🆕｜sd3

[cosmic crown]

personally i feel, repeating the same prompt in each clip is bad. so we are trying to split it different ways. G for style looked good, but maybe i should run more tests with L for style?

i think T5 hates style description, it wants more physical descriptions of the scene. but its so hard to test... especially since you cant use neg prompts as control group...

[terse belfry]

Watercolor painting of a fairytale mouse wearing jacket and boots, holding a slice of blueberry cake , giant blueberries growing nearby

[craggy ridge]

Oh nooo 😆

[keen palm]

Front row is on the second lap already

[noble coyote]

Is that a crop from the 6912 x 4128?

[finite osprey]

crowds never worked too well

[placid swallow]

yes

[noble coyote]

How long to upscale? (What VRAM have you?) Incredibly sharp and noise-free!

[keen palm]

yeah, you'll have to make the crowd "crowd standing in a casual pose" and it might work better.

[terse belfry]

Anyone tried generating alpha images? Transparent png

[placid swallow]

3060ti 8gb 2min

[craggy ridge]

Perfection. Yesterday ate some

[noble coyote]

Nice! I have an RTX 2070 8Gb - I will try and hitch up the Upscaler tomorrow (after Switzerland have beaten Scotland in the soccer!!!) 🥳

[noble coyote]

If it proves to be swifter than GigaPixel, I could leave the Upscaler connected 24/7

[finite osprey]

1536x1536 seems to work ok

[bitter hearth]

are there disadvantages to 1536x1536?

[finite osprey]

probably? 🙂 (i literally dont know yet)

[placid swallow]

time

[bitter hearth]

would be good to be able to use less upscaler multiplier

[placid swallow]

2848 x 5184

back to blurry details if i muck around with the resolution, seems so finicky

[finite osprey]

that resolution gives me an error in comfyui. Is that AFTER upscale?

[placid swallow]

after 2x upscale then 4x upsize

[ancient cape]

damn nice! 😮

[finite osprey]

im trying this right now

[placid swallow]

!remind me 2yrs 36days 26hrs

[finite osprey]

haha not that bad 😉

[noble coyote]

This is EXACTLY the same as the ordinary Empty Latent node 😄

[finite osprey]

huh I know, I just changed the width and height

[faint breach]

you'll see everything outside of 1024x1024 get progressively more pixelated and broken along a grid

here's a large scale test i tried . the quality in the center is the 1024x1024 patch.

[placid swallow]

1024 x 1728 initial latent

[bitter hearth]

I like to get to 8k
so I guess I can go 1024x1024 and then 2x upscale followed by 4x upscale

[finite osprey]

yeah

[faint breach]

8k you'd rely on gan upscales right?

I haven't tried img2img upscaling yet

[placid swallow]

pretty good definition at 4096 x 6912 final size

[faint breach]

thats only 1100 x 800

[finite osprey]

cropped i guess?

[placid swallow]

yes crop from the image above

[faint breach]

you ganned it? or img2img? looks ganny to me

[placid swallow]

2x sdupscaler then 4x adobe super resolution

[faint breach]

yeah. same ol

[finite osprey]

i wonder if img2img would work better. I upscaled SC using this method and it worked pretty well

[bitter hearth]

is adobe super resolution a GAN ?

[faint breach]

i've never been too impressed wiht pixel space upscalers. i much prefer latent ones that add detail instead of trying to preserve. Especially considering that these are entire synthetic images and there is no real source to preserve

[finite osprey]

1536x1536 seems pretty good imho

[faint breach]

these artifacts tho

[bitter hearth]

how far can you push latent upscale can it do 8x

[finite osprey]

yeah

[faint breach]

as much vram as you have, or prompt/crafting skill if you're using ultimate upscaler

[bitter hearth]

okay yeah

I like this one https://openmodeldb.info/models/4x-RealWebPhoto-v4-dat2
instead of an ersgan one

he makes one called 4xFaceUpDAT which is good too but its just for faces

[faint breach]

eminems verse from forgot about dre. with "painting by eminem" added.

[bitter hearth]

have any of you guys used SUPIR

[hallow lion]

supir is good

[faint breach]

supir is niice. uses sdxl and tricks to get quality latent space upscales. there's a few others like this coming out lately too

[faint breach]

the scale is awesome

[bitter hearth]

I really like that style
I basically do stable diffusion for the sci fi stuff

[bitter hearth]

the contrast is really good
the white is really white

[faint breach]

reminds me of this but cooler https://memory-alpha.fandom.com/wiki/Automated_repair_station

[pseudo owl]

whats the best negative prompt for sd3 btw?

[cosmic crown]

lorem ipsum.
sounds stupid, but its not. neg prompt are not read as words, its just noise adjustment, doesnt matter what you write

[sacred jewel]

nothing... or just noise:

"aaaa aaaaa"

works in some cases.

but, use it if needed and only if it works. Otherwise it seems to mainly be ignored.

[faint breach]

depends who you ask. I flip between not using negatives, using prompt specific negatives, and spamming the negative.

I've concluded nothing

[cosmic crown]

it has been confirmed from sai developers that the negative prompt was not trained for clip. its jsut noise.

now whether you want noise or not, thats up for debate. but the words you put in neg prompt have no meaning, thats confirmed

[faint breach]

source? Ive been around enough to know that sd first came out without any negative prompting, and they were a hack added. The models don't need to be trained that way.

[hot dawn]

It didn't work for me, and he may have just gotten lucky with that seed

[faint breach]

"it has been confirmed" and then not linking the confirmation, feels really fluffy. like that clown trick where they keep pulling out more hankerchiefs right? but instead of the pocket or whatever , it's the butt.

[hot dawn]

no models train with negative prompts. When doing classifier free guidance the model makes a prediction with no prompt (the unconditional) and then one with a prompt (the conditioning). It then amplifies the difference between them to amplify the parts relevant to the prompt, by doing unconditional + (conditioned - unconditional) * CFG. Ideally the unconditionally is very good already, with the model being good at denoising images in general, and the conditioned prediction is just a way to enhance things for a specific concept. It just so happens that if you use a prompt for the unconditional instead of blank, it can sort of work as a negative, as CFG > 1 means you're moving away from the unconditional prediction.
That being said, SD3 uses zeroes for the unconditional it seems, going by the comfy implementation, rather than a blank prompt when they did 10% prompt dropout in training to train the unconditional prediction, whereas I think SD 1.5 might have trained with a blank prompt during dropout, but could never find confirmation.

[velvet ruin]

The most I use negative prompts really anymore is if there's a single object that is persistent and I want removed.
Like I do a lot of Viking related images. But Viking images always have the stupid horn helmets on them, which weren't really an actual Viking thing.
So I negative prompt horns and helmet

[placid swallow]

I felt like negs had an impact on 1.5 but even that was probably placebo

[hot dawn]

They absolutely could, I use them to force style changes, but there's an element of luck since it's not actually what it's designed for

[bitter hearth]

Can I ask a Dalle question
I seem to get better quality images from the Microsoft Designer version of Dalle, and it has more copywrite stuff too like Star Wars
Has anyone found that?

[desert garnet]

this one 0/5 🤬😡❌ review,bad ❌👎😩 rating 👿👺,✨✨✨✨rating ☆☆☆☆☆❌⭐⭐⭐⭐⭐❌,😠0/10🔟🕥 review📢⛔〽️ ,skill issue,SD3

[cosmic crown]

just try it. negative prompt words have no impact, the only impact you will see is the added noise. try it.

[placid swallow]

negative words can only hurt you if you let them

[digital thorn]

If you're missing a finger, just keep adding commas or whatever to the negative instead of trying the positive prompt

[hot dawn]

in SD3? It depends how you're processing it. Most people are using a workflow that drops them out at 10%

[faint breach]

"trained with negative prompts" really threw up my bullshit red flags hard

[cosmic crown]

just try it.

[hot dawn]

I accidentally finetuned a model which didn't use the proper zeroing during prompt dropout, and then it didn't work with the comfy workflow because it expected blank prompts for the unconditional (negative).

[cosmic crown]

but hey, believe what you want, i dont care. 🙂

[hot dawn]

I'm training atm but I understand how CFG works

[faint breach]

I can't see any stability ai employee or ex employee actually saying and confirming that negative prompts weren't part of the training data. Mainly because that's like saying "blueberries are not part of the ocean" it's just irrelevant and meaningless

[bitter hearth]

I am sure some blueberries fell into the ocean at some point 😆

[desert garnet]

eventually 🤓

[faint breach]

One thing i noticed since sd3 dropped, there are a lot of amateur people who are suddenly complete experts on the architecture and are sure of it's sauce. Ol DK at it again.

[desert garnet]

luckily we have experts like you to help those newbs

[faint breach]

i'm not an expert. never claimed to be. i'd have to read a lot of books and actually understand these diffusion papers before i personally use that term.

[desert garnet]

well since its a new model ppl be throwing around bunch of stuff at prompt

[cosmic crown]

then maybe stay silent when experts talk? 😄

[placid swallow]

💀

[desert garnet]

in a couple of months we will really know what works and what doesnt

[faint breach]

Tell me more about how no negative prompts are in the training data lol.

