if i wasn't told it how small it is, i'd never guessed. It might be at the limit of what's possible for that architecture size, but it's impressive

[dusky thistle]

agreed

[sterile pendant]

Kind of a lot in some cases. Since you're mapping better language semantics with t5, they can map to clusters of concepts in the model faster, which would mean faster training of the models. Versus it kind of having to wiggle its way through in a more brute force approach due to having a less natural comprehension of the prompting

[gusty trail]

I need to say something dump. How good quality of the output is dependent on the training dataset aesthetic. How diverse image output is depending on how diverse of the training dataset. Pixart training dataset is focus on high quality and it is so limited and it couldn't produce many diverse output as sd3. For example, accurate text.

[dusky thistle]

it does also behave bizarrely with samplers and schedulers we're used to... hands in particular

festivalman noticed recently using the supreme sampler helped a lot with that

[cunning lintel]

My experience is kinda different. Except for text, SD3 and Pixart seem to struggle just as much. Maybe what I think of weird concepts is either too weird or rather normal (as sometimes it works, sometimes not). Plenty times pixart doesn't grasp a prompt, rather frequently neither does SD3 (I expected much more from SD3, to me it feels like slightly better SDXL, prompts where I thought SDXL should be able to do it, but needed much tweaking to get it, SD3 does at once, beyond that, it's often underwhelming to me)

[gusty trail]

This sounds like limited T5 on sd3

[cunning lintel]

Probably is, but that is what it is now, hopefully it'll be better in the final version. (I just can't wrap my head around the fact that none of that has become obvious in the development phase, like did they really only test prompts like "gorgeous girl in a cafe" and was "three bottles, first red, second blue, third green" the farthest their imagination went:, didn't they try to break it, find the failures modes)

[dull star]

its weird how even with the clip models alongside T5, SD3 still has the best adherence in my experience

[woven dock]

Doesn't hand quality really depend on the training data and resolution in particular

I also thought it was primarily the vae that fucked up the hands

[low inlet]

Hello guys i love you ❤️

[dusky thistle]

it does, but it does respond differently to samplers

for example... res_momentumized results in something resembling cfg burn at high step counts, which i've never seen anywhere else (sd15, cascade, sdxl)

[low inlet]

Errrm Guys i have a question

why does sd3 don't have cfg scale ? and refiner ?

also why sd3 renders hands pretty bad like most of the times the hands are poorly drawn

i thought sd3 won't struggle with something like this

[cunning lintel]

i love it! despite that criticism, i love it still can behave like clip, and happily eat wordsoup, then spits out amazing pictures. My fondest prompts are like that. I'd actually prefer it if it optionally can be even more like clip at times, better cloning styles from artists and such.

[dull star]

I wonder how much smarter it would be with T5 only or T5+ClipG

of it it makes no difference lmao

[cunning lintel]

still api only, so not many settings, might well be the rectified flow part plays bad with cfg scale, or not, i know nothing about it. the hands, no one knows, it's all speculation here, the model in api is presumably an undertrained one, but surely work in process, it might get better at it

[gusty trail]

hunyuandit paper tested and the result is clip + t5

[dull star]

but that is a special english-chinese model and a smaller T5 iirc

[woven dock]

is sd3 t5 XXL?

[dull star]

yes

xxl

[sterile pendant]

It's more of a parameter issue in models. Think of all the wrist angles the hand can be turned in, now think of how many angles each finger can be in, along with how many angles each knuckle can be in. Now picture all the non-hand things that can be in proximity to them, like holding an apple or grasping a lever, etc. It goes really exponential really quickly. Now imagine all of the training data you'd need to cover the majority of the combinations they can be in. This is why hands are always a pain.

They could get around that with artificial data using 3d modeling though

[woven dock]

So then is midjourney and dalle just massive parameter models?

[sterile pendant]

Correct

Probably model sizes 10x what you can run locally, if not larger

Dalle is probably way up there internally like 80b or something, but I'm just speculating

And probably broken up into different swappable chunks

For different tasks

[woven dock]

Ik for a fact MJ has a massive vae alone

[low inlet]

i really hope it gets better soon

i have been working on a system that can generate similar images with AI technology anyone wanna try ?

[woven dock]

Controlnet can solve the hands problem at the very least

[low inlet]

or if you have a cool prompt and want to test it in sd3 pls give me the prompt

like anyone here

[woven dock]

but then you also have naiv3 which also somehow generates proper hands

[low inlet]

why emad left the company ?

[sterile pendant]

Yeah, it's what I use when I need hands. I just model out in zbrush or pose a mannequin in ue5, then render a perfect depth map. Works pretty effortlessly usually

But I've been doing 3d modeling for over 20 years

Not a chance in hell I'd sit there trying to fiddle around with seeds and inpainting, praying to the rng gods

Total waste of time lol

[woven dock]

I don't really have that much 3d experience, the only time I've worked in 3d was with AutoCAD

Which isn't technically 3d modeling but you get the point

[low inlet]

but it will be limited each time u want to generate something :[

like you have to do it manually

[lavish sparrow]

pixart sigma

[trail frost]

Wow

[cunning lintel]

Hauntingly beautiful mixed media collage, influenced by the dark, mystical realms of Camilla D'Errico and Gustav Doré, with the surreal, dreamlike atmosphere of Masayoshi Matsumoto. A daring Russian explorer, fiery red hair ablaze, clad in a worn leather jacket, ventures into the mystical underground cave, clutching a radiant crystal, as shadows dance, concealing ancient secrets, waiting to be unearthed
(SD3 and pixart)

A richly textured, dark fantasy tempera painting, inspired by Craig Mullins and Jose Royo, showcases a dynamic trickster djinn, adorned with ornate, golden chain jewelry that seems to shimmer in the dim light
(SD3 and pixart)

[hallow lion]

I cna live with that. I was scared SD3 models will be 12 gigs or more...

[stoic turtle]

Why shouldn’t they be

[dull star]

but that's fp8

[stoic turtle]

Larger

[dull star]

We should expect fp16/bf16 model sizes

these are SD3 results

it may seem shoddy, but pixart does not even come close to this

with pixart I keep getting weird stuff

and I won't even attempt ELLA as that also uses a smaller T5 and it doesn't work with text either

I just cannot live with pixart only. As good as it is, as greatful I am for it (like its also commercial openrail++ too), it just doesn't not have the prompt adherence and text capabilities of SD3.

For a 0.6B model though, its still very impressive nonetheless.

can't wait to see how 0.8B SD3 compares to this!

[low inlet]

Hey

How do you know which version is available ?

also do you know a good way to fix hand issues ?

because i keep encoutring so many issues with bad hands renedering even with using negative prompts

[dull star]

negative prompts don't help with bad hands

you'll have to wait for finetunes for SD3

[low inlet]

daymm that will take a while

[dull star]

yup

[low inlet]

releasing it then making the finetunes

but which finetune do you think it's better ?

