🆕｜sd3

[sterile pendant]

and it then becomes statistically signficant

[scarlet wharf]

left is AYS kolors workflow, right original workflow by kijai. add AYS seem give more adherence on prompt

[craggy crest]

sure. but not very useful for the individual person. the only thing that really is useful to them is whether the results they personally get are what they want or not

[edgy kelp]

Reading this link with my tiny eyes it looked like "Im gay dot org"

[hallow lion]

Manual man, we need a manual. XD

[sterile pendant]

play around on the site, you'll quickly see why it's such a powerful tool. it pits two random models against each other while only showing you the prompt. you vote L/R or tie, then it shows you afterwards which was which. the whole point is to test how flexible a model is. on the rankings page, you can click the stats button to see how a model fairs against other specific models. Oh and the prompts are random stuff people ask for (you can prompt for something as well i think, but ive never tried)

so basically, when you thrown a bunch of random prompts at a model, you get a much more realistic rating of how well the model handles various types of concepts vs the cherrypicked BS that most papers show where they have some ultra fine crafted prompt that happens to work really well with a model, since they know the data/captions that were used to train it.

[craggy crest]

you do that on svd too - however, i'm pretty sure that while it's useful to other peopel, it's not as useful to them as just learning how to talk to the models correctly and generating themselves

[sterile pendant]

you should see some of the random ass prompts people throw at it, it's actually kind of rare to see cliche waifus and stuff.

people are undereducated and the average reading/writing level is that of an eleven year old

[low stone]

All of this is until 8b is out, so if you could get on that. 🙂

[craggy crest]

what does that have to do with working with the AI, learning how it thinks, so you can get what you want out of it every time? that's what is actually useful to someone.

[craggy crest]

you can expect 8b in about 5 years, longer if people continue to harp on it

[low stone]

I'll also drop more sugar on you, I've been using your aam xl model a lot. It's really great as a general model as well, able to handle an impressive amount of concept beyond just "anime".

[foggy cloak]

First impressions matter a lot, SD3 got off on the wrong foot

[edgy kelp]

Imagine if they keep hyping it like the 2B and in the end it's not the same as they are using in the API but it's a neutered one

[low stone]

Hopefully in dog years. 🙂

[craggy crest]

2b isn't neutered and it wasn't lobotomized

[lavish osprey]

Thanks, I'm happy you like it. There are also finetunes of that one now by other people

I like Neta Art

[edgy kelp]

My message did not imply 2b is neutered

[lavish osprey]

in general most models built on top of Animagine are pretty great

[sterile pendant]

apply transductive reasoning here: dumb/undereducated people vs highly complicated tech stuff. i'm 99.999% positive that you don't actually fully understand what's happening under the hood. you likely have a gross overview understanding, but it's just the tip of the iceberg.

[lavish osprey]

AAM XL in particular is built on top of both Animagine (indirectly) and Dreamshaper. Another reason why it's able to go Turbo without too much quality loss.

[low stone]

Agreed, anime has a lot of dynamic camera and posing angles that break out of the standard symmetrical portrait looks.

[sage burrow]

@lavish osprey I'm wondering if my glif (that I remixed from another) has anything built in somehow that isn't mentioned? I looked through it all, and it's only SD3 and Claude helping with prompting. But somehow the images come out better than just straight up SD3 on my own computer.
To see any settings, you can just hit remix. But more likely it's something glif has built in to just make everything better, OR, is claude the awesome and it's all about prompting?

[craggy crest]

i'm 999.99% positive that i understand exactly what is happening under the hood.

[lavish osprey]

Glif is using API, so SD3 Large, right? Or they allow to select SD3 Medium?

on your computer it's 100% sd3 medium, since we didn't release large yet.

[bitter hearth]

Hello

[craggy crest]

meow

[sterile pendant]

well i'm sure dunning-kruger would have some things to say, but i'm not going to be mean at this hour lol. regardless, i promise you're not prompting as ideally as you think you are, which means you don't actually understand "ai" as well as you think you do.

[edgy kelp]

Balls

[craggy crest]

i promise you that after spending more than 3000 hours meticulously prompting stable diffusion, changing as little as one character in a prompt, i know what i'm doing and that you do NOT know, though you think you do, my skill set. I also promise you that i'm a programmer and have probably got a deeper understanding of the code than you do.

[craggy crest]

shiny

[bitter hearth]

Pineapple pizza ball

[uncut river]

promise promise
dogs that makes lots of sound
typically dont bite

[edgy kelp]

MFW I'm italian

[craggy crest]

i'm not sure if that's edible or not... where's the crust?

[bitter hearth]

Inside is all crust

[bitter hearth]

Chocolate and MMs?

[craggy crest]

not on tomato sauce. On marshmellow, sure

[edgy kelp]

I'm fastly losing my health points

[bitter hearth]

You'll become American after this one

[edgy kelp]

10 health points, however 200 diabeetus points (which is like the USA mana points)

[sterile pendant]

i know plenty of people that have done some kind of thing for a job for decades and are still not great at it. practice doesn't make perfect, perfect practice makes perfect. part of growing and getting better at something is accepting that you don't know it all and that there is always something to be learned. otherwise, it's just megalomania/narcissism mixed with dunning-kruger.

[craggy crest]

think what you like.

[sterile pendant]

i will

[craggy crest]

and keep it to yourself

[sterile pendant]

i wont

[finite osprey]

doing something wrong a LOT doesn't make you good 🙂

[bitter hearth]

You will

[edgy kelp]

Dunning Freddie Kruger effect

[craggy crest]

how about a fruit pizza sphere

[bitter hearth]

I have no clue what you guys are on about, seems very dumb

[edgy kelp]

Hilarious joke, dunno if you made it on purpose

[bitter hearth]

Like the people I see in games arguing who is right

[bitter hearth]

Idk, I didn't read half of it

[edgy kelp]

I think we share the same braincell

[finite osprey]

Sharing is caring

[bitter hearth]

I share my balls half the time

[craggy crest]

now stick that inside a bubble

[edgy kelp]

Nice ball

Ball rate 10/10

[bitter hearth]

[edgy kelp]

Imagine doing this in the house of a vegan, they won't be able to eat

[sterile pendant]

nah, bros acting like he's dr. jenkins from starship troopers saying "its afraid..." like he can magically read an ai's mind or something and trying to hit people with the git gud spiel lol...

[craggy crest]

i said learn how to talk to it. learn how it thinks. it is not that hard.

[finite osprey]

why not use the square ratio for this

[sterile pendant]

you can't "talk" to something if you don't know how it "thinks"

[craggy crest]

what'd you do with the rest of the fish

[craggy crest]

well we certainly know how you think. ignoring you now.

[bitter hearth]

Rolled away

[craggy crest]

fish heads in floating spheres.

[edgy kelp]

Fishy situation I reckon

[craggy crest]

@bitter hearth

[sterile pendant]

yeah, logically and realistically... but alright man, i get it, your ego is hurt and at this point, you're probably taking this as cyberbullying or something. see how i'm "talking" to something because i can understand how it "thinks" inside?

[bitter hearth]

[foggy cloak]

How are you using SD3 with 4gb VRAM 😭

[craggy crest]

naw, just done trying to explain something to a person with a huge ego and lack of comprehension

[edgy kelp]

He's a liar (not a cat)

[low stone]

There's a block option that alleviates you from certain unpleasantries. It solves a lot. 🙂

[foggy cloak]

Ban that mf

[finite osprey]

I just realized you can go negative in the positive prompt. Like I can write "None of them are outside." and it worked.
You all probably already knew that, but anyway

[bitter hearth]

By utilizing the powers of the internet im able to use the api and be lazy in my bed while my PC is crying in the corner

[sage burrow]

They don't say, so perhaps 😉

[reef urchin]

Only their cat has 4gb of vram. They themselves have more than enough.

[edgy kelp]

Underrated joke

[low stone]

I'm always impressed by how perfectly spherical your balls are.

[edgy kelp]

Dude works out

[lavish osprey]

it's Large probably, They had that contract before Medium was released. Unless they switched to Mediun for the price

[craggy crest]

i know, the problem is that you then can't see what they are saying - and sometimes it's important that you do so

[low stone]

Does it really matter? Not really.

