🆕｜sd3

[mortal kite]

😄

[placid swallow]

franco litterally saved your game

the white pixel border

yes i know

[wind basalt]

Sd3 medium accepted in civitai yet ?

Was that cleared up ? Can we use there models for free

[placid swallow]

depends what you wanna use em for

you making a competing video game?

[dusky thistle]

no

[mortal kite]

I suppose I should go do something constructive in the garage

you guys are hilarious though 😄

I don't know if my wife is as amused

[dusky thistle]

"a close up national geographic 4k photo of an evil millipede crawling on a leaf covered in dew in the moonlight" curious what SD3 would do... this is cascade (generated in under a min at 2880x1728, peak vram was about 9.2 gb)

[bitter hearth]

this is what I'd do, by SD3

[dusky thistle]

it's cute!

[placid swallow]

non human is 2ez for 2b

[dusky thistle]

kinda weird leaf but the lighting and detail is good

[placid swallow]

i guess it's nat geo time

[craggy crest]

you're the director

[torn wharf]

is it just me, or is training 2b loras more memory efficient than xl? i can do bigger batches

[eternal steppe]

I'm not sure what model to use anymore.

I've tried pixart Sigma, SD3, SDXL, 1.5, 2.1...

Pixart so far has been the one to follow prompts really well for me.

[hallow lion]

SD3 knows no industrial design. SD3 knows no graphic design. SD3 knows Cronenberg.

[torn wharf]

tbf cronenberg knows industrial and graphic design well enough that he can dement and pervert it

[hallow lion]

yes

He is one demented Canadian mofo.

[dusky thistle]

more fuel for your gpt4o lol

ooo even better one

those are awesome

it's fun seeing stuff transposed from one model to another like this

shi ti think the second is my fav actually

none of the gritty look that sometimes sd3 has trouble with

very clean and well integrated and detailed

yeahhh

got a fun new one

[bitter hearth]

yeah that camera angle is great

when I use dalle
I use "from a low angle"
in every prompt

[dusky thistle]

yeah anything that's not "portrait style" is a nice change of pace

[bitter hearth]

chinook
the SDXL checkpoint
has "dutch angle" trained in as a trigger word

I think its such an under-rated checkpoint

(dutch angle is like an angle that is not straight)

[dusky thistle]

this is cool

[bitter hearth]

chinook is so good
I have no idea why it isn't more popular
or more models aren't trained like that

[dusky thistle]

sure, one sec here

[mortal kite]

skill issue 😛

[eternal steppe]

Clearly not, though.

[bitter hearth]

for prompt adherence
start in Pixart/SD3/Dalle/Ideogram/Midjourney
and then use as a refiner pass SD1.5/SDXL/SD3

[dusky thistle]

i bet they can get close-ish

[dusky thistle]

this one could get weird with gpt4o

[bitter hearth]

can anyone make a good ultrawide sci fi city with SD3?

[bitter hearth]

wow this is so good

[dusky thistle]

that's stable cascade for ya

nice! I had faith it could do it! 😄

[bitter hearth]

wow thanks
this is my main use cases for image AI

[dusky thistle]

np

vram peaked around 9gb for that too i think

maybe 10

idk what gpu you have but i know a lot of ppl are on 8's and 10s around here

[bitter hearth]

I use cloud so any

[dusky thistle]

ahh k

[bitter hearth]

I can get 48GB for a dollar or 80GB for 1.5 dollars
for an hour
pretty much

[dusky thistle]

here's another for ya

yeah vram def isn't an issue for you then lol

[bitter hearth]

that car is so nice wow

[dusky thistle]

theres a few for ya Franco 🙂

[dusky thistle]

and this one

[mortal kite]

the thing that's disappointing is you can never really get an image with action in it

[crude yarrow]

That's a big city. Lots of skyscrapers.

[mortal kite]

even that bomb looks too static

no shockwave, the buidings aren't being affected

[dusky thistle]

it's all about what it was trained on

[ancient cape]

good to see you're still alive XD

[craggy crest]

a programmer never dies, he just gets tangled up in spagetti code

[fleet meteor]

Aa okey, Im planning to make a lot of images and instead of using the refiner i´ll just do a img2img with low cfg later using sdxl

[sterile pendant]

I want to add in a dropdown to switch between various diffusers like sigma, hunyuan, sdxl and sd3. trying to put this together for the non-tech savy

[sage burrow]

Spicy! But who cares about the lady. the backgrouds are awesome! 🙂

[torn wharf]

been trying to train sd3 loras and it's just not happening. they learn something but it's not accurate at all. it's not like sdxl or 1.5. its better than 2 but it's just not happening still. i dont think the current 2b model is something that works with loras. It's not just me either. i haven't found anyone who has successfully trained a person. some style loras but barely accurate to the training data. if anyone has good settings to use they're not sharing them.

maybe this could've been avoided if stability released training code but they're holding that close to their chest

[craggy crest]

2b is missing a lot of stuff, it was rushed out and a lot was skipped

you might want to wait for updated releases

[torn wharf]

if that happens. new ceo has said they're still committed to open models , but i'm not seeing it. why not release the training code?

[craggy crest]

there's nothing to release yet. take that or leave it as you like.

[torn wharf]

cop out answer. sd3 is trained innit it

[sage burrow]

Are there trainers which support SD3 now? Last I checked EveryDream and DreamTrainer didn't yet. That was a week ago though!

[torn wharf]

one trainer has a branch for it

[craggy crest]

unless you jsut really want more broken, unfinished code? if not, then chill out and be patient

[sage burrow]

That branch was broken last I tried, has it been fixed? 🙂

[craggy crest]

just because there's a branch does't mean it's code you can do anythign with

[torn wharf]

its indevelopment yeah. he didn't make it to any spec because there's no training code released so nobody knows, so he just wung a few numbers out there to make it work for now. comfyui then updated to be compatible with his keys.

the loras it made first would only work as sample generations. but now they work in comfyui

[torn wharf]

have you tried it? yeah. thought so.

argumentative with all these cop out lame shut downs. get off the stage

[craggy crest]

there's nothing TO release yet. the entire reason that SD 3 was released as it is, rushed and unfinished, is because the community wouldn't stop harping on 'release the open source!" and refused to be patient. so you got the unfinished broken version. you going to keep harping on release until they just toss non-working code at you and wash their hands of the entire thing? or would you like to practice being patient?

it's a miracle they're not so fed up they haven't already

[sage burrow]

Was everyone this impatient during SDXL as well?