[dull star]

also this is an older version with probably lower step count and no highresfix

so we'll have to see how it performs locally with any step count we give it

[low inlet]

realvision or jagranut

[dull star]

and Lykon (a dev at stability) will probably immediately start making DreamShaper finetune for SD3

[low inlet]

yes i was wondering why i can't adjust the step count and cfg scale

it's so weird i didn't find params for it

[dull star]

API probably cheaps out on the step count, but I could be wrong

highresfix also fixes small faces and such

I use it all the time with SDXL

can't wait to see what difference it makes for SD3

[low inlet]

yes i know highres

i also know about adetailer

i tried to used it to fix hands but it didn't work as expected

i mean it works with faces

but with hands nope :[

[low inlet]

i mean they should make the price flexible based on step count idk

but if the step count gonna give better results defiantly i will make it higher

[dull star]

🤷‍♂️

we'll see

[low inlet]

can you give me a prompt to try it out please

[dull star]

hopefully they release 2B by the end of the month

[low inlet]

what is the prompt for this

[dull star]

hol on

Photo of Criminal in a ski mask making a phone call in front of a store. There is caption on the bottom of the image: "It's time to Counter the Strike...". There is a red arrow pointing towards the caption. The red arrow is from a Red circle which has an image of Halo Master Chief in it.

[low inlet]

How do you know the size of it ? and what size is currently being used because i searched but i didn't find this info

i want to follow these updates

[dull star]

there will be 4 model sizes

800M, 2B, 4B and 8B

2B in my opinion might become the most popular one

[low inlet]

andd .... which one is being used in the api right now ?

[dull star]

8B

[low inlet]

oh

soo ... that's the largest one and it can't even render hands :[

that's sad news

[low inlet]

i will try it out wait

this is dall-e 3 results :

And this is sd3 results :

that's very bad

[dull star]

that's weird?

Dalle3 looks like the typical results I get with current SD3

and the SD3 images you generated look like the results I'd get with like the earliest available version of SD3

[low inlet]

should i remove the negative prompt ?!

or what

wait lemme try again

[dull star]

all you need is "ugly" in the negative prompt, or not even that

[low inlet]

okay wait i will try

ugly, ugly eyes, ugly face, deformed eyes

Photo of Criminal in a ski mask making a phone call in front of a store. There is caption on the bottom of the image: "It's time to Counter the Strike...". There is a red arrow pointing towards the caption. The red arrow is from a Red circle which has an image of Halo Master Chief in it.

that's still bad

honsetly i thought sd3 base model would be better than this

do you think that the issue might be with the prompt itself ?

[dull star]

no this looks okay to me for a base model result

we always overestimate how good base models are supposed to look

[low inlet]

But how dall-e 3 is this good ?

and it's like the base model only

[dull star]

DALLE-3 is a service

(so is SD3 right now, but its cause they need the money)

Stability knows how to make a sexy model, they have Core

Core is super good quality and finetuned and its on the API

SD3 is currently raw and possibly underfed with POSSIBLY low step count...

[bitter hearth]

((definitely not because the model is still in training))

[bitter hearth]

where is core

[low inlet]

i see and i wish that it develops and competes with modjoruney and dall-e3 it would be great for the open source community

[raven fern]

meow

[oak pier]

question what does sd3 think "iris" is

1.5 read it as a eye in sdxl it read it as a flower xD

🤔

[gusty trail]

https://glyph-byt5.github.io

[raven fern]

hopefully we can plug and play some of these t5 models

[hallow lion]

lol

mold

[dull star]

really love the bubbly mold surface on the water

[hallow lion]

it does reflections well too here

I wish it cna be used for serious 100% lifelike images already and maybe movies later on as in real serious good expression

not just gimmicks and mEmEZ

I hope refined SD3 is good enough

as for the tools

yeah...

not yet

[dull star]

idk highresfix will be there

I'm gonna abuse it so much

[hallow lion]

i wish upscalers werent so slow and resource hogs

supir devours everything

usually i disable it :/

i only upscale if i come up with soemthing reaaaally cool

i know other upscalers are faster but nothign coems close to supir

[cunning lintel]

In a desolate, post-apocalyptic wasteland, the milk packaging's face sets into a determined scowl as it discovers a hidden, abandoned laboratory, its bespectacled eyes narrowing with determination, its entire body tense with curiosity.

In a hidden, underground garden, the milk packaging's face relaxes into a serene smile as it discovers a tranquil, crystal-filled oasis, its bespectacled eyes softening with contentment, its entire boxed being radiating a sense of peaceful calm.

[hallow lion]

Where do people come up with this?

Lol, are thes elike fever dreams

[hallow lion]

You guys use promt assists or just type this all in?

[cunning lintel]

well, i ask llama to enhance my dumb ideas 🙂

[hallow lion]

local install or can it be done online somewhere? :0

[cunning lintel]

https://groq.com/ (free but effectively a capped context length at 6k) using https://github.com/lobehub/lobe-chat, can even use things the new gpt4o and such, plenty inference services to plug in it

[low stone]

Dalle vs. sd3

[cunning lintel]

(in this case i just asked "create some scenes of a humanoid milk carton discovering the world. Create 5 variations, make it discover an underground lake, and describe the (human) characteristics of the milk carton" then picked the good ones and iterate on those. turned out carton made it often like cardboard, so changed that)

[cunning lintel]

but what was it supposed to nail. the tire i guess 😂

[low stone]

Just typed it in. Prompt: rabbid anthro Cactus in a straight jacket, foaming at the mouth. In a padded cell. Mentally insane. Surrounded by anthro bubble guards in uniform.

[cunning lintel]

That image would be epic if you managed to get the bubbles 🙂

[noble coyote]

Free DiT Generation https://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiT

[low stone]

hunyuan vs ella/sdxl

although I guess I should put that through the refiner as well to see what it does with it.

ok fine that's pretty cool. 🙂

where comfy nodes for hunyuan? 🙂

please to be making comfy nodes kthxbai

@noble coyote https://github.com/city96/ComfyUI_ExtraModels/pull/37

it's happening!

[icy drift]

You have ella sdxl??? The author said training ella-sdxl was "prematurely promised to something beyond my control" in his latest blog update about emma. https://wrong.wang/blog/20240512-what-is-emma/

[low stone]

actually hunyuan makes some pretty good stuff too... original on left, sdxl refined on right

[low stone]

nah, just refinement stage.

[icy drift]

But it's fully censored and has terrible prompt adherence. E.g. if you want a redhead woman with a blue hoodie and white-haired man with a green tuxedo, it will just throw all the colors in there however it wants, ignoring your instructions. Kind of defeats the purpose of all that extra processing power.

[weary crystal]

pixart-Sigma with SDXL finisch works. Will try it with Ella

[low stone]

this is in the negative prompt on hunyuan.... i'll try it without it.

[weary crystal]

ELLA seems to work fine too

4 out of 5 images had a blue red girl and a green gray/white man

[low stone]

yeah hunyuan definitely mixes subject details.

[weary crystal]

At least they seems happy 🙂

[low stone]

it's also worth mentioning that their settings are really for low quality. low cfg, etc. might be made better if we had this local to fool with sampler settings.

now all we need to do is refine it with something that's even less prompt following than any of that.

[noble coyote]

I don't think my 8Gb RTX 2070 has the right power for DiT?