[edgy kelp]

Did you shave with the razor?

[low stone]

No razor can tame that

[edgy kelp]

That's the aftermath of razor shaving (some weeks later)

[bitter hearth]

That's on a cold day

[low stone]

Clearly some shards of razor have been left behind.

[bitter hearth]

Booba

[edgy kelp]

Can you output a ball lying on grass?

[lavish osprey]

is it bad that I can tell this is 100% fireworks sd3 8b?

[bitter hearth]

I dunno

[bitter hearth]

Try it on 2b
My prompt is "A shining metallic ball with 2 arms open lying on grass"

[craggy crest]

when you've spent too much time laying in the grass

[edgy kelp]

Science has gone too far

[bitter hearth]

[edgy kelp]

Damn

[bitter hearth]

runs away

[edgy kelp]

New York Times Opinion: Is This Too Suggestive For SAI?

[uncut river]

lol! does this count ... ?

[bitter hearth]

Clearly underhanded methods were used

[edgy kelp]

Because hands are under

[craggy crest]

photograph of a woman leaning sideways on a grass wall?

[edgy kelp]

That grass looks like a... yeah lol I was beat to that

[low stone]

8b

[bitter hearth]

How sd learned hands: it grows them

[uncut river]

the side view made me rotate the image, was more hoping for a front view with flat solid green wall of grass

[edgy kelp]

Extra fingers are, in fact, fruits

[craggy crest]

try adding "front view" to your prompt then

[edgy kelp]

Imagine a mummy during that covid toilet rolls shortage

[low stone]

Kolors did a good job too. Not sure I can think of an idea for that one. Maybe in the aisle with a barren shelves with hands raised to the sky yelling why

[edgy kelp]

That's nice, maybe with a face mask would be more understandable though (I guess)

[bitter hearth]

Mummy ball?

[low stone]

Neither sd3 8b nor kolors is understanding "barren shelves". It's giving me the opposite. Maybe he's angry about too many choices.

[edgy kelp]

100% sure OpenAI will make a special denoiser (or call it what you like) for AI generate images and train the next Dall-E on insane AI generated images but will be sure to edit every one with some automation in order to avoid artifacts

[bitter hearth]

Try empty shelves

[sage burrow]

Glif? Or, it costs a a subscription fee to change your username, could be an old one lol

[bitter hearth]

[low stone]

Out of a lot of attempts, managed to get a single one with "empty" 🙂

[edgy kelp]

That's my local supermarket after I bought all the toilet paper (I have a hyperactive gut)

[sage burrow]

Ref image maybe?

[low stone]

That exceeds my laziness quotient.

[edgy kelp]

BRB gotta photograph myself in mummy bendages while I show despair in an empty supermarket (I have the budget)

[craggy crest]

try "empty shelves"

[edgy kelp]

Empty shelves is a banned phrase from the SAI datasets because it's too naughty

[low stone]

I smell a Lora. Or is that something else?

[edgy kelp]

No I was joking

[mortal mesa]

govt intervention, cant show empty shelves

[low stone]

I know, but a dedicated Lora of people shouting why at the skies in grocery stores would probably find a niche market on civit

[edgy kelp]

Gotta make a dataset with Dall-E 3 and then train a lora on it I guess

[mortal mesa]

go to walmart and take pictures for your dataset

[low stone]

Wow even dalle couldn't do it

[edgy kelp]

I'd rather not pose kneeling down with a desperate expression to be photographed

[uncut river]

so ... real life photo model for clothing branding, seems an easy task for ai.

custom desing NOEDEL brand outfits. not for sale yet!

[uncut river]

Off-grey stretchy combi outfit, top with half-long stretch skirt. Good for both parties or casual activities. Not for sale yet. Noedel brand.

sleeves not included

[uncut river]

Hey Goo Goo Gage, I hope you get banned for life

[craggy crest]

y'all are such good friends

[uncut river]

well crystalwizard, I certainly hope you missed that video which has been removed. was there for too long.

[sage burrow]

So how come Ella over the various Claude or Ollama ones?

[uncut river]

I don't care about the hands, but how can I make SD3 stop using double ll ? It's Noedel, not Noedell

[craggy crest]

photoshop it!

[uncut river]

no, sd3 should do it. back in the days using sd15 i first generated and image, then photoshopped the text over it, then do a refinement with img2img

SD3 should do text better, it's not like I'm asking for a full poem in correct layout

maybe I should ...

btw, this is my character Caitlin. Always wearing custom designed glasses, cuz she wants and can afford it.

[craggy crest]

8b, does. 2b - doesn't have that sort of enhanced ability.

[uncut river]

I just want a working imitation of M$ Word Art (tm) inside the SD3 medium model. perhaps too much to ask for a 2003 technology?

😄

sorry, that was bordering trolling. Nevermind, I switched to generating realistic images of Caitlin anyway!

[bitter hearth]

like ascii ?

[uncut river]

lol, no. ascii is 1963 tech

[bitter hearth]

[bitter hearth]

give me the prompt you used there

[mortal mesa]

have you ever looked at Comfyroll studio nodes

[uncut river]

no, what is that?

[mortal mesa]

lots of graphics and text nodes, its quite good

https://civitai.com/models/183551/comfyui-comfyroll-custom-nodes https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes

[uncut river]

Caitlin, expensive glasses for each day or outfit. Cute looks but a calculated cold character, who loves her work and does not like standing still. Her study in behavior management really helps her at the facility.

[devout schooner]

For all the talk of censorship in SD3, I find it draws women who are randomly straight up naked (but disturbingly without nipples or like anything other than continuous skin downstairs) quite often, even when I didn't prompt for it at all. Not gonna post any examples cause they're all creepy, looks like burn victims or something lol

[mortal mesa]

yes

[uncut river]

damn, it seems to be hard to get a real photo style with purple eyes from sd3

[uncut river]

yes, it seems there is still some nsfw hidden in the model, but indeed somewhat masked/mangled

it seems they have not removed all nsfw from the dataset and trained from scratch, but instead evolved existing models based which do include nsfw images in the dataset. I think sd3 is just trained not to show the more explicit stuff that is hidden within it.

[mortal mesa]

if nipple reroute to smooth plastic

[bitter hearth]

does anyone have a guide to CFG and Steps numbers for SD3

[uncut river]

maybe it's just me, but I think sd3 makes better sexy images when putting stuff like (naked, nude, explicit:1.1) in the NEG prompt

yes, go for low CFG

[bitter hearth]

there was a complex discussion on here a few days ago about negative prompts
they might not do much

[uncut river]

like between 3.8 and maybe up to 6 typically around 4.0 to 4.4 or maybe just keep it on 4.375

[bitter hearth]

I am not sure as I am just now barely starting to test SD3

ah okay thanks

yeah 2 was too low and 7 too high

[uncut river]

and if you want to go crazy, don't let the usual limits stop you

btw, I only know sd3 medium running locally

im not sure, but I think the architecture and prompting for the larger models differ

[bitter hearth]

I am using a huggingface space

but the only problem is it doesn't say the sampler

I might have to do it properly in comfy to know what the actual full settings are

[uncut river]

hm, i think samplers are overrated

I just use Euler for about everything

[bitter hearth]

its mostly that they either converge or not

and need a different amount of steps

[mortal mesa]

many samplers dont ever converge by design, kinda why i like DPM_Adaptive it picks out how many steps it needs by itself to converge

[uncut river]

Caitlin back at work, as happy as she can get! (she loves serious work...). though maybe sd3 turned her a bit asian. Oh well, it's all about behavior for Caitlin, the good looks are just her trademark.