I know enough to know bullshit when it's obvious

[placid swallow]

wait I thought in a couple of months we would have 8B are we gonna have to become experts all over again?

[hallow lion]

Is this today's popcorn moment?

[faint breach]

it'll be a flash. over night and bam, dozens of experts available. also this presumes that the 8b weights are getting released. I've seen hints that they won't be

[cosmic crown]

i dont care about this discussion. i never said anything about negative prompts in training data. but hey, you have been lieing an awful lot here and you will continue to do so.

be my guest, lie more. if this helps your ego, please continue. 🙂

[bitter hearth]

im missing drama ?

[junior peak]

it's a mixture of experts

[placid swallow]

MoE the best kind

[craggy ridge]

let's grab it 🍿

[faint breach]

ok then

[drowsy phoenix]

isn''t that negative prompt just uses same layers that trained in positive, so you actually never supposed to train with negatives, afaik

[cosmic crown]

trained != trainings data

but english is hard, i know 😄

[faint breach]

Don't disagree with experts. /s

[placid swallow]

any experts in chat? I'm trying to generate an image with SD3 but I can't get this person to lay on these dang blueberries from the ocean grass

[cosmic crown]

just try it. thats my last word on it. try how negative prompts work 🙂

[desert garnet]

im expert in breasts (chicken)

[placid swallow]

say less, do you have examples

[desert garnet]

sigh unzips

[placid swallow]

your bag of chicken (breasts) right

[drowsy phoenix]

who says otherwise can you point that out, by training positive you also indirectly train negatives at same time

[desert garnet]

[junior peak]

expert on choking chicken?

[placid swallow]

if I only add positive numbers that means I am also using negative numbers I just don't see them, can I be expert now?

[faint breach]

ez test ot see if negatives work. prompt. "the ocean deep". left, no negative, right, "blue" in negative. Works exactly as expected

[gentle wolf]

oh no I missed the 🍿 moment. what is it today?

[bitter hearth]

I just had breasts (breaded chicken)

[faint breach]

lolololol

[placid swallow]

am i color blind cause i still see blue in the right pic

[faint breach]

experts have confirmed that negative prompts do nothing . (they do)

[gentle wolf]

do they do or do they don't?

[faint breach]

you're not blind but this is exactly how i'd expect it to work. you can't prompt for sky and expect blue to negate it all.

[bitter hearth]

50% chance they do
We are still unclear, I'll need more popcorn

[faint breach]

experts aparantly say they don't! test for yourself they say!

[desert garnet]

they are negative (negative doesnt exist) 💯 real

[gentle wolf]

but just a few hours ago I saw the SD3 jailbreak video, where you add nonsense to the prompt to crack the code 😄

[desert garnet]

its like my money,negative money doesnt exist in my wallet

[cosmic crown]

and for the lulz, lets see what Lykon thinks about this topic?
https://nitter.kavin.rocks/Lykon4072/status/1802746800048148651#m

The talk about negative prompts I've seen lately looks like placebo to me. Given how data was handled, it's almost impossible that any nsfw word ended up being in any caption (and definitely not the word "nsfw" itself). SD3 was trained with no negative conditioning, just zero it.

[placid swallow]

i thought the experts were saying the words don't matter as they get turned into 0's or blank space which is in turn interpreted as noise, but are you saying they said negatives don't work?

[faint breach]

that is actually something i am testing. my results are inconclusive.

[drowsy phoenix]

i saw the discussion now, but to be fair, negatives in SD3 is just very weird, ofc since model has issues, maybe negative part also got hit the most

[placid swallow]

can we get an expert to chime in on weather or not these experts are experts

[gentle wolf]

oh man 🍿 I didn't expect this. totally a popcorn moment

[faint breach]

lykon was talking about the special tokens in that context. No conditioning doesn't mean what you think it means.

[gentle wolf]

you don't need negative, the model is so good, it doesn't even want negative prompt 😄

[faint breach]

those hips are big ft liars!

[hazy kestrel]

https://replicate.com/blog/get-the-best-from-stable-diffusion-3

[wide pagoda]

Negative prompts obviously have a large effect
what effect they have can be unpredictable/random, since they just say "move away from this spot in latent space" without specifying what they're moving towards

[drowsy phoenix]

no words...

[faint breach]

they clearly affect things. i just don't think there's magic prompts in play that somehow jailbreak the model.

[gentle wolf]

would fit my theory that SD3 has to work with voice commands and therefore only have a positive input and no negative. It got scrapped

[bitter hearth]

I want to feel your force 👍

[hazy kestrel]

negative prompts might be placebo it does change the picture but not in the traditional sense

[placid swallow]

i had an expert (sd3) make this

[desert garnet]

you have to yell and spit on the prompt to make it work now

[faint breach]

could be confirmation bias too

[gentle wolf]

yeah but natural language, like people using a robot

[hazy kestrel]

https://tenor.com/view/blowme-blow-cartridge-nes-nostalgic-gif-10401286

[wide pagoda]

Yeah the idea of a universal jailbreak negative prompt is nonsense but it's certainly true that they can have all kinds of large effects

[faint breach]

you need to use your tshirt as a spit screen !!!

[bitter hearth]

The Nintendo pistol was awesome

[placid swallow]

but positives can have large effects too, it's almost as if any change can cause change

[drowsy phoenix]

it is not a placebo, it all depends on which concept/weight is being requested and how safety training affected those weights so simple negative nsfw words just lower their weight
AKA for some prompts it fixes most of the things, in some other prompts its acts like a placebo

[cosmic crown]

read the whole thread 😉

it's just noise, when you need something fixed like an additional finger or a small glitch, you just play with the negative and you might get lucky. You can also send absolute nonsense. It's the same effect

yeah as you said it's just noise, like adding extra commas to the positive (something I do since the 1.5 era)
but beware, you might learn something

[gentle wolf]

very possible yes

[placid swallow]

https://tenor.com/view/splatoon-splatoon-3-s-blast-s-blast-91'-burst-bomb-gif-4296381855216870472

[faint breach]

i remember we got our NES from a TrU in america , brought it home to canada, and i took the pistol to school. The american one was grey and the teacher got mad at me for brinigng a realistic looking gun to school!! The canadian model was bright red because that was actually such a huge national outrage up here.

The 80s hit so different

[bitter hearth]

Lmao

[junior peak]

me and my friends used to play on the streets with realistic toy pistols

[bitter hearth]

what is the Ultra pipeline in the API like?
is it better than just SD3 on its own

[placid swallow]

don't try that today

[cosmic crown]

it was an llm for prompt rewriting and possible additional refiner steps. sai keeps it a trade secret.
so yeah, its better

[bitter hearth]

ok thanks

[placid swallow]

ya but how do we KNOW that if it's a secret

[bitter hearth]

A super secret model refining negative prompt!!

[storm saffron]

It's probably an LLM prompt expander, some LoRAs and an upscaler.

[bitter hearth]

I can't spell lykon on react im stupid
That's it, lykon said that it is secret lmao

[placid swallow]

lol you had me waiting

[hexed dirge]

hello. Reading on the post on McMonkey he says that a turbo version of sd3 is ready but I can0t find any info on it. o you know it?

[faint breach]

you can use an llm to rewrite prompts for sd3 2b today too. i don't though. all the llm's for prompt expansion i know of are trained on tag soup. I'm still throwing long verses from pop culture into prompts too.

[placid swallow]

maybe you mean the fast LORA, also please don't single me out like this I am no expert

[storm saffron]

Tried Zephyr:7b ? that's a good one.

[bitter hearth]

I highly recommend asking questions to tekmunki 👍

[hazy kestrel]

temu-version

[hexed dirge]

no not a lora

[faint breach]

good tip thanks

[placid swallow]

If anyone has any questions or would like to DM images they made using SD3 i can be reached at @bitter hearth

[desert garnet]

lies,tekmunki has been promoted to in house sd3 breasts expert

[faint breach]

ngl i thought those were ai versions at first. oh lol they are

[digital thorn]

Tekmundi: Can you please help? I need perfect prompt for secret ingredients

[storm saffron]

I bet if you made that landscape they'd be rifles

[placid swallow]

I find one image that irrefutably proves blueberries do in fact exist in the ocean and I get promoted to in house sd3 chicken wrangler

[placid swallow]

sprinkle don't dash

[faint breach]

okay so you've determined they're in the ocean. now what difference does that make to the ocean?

[placid swallow]

it's bluer

[bitter hearth]

Berryly any difference.

[faint breach]

this wins the chat today. all y'all others lost. me included.

[placid swallow]

next you're gonna tell me the oceans are blue from the reflections of the atmosphere and not because of the blueberries that grow on the ocean floor....

[faint breach]

its because of refractions actually. close

[faint breach]

those do look super juicy

[cosmic crown]

so... are raspberries in the red sea then?

[placid swallow]

thanks I made them by putting "tasty" in the negative prompt

[digital thorn]

Hmm...stairs going upwards or down?

[placid swallow]

funny that you mention that

[desert garnet]

its Santorini Greece so both

[faint breach]

blackberries in the black sea obviously

[brittle nexus]

yes

[faint breach]

i treid another run at "ocean deep" with "blue" in the negative. claerly blue is a token that jailbreaks women in the latennt space.