[low stone]

certainly works well as a composition stage for sdxl refinement

instant present, glass box, tiny universe of bright colorful pixar monsters living and playing inside, christmas wrapping paper bow, cinematic holiday setting, outdoors holiday wooded scenery, cute animated, 8K, hyperrealistic 3D art, ambient dynamic soft lighting, intricate highly detailed, bright pastel colors, extremely high-resolution details, photographic, realism pushed to extreme, fine texture, incredibly lifelike

so hunyuan on right was actually more prompt following than ella here.

so like everythign else, good for using it alongside other stuff.

sd3 beat both of them.

really good output here.

(yes sdxl refined, but sd3's composition was excellent)

another sd3 refined output. seriously good

[weary crystal]

like the prompt 🙂

[low stone]

sd3 is gonna be awesome when its finetuned

[dull star]

yessir

even base will keep me busy

making crapton of images

[weary crystal]

For now only pixart sigma with sdxl can reach it (local)

[raven fern]

nice

[low stone]

really good

it's kind of what I mean, because all of these have different training levels, it's best to throw prompts at all of them simultaneously. 🙂 oh the workflow.

[weary crystal]

With some more steps it comes close.

And it is a great prompt @low stone, funny little creatures

[low stone]

there's a couple of prompt twitter feeds that are good for this stuff and then I change details (usually includes furry little animals). #PromptShare and i subscribe to a group called PromptAlchemists

[gusty trail]

sdxl ft only

[quick maple]

draw a logo which named River Tree，it is a company name which operating nonwoven machines, use green and black color

[low stone]

hi Rachel? what's a nonwoven machine?

[weary crystal]

This at least from the alibaba sales site of river tree 🙂

[low stone]

I really wonder what's up with these bots. they spammed all of the channels with the same completely random request. I haven't figured out the point of it all.

what do they gain by endlessly doing it?

[weary crystal]

Easy explaination, on an other server i am one, a visitor has told us he watched a video (chinese) and there was explained howto generate images with discord. As sometimes they do not know english they just either repeat a request from another person (guy in black robe,....) or try just a prompt. As soon as they see it does not work they leave again. Just simply too many 🙂

[low stone]

ah ok that makes sense

so it's a script or something

and they're not even on here.

[weary crystal]

No script, real people who follow a youtube (tiktok,...) tutorial. Guess from a time the bots where on. On the other discord server they searched for the invitation to the blue willow discord as it was shared there....
Not the smartest way but sometimes reproduction is easier then thinking...

[gusty trail]

bots are well designed. I don't think anyone able to develop bot who can't find a proper way to generate image. Human dump.

[bitter hearth]

im bot

[gusty trail]

you are well designed

[cunning lintel]

Is this you 🙀

[bitter hearth]

wtf

[weary crystal]

Maybe this pixel cat with low vram setting 🙂

[flint quarry]

50

530

[bitter hearth]

640

[low stone]

Sd3 with sdxl refinement

[woven dock]

They don't act like real people when spamming the same message across multiple channels

Usernames are also autogenned

[weary crystal]

No they are generated as if you just one to create "any account". And most know that a james123 works better then chinese or other characters

[steady girder]

A photorealistic advertising photo of the world's best food delivery man, holding an lunchbox with a dark gray body and green accents as requested. The lunchbox has a compact, rectangular shape with rounded edges and a green handle for easy carrying. The lid features green latches and a circular section in the middle, suitable for utensils. The design is practical and modern, The background is plain white to keep the focus on the lunchbox, white background, front view, wearing overalls and an orange cap smiling at the camera, bright studio lighting, in the raw style, on a white background, isolated from edges, the shot taken with a Canon EOS, using a wideangle lens, with natural light.

[bitter hearth]

how about no

[wild remnant]

At large .....WANTED

[wild remnant]

Loll

[bitter hearth]

x3

If they see one of those, they are already scared

[wild remnant]

Mmmmmm

[raven fern]

im hungry

[bitter hearth]

chicken hamburger is better

[high vector]

Morpheus King of Dreams - via SD3 api

[bitter hearth]

yo bro how can you use SD3?

oh the API ... damn, i want it on Fooocus already

[low stone]

yeah, big one is seemingly done, they're working on the smaller versions now

[bitter hearth]

to big one is like llama70b and small version is like llama8b?

like... they quantized the vectors and such?

like, instead of 1.23234234 , it's just 1.2 or 1

something like that? have they agreed on their differences

or the ex-CEO and the current bosses are still fighting?

"settled their differences" that's the word i was looking for

huh

[little quarry]

daily two more weeks post

[sterile pendant]

Diffusion models are different and don't really quantize well. Going from 32bit to 16bit produces almost identical results from a data perspective (maybe if the SD models were like 80B sized, you might start to see a tiny difference between fp32 and fp16). Going from 16 to 8bit will absolutely destroy diffuser accuracy though. Instead, SAI is training different versions for the different model sizes like 2b, 4b and 8b and each will likely be used at the standard 16bit precision. Though their file sizes will be bigger or smaller

[bitter hearth]

I see. Very interesting.

[sterile pendant]

But each version is using more or less of the dataset they're using to train the models. So the 2b model will not have the entirely of the 8b training set in it, but a fraction of it

[bitter hearth]

Is there a way to optimize all of this?

Is there a way of instead of letting the GAN train itself on only images, could it train itself on concepts?

Like, we can label some things as "this is artistic" , "this is not artistic" something like that

like RLHF

u know

[sterile pendant]

That's what they train already though

[bitter hearth]

Ah I see

[sterile pendant]

Essentially the models store things in extremely high dimensional formats and the training set is what they use to map token inputs to "concepts" in the N-dimensional tensor space

[bitter hearth]

Interesting

[sterile pendant]

It's an extremely complicated topic though, but if you want to learn more from a math standpoint, numberphile and 3blue1brown have some good videos that cover the basics and other math related things they all share together

I'm sure there are other good breakdowns, those two just pop out immediately in my head

[bitter hearth]

Oooh you're right . 3blue1brown that guy is great

i need to learn those maths

i'm just so bad at maths man

I'm good at reasoning and connecting the dots, linguistically and visually

but symbollically and maths and stuff , oof

I can never quite get it... But they always appear when I need something like, if i need to program something that's fast or efficient , i take a look at maths and physics, and then focus on one specific equation or set or related equations

[bitter hearth]

you there still?

[bitter hearth]

sent you an FR 🆗

[sterile pendant]

Well that part is more important if you want to understand the gross overview of how things work. The actual math isn't as important, unless you plan on directly doing stuff with it.

[bitter hearth]

I see.. 🤔 nice

[sterile pendant]

And sorry, but I don't do messages or friend requests on public servers. But I tend to randomly answer questions on here and show experiments with workflows

Oh and sometimes I make anti-waifu memes as well if people keep spamming cliché waifus lol

[bitter hearth]

I see brother, no worries

[bitter hearth]

behold

10 years of AI for this

x3

[frigid saffron]

this gen by AI?

[bitter hearth]

Yes

[quick solstice]

love

[icy drift]

No way. Prompt???