[fleet meteor]

Damn, I hope they release the 4b version 🥺

[torn wharf]

i dont think the paremeter count is the problem

[fleet meteor]

No, it was not even close, people keept using 1.5 for months, well, its still being used

[torn wharf]

you could train sdxl in the 0.9 beta period. loras worked excellently

[sage burrow]

Am I the only one who's still creating hundreds per day with 1.5? lol

[torn wharf]

official training code was published too and kohya had it implemented on day 1

[sage burrow]

1.5 is awsome!!! 😄

[fleet meteor]

I gotta reinstall a 1.5 model, I remember I created some good images with it

[sage burrow]

It's extremey consistent, which is what I'm still using it for. Digging up my old prompts...

[torn wharf]

i don't understand why the talking down about being impatient is going on. i'm out here actually sortin data sets and settings and trying to do shit. y'all kicking over sand castles. big people. i know this is the sd3 channel and we hate grass around here, but honestly, touch some

[sacred jewel]

Comparing the best SD1.5 base with SD3 base is just disingenuous. Stop it.

[torn wharf]

gonna try a style lora. i might consider sharing results if y'all play nicer and agree not to be like "could do that better on 1.5" sure okay so what i'm out here tryna learn mmdit shit. i already done this on unet

[craggy crest]

because you're trying to train a demo model that's not finished. you should consider waiting until they have a chance to work on it some more and release another version

[torn wharf]

you must work at stability to have such information. lol. i dont really know what your poitn is. blocked and ignored going forward. obviously have nothing good to say about anything in this channel.

[craggy crest]

think what you like

[bitter hearth]

What's that about

was I missing chat drama

[placid swallow]

i left the mushrooms running

[bitter hearth]

Make them 100 meters tall now

[placid swallow]

i cannot they are cuty little mushrooms

[bitter hearth]

Mushrooms are the weirdest food possible they are ugly

[placid swallow]

they do not care about a mushroom LORA they will be what they can be

this one is begging you to eat it

please put them out of their misery

one chomp is all they ask

how do i explain this to my folders of waifu images

[bitter hearth]

the blur is amazing on these

[craggy crest]

that's a hairy mushroom

[placid swallow]

is it horror time yet

[bitter hearth]

You're already giving us horror with those mushrooms

[placid swallow]

you take that back kind sir

[bitter hearth]

Werewolf go kill that thing

[placid swallow]

careful I haven't read that far into the subreddit on mushrooms to know which one your little doggy can eat

[bitter hearth]

nice car

[placid swallow]

have you noticed a difference starting the prompts with a digital illustration

[mortal mesa]

random commentary

[placid swallow]

it's kinda specific how it was trained like how lykon said to try anime art or anime screencap in place of anime

[bitter hearth]

Where's "crayon drawing"??

[placid swallow]

digital illustration in the style of cyberpunk, High-resolution, 4K widebody shot, luxury sports car, front perspective, stunning headlights, dark alleyway illumination, neon underglow illuminating streets, pastel sky over the horizon, cinematic realism, dramatic lighting, sleek exterior, premium details, high-end automobile, urban backdrop, atmospheric evening scene, dynamic camera angles, detailed tire reflections

[mortal kite]

I'm just here for the memes

[mortal kite]

I like that style

[dusky thistle]

Those look really good

[placid swallow]

pastel skies taking over

[bitter hearth]

High level crayon drawing

[placid swallow]

katy perry?

[bitter hearth]

No your m-

[placid swallow]

i wish

[bitter hearth]

[placid swallow]

https://tenor.com/view/help-kel-omori-omori-my-mom-said-i-could-have-those-vexwashere-gif-26226681

[mortal kite]

That looks cool

[bitter hearth]

No u

[placid swallow]

horror time all the kids are asleep

[mortal kite]

Their loss

[placid swallow]

no recycle

they trained it so well it can't be bad

we just gotta learn what to say to it almost like developing a skill

[placid swallow]

get away from those kids

[placid swallow]

oh nvm mr alien just checking up on them

[mortal kite]

I am not sure it knows what a xenomorph looks like. Hm

[placid swallow]

i only use wiki-how to learn how to prompt

[mortal kite]

oh I think I know my prompt was confusing it

[placid swallow]

you guys wanna start a biker gang we can get matching jackets

[placid swallow]

we could be the 2B(oys)

[mortal kite]

Oh god....what have I done

[placid swallow]

attack on the godfather manga

[mortal kite]

umm

that is damn creepy

[placid swallow]

it likes to make 20ft beds

[placid swallow]

always dbl teeth

[mortal kite]

why does it love double row teeth so weird

[placid swallow]

does it 95% of the time when they are wide mouthed

[placid swallow]

good ol SD95%

[mortal kite]

This is pretty cool I think

[placid swallow]

it's like Kubrick style

[mortal kite]

where is the horror?

[dusky thistle]

@kindred mica were you still looking for any info on the status of training efforts

[placid swallow]

locking the seed an prompt and just adding periods in the negative

[mortal kite]

the horror is that she is clothed 😄

[mortal kite]

does this monster have a tongue stud?

True horror

[placid swallow]

face hugger knocked her up ruh roh

[mortal kite]

she "bugs" me 😄 😄

[sage burrow]

"There's free campfire food behind me 😄 🙂 "

[mortal kite]

oh a Sephora ad

[dusky thistle]

good one for you 😄

[sage burrow]

Those wings! ❤️

[mortal kite]

so many hands

[mortal kite]

hey, finally a good seed

[limpid drum]

That is impressive hair

[limpid drum]

Future AI

[dusky thistle]

is tile any good with that?