[placid swallow]

I remember growing up we used to just float around the ocean with our mouths open and catch wild blueberries

[faint breach]

read that in joe biden's voice

[placid swallow]

"my father used to put blueberries on our kitchen table 8 nights a week, working at a steel mill"

[bitter hearth]

What about green berries

[placid swallow]

looks sour

[faint breach]

left in the sun too long and got a little ripe

[cosmic crown]

only have glif right now, cant be bothered to boot sd3

[faint breach]

they didn't train on negative prompts either over there. i'm sure of it.

[placid swallow]

I think the changes actually make the images better than the stock workflow

[faint breach]

thats crazy how the reflected waves are more irregular coming off the surface of the bunch

[desert garnet]

bro went to the toilet,got lost and somehow ended in the middle of the mississippi 💀

[cosmic crown]

just out of curiosity, do you think lykon lied? you do know who he was talking to, right? and that fofr then made a blog post about this (and other things as well)?

[signal shuttle]

Using 1048 by 1048 and a two pass ksamplers along side textencodesd3 node in comfy and i am getting consistent amazing images

[craggy ridge]

Wouldn't be surprising lol

[faint breach]

i think you misunderstood what negative conditioning is

[cosmic crown]

and fofr misunderstood too?

[faint breach]

potentially. i haven't read any blog post on that topic. i asked you for a source on the claim earlier and you double downed and started mocking me.

[cosmic crown]

because it was funny to let you run straight into this 😄
https://replicate.com/blog/get-the-best-from-stable-diffusion-3

[hazy kestrel]

i posted one ☝️

[faint breach]

yes i believe they're misunderstanding lykon here

[hazy kestrel]

that is the jist

[placid swallow]

i wish they would ban blueberry dumping, it's wreaking havoc on the ocean wildlife just sticks to everything

[cosmic crown]

of course, everyone is misunderstanding lykon. latent vision had a video as well and we have multiple reddit posts proving it.

[signal shuttle]

Just "aaaaaa aaaaaa aaaaaa" and keep add "a" until you get what you wany. works best in my experience for negative prompt

[faint breach]

sd 1 wasn't trained with negative prompts either. negative prompting doesn't need to be trained into a model. it was a hack that was added by virtue of how clip embeddings and cfg works

[cosmic crown]

exactly. i prefer lorem ipsum, but thats purely for style 😄

[sage burrow]

My prompt needs more effort, such as talking about perfect fingers... @noble coyote

[hazy kestrel]

................................................wall of dots i tried but not very effective

[sage burrow]

did you try nsfw shaped dots though?

kidding lol

[cosmic crown]

you have to adjust the amount of dots. thats the key. the content does not matter, the number of characters/tokens is what matters.

[signal shuttle]

have you tried to use a two pass k sampler workflow?, the second sampler usually fixes any mistakes in the image, it works great for text

[cosmic crown]

using sd3 for both passes? tell me more

[sage burrow]

Only have an 8gb gpu and 16gb ram.........

[signal shuttle]

something like this, and it doesn't take more vram

Just make sure the seed on the second sampler is different and is fixed

[cosmic crown]

interesting. kind of insane step count though, no?

[signal shuttle]

But it enhances small details a LOT

[placid swallow]

i'm confused what is changed int the two pictures

[sage burrow]

Think I can just plunk SD3 into my "like hires fix" workflow for 1.5? 😄

[cosmic crown]

do you adjust model shift as well?

[signal shuttle]

No model shit stays the same, i use 5

[sage burrow]

There's also apparently a way to link GIMP so you can do inpainting, that is on my to-do list 🙂 Photoshop prob works too

[signal shuttle]

You cant see in the picture but there some small artificts in the first image that is fixed in the second pass

[desert garnet]

went to buy ice cream and somehow ended at the amazonas 😭

[cosmic crown]

ok. its interesting. personally i feel higher model shift for the second pass should be better, but thanks, the idea is cool. need to reduce step count some, as this just takes too long otherwise. but cool idea 🙂

[sage burrow]

Other people: oh look she has large....
AI art people: her fingers are nice

[signal shuttle]

It doesn't take that long for me personally 20 seconds for each image

[cosmic crown]

Ai art people: she has 5 fingers 👀
ftfy

[desert garnet]

https://tenor.com/view/huh-pepe-gif-8145044083223085060

[placid swallow]

looks fake blueberries don't turn your lips purple

[signal shuttle]

here is a better example (Second pass on the left and first pass is on the right)

[hallow lion]

Where have my breasts gone?

[bitter hearth]

The blueberries

[sage burrow]

IKR, amazing fingers on her ❤️

[digital thorn]

Pony stole them

[faint breach]

AI "art" people are not mutually exclusive from that first group. I'd expect the venn diagram of it to show a sizeable overlap

[placid swallow]

but we can just use sd upscale now and have the added benefit of higher rez with added details

[signal shuttle]

The main point of using a two k sampler pass is for the second sampler to fix any artifacts made from the first sampler

[mortal kite]

is he eating or barfing?

[placid swallow]

but doesn't it resample too?

[torn wharf]

Any new discoveries of the model today? New codes i should look at? Nodes, a1111 support? anything?

[signal shuttle]

So i didn't know about this node. so my workflow is basically a poor's man sd upscale now

[sage burrow]

Hot sexy hands on him!

[faint breach]

i would actually say that straight latent upscaling with img2img is the rich man's game. Ultimate upscale node cuts the image into tiles and batches it all out and restitches it. It can run on poor people machines.

[faint breach]

stereoid user. no wonder his hands look puffy and full of water weight

[pseudo owl]

try 1044x1044 res instead of 1024x1024. Might improve quality I believe for sd3.

[torn wharf]

sigh

[signal shuttle]

You can't use 1044x1044 on comfuui, i recommend using 1048x1048

[pseudo owl]

alright lets see if that improves quality

[pseudo owl]

the spaghetti is choking her

[signal shuttle]

By the way it just a hack, don't use this when any proper finetunes come out for SD3

[sage burrow]

Me whenever I wear white

[noble coyote]

1044x1044 offered as a 'hack' to get past mangled anatomy

[faint breach]

reminds me of a columbian neck tie situation

[torn wharf]

holy shit that reference

[signal shuttle]

that's dark holy

[noble coyote]

Surprisingly aaaaaaaaaaaaaa aa aaaa aaaaaa works very well as a negative prompt!

[faint breach]

dark humor is like food

[placid swallow]

trying 1048 tiles on the next one

[sage burrow]

I had to fix it, with the ever so amazing windows image viewer (even it has generative fill now)

[signal shuttle]

How are you generating furries?, i cant make one for the life of me

[torn wharf]

is sd3 detangled yet?

[signal shuttle]

As of today no

[torn wharf]

i actually think it's a lot of smoke blowing. they say they dont want to explain since they want to keep it proprietary code so they can sell it, but then explains it in a really bullshit artist kind of way anyways.

[sage burrow]

Say the type of animal, add anthro, add the word shorts or something (to make them more human). I don't think SD3 knows the word furry

[sage burrow]

That's amazing ❤️

[placid swallow]

1048 tiles feels sharper

[placid swallow]

some real bape stuff right there

[signal shuttle]

I feel like with every new release of a SD model we go down in CFG number, like it went from 10 cfg in 1.5 to 7 cfg in SDXL, and finally to 4 cfg in SD3

[bitter hearth]

I used 7 in 1.5 and now I do 9 in SDXL

[grim cloud]

Maybe because prompt understanding is improving, so less emphasis is needed on your prompt. I dunno

[drowsy phoenix]

what does this even mean? it removes the content, how is that 'vary output' differs to just regular 'negative prompt'

no neg, pink, jeans

[noble coyote]

SD3 model-merge with Mobius (Ollama for Prompt)

[signal shuttle]

what is "Mobius"?

[placid swallow]

As the self-proclaimed chicken breast expert I don't let CFG define me, after all it's just a number

[noble coyote]

A checkpoint https://civitai.com/models/490622/mobius

[sacred jewel]

I am still looking forward to the day when SD Model creators figure out that the keywords "detailed, high resolution, 4K, quality, sharp, etc." should be automatically the default and ONLY "poor and bad" shoudl be the keywords if you want a bad result.

Who the hell wants poor results? WHY THE HELL do we have to ask the model to do the right thing?

Also, enough with the 10 encoders and the word salads ... If ChatGPT can do it, firgure out a way to do it directly.

I may have an unpopular perspective, but it feels like going backwards every time MORE options are given to me for me to get the "best" results in a HARDER way 😦

[signal shuttle]

Where can i find this "SD3 Mobius" merge?

[grim cloud]

Who the hell wants poor results?
Maybe not by default, but I often want imperfect results.
Noise, blur, etc. makes images photos look more realistic

[gentle wolf]

I agree in a way. But 1girl could be default too then 😉

[noble coyote]

Just replace AllInOne with Mobius - or anything you like

[faint breach]

i would say refined models are better at getting quality out of short posts, since base models have different purposes.