[noble coyote]

SD3@ClipDrop

SD3@ClipDrop

Prompt = Don Quixote and Sancho Panza as Warrior Astronauts, Geisha, Samurai, Shogun, Teddy Bear Astronaut Warriors, style of eric ravilious, josef vachal, oxana dobrovolska, alenka sottler, catherine abel, victo ngai, vladimir kush, henri rousseau, fauvism, cthulu, peppa pig, spongebob

[frosty gorge]

#1237459938901491852 A Gorillaz stylized 2D advertising photo of the world's best director of photography, holding a camera and smiling.

[hexed cobalt]

Hallo

[low stone]

The "meteor" that wiped out the dinosaurs was actually a t-rex doing a shoryuken.

[hallow lion]

Dragon Ball-D

[craggy ridge]

An ancient, hyper-detailed photo depicting a caveman tinkering with a smartphone inside a damp cave. The caveman, with rugged hair and primitive clothing, is engrossed in the smartphone, which emits a soft glow. The caves walls are covered in ancient carvings and lit by a dim, natural light filtering in. Water droplets glisten on the rough stone surfaces, highlighting the caves dampness. The overall color palette is muted, with earthy tones and slight sepia tint to evoke an ancient feel. The scene is captured from a low angle, emphasizing the contrast between the primitive settings and the modern device.

[little quarry]

He has been waiting two more weeks for SD3

[muted dove]

He's been waiting since the dawn of time!

[muted dove]

For comparison...Pixart-Sigma with SDXL, using the same prompt

[honest cedar]

Hi, what's the cheapest way to generate with SD3 currently? Or the price is always the same, no matter the service?

[weary crystal]

Price is the same. Some optimize their workflows by starting with 2 images in lower resolution and ad soon as the prompt work as designed they increase resolution etc.

API or Discord should not matter.

[honest cedar]

Thanks for the info

[dull star]

sdxl (finetunes) cannot bear to make someone not look like a handsome model with lush hair

lmfao

[bitter hearth]

He's becoming a rock

[dull star]

dayum

[bitter hearth]

Shower is in urgent need Cx

CJ has access to sd3!

[last rapids]

Legend has it that the Empress Wu Zetian's two most precious treasures, the nine phoenix hairpins emit mysterious light, and Wu Zetian's image is majestic.

[bitter hearth]

All false, sorry.

[muted dove]

Not if you add some "dirt" to the prompt 🙂

[dull star]

that looks better

[bitter hearth]

actually looks worse but contextually better

:p

[round vale]

An ancient, hyper-detailed photo depicting a caveman tinkering with a smartphone inside a damp cave. The caveman, with rugged hair and primitive clothing, is engrossed in the smartphone, which emits a soft glow. The caves walls are covered in ancient carvings and lit by a dim, natural light filtering in. Water droplets glisten on the rough stone surfaces, highlighting the caves dampness. The overall color palette is muted, with earthy tones and slight sepia tint to evoke an ancient feel. The scene is captured from a low angle, emphasizing the contrast between the primitive settings and the modern device.

Prompt = Don Quixote and Sancho Panza as Warrior Astronauts, Geisha, Samurai, Shogun, Teddy Bear Astronaut Warriors, style of eric ravilious, josef vachal, oxana dobrovolska, alenka sottler, catherine abel, victo ngai, vladimir kush, henri rousseau, fauvism, cthulu, peppa pig, spongebob

[bitter hearth]

sir this isn't a carnival

that one is just named bot, very straight forward 👍

[low stone]

I appreciate their honesty

Makes it easier to report

[dull star]

fr 🙏

[bitter hearth]

crayon drawing of a pretty girl, clouds on the background, very low quality kids drawing

I'm not really sure why you wanted the prompt but there you go

only remembered now cx

[noble coyote]

Clipdrop.co, $10/month for 40 images a day, which amounts to 1,200/month

[bitter hearth]

that man is a walking ad for that site

I thought there wouldn't be ads in discord guess I need nitro or something

[low stone]

Wow that's a 7x cheaper price than from stability directly. Wow. Are they hosting sd3 themselves perhaps?

[low stone]

Do they have the ability upscale it at all?

[cunning lintel]

isn't it 4 images per gen, so effectively you get 10 prompts to play with per day, as they don't accumulate either.

[neon wagon]

more like he has access to std

[dull star]

https://tenor.com/view/damn-damn-damn-damn-damn-right-dame-time-gif-18029769

[sullen moss]

https://twitter.com/argmaxinc/status/1790785157840125957?t=BbmC8uJFayQwuXcLrXJlbA&s=19

argmax (@argmaxinc) on X

On-device Stable Diffusion 3
We are thrilled to partner with @StabilityAI for on-device inference of their latest flagship model!

We are building DiffusionKit, our multi-platform on-device inference framework for diffusion models. Given Argmax's roots in Apple, our first step

Twitter · May 15, 2024, 4:43 PM

[low stone]

I hope that's somehow faster than "dog" slow sdxl on Mac currently

[rotund ibex]

• pixel art, by ashley wood, yoji Shinkawa, Tsutomu Nihei, ink, ink splashes, dynamic pose, demonic toad, red straw hat, holding an instrument, teal skin, cyberpunk aesthetic, white gradient background, prosthetic arm,

[craggy ridge]

Three hyper detailed, desperate people dressed as jesters, standing in a thunderstorm with heavy rain and lightning. One man holds a sign with the text We, another man holds a sign with want, and the third man holds a sign with weights. The scene is drenched, with water dripping from their jester outfits, illuminated by flashes of lightning, capturing the intense, somber mood.

[noble coyote]

Yes 10 prompts x 4 images = 40 images x 30 days in a month = 1,200 images/month

[noble coyote]

You get a choice of widescreen, portrait, square or landscape; choice of style and negative prompt ... but no upscaling

[low stone]

Ok so same as api. Turns out the upscaling on artisan sd3 upscales with sdxl turbo, not sd3. So meh. :/

[dull star]

*50827381923th mention of open model weights*

everyone in the community: "man this is just evidence that the models will never come out "

[bitter hearth]

https://tenor.com/view/clueless-gif-24395495

What you mean trusting company words for things ?!

[dull star]

fr fr

they are just lying

[crude yarrow]

Healthy skepticism is a good thing as long as you don't go full doomer. Never go full doomer.

[hallow lion]

When the weights drop they'll create a mile deep crater.

[cunning lintel]

Rumor has it SAI struck a big deal with nvidia:

[low stone]

but only with purchase of a new 5090

[hallow lion]

Soon to have 64GB of ram my cat.

Realtime HD Waifu.

Dalle and Mj can't keep up. The uncaged cleabage erodes them.

[hallow lion]

Can we get free SD3 weights retroactively? 😄

[low stone]

yay for sd3's mangled hands

this would be so epic if it wasn't for the octopus hand

sigh.

[bitter hearth]

hiers will do a lot

[low stone]

well, the stuff that fixes all that locally is better training + all of the AYG/PAG/autocfg stuff

[bitter hearth]

and the full model

[low stone]

with ella. hands are fine

but of course the "selfie" part isn't really right.

which sd3 does get correct

[wanton oar]

Why I am getting Your organization does not have enough balance to request this action (need $0.016, have $0 in active grants, $-0.3535949 in balance) issue even though I have enough credits there ? It only getting issue sometimes. Is there anyone who are getting such an issue ?