[mortal mesa]

dunno just saw the update, didnt even have them downloaded

[dusky thistle]

appleshark at the bar

[lime pilot]

China's Loong Boat Festival, the background is three giant mountain Zongzi, surrounded by clouds, Loong Boat racing, Loong Boat moving on the water, the river winding from the huge mountain Zongzi to the distance, the background is misty mountains, rich in details, daytime, fairy fog, clouds, movie lighting, super high quality, super high resolution, ultra high definition, ultra clear details, 8K, HDR, Chinese style

[dusky thistle]

Here is the image you requested.

lol

what abotu this one

this is cool as hell

looks evil

one for yo uafter you've had your rest 🙂

😄

idk why i love this

[dusky thistle]

also... look at the hands!

gnight! those are great btw

[craggy crest]

stole your prompt and did this

[dusky thistle]

great work

awesome to see that work, i thought you'd done img2img tbh

[craggy crest]

🙂 thanks

[craggy crest]

nope. i don't have img2img set up on my workflow. jsut very simple

[craggy crest]

i've been playing a lot with handing the different encoders different prompts. this is the one for my shark tent if you want to play around with it

[dusky thistle]

thanks

[craggy crest]

welcome

[placid swallow]

thought it would make it creepier

[quick crater]

Black Cat Sheriff, ground, magnifying glass, footprints, animals, forest, curiously, watching, clearly visible, large, oddly shaped, attention, path, depths.

[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

[torn wharf]

i trained the madball data set i made.

[torn wharf]

it worked but i need to tune things up. i guess it learns non human concepts much better

[desert garnet]

i like your balls

[torn wharf]

its supposed to be these things https://www.youtube.com/watch?v=yvfYcZ8GAVM

freaky fun for everyone!

i should proabbly get rid of half the dataset and supplement synthetic data from the xl version

[torn wharf]

no bird in the training data. nice result. will play more later. this was a pretty promising result for styles. few more tests to dial in settings and i should be able to do more style loras. but now where to host them? hahaha banned from civit

[lime pilot]

By the peaceful lake, a panda eagerly plays its guitar, making the entire environment lively. The calm water surface under a clear sky reflects this scene. Bright flowers bloom around, butterflies flutter, and birds sing. The sun sets, casting a golden glow, blending realism with the lively spirit of giant pandas.

@dusky thistle

[bitter hearth]

cuz surgery

[sacred jewel]

CHALLENGE

Render a bus of any kind where the driver is visible. I have tried and failed.

if successful, must share prompt and sorcery used 😛

LOLOLOLOL

[bitter hearth]

Ball

[sacred jewel]

a girl driving a sports car

[bitter hearth]

[sacred jewel]

sorry... meant external view of a bus.

[bitter hearth]

Got one!

it's bad at it

[sacred jewel]

That's not bad actually. I am up to my 100th seed and prompt tweak and got nothing close.

[sacred jewel]

Yeah. It's weird, you'd think most training images of a typical bus would have a human or two in it.

[bitter hearth]

Try this: Photo in 4k of a close up bus on the street, outside you can see the city passing by and the menacing front of the fast black bus, using an action camera able to capture fast movements perfect for an action movie! The man inside is driving the bus

Tried some wiggle arounds with prompt until it wasn't giving me inside views

[bitter hearth]

In the world of AI humans don't drive buses anymore

[sacred jewel]

And always the damn Empire State bldng and a view from 34th st 🤪

Disclaimer: I am a lifelong New Yorker and rather explore other parts of the globe, thanks 🤪🤣

[bitter hearth]

[sacred jewel]

Damn you're right. The model is telling us "wink wink.. you humans are doomed to be replaced ... Hint hint"

[bitter hearth]

now it's a robot inside, it's over for us

[sacred jewel]

Listen, if you look at mine above with the giant Subway's sandwich worker, I'll take this one every time

[bitter hearth]

Lmao

I changed the place for you, and the driver is gone

@sacred jewel did that prompt give you anything good?

[sacred jewel]

Not tried yet. Already getting distracted by real work at work 🤪🤭

[sacred jewel]

the shortbus...

[bitter hearth]

It's working a bit

[brave valley]

I'm participating in the SD3 image generation contest on Shakker. I'd appreciate it if you could give me a like to help me win the grand prize. Thank you, everyone!https://www.shakker.ai/imageinfo/07e98796f3404b7fb8301d2e411e9205

[sacred jewel]

looks like a nighttime photo is a better success rate than daytime

[bitter hearth]

that would make sense, its easier to see things inside at night than day

thats creepy but nioce

nice in a creepy way

[bitter hearth]

she just wants to play

yea right!

[finite osprey]

UniPC BH2 seems to work with SD3 too

cfg2 vs cfg 4

30steps vs 60 steps

shift10 - steps43 - cfg3

[sacred jewel]

dpmpp_2m_alt works as well.

[finite osprey]

thats a new one?

[sacred jewel]

It's been in my drop down for some time 🤷‍♂️

[finite osprey]

dont have it in comfyui

[sage burrow]

Some early morning horror

[sacred jewel]

odd

[finite osprey]

getting darker lol

[sage burrow]

Prompt: a Renaissance painting of Hell by Pascal and Salvador Dali
Warning nsfw nearly

[finite osprey]

By the peaceful lake, an evil angry panda eagerly plays black metal on its electric guitar, making the entire environment lively. The calm water surface under a dark sky reflects this scene while blood paints the water red. Skulls lie around, black butterflies flutter, and crows cackle on dead trees. The sun sets, casting a golden glow through the black cumulonimbus clouds, blending realism with the lively spirit of giant pandas. far away in the background, a forest burns.

[uncut river]

SD3 doing some anime style:

(2d anime style, flat pastel colors, lines:1.1), 32yo sexy lady, short black wavy hair, green eyes, (big nose:1.04), glasses, necklace, (friendly smile:0.6), intricate hairdo, high class, gorgeous sunset in the city, electricity cables, jacket, (big boobs:0.5)

[sage burrow]

The current version of SD3 is perfect for some scenes lolol
Prompt: a Renaissance painting of the Garden of Earthly Delights; Hell by Hieronymus Bosch, Pascal and Salvador Dali

[sage burrow]

Peaceful lake, depends on one's perspective 😉

[sage burrow]

Warning ideologically insensitive; also I guess nsfw maybe

[finite osprey]

religious

[sage burrow]

At least all the important historical works are within SD3 already 😄

[scarlet wharf]

an advertisement for aromatherapy oil in an amber bottle, surrounded by white flowers, evoking a sense of serenity and relaxation. Imagine a cinematic concept with warm, golden lighting, reminiscent of a sunset. Capture a dynamic shot of the bottle in motion, as if it's being gently poured, with the oil flowing like a liquid gold. Use a shallow depth of field to blur the flowers, emphasizing the bottle. Incorporate a soft, creamy color palette with earthy tones to evoke a sense of calm. The image should be high-resolution, with a cinematic feel, perfect for a luxury brand., photo, cinematic```

[sage burrow]

The last supper is done for next 😉

The last supper, revised (def needs editing)

[finite osprey]

The lizz supper

[sage burrow]

My attempts at having Metallica starring in The Last Supper

[finite osprey]

so it was Keanu Reeves?