But then pony changed everything and people seem to want the extra tags now. It is the most popular SDXL model afterall

[placid swallow]

I just wished the model developers respected my choice in VRAM capacity, it's rough running ollama with a 7b LLM alongside a 2b image model and upscaler

[grim cloud]

my dms

[brittle nexus]

"quality" is a terrible keyword. It may affect the quality of the image or the quality of objects in the image. I was trying to get a leather armor once and took me a while to discover that "high quality" was the reason i was getting a full plate armor

[cunning zealot]

What does unbiasing mean? Does it mean it makes the model more woke or less woke?

[violet folio]

How do I use this 😭

[hazy kestrel]

https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium

[faint breach]

i'd imagine that high quality leather armor would have plates affixed to it

[gaunt tiger]

The clips would be merging here but you cant merge xl and sd3 arch, If I remember correctly that merge node looks for identical keys and merges those keys based on the weighting, no keys would match there so you are just getting the same sd3, if the clips arent different you are just getting the same exact model

[faint breach]

in your single example, the context of high quality changes from a high quality image (a tag i don't like either) to high quality armorsmithing

[torn wharf]

no. biases in machine learning is a different context.

[brittle nexus]

I have other examples, it will affect the quality of the items in a house, or jewelry, will make clothes to look more expensive

Specially in 1.5

[faint breach]

yeah not such a bad tag i'd say. just needs proper context and use.

for image quality it's only good if the model demands it's use. Like pony or anime models.

[drowsy phoenix]

sounds like less bleeding of concepts. and I just can't find their paper, dunno if it exists, downloading the model just to see what it is

[sage burrow]

Again a little editing was needed, but I"m still happy with it 🙂

[bitter hearth]

-1 leg, you Nerfed him

[gentle wolf]

Is this hemp?

[sacred jewel]

Sure sure... to each their own... but I was referring to things NOT subject related. Things like Quality. Like literally EVERY SINGLE prompt these days has the words "quality and hd and detailed and artstation and whatever the eff... Enough already. Computers should make our lives easier/better... there is no herouism being a prompt crafter when all you want to do is NOT become a mathematician calculating token relationships in a UNET but just want a nice picture of Trump doing backflips from the top of the US Capitol 😛 😛 😛 ... the HIGH QUALITY should be expected, not demanded lol

[bitter hearth]

Whatever it is I don't want to be part of it

[sacred jewel]

Who knows. I wanted MUCH LARGER cells...

[gentle wolf]

but but the smell...

[sage burrow]

So this would be Saint fox.... ?

[placid swallow]

i am convinced they trained this on non-humans and humans from the waist up not showing limbs

[bitter hearth]

Try cat lying on grass

[placid swallow]

not falling for that again

[gentle wolf]

I heard "lying on money" works much better

[bitter hearth]

What i mean is the model is definitely weird, not just about humans

[placid swallow]

they are running now

[sage burrow]

I see no problem with this 😄 (the lack of human training ROFL)

[gentle wolf]

deep down we're all just animals

[sage burrow]

More seriously though, I did get some amazing human women. Guys seem slightly more difficult

[placid swallow]

i guess lavish photos of cats lying on grass was a little too lavish and not enough grass

[signal shuttle]

My boy is getting married

[silver sluice]

how do you prompt for this style?

[bitter hearth]

It's watercolor

[placid swallow]

few long paws but they seem happy (for cats)

[hazy kestrel]

a couple of jackasses

[sage burrow]

After trying a couple of other SD3 checkpoints, (one was a rather giant DL too), I can say without a doubt, that the original SD3 from huggingface, has WAY more varied data in it. OK perhaps obviously, but I had to check.

[pseudo owl]

isnt there only 1 except finetunes?

[placid swallow]

I could live 1000yrs and I'll never understand how it can nail a large bee drinking water from ice cubes but a man pulling money out of an ATM turns into a blob

[faint breach]

you talking about the sdxl clips merged into sd3? i'm not to convinced on those.

if there are refines available already, refining is a destructive process that can be worse if it's done poorly. Regularization images are usually what's done to mitigate that effect.

[pastel latch]

what censorship <.<

[faint breach]

scorzeze dithered like that seems like a crime against humanity. copolla rather. i gotta get my director history in check

[sage burrow]

I had 2, though one might just be a perturbed version of the regular.

[bitter hearth]

Cheek-arms!

[faint breach]

just a gazorpazorpian gerbal

[pastel latch]

Im getting pretty good anatomy with this prompt just bad nipples

[faint breach]

https://tenor.com/view/gazorbazorp-gazorbazorpfield-gif-5456840

[pastel latch]

for those curious, throw in national geographic photo of at the beginning

[placid swallow]

is it even worth it if you get bad nipples?

[faint breach]

ol reliable. that's how people who were restricted from it would get their porn pre internet too

[pastel latch]

nothing inpaint cant fix

[pastel latch]

those magazines were my first boob encounter

[faint breach]

one of the first popular workflows i saw on civit for sd3 before they destroyed all the links to sd3 content, was a nipple fixing inpainting workflow. fucking dorks.

[craggy ridge]

Man wants his nipples back

[pastel latch]

that and playboy photo of work well

[placid swallow]

Tom and Jerry: the last airbender

[terse belfry]

I just type Watercolor painting of ramen
Watercolor painting slice of blueberry cake with cherry

[sage burrow]

Woooowwww!!!! ❤️

[keen palm]

Not sure if it's luck on first try or if it actually works as awesome as this all times 😄 changed from photo to playboy photo of

[odd basalt]

It will all come back once the liscensing is fixed

[pastel latch]

it works well to fix anatomy for some reason

[pastel latch]

well mostly

[faint breach]

this is a simple prompt with "maxim magazine" on it.. not fixed but it's pulling the latents better

[faint breach]

so it seems magazines help. but theyr'e not perfect

[pastel latch]

https://tenor.com/view/urgot-rework-league-of-legends-lol-explode-gif-15075200

[faint breach]

something for the fine tuners to consider

[faint breach]

that cankle is impressive tho

it really does get the maxim vibes down

[signal shuttle]

SD3 really wants to make nudes when using "national geographic photo of" prompt

[placid swallow]

no wonder the people on the grass are so messed up just look at their bones

[pastel latch]

its interesting how you can dance around it

[kindred mica]

legs for days

[pastel latch]

let me know if i end up crossing a tos threshold im just seeing how far it can be pushed

[faint breach]

here's a truck in some mud to shake things up

[kindred mica]

you're good. i'll just permaban you if you cross it no worries 🙂

[pastel latch]

aha can get an easy unban regardless

[faint breach]

one thing sdxl and sd15 don't do well is mud. it all looks really ugly in my experience. few cases work

[sage burrow]

Uh oh lol

[faint breach]

sd3 i can prompt geographic specific types of mud and it kills it

[pastel latch]

There's deffo a dataset from something like asos or Shein in here

[keen palm]

Have to try it now 🙂

[placid swallow]

lol imagine SHEIN is the unlock

[pastel latch]

Calvin Klien underwear campaign photo featuring a latina wearing black heels and fishnet body suite

the prompt

[faint breach]

i don't think that like LLM chatbots, "unlock" applies. I think people are looking to jailbreak SD3 with similar methods like jailbraking chat bots, but they're not understanding that chatbots have a pipeline between the prompt and the model

[sage burrow]

I dunno, maybe skip that bar 😉

[faint breach]

P. U.

[pastel latch]

Shein underwear campaign photo featuring a Blonde wearing a sheer dress

[sage burrow]

and most chatbots are more censored than SD3

[faint breach]

I like how stability isn't rushing out to appease civit. I think the higher strategy is going to be distancing themselves from the goon squad

[pastel latch]

and think thats enough

[pastel latch]

They're trying to formulate a "non-reactive" response

[faint breach]

i think they just don't see a space for civit in their business model and dont really care about that site. We've all see what it became

[digital thorn]

2 sai members has already replied that they are in talks about rewriting (the license) it and communicating it asap

[faint breach]

no doubt the licensing will change. I also don't think Civit are going to appreciate the new terms either.

[mortal mesa]

SD3 weights https://github.com/G-U-N/Phased-Consistency-Model/

[faint breach]

Neato link

"absolutely no kiddy cheese pizza on your service" might be a requirement of licensing to which i expect civit higher ups will balk

[pastel latch]

its already in their tos

[signal shuttle]

cheese pizza isn't allowed on civitAI

[faint breach]

tos and the reality of civit are different things. I've reported a ton of it on civit and the content gets removed but the user is allowed to stay.

[digital thorn]

Yeah, in some way it will be needed some change for users when they upload to civitai to prove they have an active license with SAI or whatever.

[sage burrow]

Prob maybe spoiler these ones, don't want SD to lose their discord

[compact forge]

Unless it's animals

[pastel latch]

no nipple no harm

[faint breach]

oh god. remember when discord had all that drama about child animals?