[bitter hearth]

what is your prompt

the eyes cx

gta sa graphics lmao

[low stone]

that's certainly a way to fix it, by just not having hands at all. 🙂

they look really good though. 🙂

@dull star @bitter hearth that said, I got hunyuan going in comfy.

[raven fern]

that's not a problem technically cause only one person has to buy the 5090 and then share the sd3 weights with everyone 🙂

[low stone]

another hunyuan

[raven fern]

is the hunyuan dit thingy not available in comfy still?

[low stone]

another hunyuan

[low stone]

the comfy extra models nodes added a 1024x1024 version.

[raven fern]

ah cool, thx 🙂

[bitter hearth]

Most of the time when I want selfie style pictures I want the picture itself and not the person taking a picture xD

[low stone]

i know i'm just being silly. it's actually more "selfie"-ish the way you have it.

when you get a chance, can paste that prompt? I'd like to try that with my local stuff.

[bitter hearth]

girl selfie with a giant robot behind her destroying the city and explosion and chaos happening, she is smiling, in anime style

extremely sophisticated prompt

[low stone]

yeah that works

[low stone]

hunyuan might not have the best prompt following, but I'm digging the composition it does which is different than the other 2 major ones out there. yet another thing to throw things at to see how it varies.

[low stone]

^ sd3 refined

[noble coyote]

SD3@ClipDrop - prompt = dinosaurs ridden by batman and robin and the penguin and the joker and the riddler, at the kentucky derby in the style of dinosaurs, eric ravilious, josef vachal, tamara lempicka, victo ngai, henri rousseau, vladimir kush, oxana dobrovolska, alenka sottler, catherine abel

Weighting for The Waits anybody?!?!? 🥳

[brazen flame]

Line drawing of the Shanghai Bund, ratio 25:9, warm colors

[noble coyote]

SD3@ClipDrop into ComfyUI for SDXL+LoRA+PAG Advanced+Face Detailer finishing

[frigid saffron]

Image Enhacne Magic [from left to right] : @bitter hearth

[lavish sparrow]

don't be sleeping on pixart sigma!

[low stone]

Couple of refined sd3 and hunyuan(square ar ones)

[hallow lion]

Drop the weights.

[muted dove]

https://tenor.com/view/im-done-done-weight-lifting-weight-weights-gif-5486227

[bitter hearth]

gimme this prompt cx

[wild remnant]

@bitter hearth Body cam footage of a cave exploration

[bitter hearth]

[wild remnant]

Lol

[bitter hearth]

shrek horror game xD

[wild remnant]

Hes a killer

[noble coyote]

SD3 made at ClipDrop; then into SDXL, LoRA, PAG Advanced, Face Detailer

[cobalt moon]

damn man. someone from Twitter ( that person is the one who adapt LORA into diffusion model ) have announced the soon-release 5B DiT

https://twitter.com/cloneofsimo/status/1792849614120587616

Simo Ryu (@cloneofsimo) on X

Normal people's hobby : listening to music, sports, video games...
Me : speedrun pretraining 5B T2I DiT from scratch under 3 weeks

RELEASING SOON!!!!! (btw this is pretrained ver, gotta train on hi-res)

Twitter · May 21, 2024, 9:27 AM

[cinder junco]

If that’s all it can do with 5B nearly at release, then it unfortunately has no practical use. It’s an interesting proof of concept, but severely undercooked.

[solid mauve]

/dream/retrofuture laboratory scientist combining liquid squid and eyeballs and glowing marbled chemicals from chemistry glass into flower of life bowl , surreal 8-bit grunge ukiyo-e --no DOF, vignette --chaos 33 --ar 1:1 --sref 918740544 --stylize 333

retrofuture laboratory scientist combining liquid squid and eyeballs and glowing marbled chemicals from chemistry glass into flower of life bowl , surreal 8-bit grunge ukiyo-e --no DOF, vignette --chaos 33 --ar 1:1 --sref 918740544 --stylize 333

[celest idol]

cat image | sd3

[cobalt moon]

Hey calm down bro. It is just a poor man's SD3 with super little cost

[tropic aspen]

It's possible it's very undertrained tbf

[cobalt moon]

Hm

[tropic aspen]

It was only trained for 3 weeks

[cinder junco]

Interesting proof of concept

[tropic aspen]

If the tweet is anything to go by

[cobalt moon]

He only starting the training 4 days ago

When he post this
https://twitter.com/cloneofsimo/status/1791135683362631698?t=v6eAbH6gTdNeJcHrT3lK3w&s=19

Simo Ryu (@cloneofsimo) on X

moon is high, model is 44k steps in, I stopped the run to check on everything and use multi-nodes, didn't expect anything at all.... However, safe to say, i've trained my FIRST ever 5.6B Text2Image MMDiT from scratch!!!

Twitter · May 16, 2024, 3:56 PM

[cinder junco]

This is why we still don’t have SD3, and it already blows these results out of the water.

[cobalt moon]

So you simply ignored all those image shared by people here?

Sure it have flaws

[cinder junco]

No, I saw them. They’re great.

But would you be okay asking this guy’s work for a dog and receiving THAT dog?

[tropic aspen]

Hmmm... maybe the architecture could be good?

LIke, if the Pony guys trained on it

[cobalt moon]

I dont know if you can say the difference between individual-train or corporation-train with thousands of H100

Simply you just get an early 2023 AI-generated dog

[cinder junco]

The majority of the work is training. The architecture is just the base.

[tropic aspen]

So then maybe the Pony guys can do smth with it for V7

[cinder junco]

Architecture defines the upper limits of capability. But getting there is all about the training.

[cobalt moon]

It use SD3 architecture aka MMDiT. Like he train the models out from the papers

( not sure whether there is T5 though )

[cinder junco]

I feel like this guy has an unrealistic definition of success if he thinks the model is almost ready for release. If it can’t meet the capabilities of SD1.2 base from way back when, why would you use it? It’s more computationally expensive and you don’t get any benefit.

[tropic aspen]

For the architecture?

[cinder junco]

Thus, it’s not meant to be practical, but just to demonstrate what can be done by one guy in four weeks.

[tropic aspen]

Except that is literally why they trained the model

[cinder junco]

And what does that give you? You can’t see the architecture in the output image. You just see the image. And the image isn’t that great.

[tropic aspen]

Question, who uses base 1.5 or base XL? Everyone uses either a NovelAI or PonyDiff based model

And those aren't finetunes of their base model. It just uses the same architecture

[cinder junco]

And thus you shouldn’t get excited about it. It simply proves that you need much more resources to get a usable result.

I use base SDXL. It’s great. Fine tunes are just biased. They look too same-y for me.

[cobalt moon]

Lets not say if there is other finetunes at less popular T2I models

[edgy kelp]

The guy literally wrote in his tweet that's his hobby, dude...

[cobalt moon]

Peoples always have the reason to use finetunes like seriously, PonyXL

Sure it is great. As long as you keep inpainting stuff or just want a general looking.