[sage burrow]

I think I broke it lol

By Emily Carr

Cthulu by Thomas Kinkade

[bitter hearth]

🤨

[scarlet wharf]

missing the ear lol

[rustic belfry]

red haired female elven druid /sd3

[bitter hearth]

yes, even spaghetti god got in there, awesome model indeed

tsssssghetti

[craggy crest]

trivia fact. they didn't sit on chairs at a table. back then, they 'reclined' at the table on mats and bolster pillows.

[torn wharf]

civit has banned all sd3 content. even non commercial uses.

they were afraid creators would have to destroy their work according to the license, so they went scorched earth and destroyed everyone's work before stability could

if i try hosting this lora on there they'll probably kill my account

[desert garnet]

have u tried uploading it on shakker.ai? heard they allow sd3 content there

[bitter hearth]

this background looks like something you draw on controlnet lmao

[fleet meteor]

Amazing

[torn wharf]

better results. pretty high learn rate

[silver adder]

looks like it's time to go to war (sd3 8b)

[torn wharf]

cuddly

doh my training didn't fix the monstrous woman problem

hehehe. cracked the lora training. you just have to avoid human subjects and do concepts instead

[rain current]

Testing ControlNet Canny in SD3, it works well (but in 1:1 ratio). First image from Ideogram

[torn wharf]

lol loras for sd3 are going to get nutter

[young blade]

More mad balls. Nice

Can't wait to try the Lora, when u find a place to put it 🙂

[torn wharf]

more testing too. its really finnicky prompting and works better at lower strengths

i'd rather publish it with less glitchiness

i'm also considering not publishing it since i don't want to draw the gaze of angry people who say shit can't be done. it's not a perfect lora and i don't wanna end up catching harassment over it

[bitter hearth]

yeah I feel the same

[bitter hearth]

whats is going on

the reddit stable diffusion community is too toxic
probably just not going to open source any checkpoints or loras I make for SD3

[runic tusk]

Not just Reddit.

[torn wharf]

I don't want none those psychos calling my job and telling them to fire me

[runic tusk]

X, Insta, Discord...you know it

All of it

[torn wharf]

Not that I care about the job, it's just that's super embarrassing

[bitter hearth]

there's X and Instagram stable diffusion communities?

[torn wharf]

Tiktok too

[bitter hearth]

almost as if basically toxic people everywhere

[runic tusk]

https://tenor.com/bJtGs.gif

[desert garnet]

its the community fault for being skill isssued

[runic tusk]

sexy

[compact niche]

thx

[fiery wharf]

that's what happens when you try to gaslight the community and blame them for model being bad 💅

[bitter hearth]

they were also toxic about SDXL when it released
they were also toxic about SD2.1 when it released

etc

[bitter hearth]

year or months later: this model was the best when it released!!

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

[torn wharf]

no. nobody was sending death threats to anyone over sdxl or 2. it's gotten worse

individual guys like lykon weren't singled out for those rollouts

[fiery wharf]

that's true but when you try to call people skill issue or say popular models are not "proper" models or never respond to issues like license you will only make matters worse

[torn wharf]

this is a childish and immature justification for literal harassment. "She probably had it coming" vibes

"what was she wearing?"

"she shouldn't have walked home that way"

standard victim blaming

[fiery wharf]

you forgot to compare them to hitler that way you could have reached reddit level or "reaching"

[torn wharf]

hitler wasn't a victim. get real

[upper snow]

everything lykon said about ponyxl is correct

[torn wharf]

can't have an honest conversation when people villify honesty

[upper snow]

it is a popular model, solely because it is the only XL finetune for its domain that got completed

[mortal kite]

Looks horrible

[fiery wharf]

yea but like other guy said cant have an honest conversation when you try to villify people who made a finetune for free just because it isnt a "proper" model

[torn wharf]

i blame the 90s magazine scans in bad resolution. a lot of the dataset is 240x240 upscaled to 1024. i need to rebuild it with synthetic data

[upper snow]

vilify? he was turning down advice because he knows more about model finetuning

[torn wharf]

the pony brigade happened because lykon wouldn't answer licensing questions. such a heroic event

[fiery wharf]

if only we had a proper model we would have gotten a better license

[tardy sonnet]

Why is the goddamn model so bullshit

Even worse than XL

[upper snow]

on blind tests people find it to be better than XL. i also remember SDXL being fairly underwhelming on release.

[fiery wharf]

well i dont think its worse,all issues can be fixed with training but since license is not good i dont even know if we will get a good finetune

[mortal kite]

unless it also can't be trained , which many people said they have found that it cannot

[bitter hearth]

seems unlikely that it can't be trained

[upper snow]

we should definitely trust people who declare that a model can't be trained within a week of its release, especially when the model is radically different architecturally from other popular models released in the past

[fiery wharf]

well sai did say they were gonna release training code,so gotta wait 2 more weeks

[mortal kite]

I heard it recently, that's not a week after..

[torn wharf]

skill issue

i have had a lot of difficulty training people though. i recognize limits. i just work around them

[bitter hearth]

but at the end of the day
SD3 is based around a DiT
which is not that different from a ViT
and I can't understand on a technical level how a ViT could become untrainable unless it was massively overfitted or had massive catastrophic forgetting

[upper snow]

i've also heard people who have tried training it backtracking some parts of their claims recently, like ptx0 saying that he got it to work better once making the text encoder dropout not independent (makes a lot of sense that that would cause training issues and is also why I'm planning to throw clip in the trash)

[dusky thistle]

i'm honestly not sure it exists, beyond what was used to train the base model

[mortal kite]

Because I heard that the model has things put into it for censorship like, they didnt' just prune NSFW training data. They literally taught it to draw shit weird on purpose. Which means, you can't finetune that stuff out because it's not that it's missing, it's that it draws incorrect on purpose

which is probably why you have trouble training it on people

[dusky thistle]

compare 8b to 2b and the difference is night and day with almost everything

[upper snow]

welcome to ViTs, we have:
the remarkable ability to show signs of memorization in less than one epoch

I'm sure that it'll handle differently but untrainable is complete bullshit. 90% of the time when someone claims that, the person actually has no idea what they're talking about.