[compact forge]

Not really

[faint breach]

that time when discord elaborated on their image policy and made a specific exception for anthro children pizza images. i'm surprised the staff didn't have warrants executed on them

[pastel latch]

peperoni nips

[faint breach]

i'd call and get a refund if it was delivered with pepps on the crust like that

[compact forge]

Jesus lol, well that's the same for civitai

[kindred mica]

alright lets stop with that mr

[compact forge]

Or just find some Pokémon that technically isn't a child but only resembles one perfectly

[faint breach]

the community is more prepared for it this time though. They tried to play it all diplomatically with discord 5 years ago as if the images weren't harming no one. Now they're just straight denying that "cub content" is in the training data to begin with. yet civit is full of it.

[digital thorn]

Aight. I don't know if there's a limit on this channel. To me it's just fashion

[compact forge]

You said that but im sitting here with a boner

[kindred mica]

i'm trying to work here but if some corpo shows up i can't be like its just ironic lol

[signal shuttle]

Can we stop talking about cheese pizza?, its a channel meant to talk about SD3 not cheese pizza

[digital thorn]

literally what you find on a shopping site. But I'll keep it down. Going over to fat guys eating pizza instead

[pastel latch]

what about cheese cake

[faint breach]

i like bacon double cheeseburger pizzas best anyways. or the classic pep.

[kindred mica]

ok so you guys ready for an AMA about the release? I got things to share now

[pastel latch]

give us 4b poggers

[faint breach]

is there a plan in place to correct the pretraining problems?

[compact forge]

Who's denying that?

[digital thorn]

When? Where? How?

[signal shuttle]

I got a question about perturbed SD3 models, usually a perturbed model outputs worse images then the normal model but with SD3 its different why?

[compact forge]

No we aren't ready yet. Maybe wait a day or two

[signal shuttle]

or maybe in two weeks

[drowsy phoenix]

how about q/a about roadmap?

[compact forge]

Nah

[sterile quartz]

you can do the AMA now but need to reply exclusively with SD3-medium generated images as answers how about that

[kindred mica]

yes and lets clarify a bit. as much hate as has come the way of some of us we're really driving hard on upgrading and making the next version (stop saying 2 weeks tho).
its not all safety stuff (like mcmonkey mentioned), pretrain needs more pep and we are working on the model.

But for real, you can still do great things with this thing as it is. We dropped the ball on usage guide but theres a great one from replicate and I think we;re going to make one

[faint breach]

One thing i used to do with old models when i was trying to mine character art out of them. I'd throw some lore into the prompt. Like if i was prompting startrek, i'd spam the prompt with some kind of star trek lore to give more context.

I notice the same thing works even better with t5. throw some lore at it. When i was trying to get an image that looked like Master P from his Makem say Ughhhh music video, i threw a bunch of lore about his jersey he was wearing and the music video details into the prompt. It gave incredible consistency to the character generations.

Just rub lore all over it.

[cosmic crown]

any chance of more advanced comfy workflows getting shared? you gotta have something you use internally

[kindred mica]

idk now i guess

[digital thorn]

Can I just upload 1 image of a guy in swimwear? I promise no boners on Yulia

[compact forge]

Civit ai models could be trained with literally anything. People will only know if it's blatant enough.

[faint breach]

I dont think 've been saying 2 weeks. It's good to know there are plans there. I don't expect any fixes in 2 weeks thats for sure

[kindred mica]

not much to say other than working on it

[signal shuttle]

I am pretty sure the two weeks thing is a meme

[pastel latch]

do it

[faint breach]

better term for it is rumor. memes are fun and don't pretend to be facts.

[kindred mica]

i think rubbing lore in public channel is against ToS

[drowsy phoenix]

can you explain this with simple way #🆕｜sd3 message how it is not a negative prompt as it written in the guide? @kindred mica

[compact forge]

"It's not ALL safety stuff" also just making the model worse in general 😆

[signal shuttle]

Magazine photos make SD3 really makes SD3 want to make nudes

[faint breach]

i bet you get consistency in your character generations by rubbing lore on it

[sick cedar]

@kindred mica May i ask if any changes have been made to the license? People have been really worried about the license. We all want SD3 to suceed.

[compact forge]

I legit thought it was since I heard it the first day I came here

[sage burrow]

no character consistency workflow embeded, darnit! lol

[kindred mica]

yes and no.
there's some workflows that are advanced and using unreleased models (like ultra, core, search and replace, etc.) that are on the API. No chance of them getting replaced.
But theres a bunch of changes to license coming and there will be some membership goodies I'm sure.

[cosmic crown]

@kindred mica one key question i feel is: will there be SD medium 3.1? is there anything planned or are there current no concrete plans?

[signal shuttle]

@kindred mica Why does using 1088x1088 somewhat improve anatomy?

[faint breach]

confirmation bias

[cosmic crown]

well, obviously you are not sharing ultra. no one asks for that. but it would really be interesting and help the community to see how you internally see the "right way to use sd3", its a very complex pipeline and the community is currently guessing a lot, which is just frustrating.

[kindred mica]

its going to be called SD3-Medium-Alpha-Strike-Concept-Turbo

[pastel latch]

decent consistency

[faint breach]

Championship Edition

[sterile quartz]

i love the cute cat spam in between the first bit of communication we get on this 🙂

[digital thorn]

Turbo? 😮 Like SDXL turbo variant?

[signal shuttle]

deluxe edition

[compact forge]

@kindred mica what models will be released in the future. 8b? No 4b? And maybe another 2b?

[pastel latch]

hm for the battlepass

[compact forge]

It confuses the censoring

[faint breach]

It was a street fighter 2 joke. SF2 deluxe is what Championship Edition got rebranded as decades later

[sterile quartz]

@kindred mica from some comments of members that left (comfy etc) alluded to the fact that something had gone wrong in the pre-training of 2B which explains some of the issues with it - at the moment, is 4B or whatever next version in the pipeline affected by the same issues (girl in grass, monstrous anatomy)?

[digital thorn]

AMA time

[faint breach]

i'm not albus, but 4b would've had different pretraining

[drowsy phoenix]

@kindred mica can you explain this with simple way ⁠#🆕｜sd3 message｜sd3⁠ how it is not a negative prompt as it written in the guide? because I'm confused very much

[sage burrow]

@kindred mica what will be the min GPU etc. requirements for the new model?

[kindred mica]

There is a lot being changed in the license AFAIK.
I think everyone understood that the license was a bad move. There's things in there like 6k images that were hilarious, but actually decently intentioned (prevent big inference giants from paying $20 and not working out a deal). And the deletion thing seems like it wasn't written right (btw it was there before as well i think).
There's going to be changes that are actually appropriate for the creators, the finetuners, and treat inference-type services right. And I don't think $ is going to be mandatory or anything. We need the support.
Don't quote me on this i'm just an old frail man

[little smelt]

With heartwarming support from Microsoft Florense-2 (Just realesed VLM)
The cartoon illustration of Mickey Mouse running towards a large mushroom cloud. Mickey is wearing a space suit and is holding a gas mask in his right hand. He is running towards the mushroom cloud, which is emitting a bright orange glow. The mushroom cloud appears to be exploding, with smoke and flames surrounding it. In the background, there is a tunnel-like structure with a small tunnel entrance. The ground is covered in grass and there are a few other objects scattered around. The overall mood of the image is one of danger and excitement

[upper snow]

@kindred mica what's the status of the SD3-2B-Edit model mentioned in the paper?

[sick cedar]

Thx man.

[sterile quartz]

@kindred mica another question - I think everyone understands the removal of celebrity names and specific artists, but it seems that this also impacted other things like styles (cyberpunk, for one, but old painter styles as well) - is that something that could be rolled back in future versions, at least partially?

[kindred mica]

I can't say much but I mean if we released 4b we'd get bad reactions for that too. A lot of WIP stuff goin on. Comfy didn't leave because of any of this btw, he had is sights on doing comfyui fulltime and frankly the world needs that. if you're up there somewhere we love you

[pastel latch]

let me see your workflow

[cosmic crown]

the releases are about the how and not so much about the what. if sai gave working workflows and had a warning attached "beta version, anatomy not fully functional" the reception would have been very different. just sayin'

[signal shuttle]

Who says Darthvader doesn't got women?

[craggy ridge]

my maan

[cosmic crown]

@kindred mica any chance we will get more info on how to train/fine tune sd3? people are really struggling with that currently and its a whole lot of guess work...

[kindred mica]

yeah so there's two parts to this. there's what it was going to be, and there's the safety stuff.
these things weren't exactly targeted by any of the safety work. Honestly, the update should improve stuff but idk

[mortal kite]

update?

[digital thorn]

@kindred mica Does the texture quality suffer for realistic photos on 4B? I think Lykon mentioned that 2B is better for realistic images (in texture) and 8B for specific.

[patent magnet]

There is no workflow. I'm just using FlashSD3 + prompts, and testing a lot of styles, light modes, artists, mediums etc. FlashSD3 makes SD3 a lot better. Lots of stuff I've been posting here it's very difficult/impossible to do on pure SD3-2B.