[bitter hearth]

I've always used DPM++ 2M Karras, DPM++2S a Karras or DPM++ SDE 2M Karras

cos that gets you an ancestral one and an SDE one

as well as just DPM++ 2M Karras

Interesting

[uncut river]

very similar prompt in SD15 - RealisticVisionV60B1 - for reference

you don't see it, but there was color bleeding all over, so had to cherry pick this one. SD15 might be able to give nice(r) results, but at the cost of some rejects

[craggy crest]

prompt: hdr photograph, head and shoulder shot, a man,1960s hippy

[bitter hearth]

hdr photograph, head and shoulder shot, a man, cyberpunk 2077 hippy

[craggy crest]

i just thought it was funny that when i asked for a 1960s hippy, i got a jim croce look alike

[uncut river]

makes me think of Frank Zappa

[craggy crest]

yeah, sort of a mix between them I think

[muted dove]

https://tenor.com/view/adam-driver-snl-kylo-ren-thumbs-up-okay-gif-16158853

[bitter hearth]

sadly the bartender triggered the anatomy problem

ok so putting man in the negative is the way to go

[torn wharf]

my friend challenged me to train this barbie bimbo instagram model on sd3 and said it was impossible and could't be done. i'd had difficulty getting people's likenesses but i thought i could show him up. so i think i trained sd3 to do this megan millions girl using pics scrapped off her instagram. it does a lot of selfies mostly but it works.

||https://ibb.co/album/F5cDXw|| may be slightly pg13. these are all from sd3 with my trained lora. she's not my type but i love giving a good ol "in yo face" to haters.

if i could do this with 100 images over 100 epochs. minute per epoch. i believe in fine tuners

[edgy kelp]

Results are good, but damn a lora on 100 images for a single subject...

[lavish osprey]

negative works

[torn wharf]

maybe that's where i went wrong. other expert lora trainers are using 80000 alpha

[edgy kelp]

No I didn't mean you are doing it wrong, if it's working it looks like an issue of the architecture of the model. I'm used to a max of 20 images for the best likeness

[torn wharf]

i was joking a bit lol. i have no idea what i'm doing. i'll try with less images.

Those results are VERY cherry picked i should say too. All the underlying model problems are still there

[bitter hearth]

depends on the character or whatever

if you need more or less images

[edgy kelp]

I think the average instagram model should be easy enough to train on 10 images, you'd have more difficult time with videogame aliens characters

[craggy crest]

i think you could just prompt that character without training anything and get it without too much trouble

[desert garnet]

yea its easier to train on photorealistic than on highly stylized imgs like 2d

[edgy kelp]

Well... if you have to train nudity on SD3 maybe you might need some thousands of images for a lora, as Cat with 99999 gb vram implied

[bitter hearth]

DO NOT make us use our vram

thats just asking for it

[torn wharf]

sd3 training is more efficient or maybe i'm just stupid. I can barely do batches of 2 on sdxl but on sd3 i can do batches of 10

wth room to spare

[edgy kelp]

Do not make me come here and use my VRAM, yeah?!

[bitter hearth]

🫣

[edgy kelp]

Me: clueless

[torn wharf]

like, totally

[edgy kelp]

I have no idea either

[torn wharf]

https://tenor.com/view/rollin-homies-batman-robin-head-nod-gif-5440088

[edgy kelp]

BALLS

[torn wharf]

https://www.youtube.com/watch?v=EZNFo5lL4iw has this energy to it

[bitter hearth]

oh goodness!

ok so it can make a bartender correctly
but only if the bartender is batman saying hello

[craggy crest]

thosea are some interesting chairs

[bitter hearth]

lol yeah

art deco in the prompt does that

if you cherry pick you get much better results, this one is fantastic

its very inconsistent

[torn wharf]

https://tenor.com/view/scary-movie-strong-hand-gif-5122498

[bitter hearth]

hello

hmm the image quality goes up if you stop prompting for super heroes

same prompts but with bartender instead of batman or superman

the fashion

[craggy crest]

add rococo to your prompt

[placid swallow]

can someone DM me the 3.1 2b plz

[bitter hearth]

you have to prompt for it

if you cant I think its a skill issue

[placid swallow]

no way jose I copied all the skillz i needed from the hugginface

[craggy crest]

huggingface called, they want their skills back

[sage burrow]

A narwal anthro 😉

[bitter hearth]

funky

"inverted colors" in the prompt

[hallow lion]

those are some magnificent cats

[hallow lion]

hey its the sd3 wizard form the promo banner, he's back with a gooder license and the promise of an even gooder model to fix and rectify the universe. SD3.5 will be great just like SD1.5 just like poetry- it will rhyme.

[regal hollow]

how to use?

[hallow lion]

thats a very vague question...

if u mean sd3 then u got the API and/or comfyui but youll have to dig a bit for the models since civitia banned them

[regal hollow]

i want to use like a midjouney in discord

[hallow lion]

then go to the midjourney channel

[regal hollow]

i cant?

nono i want use sd. but use direction is like a discord usable

[hallow lion]

well there is Artisan

[regal hollow]

thank u

[finite fractal]

At dusk, a muscular man riding a bicycle at 120 KM/H on the highway, dramatic lighting, intense motion blur, dynamic pose, cinematic atmosphere, high-speed action, detailed muscles, realistic style

[sacred jewel]

That is the 8B dollar question.

[bitter hearth]

I am bored... so its time for more balls

[torn wharf]

balls!

[bitter hearth]

No one balling

[bitter hearth]

[placid belfry]

made with the medium opensource version

[sullen moss]

What about to lay down? 😉

[edgy kelp]

Neat, I woke up to some balls!

[sage burrow]

There seems to be 3 versions of sd3 for dl. 1 w/o clips, the 10gb one w clips, and the 15gb one w clips. Does the w/o clips version require less vram? Does the 10gb one require less vram than the 15gb one?

[muted dove]

The largest includes clips and the T5. You can load everything individually, or all together. The different sizes just give you flexibility on which parts you want to use/load.

[bitter hearth]

I like SD3 now

didn't really at first

but making stormtroopers invade 17th Century France has been fun

[sage burrow]

Thank you 🙂
I have the 10gb one, tempted to get the 15gb instead. They both have the clips.

I love it when software "shopoing" is free 😄

[sterile pendant]

The most flexible way is to get the smallest sd3 model(doesn't have text encoders) and to then download the three individual encoders. Within comfyui, you can then use a single, double or triple clip loader to pick which ones you want

And it will take up the same amount of storage space as if you downloaded the largest sd3 checkpoint that contains all three encoders with it

[static cedar]

one tree during winter, reflection from lake, all white --v 6.0

[sterile pendant]

oh and dont worry about the fp16 version of the t5, it's mostly pointless. you could run a million A/B blind tests and would likely see them both within margin of error of each other, in terms of voting

if you have the ram/vram for it, sure, go for it

[sage burrow]

I can barely run sd3 to begin with lol (only 8gb vram)

[sage burrow]

It knows quite a few painters from that period as well, I'm case you ever want to get really specific

[bitter hearth]

ah okay that's good

midjourney is the best at knowing artists I think

[edgy kelp]

Dall-E 3 very likely knows even more, but the prompt expansion might fugg them up... and not to talk about the filters lol

[bitter hearth]

I don't know anything about anime but this is my sci fi anime attempt

not even sure which anime that style comes from

[limpid thunder]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[sterile pendant]

I run SD3 on 8gb just fine, still using an old RTX 2080 FE on this PC

[hallow lion]

the one thing emad delivered on: it is PC resource friendly.

Thanks emad. Say thanks kids.

[sterile pendant]

well i'd also give a HUGE shoutout to comfyanon and all the others that work on comfyui. that's where the real performance comes from. his system for automatically handling model offloading at various steps and stages of the process, is what keeps the vram usage down.