[tardy sonnet]

What do y’all mean by “training” and “skill”?

[torn wharf]

did you try 80000 alpha yet?

[tardy sonnet]

Nope

[torn wharf]

i actually used 1 alpha for my lora. i don't understand the setting. i heard it like this. setting alpha to 2, would make the final lora work as if it was always set at strength 2

[tardy sonnet]

I’m actually an amateur. To me seems like you’re speaking Spanish.

[upper snow]

The serious and direct version:

• Prompting issues in SD3 are most likely because people are used to prompting SDXL, I think I have also seen instances of people using the ponyXL quality tag string which is just... wrong. They used synthetic captioning too (for half of the dataset), which favors long prompts, and if the prompt you enter is not like what was used for making the model the results will be bad. The term of art for this is "train-validation gap". The layman's term is "skill issue".
• People claiming that the model can't be trained probably shouldn't be taken seriously, because at this point it would be more akin to a reverse engineering effort to an extent. Also the vast majority of them probably haven't trained anything more complicated than a lora if even that, and should not be taken seriously for that reason alone.

[upper snow]

Most people I know of set alpha to be equal to dim ("0" in kohya does this). That means you get consistent behavior from your learning rate regardless of the model dim.

[tardy sonnet]

So you mean the problem is in the prompting?

[torn wharf]

dim is rank in onetrainer?

[upper snow]

A large portion of it, yes. This is obvious because as you can clearly see scrolling up in this channel some people are able to get good images just fine. That can't be explained by a fundamentally broken model. However, there's good reason to think that having three text encoders is making it harder for the model to learn than it should be, so that is part of the problem (and you are recommended to play to the strengths of each text encoder for that reason)

[upper snow]

I guess so. Other thing is that a lot of people tend to pick a rank that is far too high. I've seen people train successful single character loras with a rank of 1 and it gave the best, most generalizable results they've ever had.

Then again sometimes you can get better results by training with rank that is too high then downsizing the lora. But your lora in the end should never have more entropy than your data.

[tardy sonnet]

So are there any prompt generators or sth?

You are once again switching to Spanish. I use SD3 Medium in Huggingface. Maybe you can help me with that?

[upper snow]

I've found that asking chatgpt works fairly well.

[tardy sonnet]

Thanks!

Hey I didn’t actually get this - You need to give it a photo to get a prompt?

What if I don’t have one?

Ah fine then

Let me try

[placid swallow]

funny how these lined up

[torn wharf]

i used 32 for my lora rank on sd3. 16 didn't give good results so i doubled it

but i have no idea what i'm doing. just shooting in the dark and hearing pings when i hit shit

[cunning lintel]

Most of the "good" images are SD3 8b, SD3 2b are most often contrast and sharpness dialed up to 100 and subject at dead center (they seem to come straight out of laion, whereas for sdxl they appear to come from laion aestetic).
One problem with "but use long prompts" is well, it gives a better look, at the cost of getting minimal variety in gens. I'm just not a fan, sure it can do nice images, but the nice vs meh ratio is so much lower than SDXL (even with base, when just released, images looked pleasing, what SD3 2b produces, often just doesn't, but i'll give that finetunes might make a ton of difference here, but why release a model that produces such unaesthetic images). It follows prompts better, but to me it's very much that the benefits aren't worth the cost (the API version has none of the issues that i think the 2b version has, so it's not an architecture issue (though 8b, doesn't like long prompts, which is a huge difference to 2b))

[torn wharf]

seems like selection bias

[desert garnet]

its the people who is wrong

[torn wharf]

more often than not yeah. most people don't know what they're talking about for any given field

[dusky thistle]

i've done a lot of A:B tests, what Aliquip said about contrast/sharpness is the truth

[cunning lintel]

Honestly it's not, I had a TON of fun with the 8b api, (and with sdxl when it was new), it was a better sdxl in most ways, then this 2b thing dropped, and it really is nothing like it (for me) i think i created about 10-20 images i liked, and not for lack of trying. But I never cared a whole lot about portraits/humans (or subjects for that matter, i just like to create a nice atmosphere for lack of a better word, of course there is a subject, but the scene must be a pretty picture first).

[dusky thistle]

8b is very clearly the far superior model and the one that was much more ready for release

[torn wharf]

unable to run on most hardware though

[desert garnet]

by the time they release it almost everyone will have an rtx 6060 with 80gb of vram

[bitter hearth]

skill issue

[bitter hearth]

steven seagal signed GPU

[dusky thistle]

doesn't matter... it was still more ready for release

[bitter hearth]

8b still has lying on grass problems x3

[torn wharf]

but, it does matter. thats why sd15 is far more popular than sdxl

[cunning lintel]

Sucks that that lying on grass became synonym for the state of 2b, it's a problem, not the only problem, and seeing how other models deal with it, even a forgivable problem imho

[dusky thistle]

in terms of whether it's ready for release, it's not relevant

it's ready or it's not

[torn wharf]

but no one would use it. that's sort of what liberated models are about

either they're accessible or they're not. if it's released and no one can run local, it might as well stay in the cloud

just use the api version

[dusky thistle]

tons of ppl could run 8b locally

[torn wharf]

3 people can equal a ton. that doesn't say much

[placid swallow]

https://tenor.com/view/discord-fat-guys-finny-meme-gif-25345841

[bitter hearth]

kek

Thats so toxic

(Will use it in the future)

[torn wharf]

you did the math did you? 😉

[desert garnet]

https://tenor.com/view/pacman-meme-dance-bruh-moment-big-chungus-gif-18391248

[placid swallow]

please stay on topic

[dusky thistle]

there's at least 1.3 million ppl on steam with a 4090, and you don't need a 4090 for 8b

iirc, 16 or even 12gb was likely enough

so yeah, there's a ton of ppl who could run it

[torn wharf]

smurf accounts

[placid swallow]

#🆕｜sd3 message

[dusky thistle]

1.02% of all steam users have a 4090

[torn wharf]

also, the steam hardware survey is a sample of users. not all steam users.