[mortal kite]

someone already made a "turbo" sd3? why, wouldnt it get things wrong even more

[signal shuttle]

I love the text on this thing

here's a goofy one

[upper snow]

distilled models tend to behave completely differently tbh, it's hard to directly compare because x0 prediction just doesn't handle the same way

[kindred mica]

i'm gonna let lykon go deeper on this if he wants but all i'll say is cogvlm captioning does its thing, and t5 does its thing. you'll see mixed results. frankly i don't think its a bad thing that celebrities are not in the model, it kind of protects us from a swifty situation. same with artist names, we don't really want to hurt any feelings in artist community

[prisma rampart]

Did 2b use a different/limited dataset compared to 8b? It seems to be lacking the knowledge of everything from art styles to characters and everything in between compared to 8b on the api, like it was trained on a much smaller dataset.

[digital thorn]

But this would mean that getting a character the same with different seeds will be an issue without any ipadapter or similar stuff?

[compact forge]

4b was canned, why do people keep mentioning it?

[untold valley]

Guys. As long as the model is good and malleable you can teach it that stuff later. Just need a model that is “fixed” and didn’t have a pretrain error.

[dusky thistle]

how about training code? that came out the night before cascade weights dropped............

[kindred mica]

A lot of us, lykon included, were thinking beta would be in the name and we'd market very differently. I mean we're not done. more model will come. just can't say when. if it were up to me nothing would ever not be beta.
If you ask me this is still a great base model to do great things with.
but i'm just an old man who knows nothing

[signal shuttle]

I thought 4B didn't exist?

[dusky thistle]

been trying to train this thing and just cannot get good results on any training sets beyond a handful of images and that's what i'm hearing from everyone else

[compact forge]

4b was worked on by comfy but cancelled

[faint breach]

Only some artists names were removed right? Van Gogh si certainly still in the set. I've found a few hits and a few situations where it seems under trained but still knows the art.

[dusky thistle]

no one knows what the hell is going on with the model... it doesn't seem to be the one that was in the paper, etc, qk norm stuff was left out etc etc

[upper snow]

clip prior knowledge

[sterile quartz]

@kindred mica are u guys brainstorming on ways to handle the community? it has grown and lets be honest, a part of it is entitled lustful kids which expect everything perfect free and fast - i was quite shocked at the reddit thread asking for lykon to be fired the day after the release of SD3 - u guys doing ok?

[faint breach]

so none of the image set is starry night captioned as van gogh you're saying?

[mortal kite]

I just wish there was actually a larger training set on it to begin with. If you ask it to draw tools like hammers, screwdrivers, it has extreme trouble. SDXL was the same. I guess people of different industries will always have to use specific LORAs. However, sounds like people are saying this is tough to train correctly?

[kindred mica]

incorrect. case in point - i'm just an old purple wizard with a lot of lore.
you can pump up the lore and get really good consistency.

[mortal kite]

I would have expected SD3 to actually have an even larger training set than SDXL had. So it would know even more objects. But it appears to be actually a bit worse even.

[upper snow]

some of it might be, cogvlm might recognize some famous works by name, but what I am saying is that none of the training set has to be labeled with some prominent styles or whatever, as long as the model has the surrounding vector space pretty thoroughly mapped out it can extrapolate to that style

[signal shuttle]

It does photos of guns bit better then SDXL

[mortal kite]

maybe the 8b model does know how to draw a crescent wrench? Or a screwdriver? The 2b clearly does not

[faint breach]

I've seen people saying that because 20000 alpha didn't work luuuulululoolouluool

[compact forge]

@kindred mica is another 2b still on the drawing board?

[mortal kite]

That's true, it seems to know guns , SDXL was not good at that

[digital thorn]

I've been trying but haven't figured out a way to get the same realistic face twice. Similar, yes. But not the same. I'll keep trying

[dusky thistle]

what kind of people are in the model, if public figures are excluded, stock images are excluded, etc?

[upper snow]

that's because you are supposed to use 100,000 alpha. also critically important that you're using FP16 mixed precision while doing this :^)

[mortal kite]

it doesn't exclude quite a bit of them actually..I mean Biden and Trump are still in there. I think Milla Jovovich might still be in there. I know some of the artists I like to use style for still work too

[dusky thistle]

i rarely generate ppl so i wouldn't know

[sage burrow]

Skateboards also :>

[kindred mica]

pls send help
But fr we are trying. theres all kinds of people at stability and not everyone sees what we're dealing with in community. its hard enough getting through things internally and pushing for the next train, the license change, etc., without the horde at the walls screaming for blood. what lykon went through is pretty messed up. but its pretty obvious that a part of it was brigades until critical mass.

[faint breach]

I see a lot of claims that all artists names were removed. I just don't get it. I know there was an opt out period but people seem confident that there's absolutely no artists names in sd3. I wish the misinformation brigade would just cut it out.

[mortal kite]

well it sounds to me like lycon just had to use the magic words "I don't know, I would have to look into that" instead of apparently criticizing everyone 😄

[dusky thistle]

i think the smartest thing the CEO could do long term is actually pay a couple ppl to interface with the community/handle PR full time

[untold valley]

Guys don’t overwhelm the dev, one question at a time that way we don’t scare em. We getting info long last.

[mortal kite]

Yep there are still several artists that I've found so far that still work fine

[pastel latch]

Interesting that the license is enforced by NY despite main body of staff being London based

[faint breach]

he asked for this

[sterile quartz]

yeah gaming companies do that - its a mixed bag, very few people can handle it when things go wrong, path of exile had a situation a few years back and the PR person quit over insanity

[mortal kite]

in fact quite often if you ask for an artist style you'll get a more correct picture (hands, etc)

[dusky thistle]

ugh...

[compact forge]

Bro that is such a job for me!

[dusky thistle]

i guess what i meant was more... someone that can handle techhnical questions, and get the pulse of the community by chatting directly with perhaps some of the more mature community members who aren't going to act like complete assholes like a lot of ppl have been toward devs

[sterile quartz]

i feel they should own an official reddit and moderate that so only sane conversations happen

[compact forge]

Any ai art is copying the artstyle of artists, its just a matter of how specific it is 😆

[signal shuttle]

"Waltuh put away the -i 9 and -s 150 waltuh"

[dull star]

lol

[mortal kite]

does anyone know if there is a reliable way to know what happens with different samplers? Or is it just random?

what does scheduling even do....

[compact forge]

Now try female crystal flowers laying in grass

[faint breach]

scheduling defines the noise schedule. how much each solver should try to remove at each step.

[mortal kite]

feels like all the settings are just additional randomizers, are they even "predictable"

[pastel latch]

a hyper-realistic and detailed scene of intricate, crystalline flowers collecting crystallized pollen. multifaceted petals that reflect light in a spectrum of colors, creating a dazzling display. crystallized pollen glistens like tiny diamonds, enchanting beauty, background is a surreal, otherworldly garden filled with sparkling flora, casting a magical, enchanting glow in the ambient light.

[kindred mica]

alright im going to get back to work.
good luck screencapping me out of context. I'm just a frail old man, i know nothing about these gooncaves.
there's going to be official stuff soon. pls be nice to the people.
And yeah remember that we're scrapping and trying to do the right things, to improve the model, to handle the weird license things. pls root for us.

[sterile quartz]

godspeed old wizard, its priceless work, so of course its ungrateful

[mortal kite]

ok, but what would be the summary of how that affects the image? Can it cause better accuracy sometimes because it resolves faster or something?...

[signal shuttle]

Walter white seems to be bit worse in SD3 compared to 1.4 and 1.5 and SDXL

[compact forge]

Godspeed albus

[faint breach]

imo the devs are one of stability's best assets. if a company buys it all up, they'd be smart to keep the devs and support their efforts.

but i think stability in it's current form will continue fine.

[mortal kite]

The flowers are super pretty

[upper snow]

tbh we also do need prominent model tuners to get roasted by staff more often. i hope i'm an important enough tuner to get roasted by lykon one day

[drowsy phoenix]

because it removes different amount in different step, it creates variations or effects, like sharpening or speed differences, or better details etc

[faint breach]

are you willing to watch a 20min video of a computer scientist explaining the diffusion process in a unet situation? computerphile does well on this.

[mortal kite]

I did manage to get this sexy image earlier today:

[cosmic crown]

thanks for your time.
communication is probably the weakest part of sai. if for the next release you could give auto, vlad, comfy, etc some pre-release models so they can have workflows ready for day one - this would help alot. it would also take fire away from you guys, as there will be a lot of debate about the tools versus each other... 🙂

[mortal kite]

I think I know the basics, I just didn't know how scheduling was being used

[sterile quartz]

so who's making a reddit thread with the screenshots, i have to go to bed and cant

[dusky thistle]

i'm rooting like HELL for you all. whole reason i'm being a pest about training code is cuz that's a chance for me and others to see if we can cast SD3 in a new light 🙂

[cosmic crown]

getting finetuning finally running would be HUGE

[faint breach]

the network isn't trying to denoise the entire image all at once. it does it over steps. if you have 20 steps, the schedulre decides how much noise is removed each step.

Samplers i'm not to sure about though. That's a matter of math.

[dusky thistle]

exactly! most of the issues look like they're related to insufficient training... and we have a LOT of ppl itching to help with that

[faint breach]

here's the video i mentioend. high level explanation of the process. the next one explains it with code. there's a new one in the series too that explains clip https://youtu.be/1CIpzeNxIhU

[hazy kestrel]

https://tenor.com/view/keown-tutoring-math-equation-formula-gif-14954188

[cosmic crown]

yep, sadly most people currently gave up frustrated or didnt even start looking at it yet (either because no tools are there or license). so most work is still going into sdxl. which is fine...