[hallow lion]

yes comfy is amazing, it is as fast as foooocus but without the quality loss

[cobalt moon]

6GB also run just fine

although may be a little bit slower by community standard

( 40 sec per 28 steps 1024x1024 )

[hallow lion]

i started with 4GB now that was slower than standards

took about 2-3 mins for 512

[cobalt moon]

that is fast at my standard

[hallow lion]

4-5 for sdxl 1024

[sterile pendant]

nah thats about average if you're using an older gen card. my 2080 spits out 35 step 1024x1024 sdxl images like like 30 seconds i think

[cobalt moon]

lmao

[hallow lion]

🤔

[cobalt moon]

it is a Laptop 4050 so yeah

[sterile pendant]

if you know specific resolutions you like to use, i highly advise going the tensorRT route for experimenting. like if you don't need to use cnets, loras, ipa, etc, you can get like 75% reductions in generation time. it's great for exploring (they might be able to work some of those features? not sure, never tested to actually see)

like for me, the only resolutions i use are 1024/1024, 1152/896 and 1344/768, or reversed

[bitter hearth]

I like ultrawide

more than I even use square LOL

[sterile pendant]

yeah, but the problem is that most models are not trained for extreme aspect ratios like that. sure, they'll work sometimes, but think of the dataset used to train the models. there is likely VERY little content involving things like 21:9 aspect ratios. 16:9 would is likely the widest they naturally want to go, which aligns pretty well with 1344x768(still roughly 1 megapixel, so well within a typical model pixel range, and also, both numbers are divisible by 64/32/16/8 evenly). i use 1344/768 a lot because if you do a NN latent upscale by 1.5, it puts it just a hair of 1080p and is super easy to crop/resize to size

[deft briar]

Create a highly realistic and dynamic image of the Indian cricket team celebrating their victorious moment after winning the Champions Trophy. The scene should capture the exhilaration and joy of the players as they celebrate on the cricket field. Use vivid colors and sharp details to portray the players in their blue uniforms, some holding the trophy high, others embracing, and some jumping in joy. Include elements like confetti raining down, fireworks in the sky, and a jubilant crowd in the background. The expressions on the players' faces should reflect pure happiness, pride, and excitement. Ensure the setting is a well-lit stadium, with bright floodlights, a lush green pitch, and the Champions Trophy prominently displayed. The image should evoke a sense of triumph and national pride, making the viewers feel the energy and emotion of this historic win.

Specific Details:

Players' Emotions: Capture various emotions like shouting with joy, tears of happiness, and players lifting each other in celebration.
Team Unity: Show the players in a close group, arms around each other, symbolizing team spirit and camaraderie.
Trophy Display: Ensure the Champions Trophy is clearly visible, being held by the team captain or a group of players, reflecting the significance of the win.
Background Elements: Include a cheering crowd, waving Indian flags, and banners with congratulatory messages, adding to the festive atmosphere.
Action Shots: Some players could be shown spraying champagne or doing victory laps around the field.

[proven cipher]

when i try to create an image with SD3 1368px * 2048px it just fills top and bottom with nonsense and process a square. Is there any workaround ? SDXL works fine in comparison.

[signal shuttle]

That resolution isn't supported in SD3, try creating a 1:1 image or a 16:9 or a 9:16 image

[proven cipher]

9:16 i have the same issue 🤔

[sage burrow]

Nvidia 2060 here. It runs fine with simple workflows, but not with any fancy ones.

How is 2080 old?

[signal shuttle]

how much is your cfg?, and what sampler + scheduler are you using?

[proven cipher]

cfg 7, sampler dpm2, sheduler normal

[signal shuttle]

try lowering the cfg to 3.5 or 4 and use heunpp2 with simple or normal and set your steps to 28 or 30

[sage burrow]

None of that fancy triple prompt workflow stuff though 😦 lololol

[signal shuttle]

SD3 works best with low cfgs, higher cfgs make the images look burned

[sage burrow]

I need that fb care reaction lolololol

[low stone]

Sadly sd3 2b looked burned a lot of the time even at 4 cfg. Rather frustrating.

[signal shuttle]

Weird, i never have a problem with burned images with SD3 2B

[low stone]

This is with cfg 4. Seems harsh lighting.

[signal shuttle]

are you using heunpp2 with normal or simple?

[edgy kelp]

I generally think that's an issue of prominent synthetic data in the pretraining

[low stone]

Euler. Heunpp2 is massively slower and doesn't yield better quality for me. I, doing 50 steps though which did make a difference.

I'm using comfy's workflow that he published.

It the regular one

[bitter hearth]

they don't say how many images of each aspect ratio were in the training data, but this is the table from the SDXL paper and it has 16 9 in the form of the resolution of 1536 x 640 in particular. It apparently has some images at 2048 x 512 which is crazily wide.

[low stone]

Not the regular one.

[proven cipher]

still questionable output oO

with sdxl i got a lot better results

[low stone]

Can you paste the prompt? I'll try on mine

[proven cipher]

here are more (questionable) SD3 outputs

[cobalt moon]

today I just try SD3

on my another 4050 laptop

[proven cipher]

tried just a simple prompt:
positive: top view, Music Festival, anime style, key visual, vibrant, studio anime, highly detailed
negative: photo, deformed, black and white, realism, disfigured, low contrast

[signal shuttle]

Negative prompts don't do much in SD3, using ...................... or aaaaaaaaaaaa has the same effect on the image

[low stone]

Based on comfy's workflow, I've found the optimal res to be 1024x1280 or 1024x1344. I don't use anything else anymore.

[proven cipher]

i need something like 1368 x 2048 as smartphone backgrounds

i will try with 1024x1344

[low stone]

Yeah, if you need higher, you'll have to use upscaling methods. Sd3 won't full had resolutions directly.

[proven cipher]

can you tell me your settings, or show / send your workflow.

i get heavily nonsense with SD3, idk xD (same prompt as told above)

[low stone]

https://comfyanonymous.github.io/ComfyUI_examples/sd3/

I use the workflow on the woman picture on that page, at 50 steps

[hallow lion]

give it to the cat with 4GB vram, he needs help

[edgy kelp]

Do not make me come here and use my VRAM, yeah?

[sage burrow]

I've seen Cat's images, no way they are from a bgb system only lololol

[edgy kelp]

Dude's Ballz (TM) are amazing, I guess it's not a neutered cat

[mortal kite]

I have to say tho that all the low vram users I think limit the model. Can we really get an excellent model if it has o fit in 4GB?

[bitter hearth]

I put mask in the prompt to avoid face issues
and it decided to do a lace mask

its quite clever at adapting to the theme

when I put friendly witch in the prompt it added flowers and a vase to also make the background more friendly

hmm same prompt but now they have covid mask

[sage burrow]

4 I gon't think is actually possible. Even 8 is pushing it (no advanced workflows). 12 is recommended.
I'm still surprised someone managed with only 6gb vram!

[hallow lion]

He uses the API

Allegedly. Meow.

[sage burrow]

The catch with the API (and glif and huggingface) is that we won't be able to add checkpoints and loras to it when they come out I've been eyeing up some cloud systems!

via glif SD3 large with CLaude helping

A werewolf sphinx lol

brb, Imma post this on some conspiracy groups 😉

[urban arch]

Prince Thun, of the Lion Men.

[sage burrow]

A Ctulu centaur lol

[sterile pendant]

the 2080 came out almost six years ago and was a top of the line card that cost in the 700-800 dollar range. now, it's roughly on par with a 4060 that costs less than half that and that draws less than half the power. in the tech world, six years is a long time.

[sterile pendant]

yeah that is pretty wide, but again, they are likely a very tiny percentage of the actual dataset. wouldn't surprise me if all of the data beyond maybe 16:9 or 9:16 made up less than a few percent. this is why a lot of models will list recommended ranges and warn you not to stray too far beyond the recommendations. and when you do, you end up with the weird duplicated bodies and stuff like that

[foggy cloak]

2080 is ancient in AI terms, it’s still a decent card mind you but saying it’s not dated is a stretch

[bitter hearth]

yeah its likely only a few percent of the actual dataset

I tend to use hidiffusion or deep shrink for generations like that

[sterile pendant]

Yeah it's still decent for 1080p gaming and other 3d related stuff, but Nvidia is two whole generations ahead, about to be three when the 50xx series launches.

Point was that an entry level card of the current generation is on par with a top of the line card from the 20xx gen

But it also meant I got a ton of usage out of it before needing to upgrade again, so there's that

[mortal mesa]

mmm its better than some, i rock a 2080 TI

waiting on 50 series pricing, when i get up off the floor ill probably get a used 3090

[sage burrow]

Hmmm, cloud computing options seem to be less than monthly computer payments, for a computer which will be obsolete in 2 years, hmmm

[torn wharf]

not obsolete. the 2080 is 6 years old and still useable. just ancient

obsolete has a specific meaning

lots of old tech stil has uses

[mortal mesa]

landfills!