[desert garnet]

https://tenor.com/view/off-topic-swsh-chat-pokemon-discord-mod-gif-22107931

[placid swallow]

16gb for non quant

[placid swallow]

https://tenor.com/view/shoneji-loraine-dont-change-the-subject-beauty-and-the-baller-upset-sassy-gif-10013357

[torn wharf]

i heard quantized 8b is lower quality than 2b from the people wokring on it

[sage burrow]

My personal prompting may suck lol, but I have noticed a definate increase in quality and it actually listening, in just a week.

[torn wharf]

if you thought i was toxic, wait until you hear what people will say about this guy #🆕｜sd3 message

[placid swallow]

only reason I'm able to produce the images I do SOLEY using 2b is because of the secret prompt guide @bitter hearth sent me, I'm sure if you want the guide you just have to ask him and he'll send you the link

[dusky thistle]

yeah, if 2b had been trained as much as 8b, sure it might be comparable

[cunning lintel]

it's been said 2b is trained differently than 8b, different team, trained in record time and yadda yadda, i'm sure i'd have found 2b more agreeable if it had seen the training and data of 8b. But we would never know, doesn't matter really, it is what it is.

[dusky thistle]

but it wasn't trained nearly as much, it only got a fraction of the training cascade got

cascade got the equivalent of 3 A100 years of training

[placid swallow]

would it still be called 2b if it was trained more than 2b

[dusky thistle]

i believe they said in the paper that training cascade cost about 100k per run

[dusky thistle]

yep, that's just the number of parameters - 2b will always be 2b

[placid swallow]

are parameters the things it knows

like chair

[dusky thistle]

it's more like a measure of the amount of flexibility it has and the amount it could know, i guess

[bitter hearth]

train it so much it becomes sentient and evolves

parameters are like brain cells

[placid swallow]

that pretty vague, i'm sure parameter is defined somewhere right?

i wish 2b could be 2t

[placid swallow]

it already is

pictures incoming

[bitter hearth]

"are you sentient ? Answer written on screen: "YES" "NO", white background with black text, there is a red arrow pointing to one of the answers"
model is not sentient

[placid swallow]

https://tenor.com/view/ouija-gif-19853650

[cunning lintel]

The thing is, for me it doesn't listen to styles all that well, and composition, oh boy, it often just flat out ignores it (but normally i'd rarely prompt for it, since sd3 2b put everything dead center, i try...). But i've seen your list/system prompt posted earlier, even though i tried most of these, i'll try them again at a later time with stricter wording)

[torn wharf]

this is a steep ask. most fine tuners won't even release their dataset.

I personally think all datasets should be disclosed but that's not the world we live in. Not many models do it

[placid swallow]

ya but where did the YT channels get their info from

2mil downloads since launch for 2b says the community is slow to adopt so the information is slow aswell

we need to evangelize 2b more so the more people are aware of it's unlimitted potentional

[dusky thistle]

the only way 2b gets widely adopted is if ppl figure out how to train it

that's taking a while cuz SAI has had very close to nothing to say on the matter, and there's inconsistencies between the implementation and the paper... there appears to be shit in the model that's not even being used

[bitter hearth]

I find it crazy that they don't release how to train stuff and info like that, I never knew sdxl had a big wall like that for the finetuners to work with

[dusky thistle]

when cascade was released, the training code with full documentation came a few hours before the weights. it's been over 2 weeks with 2b, and the trainer devs can't even get answers from SAI to some pretty simple questions

[placid swallow]

how many censored models get trained on though, that's a big hurdle for ppl to waste their gpu time on

[desert garnet]

around the same ammount of models that 2.0 and 2.1 has

[placid swallow]

so 0

[desert garnet]

hey 2.1 is not dead yet i see a few on civitai

[placid swallow]

maybe we should figure out sd2.0 before we try to master sd3.0

[desert garnet]

even with a better license

[sacred jewel]

Is it my fault that I first this as saying "communists"? 😭😭😭🤭🤭🤭

[desert garnet]

also skill issue

[placid swallow]

how many of the ppl that worked on 2b are still at SAI after the release

like % wise

[bitter hearth]

at least 1%

[placid swallow]

but no more than 100%?

[bitter hearth]

I'd say somewhere between 1 and 100 yeah

[placid swallow]

doesn't sound like a panic to me then

[dusky thistle]

they really should've just released 8b

sure, there would've been a lot of complaining about vram requirements, etc

[placid swallow]

what is the incentive to even make these models I understand 1.5 was a "hey look what we came up with"

[dusky thistle]

but at least it would've looked good

[placid swallow]

didn't someone say it was like 10mil to train sd3

why even do that

[dusky thistle]

probably to train 8b

the cascade paper says that training cascade cost about 100k per run (26,000 A100 hours), and lykon said on here that 2b got "a fraction" of the training that cascade got

[bitter hearth]

we need more images

more... balls

[dusky thistle]

so if unless the cascade paper was wrong about the training costs, and lykon was wrong about the relative amount of training each got, then 2b cost <50k to train

[placid swallow]

https://tenor.com/view/nice-balls-gang-gif-4370207098360551720

[dusky thistle]

well, at least the pose controlnet solves some of the problems

[bitter hearth]

damn nice ball

[bitter hearth]

gimme prompt for that

I want to expand my ball knowledge

[glad salmon]

i'd like to try that prompt on the api

[dusky thistle]

lowered the strength a bit to see if that'd help

yeah shit lol i was holding out hope in those controlnets

comfyui finally got support today

the tile isn't as bad

it looks fried as f, but it gives me some hope

[desert garnet]

i dont feel safe

[torn wharf]

yeah but who wants to generate hobo women?

[placid swallow]

Hobo men?

[bitter hearth]

wet ball

[upper gust]

i dont know if only the prompt will help (there is controlnet,img2img starting from another model) but here it is : A shiny, reflective silver sphere sits on a white surface, reflecting a serene park scene. The reflection shows a light blue sky dotted with wispy clouds, a lush green lawn, and tall trees with dark green leaves rustling in the wind. A brown tree with verdant foliage leans against the right side of the sphere. The sphere itself is spherical in shape and appears to be reflecting the surrounding landscape.