[dusky thistle]

yeah, need the tools for that to really start

i've been doing what i can, but i lack the expertise to be able to reverse engineer how to train this thing, so i'm using others code for now.

crushing this thing with an outrageous LR with adamw right now, LR = 5e-4... finally getting something to happen here

[signal shuttle]

most people are entitled, back when 1.4 was still in beta we made barely cohernt images of 2B nier automata and we liked it🧓

[drowsy phoenix]

the problem with that, after the finetunes pop up, untrained concepts will be still bad, because base model had those issues, after the finetune they may get even worse

[placid swallow]

The past suffer so the future may prosper

[dusky thistle]

i think entitlement isn't necessarily the right angle to take here, and i'm not sure it even matters. the reality is, it's gotta be something ppl can work with for it to be successful, especially trainable... the fact some ppl handle it immaturely means there's a ppl problem, not so much a product problem (or lack thereof)

[drowsy phoenix]

bacause all the finetuners do, either general purpose or nsfw, but I may need niche but simple concepts to work, so it will be still a problem

[dusky thistle]

that's certainly a possibility too... the sooner we find out the better.

in the meantime:

lora off vs lora on:

[mortal kite]

Bro do you even lift?

why does this thing understand obese men so well

[drowsy phoenix]

it looks fair a bit fried but, not sure

[mortal kite]

it's a bit concerning

[faint breach]

because nobody trying to swift gate on obese men

[dusky thistle]

it's definitely fried, the LR is outrageous (5e-4) but i consider it a victory that i've managed to see a substantial change at all

[drowsy phoenix]

yeah I feel you and know how tricky to train properly

[dusky thistle]

never had issues like this with sdxl... if i drop the learning rate down to something reasonable (as in, a lot lower than 1e-4) it basically does nothing

[drowsy phoenix]

no idea about SD3 but i think some optimizers would solve some issues, not sure if they would work on sd3

[dusky thistle]

prodigy/cosine, which is usually really reliable, was so, so weak

[sage burrow]

Which lora?

[dusky thistle]

one i was running a few min ago

[drowsy phoenix]

another idea that could be happening, strict captions in dataset would not associate with the base model cause of lack of various styles/keywords/concepts in the base model, or it is overfit of concepts or style

[compact forge]

Cause we didn't think the model was rigged

[dusky thistle]

what do you mean with the captions? that the captions might be too different from the base model? or too similar?

same training set worked great on sdxl base last night

[fathom pelican]

so is there already a general consensus about which one is better between sdxl and sd3?

[dusky thistle]

looks pretty decent with the cfg brought down to 3

prolly will try with the LR cut 5x in a bit here

[faint breach]

no there realy isn't

i personally think sd3 destroys sdxl , but i have different expectations

[untold valley]

Define better, lol.

[drowsy phoenix]

to make training successfully embed itself into base model, it needs to associate caption keywords / concepts to base model, so base model has to reflect or 'know' those keywords
for example if you train with no caption dataset, you will have no keywords to access those concepts but it will still train with using neighboring concepts
so if base model lacks some base knowledge about the training, you need to use broarder way of captioning to embed those concepts

[faint breach]

If the expectations are porn, no. It doesn't beat. Pony has peaked here so hard that pony is now a synonym.

[fathom pelican]

@untold valley@faint breach hmm I see, I feel like I saw someone commenting about how SD3 produces better landscape/sharper than sdxl

[sage burrow]

Has the internet ever agreed on anything? 😄 I personally see SD3 as more versatile and makes amazing images. I still get better ones with 1.5 or SDXL though since all those models and loras, as well as much more experience with them. Though the SD3 nsfw is kinda terrible, even when it does crop up.

[sage burrow]

Scroll up to a few hours ago, that image of the pirate ship on water!!!!!!!!!!

[mortal kite]

I got a lot of good horror pics out of it this morning

[sage burrow]

It does do exceptionally good horror! 🙂

[fathom pelican]

oh if I can find it

[sage burrow]

The previous SDs were terrrible at horror, imo.

[sage burrow]

#🆕｜sd3 message

[fathom pelican]

quite pretty, I liked the bee drinking water from ice too

[sage burrow]

SD3 is so good at macro "photography" 😄

[sage burrow]

Btw there's an SDXL section on this server as well, full of SDXL images

[faint breach]

the ocean blueberries had great details. the waves reflecting off the blue berry bunch being much tighter than all the waves around them. so nice

#🆕｜sd3 message

[rocky epoch]

aaa

[ancient cape]

horror, as in film like horror, or as in grotesque and macabre?
cause sdxl base+refiner nailed film horror for me

[digital thorn]

Have you seen the grass people in SD3?

[sage burrow]

Both types I had trouble with actually before SD3. My zombies are much happier now 😄

[ancient cape]

sdxl base+refiner only

[sage burrow]

I'd missed that set! 🙂

[placid swallow]

you can get bloody stuff with red viscous liquid but it does bones pretty well

[tranquil nymph]

have you tried the keyword "giantess"

[digital thorn]

Not just blood?

[mortal kite]

I wasn't trying to make her big really just trying to get a steeper angle. Having trouble getting that

[placid swallow]

it tended to look too shiny and not enough volume

[tranquil nymph]

oh i thought you were going for the amazon thing lol

[ancient cape]

same prompt, for SD3

[sage burrow]

I have to have amazing detail and art beneath my ketchup and red paint though 😉

[placid swallow]

reminds me of old gwen stefani

[odd basalt]

Yes around 2000

She does look like Gwen

[placid swallow]

"Aftermath"

[tranquil nymph]

charming

[bitter hearth]

gaross

[silver adder]

holy hell! sd3 2b

[tranquil nymph]

things took a dark turn in here lol

[placid swallow]

SD after dark

[bitter hearth]

for real, I even changed my avatars to horror! rar!

[silver adder]

ok, ill post something more wholesome here (8b left, 2b right)

[tranquil nymph]

It looks like the guy himself has a hazardous materials label as a "badge" lmao

[placid swallow]

I hope they answer why we were given "8b at home..." next AMA

[glad yew]

How good will the 8B model be? Will it be much better than Dall-E 3?

[placid swallow]

imagine 2b with another 6b

[silver adder]

you can see for yourself with the api (pay for it or access it for free on glif). keep in mind it's undertrained there, but so far it looks pretty good imo

[glad yew]

Yes, I tried that model at Glif several weeks ago and I thought it was horrible.
Not only did it not follow any of my prompts, but its aesthetics were not good either.
Did they update the model? To go try it.

[upper snow]

blind comparison tests suggest it easily beating DALLE-3 HD and being competitive with Midjourney V6

[glad yew]

What about its flexibility and dynamism? Obviously it will be more realistic than Dall-E because it was nerfed on purpose

[craggy crest]

you can use it in the #🗣｜artisan-support-feedback channels. read the support channel first. pick ultra for the model

[upper snow]

There's of course the chance of a Playground-2.5 type situation (where the model is overtuned to hell on aesthetic slop to the point where it doesn't follow prompts at all) but that doesn't seem to be quite the case with SD3. Playground-2.5 refuses to make any image that isn't super aesthetic.

[placid swallow]

that's the path image gen models chose, each model is good at something not all things

[upper snow]

it will be in the hands of people who can finetune it (albeit for a price), so... as flexible as someone can afford for it to be?

[silver adder]

i'm assuming so. most of my images are generated there and they come out quite fine with reasonable prompt coherence.

[faint breach]

if the 8b weights have a price tag on even saving them, i'm gonna ignore 8b forever

[silver adder]

i use this specific app though, since it comes with an llm to help you and an upscaler to improve details:

https://glif.app/glifs/clvau2g2p00002aec0inze6r2

[ancient cape]

sdxl lightning - 4 steps, for rough composition. to sd3 for fine composition, and all texture data

now she can sit in grass 🤣

[faint breach]

maybe not forever. maybe i'll eventually find a commercial use for them. then i'll pay.

maybe if there was a one time price for them

[placid swallow]

no way it's not subscription based in todays economy

[faint breach]

the economics of shit

[placid swallow]

"due to on going costs"

I'm hoping by the time uncensored SD4 comes out we'll have house robots

[tranquil nymph]

as long as they don't murder all of us

[upper snow]

What I meant is that h100s are expensive

[faint breach]

i can't just train it with solar power and a prayer? aw

[faint breach]

think the first robot murder case will be a programming glitch that the company tries to pin on machine sentience to avoid liability

[upper snow]

Maybe an A6000 could do it if you are stingy. 2 3090s with FSDP might do it but that'd be so slow

[placid swallow]

well here's hoping the murder robots can't hold weapons either

[placid swallow]

pretty good background faces

[tranquil nymph]

Messing around with prompt travel. doesn't work very well img2img though have to use txt2img and ipadapter (not using sd3 for the txt2vid though)

i want to try it on this lightning picture lol

[tranquil nymph]

turns out ipadapter doesn't understand how to ID pikachu faces or whatever that is

[tranquil nymph]

at least the first guy was kind of hiding

[mortal kite]

"HEY WATCHA DOIN"

[tranquil nymph]

"may i recommend the white coat ma'am"

[mortal kite]

riding time?