[torn wharf]

e waste is a useless thing

[bitter hearth]

Give me that useless and obsolete card

[mortal mesa]

who doesnt want more land

[sterile pendant]

Yeah exactly. But I'm holding out for probably a 5070. I'm not THAT big of an AI junkie, but I still game quite a bit. I don't care about shit like 4k 360fps style gaming though, so the xx70 model cards are more than enough

[silver sluice]

you should turn this into a where's waldo

[sterile pendant]

I'm assuming the 5070 will have 16gb vram, so it's good enough

[torn wharf]

i game tons so i'm always getting a new gpu. AI was the cause of me switching from AMD to Nvidia. I probably should've switched around the 1080 generation instead. Nvidia really started to shine so hard then.

[sterile pendant]

I have a 7900xt in our other PC, it's dawgpoop for AI related stuff, or at least it was the last time I tried it last fall

[torn wharf]

new card every couple of years but i try to find purpose for my old cards and don't just landfill them. at my old house i'd have them on my wall but they were ugly. i might make a knolling case for them next

[mortal mesa]

thanks i learned a new word today

[torn wharf]

was actually going to get the 7900 on launch week but it was apaper launch in Canada. I couldn't find any places anywhere between edmonton and vancouver that got any in stock. amd fucked around that launch hard. that was the other major factor that made me switch to nvidia

[torn wharf]

i actually learned that when sd2 came out and someone had a knolling case lora for it that was beauty

https://civitai.com/models/1203/knollingcase-embeddings-sd-v2-0 mb it was an embedding

[mortal mesa]

nice

[sterile pendant]

Bruh the term knolling case is so redundant... Almost every case you'll ever see or uses 90deg angles, even if there are curves, the base is flat or the object being held is kept perpendicular to the surface the case is on

[low stone]

Sd3 2b. I asked for inside the brain of Cookie Monster.

[sage burrow]

Never ladfill computers! in my city there's this for eg., as well as another~~ one~~about ten or so which donates them to people to broke to buy their own computer. Probably similar in every city https://www.rebootcanada.ca/#:~:text=Supporting reBOOT Canada is simple,virtually anywhere in the country.

[bitter hearth]

[desert garnet]

i will gladly accept your computer

[mortal mesa]

gotta save room for the wind turbines, they ar HUGE

[sage burrow]

I gave away some from the 90's earlier this year, are you sure? 😉

[bitter hearth]

Becky is one step ahead of me

[desert garnet]

https://tenor.com/view/free-money-meme-fast-gif-gif-16935040457953192433

[sage burrow]

Don't get too excited, the tech they give away is generally 6 years old at least. Also people have to prove that they really well below the poverty line.

[mortal mesa]

used to pick computers on garbage days and Frankenstein them into something, dont see much on the curbs like that now

[bitter hearth]

My gpu is 6 years old

[torn wharf]

when i was a kid and couldn't afford a pc that could play quake, something like $4000. I had a 486 given to me that kind of ran windows 3.1. it had a 1 speed cdrom and the discs had these jewel cases i had to put them into before popping that whole case into the drive.

I played the shit out of lemmings on that beast and learned so much

[sage burrow]

and none with GPUs I'm pretty sure lol

[bitter hearth]

I think.. it might be from 2019

[bitter hearth]

[sage burrow]

I have a dell workstation from 2018, and it has no GPU, darnit!

[edgy kelp]

BALLS!

[bitter hearth]

Eat my balls

[edgy kelp]

Haha

[low stone]

A colorful, swirling vortex of half-eaten cookies and crumbs inside a fuzzy blue brain. Chaotic synapses shaped like chocolate chips firing erratically. Dark, shadowy corners filled with forgotten vegetables. Frenetic thought bubbles containing jumbled letters spelling "COOKIE" in various fonts. Tiny, worried-looking Sesame Street characters trying to navigate through the cookie debris. Flashing neon signs reading "EAT" and "MORE" scattered throughout the brain tissue. A distant, echoing laugh track playing in the background. Cracked mirrors reflecting distorted images of cookies and milk. Pulsing veins carrying streams of cookie dough instead of blood.

[wide pagoda]

Is that an ai generated prompt?

[bitter hearth]

I always use GPT-4o for prompts its pretty good

his prompt kinda looks AI

[edgy kelp]

I think SD3 should work best with AI generated prompts as its training captions were made with CogVLM

[low stone]

It's Claude 3.5 sonnet. I was using gpt4o but I just started using Claude and it's SO much better for creative prompts.

[bitter hearth]

AI prompts, what's next ? AI images??

[edgy kelp]

Ah darn, these new things made by THE DEVIL, in my times we had Dall-E 3!

[torn wharf]

it doesn't inherit some greater compatibility with the model that created the captions. cogvlm generates natural language captions. the inherent compatibility is then natural language prompts.

for a lot of people this means using an LLM

[rich tartan]

How to use?

[kindred pumice]

Hi

[edgy kelp]

Not sure, natural language made by LLMs is not really natural

[torn wharf]

i believe you're stuck in magical thinking and don't have any evidence to support your hypothesis. cogvlm captions don't lead a model to understand LLMs better than people prompts.

You have a skewed understanding of what natural language is.

[edgy kelp]

I mean I'm not entirely sure, there are many systems that recognize "synthetic" sentences and long forms, but that's another thing, when I said it works best with AI generated prompt I didn't mean necessarily in contrast to prompts written by humans but rather "it's the intended way"

[torn wharf]

natural language was intended. not a stronger compatibility to LLMs. LLMs can produce tag style prompts too

[winged grail]

hey im having problems installing ReActor, it's not showing up when i installed it. Any way you guys could help me out?

[torn wharf]

consle probably says an error, something about no insightface installation. this is tricky on windows and i often elect to use a precompiled version of insight face. precompiled are often not ideal since that's how viruses can easily spread, but in this case insightface is popular so it's kind of easy to find a reliable one. https://github.com/Gourieff/sd-webui-reactor?tab=readme-ov-file#viii-for-windows-users-if-you-still-cannot-build-insightface-for-some-reasons-or-just-dont-want-to-install-visual-studio-or-vs-c-build-tools---do-the-following

the docs have a wicked troubleshooting section

[bitter hearth]

I actually think OpenAI did the right thing by providing an LLM that was finetuned to prompt their model
I wish every publisher of a diffusion model did that

[torn wharf]

After seeing what omost can do, i'm ceratin that LLMs to prompt the model is the future. But thats for reasons outside of "cogvlm captions mean it understands LLMs better than something human written". I mean, LLMs were trained on human material to begin with

[winged grail]

yeah its the insightface

[torn wharf]

if you don't use a precompiled version, the install requires all sorts of visual studio build tools and gets hairy to set up.

[bitter hearth]

the original cogvlm is not too smart though
its not like GPT-4o
there are patterns in its captions

like repeated mistakes etc

[winged grail]

@torn wharf i sent u a dm

[bitter hearth]

[torn wharf]

you see that ninja scrolls is getting a remastered theatrical release?

[bitter hearth]

I'm afraid I know nothing at all about anime

[sage burrow]

I was in best buy today picking up plugs. Stopped by the computer department for fun. I looked at the specs of the ones on display and said out loud "is that ir?!"

The person working there explained that the second any high end computers are released they are bought immediately. There's a limited amount if gpus apparently.

This was in Vancouver, Canada, a rather large city.

Not sure how true that us but Nvidia owner us buying a few more yachts I think lol

[bitter hearth]

all I know about anime I learnt from looking at stable diffusion generations lol

yeah nvidia is most valued stock right now

[sharp stag]

Hey, need someone's help with gen. that has stable installed on a local machine

Long story short im away at work and will be home in like a month. Need someone with "stable" running locally (non XL, can be makeayo). Had been bored and wrote a prompt in my spare time. Wanted to see how well it performs, but it needs to been locally since its long/weighted and i wont get the same results with cloud based gen. (like never). Side note its nsfw(nothing hardcore or lolicon). I would send both positive and negative privately. Thank you

[torn wharf]

ninja scroll from 1993. it's legendary like akira. back when it was all hand drawn and they still took detail to unseen heights. I dont know anime too much but i respect animation as a whole.