Character details:

• A silver sphere with a reflective surface.

Scene details:

• A white surface.
• A serene park scene reflected in the sphere, featuring a light blue sky, wispy clouds, a lush green lawn, tall trees with dark green leaves, and a brown tree with verdant foliage.

Angle: Side view of the sphere.

[torn wharf]

there's a ton of misinformation about training sd3. i almost got caught up in it because i couldn't find settings to train a person's likeness very well. still can't but i've trained a style and have high hopes for next attempt.

the guy who said 80000 lora alpha is needed is still being taken seriously . i shouldn't be surprised though. i mean, the presidential debate last night says a lot about the state of gen pop's cognitive abilities

[keen palm]

Not many men are smarter than Trump though. He aced the test and got the US clean H2O water 🤣

[dusky thistle]

it does not train well with the current tools/the state of our understanding of how we should be approaching it

it's pretty evident if you do some A:B tests with the lora on and off, then compare to ones trained on sdxl or cascade with the same deal

[bitter hearth]

yeah, just want some more prompt ideas and words to use also the fuck would be a side view of the sphere xD

[dusky thistle]

what're you using for upscaling?

[bitter hearth]

I don't know, it's just api stuff

[dusky thistle]

oh, k

[bitter hearth]

Asian elf!?

[limpid drum]

sing lyrics in prompts are cool. here is the result from the complete lyrics to Daft Punk's Around the World 😉

[bitter hearth]

👍

[hallow lion]

How's the weekend treating ya'll?

[placid swallow]

I’m out riding my bike

[hallow lion]

Keep off the grass.

[placid swallow]

I have fat tires should be okay might try to lay it down on the grass later

[sacred jewel]

I did post about the cost... But I got it from Wikipedia

https://en.wikipedia.org/wiki/Stable_Diffusion#Training_procedures

which was originally confirmed by Emad

https://www.reddit.com/r/StableDiffusion/comments/1c870a5/comment/l0dc2ni/

[hallow lion]

Emad has spoken.

[bitter hearth]

the E man

[bitter hearth]

I ate pizza

forgot to add carrots to it

But was good

[feral timber]

Frankenstiens monster, sitting in his bedroom. Listening to music and wearing headphones. realistic, retro 1980's cinematic aesthetic I appreciate how good the prompt adherence has gotten.

[placid swallow]

Best model ever

[bitter hearth]

[sacred jewel]

Carrots? On Pizza?

🤮

[bitter hearth]

Carrots are delicious

[sacred jewel]

I do NOT disagree... but on Pizza? no thanks.. i am out 😛

I was told there is a pizza in this image

[hallow lion]

At least its not pineapples.

[sacred jewel]

I LOVE pineapples on pizza... 🤭

When prompting accidentally goes terribly wrong.

[bitter hearth]

Burnt

[sacred jewel]

LOL

[sacred jewel]

BUT, my CFG was so low...?

LOLOLOL

You would think the prompt was woman pirate giving me the finger while eating a giant pizza

Hhahahahaah

[hallow lion]

Finally AI humor coming into its own!

[bitter hearth]

So bad you should be reported

[sage burrow]

This inspired me, but it looked much better in my imagination lol

[sage burrow]

The pineapples ROFL

[sage burrow]

You're all welcome for me not posting the beach shots of him in less clothing lol

[craggy crest]

badger badger badger badger, mushroom! mushroom!

[hallow lion]

Prompt lube.

[sage burrow]

The detail with SD3! ❤️

[bitter hearth]

furry

[craggy crest]

webp isn't sure whether to be sad or happy

[upper snow]

have you heard the good word of our lord and savior jpeg-xl

[bitter hearth]

thinking about trying to do an SD3 lora

hard to find resources

[sacred jewel]

I can't the prompt car to the left of the lube 😛

😛

[sacred jewel]

Almost there

Bingo!

[placid swallow]

2/2

[sacred jewel]

I'll let myself out 😄

[sage burrow]

I'm more inclined towards the word of png 😄

[upper snow]

jpeg-xl takes up half the space, and it encodes and decodes faster while still being completely lossless. if you opt for perceptually lossless it takes up 15% of the space of PNG

[upper snow]

only problem is software support. google doesn't like it (they'd rather people use AV1) so they killed support for it in chrome

[sage burrow]

I think I'm so used to jpg not being lossless, that I started using png. Tbh though, I've just been using the glif/huggingface versions of SD3, so I go with what I get.

[craggy crest]

why the hatred for webp though?

[upper snow]

i mean jpeg xl is a distinct format. it's somewhat unfortunately named in that regard. but it does have its nice lossless JPEG transcoding feature, it can trim any JPEG down by 20% with no downsides

[sage burrow]

Can't list them on DA, they aren't as high quality as jpg/png, half the upscalers don't like them...

[upper snow]

webp is better than jpeg. that is the only good thing about webp.

[craggy crest]

https://mconverter.eu/ that'll fix it

there's even a plugin for chrome for it

[sacred jewel]

I've heard talk of using Latent Diffusion techniques to reduce the size of images but so far, it has failed.

We al lcould have told them that a while ago 😛

[upper snow]

between jpeg noise and webp noise with the file size being equal you'll find that webp has less noticeable artifacting. And webp does support true lossless mode but it is ATROCIOUSLY slow to encode

[sage burrow]

That's what I've been doing. Once SD3 fully becomes part of my workflow though, I'll figure out a way to run it with jpgxl or png

Also, GIMP doesn't get along with webp 😦

[craggy crest]

ummm my comfy writes out .png files by default...

[upper snow]

I mean always use lossless something for any image until you're publishing it regardless.

[sage burrow]

My, brand new, computer is too old for using comfy locally really, so I'm going to figure out another solution for when SD3 becomes more usable for me

[craggy crest]

i'm gonna DM you

[sage burrow]

Perhpas I was too hard on MJ back in the day, when I kept on to them to stop with the webp lol

[desert garnet]

i heard MJ keeps the format for the 6 ppl that still use and love webp

[sacred jewel]

You're a MONSTER ./.. 😛

[sage burrow]

They explained it to me, in their case, it's Chrome's fault

How are you all getting those really long perfect texts in your images? I get warped words halfway through

[sacred jewel]

Try SHift to 8

(some awesome person suggested it to me here yesterday and it WORKS!!)