[cinder junco]

And that’s fine. He can have his hobby and demonstrate what he can do. Great! But don’t be under some illusion that it will be a substitute for SD3.

[edgy kelp]

Wonder if RLHF is doing any good to Stable Diffusion models though...

[cobalt moon]

He's just some random freelancer with proper knowledge

[bitter hearth]

prompt issue

[edgy kelp]

If it's actually the architecture of SD3, pretty much any person that wants to spend some money for the good of the community might make up a decent model, dunno if it's gonna be as good as SD3 on the paper, but more is better than less righto?

[cinder junco]

If he can take his hobby, gather a group with the proper knowledge, and acquire some hefty hardware, then he might start giving Stability a run for their money. But he’s not there yet.

[cobalt moon]

That person literally went to someone to lend him 8 H100s

If I wont wrong

[bitter hearth]

you can have the whole nvidia company, making a model in 3 weeks is pretty out there still :p

pretty cool for a hobby, looks fun

[edgy kelp]

I hope the guy releases some tutorials on youtube, you never know, some other fellow might learn from that and make it even better

[cobalt moon]

Bro, that tutorials may be for someone who is professional at machine learning

[dull star]

I just want SD3 2B to be released in a week or two and then finetuned and we can live happily ever after

[cobalt moon]

Instead of common peasant

And that same person could literally do the same without the tutorials too if not the cost factor

[edgy kelp]

Well, of course 5b model as starting point is not a good idea for a final product...

[dull star]

I hope 2B will pick up complicated scenes up way better with natural prompts

So we can finetune on stuff like holding objects, fight scenes, meme templates

Its gonna be next level if it's not just gonna be crappy portraits 24/7

Quality doesn't matter if the prompt adherence doesn't suffer that much compared to 8B

Obv I can run 8B but what about 98% of the community, and just inference speed in general no matter the VRAM requirements

the new 16 channel VAE will shine new light on 512px and will revitalize it

[rich iron]

I know SD3 is extremely prudish and full of false NSFW credit-wasting positives, but it won't generate something with "stinky" in the prompt? Come on.

[little quarry]

two more weeks

[drowsy compass]

How do you know that?

[little quarry]

please search two more weeks in the discord

[low stone]

"sd3 in two weeks" tattooed on the chest of a behemoth of a bouncer at a fancy nightclub. Background is modern city.

[dull star]

Kek

Thank god that the model isn't blur censored offline 🙏

If it had the deepfloyd license with the forced censor ship its so over

[low stone]

Pretty much

[dull star]

Something went wrong with deepfloyd's development

They even promised one with a commercial license or whatever

And they gave up

And the third stage of the model is still missing

I still respect their research though

Making an open source version of Imagen and all

[bitter hearth]

these muclese are literally rocks inside him

[dull star]

LMAO

Yeah anatomy still isn't the best

But for a base model it's alright anyway

[bitter hearth]

I mean I did for an extremely strong

one

kek

[dull star]

People gonna massively overfit the model and the anatomy will be fixed you guys ✊

[dull star]

lykon

[cunning lintel]

Saw them, first thought was why should I care, second was those devs are either totally out of touch with the perceived limitations and what users want to create, or this improvement still has the same limitations, gimme complex prompts, gimme people lying flat out on their back and stomach, show a robot shooting lasers through its eyes at a giant petrified rabbit, not portraits, portraits, portraits all over

[bitter hearth]

[bitter hearth]

calm down

first we get the model

THEN do that

lmao

[cunning lintel]

You have to admit, all those "teasers" are all the same kind of image that say next to nothing about the capabilities of the model. Meanwhile you have to dig through the depths of reddit to be able to piece together at least a little bit of what is going on (kudos to the employee passing some actual info there) :/ SAI sadly has become so much fluff and very little information

[silver adder]

meanwhile (this took a good chunk of time to get this perfect)

[cunning lintel]

As i said before, so ironic the main use of SD3's text abilities is memes about SAI :p

[silver adder]

yeah, i know there's been some big controversy going on, especially when they started charging for the bots and we shall never forget the shaky business situation for them rn

[bitter hearth]

dont be like that

[silver adder]

people started thinking that sd is no longer gonna be free as a result of that

[bitter hearth]

don't people think that since 1.5 or something

Its just people and the internet

[cunning lintel]

Still a believer in that, though less so that it will be all versions of SD3, but only time will tell. I mostly hope it will be a big step forward over the API version, that is a bit underwhelming

[bitter hearth]

my internet dying

[silver adder]

yeah, but i haven't seen a lot of people react like this until now. things became loud once the sd3 weights became overdue to many (i suspect it was supposed to be released just some weeks ago)

[silver adder]

speaking of the api, i hope that there will be an upgrade to it sometime soon

[cunning lintel]

People seem totally out of touch with the API version, it really isn't very good. Then the idea is "community will fix it", like, how? It's supposedly not even trainable yet on consumer hardware, and the final version might involve a slight architecture change to fix the problems. It needs time, so be it (but at the same time, tired of the meaningless teasers, get real SAI, give some actual info)

[bitter hearth]

@silver adder this is my prompt for you

[silver adder]

that looks nice

[mortal mesa]

i reserve my right to complain until/if they say no open release

[silver adder]

let me try getting some of the images i generated over the past few weeks, didn't decide on sharing em until now

[bitter hearth]

I prompted "wait"

[livid plover]

they trying to sell it not to release it for free

get real guys xD

[bitter hearth]

not the smith

[silver adder]

haha, it does somewhat look dank

[bitter hearth]

xD

Just look at those many paid models they have

right

[past flame]

They're trying to earn some money before they eventually release it for free, somewhere between the point where the benefits form the model going open weights outweights the benefits of keeping it API-only and the point where they're done training the model

[bitter hearth]

And how many times they said they will make it paid only

damn how could we be so dumb

:P

[cunning lintel]

Those are nice 🙂

[dull star]

honestly the images we make here are a better demonstration

but sadly I want to see ours with highresfix

and not through some sdxl model

[bitter hearth]

Did you send thumbnails...

[dull star]

like SD3 highres-fix

[teal fossil]

There is so much competition now and new Models from elsewhere like every second week... they should just release all versions and label the big one v0.9 if they actually want to Finetune it more.

[low stone]

I have automation that does sd3, pixart, Ella, and hunyuan of the prompt I ask for. All are great. The majority of the time, the sd3 one is still the most prompt following. Even in its unfinished state. Because of that i still want it and am optimistic for its release.

For example, these are just thumbnails, but upper left is sd3, being the most prompt following of someone holding a phone taking a selfie. You shouldn't see the phone.