[torn wharf]

am on vancouver island

no fires yet thats later in the year

[sage burrow]

Very awesome 😉

[torn wharf]

balls

[sage burrow]

Where'd you buy your system? Prob newegg or Dell or something

[torn wharf]

sort of built mine. Bought a second hand alienware area 51 pc and i've replaced most of the core components at this point. went from amd threadripper to an intel alderlake

this one was a 3 paragraph long prompt instead of one i wrote. its okay too i guess. still balls.

ice cold

[bitter hearth]

top is SD3 and bottom is Kolors

not a fair comparison since SD3 a base model

but I hope it can get to Kolor's level

[bitter hearth]

wow nice

looks festive

this is sd3

sometimes it does great

[low stone]

Kolors is excellent. Sadly it doesn't know Cookie Monster or Elmo particularly well

[low stone]

Is that Death Star fully operational?

[gusty trail]

Just train a lora

[low stone]

Did they bring out training stuff for it yet?

[gusty trail]

I managed to train it with my own script. It just replace the encoder to glm

[bitter hearth]

[gusty trail]

https://civitai.com/models/571029/kolors-cotton-doll-lora-trained Here is the lora. Use it with lora loader. But you need this plugin first. ComfyUI-Kolors-MZ

[bitter hearth]

wow thanks this is awesome

will try to make some kolors loras

what is kolors

in dumb cat terms

[edgy kelp]

Also in BALLS terms please

[edgy kelp]

T5 moment

[bitter hearth]

anything becomes cool if you add 4k background on it lmao

[edgy kelp]

So many balls to witness

I'm going emotional

[bitter hearth]

you might not be able to handle this one @edgy kelp

[edgy kelp]

🥲

[bitter hearth]

not round enough

[edgy kelp]

Oh no, unballed

Sad stuff

[bitter hearth]

ball factory

[edgy kelp]

I work there 🫡

Notice also that the emoticons are BALLS

[bitter hearth]

true, everyone loves using balls pretty much

[edgy kelp]

If-you-know-what-I-mean

[bitter hearth]

beans are horrible

[mortal kite]

I always run Foooocus with SDXL on 6Gb before I got my 3090

[bitter hearth]

were you able to hires

[torn wharf]

bean corn

[pseudo owl]

should be possible, you just need to enable sliced vae decoding and vae tiling. i have no idea how to do that with fooocus or comfy but in diffusers its pretty easy.

[runic tusk]

How about having some delicious bean popsicles to cool off after that?

[craggy crest]

with comfy you'd probably have to write a node - i don't think there is one that exists yet

[bitter hearth]

delicious toothpaste makes you smile after eating the pizza right

[runic tusk]

@bitter hearth

[low stone]

now with cat meatballs

[bitter hearth]

[strange grotto]

https://tensor.art/models/744985347894651153/SD3-universal-beta_a

anyone tried this?

sd3m finetune

trained on 50000+image

[bitter hearth]

We could run it real quick

[sacred jewel]

Where can we download it locally? Otherwise, I am out 😉

[bitter hearth]

"An ancient castle draped in ivy, looks even more majestic under the setting sun"

Not sure it is worth running

Should be free on Tensor

[pseudo owl]

Hmm yeah it’s a lot more different than the examples, I think probably a sampler issue?

[bitter hearth]

I am not sure. That is a good question.

[sage burrow]

Can't sell the images on DA, damn

[torn wharf]

brace yourself

[sage burrow]

Someone needs to create these in rl, then sell them on food trucks! 🙂

[torn wharf]

customers would be like "how are we supposed to eat these" and you'd just shout "BALLS" then closeup and go to a different location.

so then eveyrone in line would be like "why would you ask!?"

balls

[bitter hearth]

I come back and so many balls

[torn wharf]

https://suno.com/song/cb2cff79-afd0-4cf2-8d5d-a7a3e05365c8 song about dem balls

[low stone]

auraflow

[hallow lion]

so many balls

[hallow lion]

I don't mind pizza being a ball but the pineapples are a disgrace.

[sage burrow]

at least he didn't mix pineapple and anchovies 😉

[torn wharf]

sounds kinda good

[cobalt moon]

hmm, one 25steps 1024x1024 required me 20 minutes on my 2GB VRAM setup

it is painful ofc

[hallow lion]

You are patient sensai.

[bitter hearth]

why would you not use lightning in that situation? LOL

[brittle dragon]

A futuristic, abstract representation of a human brain fused with circuit boards and digital neurons, set against a deep space background. The brain should be partially translucent, revealing pulsing energy and data streams within. Incorporate vibrant, electric blue and purple hues to represent cognitive activity. Add subtle, glowing lines connecting various parts of the brain, symbolizing neural networks. The overall shape should resemble the letter "C" for CogniZone. The style should be sleek, high-tech, and slightly ethereal, conveying the concept of advanced artificial intelligence and cognitive computing.

[dull star]

is this auraflow

[low stone]

What? I have no idea what you're talking about. 🙂

[dull star]

oh, I understand

I made a mistake

this is absolutely sd3

[muted dove]

https://huggingface.co/fal/AuraFlow

[muted dove]

Not bad for an early beta with regular updates promised.

[fleet meteor]

Lmao nice

[low stone]

I haven't been able to find anything it can't do.

[bitter hearth]

guys the heun ones seem good

much better than dpm ones

ok I did more trials

euler heun heunpp2 dpmpp_2m uni_pc uni_pc_bh2 were good

with

sgm_uniform or simple

however ddim_uniform gave more "baked" results

which sometimes was fun

heun heunpp2 were better for realistic people than euler overall

[dull star]

sometimes images look like something out of ideogram

[low stone]

I'm pretty sure that's the majority of what it was trained on.

[muted dove]

Try guns 😉

[low stone]

Compared to Kolors which was midjourney, so it often has a very over stylized look to it (I still love it too)

[muted dove]

I'm surprised nobody has tried "girl laying in the grass" yet 😄

[muted dove]

Aura? Maybe it's just revolvers it doens't like...? #✨｜sdxl message

[muted dove]

https://tenor.com/view/chuck-norris-approve-approval-nice-not-bad-gif-8061938

[low stone]

Yeah it could use more training on revolvers.

[edgy kelp]

AuraFlow's text encoder is a "pile-t5-xl", I assume that being trained on The Pile dataset it can understand and learn NSFW, for anyone interested

[low stone]

There's no question it's not censored.

[edgy kelp]

🤷‍♂️ I didn't test it enough to tell, but the "issue" would be also what the transformer was trained on and for how long

[fleet meteor]

Any idea how much vram it uses? I wanna try it 🥺

[low stone]

the model itself is 16 gigs. This is what Lykon means when he talks about the 8b sd3 being larger than most people can deal with.

that said, it's all one big file right now. I don't know what the possibililties are in terms of breaking that out into image and text encoder components at some point.

[edgy kelp]

AuraFlow model has 6,8B Transformer but has ONLY the big text encoder (T5), I think if you use the 8B SD3 with only the clips you'd have a different "scale" of GPU use

[low stone]

The idea of using the clip encoders instead of the t5 seems like a horrible waste.

[cobalt moon]

why not use both

[low stone]

Because clip is awful.

When you have a t5 llm there, there's no reason to use anything else.

The clip is only when you're trying to shoehorn it into small cards.

[edgy kelp]

I think if you don't use the clips you won't be able to use 99% of the Loras though

Not sure though

[low stone]

If the Lora's are trained on t5 then it's not a problem

[edgy kelp]

That's what I meant

I think most people won't train loras on the T5

But I have no idea haha

We'll see

[low stone]

Why not? We're already doing it with pixart.

There's always going to be a market for small card technology, but there's little progress if we keep holding onto outdated stuff.

[cunning lintel]

Still not convinced of that, models with clip seem to know more, both styles/artists and obscure characters. Of course might just be the training of new models on synthetic stuff combined with current vlms being pretty bad at describing style and artists being stripped. Still feels like a regression sadly.

[strange grotto]

aura

[fleet meteor]

Finally!

[low stone]

I'm certainly curious to find out. I get the impression that most of the latest models aren't particularly trained on artist styles because of copyright issues. I think cascade was the last high quality model that still has them all.