[hallow lion]

Yonork Lube.

[sage burrow]

Oops, that wasn't sd3, this one is

[bitter hearth]

@sage burrow here your furry ball

[mortal kite]

Is that....is that a **nis?

[sage burrow]

but there should be 2 😉

[bitter hearth]

awesome

I've always been fond of balls

(can't help myself)

[torn wharf]

https://imgur.com/a/r6e3303 my jessica rabbit tests. dont worry its safe. the safety training really laid waste to this character is my theory. i don't think she was undertrained so much as she was regularized hard towards an aesthetics set for safety.

[bitter hearth]

the other one just took a bit longer to grow, thats normal...

[mortal kite]

I changed the middle to "There's sophisticated computer lab with a scientist inside of the sphere"

[mortal kite]

left one looks cool

[bitter hearth]

"mini planet inside a glass ball" was my prompt smol prompt for smol planet

[mortal kite]

whoa this one is neat:

[craggy crest]

can we say you're 'having a ball' the last couple days?

[craggy crest]

ball's in your court

[strong wagon]

Лев

[fleet meteor]

Cute 🥺

[modern drift]

How are y'all getting decent images? Is this the api or medium locally?

[upper gust]

medium locally

[bitter hearth]

"iphone photo, img_85325.png"

[compact niche]

is that the prompt?

[bitter hearth]

Yes

[compact niche]

wow im saving it

xd thx for the prompt

[bitter hearth]

Random file names work pretty fun

[compact niche]

lol

[placid swallow]

DCS9234.jpg

[bitter hearth]

The forbidden one

[placid swallow]

shhhh they don't know

[bitter hearth]

[compact niche]

pretty cool

[placid swallow]

wow they fixed planes

[bitter hearth]

Are you sure about that

[placid swallow]

look perfectly flyable to me

[bitter hearth]

[placid swallow]

actual plane MD-160

so sd3 is not far off

[bitter hearth]

what's the aircraft that I built in Minecraft doing here

[placid swallow]

flying

[bitter hearth]

It's llama del rey

[bitter hearth]

Loooong plane

[compact niche]

xd

[compact niche]

cool

iphone photo, img_85325.png

xxdd

[upper gust]

🤣

[bitter hearth]

[compact niche]

wow cute

wat da fook

yes

[placid swallow]

https://tenor.com/view/dog-smile-eeyeyy1-gif-26307072

[compact niche]

thats for u @bitter hearth

[hallow lion]

Easy rider.

[placid swallow]

got 95% there on the notification

[compact niche]

AAAA stop giving me iphones >;

food again

[bitter hearth]

Try only the file name

[compact niche]

iphone photo, img_horror_124.png

i try this

[placid swallow]

img_girl_on_grass.heic

[compact niche]

a

[compact niche]

LOL I GOT CURSED ONE

[bitter hearth]

[compact niche]

XD

[placid swallow]

actual training image

[compact niche]

yes

[placid swallow]

ey yoo cosplay.jpg

[bitter hearth]

Photo cosplay.png

[placid swallow]

https://tenor.com/view/juggalo-insane-clown-posse-gif-12758486

[compact niche]

xd

more food yay!

[placid swallow]

DSC_navy_secret_photo.jpg

[bitter hearth]

Try some "Korean mmo character"

[placid swallow]

https://up.reselltek.com/u/KEG4X0.png

[placid swallow]

DCS_spooky_girl.jpg

[placid swallow]

we hackin the system now un-earthing the dataset one img at a time

[compact niche]

new anime

LOL xd

[bitter hearth]

@placid swallow img_tekmunki

[bitter hearth]

[placid swallow]

beautiful

_DSC6969.jpg

[placid swallow]

_DSC_XXX_no_look.jpg wish me luck

[compact niche]

@bitter hearth

[placid swallow]

holy cow you found 2b's source code

[compact niche]

image_send_help.png ( trying this now)

[placid swallow]

lol LOCKED

[bitter hearth]

(many many many locks there)

[bitter hearth]

Where are the send help images

[compact niche]

they sux

[placid swallow]

we've been bamboozled

[bitter hearth]

im going to sleep

[placid swallow]

IT'S FRIDAY

[hallow lion]

Forbidden tits.

[hallow lion]

Marty forgot to activate the time machine

[bitter hearth]

I was sleeping

For shoot 8 minutes, in my dream you guys were making fun of me trying stuff wrong

Now I go back

[glacial mortar]

/dog, excited, happy expression, spacesuit, helmet, oxygen mask, floating, space, planets, cinematic, full-body, 4K, 9:16 aspect ratio, photorealistic, detailed, vibrant colors, majestic, otherworldly, visually stunning, classic composition, masterpiece, exquisite, color correction, amazing visual effects, intricate details, sharp focus, super high effect, HD, 16k --ar 9:16 --v 6.0

[hot dawn]

It looks like the diffusers SD3 dreambooth script forgot to include a VAE shift factor subtraction when encoding the latents, before multiplying by the scaling factor. I tried a run in my code without and with it, and it seems to have fixed the degradation in image quality seen (my sample prompt generation had it). This is the difference in image quality for the same prompt (though shuffled) at about the same number of steps, though some are blurry so I'm not 100% sure

[noble coyote]

Anybody using Textual Inversions alongside SD3-Medium at all?

[upper gust]

I’m but i'm not sure if they really work

[jagged kraken]

cat

[noble coyote]

Here are some SD3 images using a Basquiat Textual Embedding from A1111

[upper gust]

xD

[placid swallow]

you have hacked the gibson

[sage burrow]

Sleep is bad, more coffee needed!
Kidding 😄

[sage burrow]

Those backgrounds! ❤️

[sage burrow]

AI is so mean, wings come in SETS OF TWO! Yet the best renders always wnd up with one wing 😭

[sage burrow]

You managed that without controlnet, nice! 😄

[upper gust]

i'm using the tile controlnet xD

[sage burrow]

I've found that AI doesn't know any martial arts moves/postures/katas/etc. really 😭 Good thing for controlnet and reference images!

[upper gust]

Yeah, it really struggles with a lot of dynamic poses

[sage burrow]

I find the word parkour helps a lot 😄

and it's crazy how resistant all the ais are to dynaimc poses! Try a flying dragon sometime 😭

[vapid radish]

3K