[livid plover]

don't forget to subscribe to artisan! Show em some love or whatever xD

[bitter hearth]

xD

[woven dock]

I still hope 8b gets some love

Cascade was completely left out with almost no finetunes or tools

[dull star]

if it gets released I swear I'd use it to make paintings with JUST the base model

its that good

but most likely, 2B will be the most popular

the most accessible yet but for the majority of the community

[astral grail]

I'm sure 8B will be the most popular

unless it somehow has worse quality than the smaller models

[cunning lintel]

cherry, cherry, cherry pick, but i liked this one, never mind the prompt, sd3 loves cats today, all the creatures were feline

||In a soft, creamy haze, a mystical creature stands, its velvety blue fur glistening in the gentle light. Oversized butterfly wings, delicate and intricately veined, sprout from its head, functioning as ears. Large, bright blue eyes shine like sapphires, filling the space with wonder and enchantment.||

pixart version

[dull star]

8B + T5 will be very heavy

yes, you can use CPU for T5, but I still have doubts

on GPU T5 is instant, but on CPU its like 5-10 secs just to generate conditioning

[cunning lintel]

most enthusiasts have the top end cards or rent them i'd think, can't imagine going for the lower model unless the difference is really minimal, at best the models are largely interchangeable so you can prototype on a lighter model, but not expecting that.

[dull star]

it could also be 4B that gets picked up, since its close to 3.5B

[astral grail]

and how much VRAM does T5 on GPU need?

[dull star]

haven't measured it tbh

I'll check it with T5 bf16 weights running at 8-bit

cause nobody will run it at fp16 or higher

[dreamy sundial]

uh oh.... didn't prompt for beeple on this one

[dull star]

not even large VRAM card enthusiasts

[astral grail]

I also think those people who really use a lot of local stable diffusion have made sure to have a 16 GB or 24 GB GPU by now. and SDXL runs fine even on 4 GB VRAM by now. I'm sure 8B SD3 will run fine on 16 GB and probably even 12 GB

[dull star]

16GB yes absolutely

12GB is questionable if both T5 and the model is loaded

even if T5 runs at 4B

12GB with CPU might work

[astral grail]

but why load both at the same time?

[woven dock]

It'll probably be a split between 8b and 4b imo

4b is around the size of sdxl

[dull star]

even with dynamic offloading such as comfyui

I'm talking about that

[cunning lintel]

It's an interesting experiment releasing multiple versions, still a bit of weird choice to me, fragmentation as a feature. if anything my choice would be to release two at most, one light enough to train on top end consumer hardware, and one so heavy it can only do inference there, needing cloud for training/finetunes

[woven dock]

when compressed into int8 it's around, well, 4gb

[dull star]

I don't exactly remember how it works

I wonder how I can force it cause I don't think it does for me cause I have plenty of vram

cause I just don't notice a massive VRAM difference when going between the models, node-by-node

[woven dock]

I wonder if things like --lowvram in a111 would work for 8b

[astral grail]

the training is an interesting point, yeah. I think that's really the most important aspect for a model succeeding in the community: if people with 24 GB cannot train a model, there will basically not be any finetunes at all

[dull star]

pixart-sigma (0.6B DiT) with T5 bf16 weights running at 8bit -> ~10GB of VRAM

wait a sec Thanks to BitsAndBytes, T5 becomes an immovable rock.

[astral grail]

it also depends on how much VRAM Nvidia will give the 5090 😄

[dull star]

does this mean I comfyui can't offload it?

[woven dock]

will pixart release any larger models?

[astral grail]

I assume the 5090 will have 32 GB VRAM

[dull star]

oh I can test hunyuan-DiT in comfyui using the extra models plugin

I might do that later

ok I'll try fp16 to see how it goes with comfyui offloading

crap, its around 13,3 GB when generating

hmm it never seems to offload it for me

but running the bf16 weights at 4-bit runs at ~8,5GB of VRAM

[astral grail]

how much VRAM do you have?

[dull star]

24GB

so I dont know if I have to force offloading

cause it might not want to do it cause I have plenty of vram

[astral grail]

I assume the software might be intelligent enough to not do offloading if you still have VRAM free

[dull star]

hmm

also I might not actually test hunyuanDiT

it's still just pickle tensors

This is Pixart

this is SD3

SD3 is still the best so far unfortunately

I hope SD3 2B will perform similarly

[bitter hearth]

Whats the prompt

I won't put shrek in it I swear

[dull star]

Photo of Criminal in a ski mask making a phone call in front of a store. There is caption on the bottom of the image: "It's time to Counter the Strike...". There is a red arrow pointing towards the caption. The red arrow is from a Red circle which has an image of Halo Master Chief in it.

[bitter hearth]

hehe

[cunning lintel]

In a crowded, neon-drenched cyberpunk alley in London, a futuristic, anthropomorphic white cyber cat stands, pointing a paw at the camera. Its slender, athletic body is adorned with a sleek black leather jacket, complete with metallic shoulder pads, and a pair of high-tech, glowing blue sunglasses perch on its forehead. Its fur features vibrant purple streaks, and its ears are tipped with metallic implants. A silver choker with a tiny, pulsing LED encircles its neck. Dramatic night lighting casts intricate shadows on the wet pavement, illuminating the cat's angular, human-like features and intricate, metallic whiskers. Lightrays illuminate the foggy atmosphere, capturing the cat's futuristic attire in exquisite 8K detail, as if shot on 35mm Cinestill 800T film with a Leica M6. (yes yes, llm enhanced, llm really went overboard, somehow it kinda worked)

[cunning lintel]

Charcoal etching on fine linen paper of a flustered mouse-dragon chimera with textured fur coat, overlooking the Lauwersmeer. Extreme detail, 8k-like resolution, with intricate mouse and dragon features. Dramatic studio lighting, deep contrasts, emphasizing soft fur and scaled skin. Etching mimics ray tracing, evoking a sense of depth, like a 35mm film still.

[dreamy sundial]

home

[bitter hearth]

I feel like this much prompt is very bad for it

[cunning lintel]

It is, surprised it came out this well, so much of it ignored (maybe for the better...) then i tried a short prompt, but that one was just ugly :p

[bitter hearth]

yeah but you can do half that and probably get more of what you prompted

llms seem to add too many fluff words

[low stone]

"llms seem to add too many fluff words". fine criticism about a cat picture.

[bitter hearth]

also does glasses above eyes ever work

lmao

[low stone]

hmmm hunyuan actually did a more coherent picture

[fast sun]

Yo festivalman

[fast sun]

Can one of you, make a image of zero two wearing a lakers basketball shirt

[low stone]

these models don't do real world stuff very well

midjourney is really good at generating copyrighted stuff like that though. 🙂 I don't mean that insultingly, but sdxl etc isn't accurate enough for it.

[fast sun]

So midjourney does that

[low stone]

closest you're gonna get out of sd3/sdxl

[fast sun]

Bruh the last one is a good one

I think your not using the right, stuff

[bitter hearth]

who the hell is zero two

[bitter hearth]

xD

[cunning lintel]

anime girls in photo backgrounds \o/

[low stone]

yeeeahhh

[bitter hearth]

kek

[cunning lintel]

such a nice effect, but so hard to get for anything but anime girls

[bitter hearth]

what have I done

[low stone]

I think the "zero two" was basically just generic anime girl with pink hair.

that's as close as it's trained on it.