[cunning lintel]

Yeah, i hope things like IP-adapter will allow for style transfer in the future

[sterile pendant]

and you're not going to get them again from a major company's public release, due to the wild wild west hay-days being already being over with. lawsuits and threats of lawsuits galore, shut that shit down fast. from here on out, newer models are only going to contain what they are legally allowed to contain and won't be able to include things like named people and artists, without their consent. so basically, it will mostly just be a bunch of copyrightless datasets and any artists that are okay with their stuff being used will have to opt IN and not opt OUT now. so if you want stuff like that again, people will have to risk potentially being sued to train loras/models with stuff they want, until governments step in and regulate that part as well(won't be long, two years tops for pretty much all modern countries).

[dull star]

oh hell yeah I can run it offline too?

hope it improves

[verbal epoch]

AuraFlow

[mortal kite]

They could have given a 4b though

[craggy crest]

no they couldn't

[mortal kite]

fine, whatever

tired of every single thing on the internet being an argument

[craggy crest]

just learn how to create loras (et. al.) and train a fine tune of what you want to use yourself.

[craggy crest]

then don't repeat somethign that's been beaten to death, on the forum it's been beaten to death on.

[mortal kite]

ugh bye

[low stone]

I say this sincerely, but I think this is where sd3 medium will shine. They apparently considered a 4b but it was decided it wouldn't benefit enough people, so 2b and 8b are the ones they're focusing on. I think once sd3 medium is "fixed" as per their press release, it'll be really great.

[cunning lintel]

When SD3 2b works, it still gives the crispest cleanest images by far. So if it starts to gen images in a much wider and dynamic range, it can be be amazing. If 2b had been any good, all those new models would hardly gain traction.

[low stone]

agreed. claude expanded prompts (even better than gpt4o) has really made the current sd3 2b shine.

[cunning lintel]

Noticed the same, claude is way more creative

[low stone]

gpt4o just kind of expands slightly on what you type. whereas claude adds all sorts of elements including text banners and signs that really enhance things.

which sd3/aura are both really good at.

[craggy crest]

it's the only version of Stable i've used since it released. and i've generated a LOT of images. it works every time as long as you learn how to use it

[cunning lintel]

Sure sure, that's why SAI says it's a beta model and they're fixing it, cause it works every time now 🤡

[craggy crest]

shrug. think what you like. no generative AI model out there turns out perfect results every time, but if you learn how to use it, you can get the results you want, (the first time you generate, not the 100th time) EVERY single time. 2b is no different 🧌

[mortal mesa]

dalle ect, no that i like it

[low stone]

Pulls off crystalwizard's mask in scooby doo episode "It was old man Lykon all along!"

🙂

[mortal mesa]

its Ella

[craggy crest]

heh

[mortal mesa]

ok let me navigate around the glaring issues of SD3 and make some stuff

[craggy crest]

points you at this channel and suggest you scroll through and look at all the really good stuff made by 2b first

[mortal mesa]

i can just look at stuff on my computer, no scrolling

orange you glad im here

[low stone]

i heard you liked oranges and put oranges in your oranges

[mortal mesa]

i remove this one orange

[low stone]

off with his rind!

[lavish osprey]

Pixart 800m uses kv compression. It's essentially like a much larger architecture, but compression has drawbacks

[frail shoal]

prompt ?

[bitter hearth]

Blue girl, city background
Surely

[mortal mesa]

A dreamlike, ethereal portrait of Stability AI, its digital form dissolving into swirling clouds of iridescent gas. Glowing blue lines pulse through its translucent body, as if infused with an otherworldly energy. Its face, a blend of human and machine features, appears serene yet intense, with eyes that seem to hold the secrets of the universe within their depths. The surrounding environment is distorted, with buildings and landscapes warped into impossible shapes, reflecting the AI's ability to manipulate reality itself.

[torn wharf]

dudde did you teleport the bread?

[mortal mesa]

lol pretty much

[bitter hearth]

Rawr

[torn wharf]

monstrous covid

[mortal mesa]

are those SD3? your getting cool mirrored patterns, i like it

Loads Sacred Geometry LoRA

[frail shoal]

no this datavoid finetune of pixart sigma

[low stone]

could get it to do all the text quite right.

[bitter wadi]

Wen SD3 ?

[hallow lion]

Oh noes... Emad. you OK?

[torn wharf]

looks like a madball face

[mortal mesa]

Prompt challenge, wordsmith this into something usable: text of "always coming from take me down" reflecting the text of "never going to give you up"

[dull star]

that sounds impossible

[mortal mesa]

yes 100's of gens later

i mean maybe not but ya haha

[dull star]

https://www.reddit.com/r/StableDiffusion/comments/1e1ktdh/auraflow_sure_does_like_making_the_ideogram/ @low stone

wanting an offline ideogram comes at a price

[torn wharf]

the scene is at night. A still lake fills the image and the shore is in the distance. 3d text exists physically over the water, reading "always coming from take me down" and the reflection of the text in the water reads "never going to give you up" closest i could get with this prompt

[torn wharf]

lol they probably mined ideogram for images and didn't bother to filter their dataset at all

[hallow lion]

Maybe not safe.

[dull star]

yeah its funny

idk how they didn't think of this oversight

but hey, its a free model that's in like 0.1 state

an actual free model with apache 2 license

I didn't test too much but 2 character facial expressions work better

like I ask for the person on the left to be scared or crying and the one on the right to be shouting and angry, it gets it right

[torn wharf]

likely they used the prompts to make ideagram images as the captions, so in training those unrelated captions will learn that cat, when that cat has nothing to do with whats being captioned.

[dull star]

this is sooooo ideogram its painful

but I still like how it looks though

[hallow lion]

ideogram suez SAI coz why not XD

[torn wharf]

SAI didn't make it

[dull star]

SAI making an apache 2 model?

https://tenor.com/view/thanos-impossible-marvel-shocked-gif-15104180

even older models are at least openrail++

[torn wharf]

it was made by the guy who brought loras to image models

[dull star]

yeah its wild

[torn wharf]

as far as i could tell, ideagram's terms don't limit using images as base model training

[hallow lion]

we went out took photos of real world stuff and trained our model on that.

[mortal mesa]

the guy that implemented lora wanted to try to train a mmdit from scratch, it worked they released - the story

[dull star]

its crazy

[torn wharf]

its' not just stability's mmdit architecture though. they modified it to be more efficient

[lucid swift]

they will remove that from the next training batch and its gone lol

[dull star]

hopefully

[torn wharf]

yeah it's v.01 and that's an obvious training data fix

[cunning lintel]

I saw the cat thing, tooo sad

[torn wharf]

oh i mean v0.1

[cunning lintel]

But pretty hard to get and even if you manage it's only sometimes :p

[torn wharf]

i am looking into running it, but it's nearly 7b parameters and i don't think comfy is optimized to use it in 16gb of vram

[cunning lintel]

And they use the old sdxl vae, which sadly shows :/

[torn wharf]

it's a rushed out v0.1 release yeah. They can change that. I'm sure stability had many versions of sd3 and other models that would've counted as 0.1 versions, but never released them. This guy released his and it's the ugly side of the development process that many people aren't used to

[mortal mesa]

the we made something that works stage

[cunning lintel]

fall even released a 16 channel vae, it's surely coming

[mortal mesa]

i made some bad cats on a 2080 TI or i think it all fallback to CPU

[torn wharf]

minimal viable product . on the blog it says the intention is to kickstart community engagement

[mortal mesa]

im cool with it, i dont care if its not like starbucks, i dont even like starbucks

[cunning lintel]

Funny when SAI seemed to be close to falling apart and rushed 2b, it seemed open generative AI for images was going to be a black hole, now new models are planned or released in various places 🎉 (and SAI seems back in business)

[torn wharf]

starbucks isn't even that great at coffee anymore. least the ones in my podunk hillbilly region.

[torn wharf]

i dont think they were ever falling apart. that was just yellow journalism sensationalism from people farming engagement and loving that scheudencfreud. stability's got legs for days. they wouldn't have attracted sean parker if they were falling apart

probably a little bit of social engineering from competitors

[cunning lintel]

i feel if all was fine 2b would never been released as it was, but will never know

[torn wharf]

there's a wide spectrum of diverse situations between "falling apart" and "all is fine"

[mortal mesa]

they started pumping expectations too early

[torn wharf]

spectrums?! too woke tooo woke

[hallow lion]

They were trembling...