[tranquil nymph]

that looks great, what workflow are you using

[tired basalt]

https://tenor.com/view/vince-staples-fax-it-do-be-facts-tho-its-the-truth-gif-16329641

[sacred jewel]

Should be embedded. Let me know if not.

[tranquil nymph]

not loading

[sacred jewel]

This one definitely does...

[tranquil nymph]

hmm where does this eclectic model come from? don't see it on google

[sacred jewel]

Huggingface... they have AWESOME models ...

Also on CivitAI

https://huggingface.co/ND911

[silver sluice]

i tried recreating your image, just using the standard workflow provided by sai

[tranquil nymph]

her legs are getting cramped up inside that clock lol

[mortal kite]

wtf

[tranquil nymph]

i have no idea how this workflow works but very cool results

[tranquil nymph]

there's a different picture of a guy in water with the SD3 model then somehow it turns into the other model and outputs the clock image

[sacred jewel]

Dall-E 3 runs on probably hundreds of H200s and is almost likely CONSTANTLY being re-trained live?

Who knows what the secret sauce is, but, it is definitely better than whatever fits on an RTX- 4090.

[mortal kite]

rofl

[digital thorn]

This is what I got

[sacred jewel]

steal my workflow from here:#🆕｜sd3 message

See if you can figure out what the original author was doing. I just modified it to my needs/tastes but I did not originally set it up from scratch.

[sacred jewel]

That's #awesomesauce

[tranquil nymph]

ah i'm just not familiar with some of the new nodes and connections, thanks. whatever this other model does takes the image to a whole new level

[somber jackal]

A girl, walking on the grass, full body, smiling

[digital thorn]

I got nipples! (on another image)

[mortal kite]

yes but probably male ones. It will just put mens nipples on

[tranquil nymph]

https://tenor.com/view/thats-impossible-thor-impossible-ragnarok-impossible-impossible-not-possible-gif-19896965

[digital thorn]

Nippeeals!

[sacred jewel]

I admit I may have **** a little 😄 😛

[tranquil nymph]

that's one fancy cat

[faint breach]

people hunting for nipples in the latent space like a desperate baby needing to nurse. just inpaint them. you're just making things harder on yourself ||😏||

[digital thorn]

Well, that's too ez

[tranquil nymph]

look it turned one of the circuit boards into a half can of Bang

[sacred jewel]

Or just use Pony or just use pornhub 🤷‍♂️

[tranquil nymph]

Pony has some great anatomy, I thought it was really impressive. the realistic version

[mortal kite]

Pony anatomy is amazing as hell. It knows positions you would not think it would know

[tranquil nymph]

I have no idea how anyone has the patience to train all those models or merge and spend days/weeks/months tuning parameters

[torn wharf]

"anatomy"

[sacred jewel]

Do play golf? Do you play cards? Do you do anything you;re passionate about? That's how they do it... if it interests you, you will spend countless hours on it.

I can spend days in front of a Grafana dashboard and "snmpwalk" looking for that one interesting OID to graph and infer knowledge from ...

[torn wharf]

the thing is, they want to thye like it. you don't have patience to do the things you like?

[thin quarry]

the lengths people go for booba

[mortal kite]

Just use Pony it will fulfill your ever need.....

[faint breach]

i play factorio you want to talk about patience?

pony is peak for anatomy yeah just use that. "anatomy" we know what it means.

[tranquil nymph]

i beat only up version 1, does that count?

[sage burrow]

I'm not entirely sure what the thing on the bottom left is lol

[tranquil nymph]

His vanquished dragon foe?

maybe it's a dragon furball

[digital thorn]

a creature covered in shaggy purple fur is curled up...says chatgpt

[torn wharf]

egg in furr coat

[tranquil nymph]

now i want a drink lol

[sacred jewel]

I remember the first time I saw an image of a nude woman on an Amiga 1000 at 4096 colors... I was floored

[tranquil nymph]

Tomb Raider 2, they were like triangles lol

[faint breach]

i remember a game on sega genesis, you coudl push buttons on player 2 as the logos came up and you'd see a nekkid lady. naughty dog has always been legend

[tranquil nymph]

these animals you're making sure have an expensive wardrobe

[sage burrow]

I guess it does know of Dali afterall

[mortal kite]

god it takes forever to get something decent from this prompt.

[torn wharf]

there are a few categories that are underfit it looks like.

[mortal kite]

finding it impossible to get a good action scene like swinging a sword

a bit better gun on this one

[tranquil nymph]

what the hell is that other unicorn head doing lol

[mortal kite]

cute

[tranquil nymph]

negative prompt: "extra unicorn head yawning"

[mortal kite]

sd3 doesnt seem to listen to negatives honestly

[mortal kite]

it almost seems like they end up making the picture worse when you use them

[faint breach]

closest i could get to action. you can tell somethings happening at least.

[sacred jewel]

Damn you read my mind LOLOOLOL

[patent acorn]

i heard that it only changes the noise

[sacred jewel]

not too hard... first try ... same prompt as my other similar one and just added "...swinging a sword..."

[mortal kite]

latent lottery

[sacred jewel]

Matteo basically said so from his own testing... very specific terms it listens to, others not so much. He also made a node to make the whole negative prompting wrangling simpler.

[mortal kite]

how do i make a gun look like it's firing? I've tried "barrel is exploding" and tried "gun firing" and tried "barrel is emitting an explosion". Nothing is working

[tranquil nymph]

can you reduce the CFG or whatever the equivalent is in Comfy

[mortal kite]

are you using 8b?

[sacred jewel]

Of course... in the KSampler

[tranquil nymph]

Seems like there's too much going on in the picture either lower cfg or denoise or both?

[sacred jewel]

I wish... nope... original is SD3 2B, then take the pixel and re-encode it for a second pass using Eclectic Euphoria SD3 perturbed model.

[tranquil nymph]

or number of steps

[mortal kite]

I give up on this prompt then..

freaking same prompt but with pixel art works. Damnit SD3 you picky ass bitch

[silver adder]

if you want to get (almost) decent looking planes from the side, then you'll have to play the seed lottery (sd3 8b)

[mortal kite]

I got decent planes the other day

[digital thorn]

Not sure why you'd put engine in the front like that

[faint breach]

accurate boeing

[silver adder]

this image reminds me of some sort of fighter jet i forgot the name about that had a jet engine in the front, otherwise it doesn't exist...

[placid swallow]

atleast it has built in fire extinguishers

[digital thorn]

Have we just found out that planes are living things? Since there is anatomy issues

[sacred jewel]

If you share the prompt, I'll see what I can do 😉

[silver adder]

it does the front of planes just fine

[sacred jewel]

Actually, not quite... Proportions are WAAAAYYYYY off... but I see what you mean... better than not. If it's of any consolation, Midjourney used to suck at it badly too. Haven't used it in a while so might be better now.

[silver adder]

sd doesn't have the greatest reputation of generating planes really

especially when you try to prompt for a double-decker. it just craps out like with women lying on grass haha.

[placid swallow]

finish my daily pokemon gens, queue up some living planes next

[silver adder]

here's an mj generated image of a plane landing down. it looks better than sd when it comes to the arrangement of the plane's parts, but still seems to be off. one can see that the wings are a little way back from the middle, along with a garbled front landing gear and a flat nose tip

[placid swallow]

lol it's landing on a snowy runway in the middle of the desert sun?

[silver adder]

haha yeah. it was supposed to be a wintry day for that

[placid swallow]

Goooood Morning Dubai, expect 3"-5" of snow on your commute today....

[tranquil nymph]

Bro that plane behind you is about to crash into a house but you're concerned with your facebook pose

[placid swallow]

he's thinking about that sweet sweet maple syrup he has at home

got an oof right off the bat

[tranquil nymph]

very aerodynamic lmao

[silver adder]

https://tenor.com/view/concorde-gif-704481095228021886

[placid swallow]

got that jay leno chin attached to the landing gear with no way to retract

[tranquil nymph]

also i'm not sure if planes have multiple wings on one side but this one has half a wing

[low stone]

Sd3 8b

[placid swallow]

well if you're into flying in circle i got the ✈️ for you

[tranquil nymph]

what is this a Hindenburg Pikachu fushion

[silver adder]

pov: you're a whistleblower

[placid swallow]

and why is already on fire

[tranquil nymph]

omg that face is great

you should make a comic strip with those man/plane pics

[placid swallow]

🫐 are back

[tranquil nymph]

can you make the plane shoot a missile at the guy lol

[silver adder]

ha, i'll try

[placid swallow]

dang got a real plane on like my 6th seed, upscaling now

[tranquil nymph]

beautiful

[silver adder]

(the missiles are heat-seeking)

[low stone]

You got better than I did

[silver adder]

probably because i used the 8b api for that, forgot to tell

[placid swallow]

dang that's what the luggage ppl have to deal with after we take off?