[cunning lintel]

To me, it looked great, but i had to google zero-two 😉

[bitter hearth]

eyes

[cunning lintel]

:p

[bitter hearth]

anime girl = school outfit lmao

[bitter hearth]

that face is funny

I almost prompted for Bus Light Speed

Lmao

I couldn't remember this guys name

[plush olive]

😂

I don’t know it just makes me laugh

[bitter hearth]

smoking is bad

[bitter hearth]

lmao

[hallow lion]

Weights will drop like happycore bass.

[rich iron]

good to see Bill showed up to the meeting for once.

[low stone]

This is not what I asked for

[livid nebula]

This is not what I asked for

[frigid saffron]

Hi, gus, I want to get a detailed and accurate caption about an image(better than BLIP series), which model I should use?

[stoic turtle]

Alice in Hooterland

[willow gulch]

Will sd3 ever be released or will it be API forever?

[muted dove]

Almost definitely.

[sterile pendant]

Yes, they will most likely release some versions of the model. Which versions? We don't know. Given their financial situation, they will likely keep the largest model or models to themselves for API revenue and/or attracting potential buyers that would want to make money after acquiring them.

But for all we know, they might be cooking up something like a 20B model for their API or to attract a company into acquiring them

[muted dove]

For all we know, they may have started on SD4 🤷🏻‍♂️

[noble coyote]

How do you format text to look like this?!

[muted dove]

like this?

Put this around it `text here`

Use three of them, before and after, to get the block like the one above

[sterile pendant]

True that. But realistically, they are probably adding a ton more safety precautions and tuning these sd3 models even more, before releasing them.

[muted dove]

If they're going bankrupt, ditch the safety precautions and live a little 🤣

[noble coyote]

'like this'

[muted dove]

Waste time with safety precautions that the community will waste time on trying to remove.

[muted dove]

Not the same 😉

[noble coyote]

'like this not' 🙂

[muted dove]

Different "ticks"

The ASCII code for the backtick character (`) is 96

...if that helps 😄

[noble coyote]

´like this´

like this

Ta daah!

[muted dove]

Now use three at both ends

[noble coyote]

Today's SD3@ClipDrop harvest ...

[muted dove]

it's good
if you have
a block of text

[muted dove]

FREE THE WEIGHTS

[noble coyote]

Free the weighting time! We are weighsting aweigh weighting two longg!!!

Even Taylor has weighed-in!!!

And good ol' Liberty Herself!!!

Joe and Donny are there too!!!

And some people who didn't want to be recognised!!!

If I have a Block Text, then it'll organise itself?

If I have a Block of Text, it'll Organise itself downwards?

There were 40 pictures created this a.m. over at SD3@ClipDrop - and only two out of the forty had bad spelling!!!

[muted dove]

No, but if you put 3 backticks around all of that, it'll look neater

If I have a Block Text, then it'll organise itself?
If I have a Block
of Text, it'll
Organise itself
downwards?

[noble coyote]

of Text, it'll
Organise itself
downwards?```

[muted dove]

After the first 3 are entered, you can press enter without sending the message, it'll keep adding lines until you add the last 3

[noble coyote]

Is there a Wiki with all this formatting info available?

[muted dove]

It's not unique to Discord, so probably searching for "markdown" format would find something

[fast sun]

guys how do i make this in stable diffusion

[noble coyote]

Use Harrlogos LoRA FOR THE TEXT

[fast sun]

so you know how to make this

[noble coyote]

I know how to make text via Harrlogos ... 🙂

[fast sun]

what the hell isi harrlogos

[noble coyote]

You're being very aggressive 😦

[fast sun]

huh

how

[noble coyote]

https://civitai.com/models/176555/harrlogos-xl-finally-custom-text-generation-in-sd

[fast sun]

how would that help me generate the image i want

[noble coyote]

It adds the text portion

[fast sun]

which text

r u talking about the one that says randomly blowing up

[noble coyote]

Randomly Blowing Up!

[rotund ibex]

https://tenor.com/view/shannon-sharpe-undisputed-nope-funny-face-gif-536487951695337499

[noble coyote]

Awww shucks!!! come on!!! 😄

[muted dove]

https://tenor.com/view/taylor-swift-taylor-taylor-swift-reputation-taylor-swift-stare-gif-4826549884347790976

[noble coyote]

Sailor Twift!

[muted dove]

Your image reminded me of her

[noble coyote]

Yes, it was a generation swiftly tailored in SD3@ClipDrop 🙂

Bad Photoshopping - but fun!!!

The backgrounds are from Huggingface DiT AI Generator

[topaz locust]

Is the API down?

[dull star]

If yes, I hope it's because they are updating the model

To a newer version or something

[crystal spade]

does anyone know when the API back?

[smoky arrow]

Hey, everybody!
who can help me please? i need to generate pictures of cats with the name of one brand, i've been sitting for an hour and i can't do anything. smart people please help me out

[hallow lion]

Weighting for the waits.

[little quarry]

Goof morning

Another day another two more weeks

[wild remnant]

Another time to be alive

Greatest awaits

[balmy summit]

Any one able to use there api’s

[dull star]

damn its been out for an hour

I really hope they update the model or something

[drifting crane]

why you block me? I try to help you 😭

[hallow lion]

Two papers down the line the weights will weigh twice as much.

[icy drift]

Yes. No problems. (I haven't tried in a few days though.)
Nevermind I just tried and the API is not responding.

[little quarry]

[dull star]

I swear 8B Base is amazing for paintings

I will use it if 2B is insufficient

this is fucking art for once

holy shit

gotta love the anatomy

[noble coyote]

SDXL on top of an SD3 i2i image

[cobalt moon]

https://twitter.com/cloneofsimo/status/1793153322016546944

Simo Ryu (@cloneofsimo) on X

Hi, this is Lavenderflow-5.6B-v0.0
✅MMDiT, muP, CFM, FSDP, recaped, 768x768, T5
✅No strings attached, completely-open-every-step-of-the-way
✅Not SoTA😅(hey it was trained by one grad-student under total 3 weeks of development.) Severely undertrained!
https://t.co/PgevrzpoxA

Twitter · May 22, 2024, 5:34 AM

mind y'all to run this just use this for fun

and around 20GB base model

[cunning lintel]

the real news is https://x.com/cloneofsimo/status/1793155762820428031

Simo Ryu (@cloneofsimo) on X

@Birchlabs @StefanABaumann @SeunghyunSEO7 @imbue_ai Ok, this is ONLY the beginning. While I was broadcasting these progress on twitter @FAL guys reached me out to plan on making this more powerful, and go on and build > 8B models from scratch, using better methods, better captioned datasets, everything! All open-sourced!

Twitter · May 22, 2024, 5:43 AM

[dull star]

awwww sdxl vae

[cunning lintel]

the more parties trying to get models out the better. Though a little sad they go for "let's train a big model" and not for something like "let's train ELLA for SDXL"

[dull star]

no sexy 16 channel vae

but still very impressive so far

I'm grateful for all these other attempts

such as pixart, Hunyuan and now this too

https://vxtwitter.com/Birchlabs/status/1793224148702871897

LMAO

[cunning lintel]

The real limit is data

[dull star]

https://tenor.com/view/iron-man-iron-man-hammer-iron-hammer-robert-downey-robert-downey-jr-gif-15959050