[torn wharf]

seems like an over dramatic metaphor. they were hung over after emad partied too hard

[hallow lion]

it was a solid scale 7 on the richter scale

lol

emad is SAI SAI is Emad, he is the mascot

[torn wharf]

hangovers take a couple days to get over

[hallow lion]

i hope they will recover whatever the future their contribution to open ai was awesome

Emad was the vision man tho

without vision its hard

[hallow lion]

It's a serious Friday if you need a couple of days to recover XD

You weekended the wrong way

[cunning lintel]

well, his vision is now creating magic ai money in the blockchain to fund true open ai, call me skeptical

[hallow lion]

its like Disney without Lucas.

Maybe Sai will end up like that o without Emad

just hopeless atmepts at cahs grabbign bot no real vision or integrity

i hope not tho but ye

[mortal mesa]

they will be joining the WEF soon

[hallow lion]

Someone email Musk

Musk is the eternal whimsical clueless billionaire who can save us all coz pockets deep enough

Help me Papa Elon, you're my only hope.

[mortal mesa]

is trust and safety trustworthy and safe if they hide things like that

.1 shift 6 CFG, were glitching

[hallow lion]

glitch art - never fails.

[lavish osprey]

No idea what's going on with model creators training on Ideogram outputs lately (isn't pixart also trained on Ideogram data?)

Not only it borderline violates ToS (sure, you can claim you downloaded them from HF, but come on), but no way in hell an Ideogram image is better than a real photo or art, and its prompt following and text accuracy is never going to better than what a VLM can caption from a real image (that can also caption very small text).

And forget doing that with a 16ch vae. Sure, pixart and auraflow use SDXL vae, but even that is good enough to not be a bottleneck compared to synth data training that hasn't been at least refined.

[mortal mesa]

FOMO

[lavish osprey]

that being said, this is the #sd3 channel, please keep it on topic if you can 😄

[craggy crest]

fad?

[lavish osprey]

it's probably just very cheap, since you only have to scrape, no need to (re)caption (which is very expensive on large data)

MJ/Ideogram images come from prompts, so you can get free image+caption (again, violating the ToS)

there are a bunch of datasets like that on HF

making models is very expensive. I wonder why some companies ask you to pay for commercial use over a revenue threshold (or not, like Kolors)

[craggy crest]

maybe. they seem to have hidden their TOS page

[hot dawn]

against my expectations, the non-attention parameters of SD3 seem to be more important for training than the attention blocks, despite that I froze them earlier in training it still seems most of the learning is there. It's possible that whatever censorship they did was semi hardcoded in the conditional pathway to remap the embeddings away from certain areas, though I doubt it since characters and styles are also primarily learned in the non-attention blocks.
if there any vram savings to be made for finetuning SD3, it might be best to stick to training the non-attention parameters, which are probably much smaller too

potentially training with those non-attention components frozen from the start would give different results, perhaps the easier training went there but it may not be ideal

"roman soldiers in formation, covered in mud and blood" did get much worse with the non-attentions only, possibly because there were only a few sketchy examples of roman soldiers in the dataset, and most of the learning has gone into those components, good and bad

[lavish osprey]

It's possible that whatever censorship they did was semi hardcoded in the conditional pathway to remap the embeddings away from certain areas
That's just a silly rumor.
AuraFlow has the same issue, right? It's in the "llm nature" of the architecture. At 8b params it scales very well, 2b is an attempt of having the same tech run on local hardware.
(Also not sure why Aura fails it too, since it's almost 7b params)

[hot dawn]

I was just assuming there was some since there was a screenshot of somebody from SAI saying that "something" was done to the weights before release by the trust and safety team

it seemed a bit iffy since I did see your post about how the issues with the grass pose were there earlier

[lavish osprey]

yeah but that just covered up nudity, did nothing to anatomy

if anything, some biased filter on the dataset might have amplified the issue

but nothing done in SFT under my watch had any effect on that

[hot dawn]

there's not much nudity in my dataset, but the ability to generate nude people seems to have been learned in the non-attention blocks as well

[lavish osprey]

also 8b was trained on the same "filtered" dataset and has no issues with this use case

can make it flawlessly

you can try it yourself on API

[hot dawn]

I suspect 2B can learn that pose if focused in the dataset. I've always found that lying poses are the hardest to finetune and didn't assume it was a conspiracy by SAI that they didn't work well in the base model

[lavish osprey]

(even if I still think that WHEN sd3m manages to get this prompt right, it's the best)

[runic tusk]

She's got a little something in her "pocket".

[lavish osprey]

the hands are wrong, but look at the details. Native gen, no upscaling

SFT on sd3m was done very nicely

too bad the base wasn't perfect

[hot dawn]

I didn't caption text in my dataset so I'm surprised this wasn't broken 😅

[brittle nexus]

Part of the problem was including feet as sex organs

[lavish osprey]

but this at least ensures that simple pictures come out amazing and that the model is a very (very) good refiner

[hot dawn]

there's no women laying on the ground in the grass in my dataset, the model just worked it out from general pose examples

[lavish osprey]

also please label "sd3m" when referring to medium 😄

[hot dawn]

fair enough

[lavish osprey]

it's gonna be a mess when we release Large 😄

[mortal mesa]

give it a name and refer to it like that, Gigantor

[hot dawn]

using base SD3M as a refiner

[hot dawn]

SD3XL 😅

[lavish osprey]

lol it actually fixed the text. Maybe a bit too much denoise anyway

[bitter hearth]

No balls in sight

[lavish osprey]

I usually use 0.15 denoise with sd3m

but to fix text you need to go harder

0.15 denoise on sd3m is equal to about 0.4 on sdxl

[hot dawn]

that was switching back to the base model from step 10 of 28

[brittle nexus]

The censored feet

[lavish osprey]

with steps I use 28 total and start from 20

[hot dawn]

using a negative prompt for the first stage

[lusty oyster]

Just asking for a piece of advice should I use SDXL or SD3

For better images

[bitter hearth]

Yes

[hot dawn]

depends what you're creating, SD3 is amazing at a lot of stuff, but bad at people from the waist down

[lavish osprey]

it's just bad at doing them, like with hands (feet are a little bit easier but also rare in datasets)

[lavish osprey]

use SD3 Large for most things that are not anime

use a SDXL finetune for the rest

SDXL turbo finetunes are also crazy fast

[lusty oyster]

Where do you get SD3 large?

[lavish osprey]

4 steps of dreamshaper xl turbo-lightning for gen + 3 for upscaling + 0.15 denoising (or 8/28 steps) of sd3m for refining

[brittle nexus]

API literary censor feet with blurred images. There's no feet images because they were ripped off in 2b

[bitter hearth]

Lykon do you know if prompting for "8" or "eight" is better or something

[lavish osprey]

only API for now. Still working on it

[hot dawn]

Lykon adding to my wishlist of stuff it would be great to know from SAI, it's not clear if the first timestep (1000) should be finetuned or not. I think in SD1.x it wasn't which led to greyness problems, but after that I never really kept up

[lavish osprey]

depends what you're trying to do.
Counting past 4 is always gonna be hard and random for models

models can't count sequentially

they count "at a glance"

and even humans have issues past 4

[bitter hearth]

[lavish osprey]

by the way, this is also one of the reasons why hands are a hard problem

and why a huge ass 8b model is better at them compared by small-brain 2b cousin

[hot dawn]

I was really hoping it was the 4 channel VAE but it seems not 😦

[lavish osprey]

nah

hands are just very hard in general. Too many visual permutations

[hot dawn]

with attention / non-attentions. Think I've found the cause of my sometimes all-white images. Potentially merging layer by layer could find the problem

[lavish osprey]

these are all "valid" hands

but they all look very different

[hot dawn]

yeah I've been paying more attention to hands im photos and have realized how utterly insane hands are

so often a finger just can't be seen due to being bent the right way

[lavish osprey]

most models that are decent at hands basically overfit on small data and fewer hand positions

or take the shortcut of using only anime style or only realistic style

(or are huge ass like 8b)

[hot dawn]

well SD1.x was fighting against the VAE not even being able to encode and decode hands below a certain size threshold without introducing new lines etc, so I was optimistic a more powerful VAE would lead to a huge improvement