🆕｜sd3

[bitter hearth]

https://media.discordapp.net/attachments/733762994525306972/899670992023076926/image0.gif

[sterile pendant]

What tolerance ranges are you using? That's what truly determines the quality with that ode sampler node. Like there's a big difference between 3/4 and 4/5, or 4/5 and 5/6. Literally 10x the precision difference since it's a whole extra decimal point. But it will make compute times go up by a ton

[torn wharf]

yeah i determined the problem is because that sampler it's only doing 1 step and calling it done. i don't have the workflow set up correctly somehow and just gave up. i was trying all sorts of ranges while playing

[sterile pendant]

3/4 or 4/5 work well, any higher and it's a waste of electricity usually. For the adaptive ones, just set the step count to like 200 so it doesn't abort early, it acts kind of like a jeopardy music timer

Oh and bosh3 is the only one that's really worth it for SD3

[mortal mesa]

think anyone else saw the face

[low stone]

I'm finding auraflow to be more prompt following than sd3 2b at this point. Sometimes I get more elements on sd3 but It feels sdxl'ish to be combining subject details on sd3 like that. The prompt: A chubby, bearded man with a shocked expression looks up at the ceiling, his mouth wide open. Colorful, fluffy creatures with angry faces pour out of a hole in the ceiling above him. The man's face is wrinkled and red, with sweat beading on his forehead. Bright, twinkling Christmas lights swirl around the room, casting a magical glow. A huge robot reindeer towers over a tiny Santa Claus, who gazes up at it in wonder. Snowflakes dance in the air, creating a whimsical atmosphere. The image is incredibly detailed, like a high-quality photograph taken with an expensive camera.

[bitter hearth]

maybe was it adetailer or just a random gen lmao

[mortal mesa]

like 20th gen, i do 4 at a time, its Bill Cosby if its not clear enough

[torn wharf]

yay christmas morning! i'm prompting for nintendo 64s and i'm getting all these jank looking super nintendos though.

[torn wharf]

you know, sometimes i think sd3 has no idea what an n64 is

me thinks lykon was a sega kid and is keeping the ancient console wars alive. kind of sus

yeaah has a pretty good idea of what a sega draeamcast is but no idea what the n64 is. right okay. that's not sus at all.

[junior dune]

Is there a way to train a lora with sd3 as the base model?

[torn wharf]

onetrainer sd3 dev branch, or kohya scripts sd3 dev branch

THIS IS PRETTY SUS THO. N64 DESERVES GREATER REPRESENTATION

i didn't prompt no gamecube but thats what its tryin to do clearly. that aint no n64. this safety training is bohshit

closest i could get but they erased the nintendonger

[deft whale]

/help

[torn wharf]

/balls

[bitter hearth]

[torn wharf]

heh

[craggy crest]

why does this remind me of a game controller?

[bitter hearth]

[mortal mesa]

unsettling

[edgy kelp]

I really wonder why SAI didn't even try to blur just a little the Midjourney-based dataset images, the artifacts in SD3 medium are so annoying...

[tidal spire]

Dress

Dress

Change dress

[bitter hearth]

Changing dress, please wait 2 weeks for it to be done.

[tidal spire]

Different poses

[bitter hearth]

the image I made before was heavily stylized so its personal taste really

here is an image that is just photographic with no stylization:

that is at 300 steps of order 4 implicit adams

for comparison here is the same seed with 20 steps of order 1 euler

but there is much more potential in samplers
Dr Head's node is awesome but
the implicit adams is only order 4
and it is fixed step not variable

higher order variable step implicit solvers would be a big quality increase over this

explicit solvers (euler, heun, bosh, RK4, fehlberg, dopri etc) won't work well with stable diffusion at any order higher than 2, or maybe 3 at best

due to stiffness

but implicit solvers allow us to push the order higher

🙂

[cold osprey]

Please who can help me put this in a space background and put on the monkey with a space helmet

[hollow apex]

Any good SD3 fine-tunes already out?

[bitter hearth]

yeah there is one that makes balls

its awesome

its only lora though

no big checkpoints yet

[icy drift]

Any news on when Civitai will allow SD3 again? Since we got the updated licence.

[lusty oyster]

how to reboot olama?

After closing

[bitter hearth]

yeah on another disord, civit staff said that they are going to read the new contract this week

[sacred jewel]

https://tenor.com/view/spit-take-laughing-cracking-up-lol-rofl-gif-20988052

[bitter hearth]

IDK what civit are doing

it never made any sense to me

I think I misunderstood what civit actually is

I thought it was just a model hosting site but maybe they are trying to be something else

like an open source software foundation or something

[heady wolf]

I read that they've been in communication with SAI about a commercial license. My understanding is that once you make over $1mil, then you work with SAI to make a new bespoke license agreement. That's probably the "new contract" civit staff is reading, not the one that was released to the public weeks ago 😛

[silk solstice]

Just love how colorful and fun this is

[sterile pendant]

I like the cotton candy cloud on the right

[silk solstice]

Me too!

[bitter hearth]

SD3 colours are so good

[bitter hearth]

[torn wharf]

Lol no. They're a venture capitalist start up that hasn't cracked profitability yet. Model hosting is just their hook

They're not a non profit like the FSF

[orchid crypt]

so, sd3 maked from screen shots? this is ONE image generated with clear sd3 medium...

[craggy crest]

yeah. it's a case by case basis with the contract worked out to be what you actually need for your business, not just a cookie cutter that everyone tries to fit into

[bitter hearth]

Balls

Good morning

[marble plinth]

hey guys, any hints or tips if i would like to create img2img in the style of marco grassi? I'ver trained a model with several pics of his work, and i would like to transfer that style to some pics of my family. Tried ip adapter and t2i style, but not succesfull... any tip is highly appreciated. Im using A1111 btw!

[torn wharf]

good morning ballshine

[craggy crest]

that would be the Phoenix sun

[torn wharf]

https://youtu.be/Rvvsw21PgIk?t=13

[tough viper]

how come civitai isn't hosting sd3 models?

[torn wharf]

they're banned until civit can negotiate commercial terms with stability

[desert garnet]

skill issue

[torn wharf]

They said it was because they didn't want creators to suffer and stability could require all models be destroyed. So civit deleted everything that had been uploaded and indefinitely banned sd3 content to protect creators.

|| it was really about protecting their business interests and they were only spinning it as protecting the community ||

[craggy crest]

they deliberately - like most other people - read the license wrong. they want to create their own fondation model anyway

[sage burrow]

Caus they're cheap

So I've been merging models for fun and experimentation. If you add SD3 into the mix at all, there is always missing || nipples || and 🍆 lolol
In some cases the images are turning out better, and in other cases they are worse. I guess some older checkpoints merge better with SD3 than others.

[foggy cloak]

As they should, why wouldn’t they protect their business interests if it also protects the community

Civitai has made some terrible decisions lately but that one seems like a win-win tbh

[craggy crest]

SD3 image generator, to SDXL face refiner, to Reactor Faceswap workflow

[craggy crest]

it's the fact that they lied about what they're really doing, mostly.

[desert garnet]

everyone lies even the napster ceo lied

[bitter hearth]

Balls

[desert garnet]

https://tenor.com/view/low-res-emoji-gif-15431116145489883202

[torn wharf]

How are people not sharing for commercial reasons, helped out?

Civit have never demonstrated a capacity of dealing with issues maturely. This is more of the same. Nobody was at threat of needing to destroy models and a judge would've never held that up if it was even attempted at being enforced. The whole reasoning is a joke. They're likely using it more as a negotiation chip than any kind of good will.

Ulterior motives always

Telling the community they're banning all models until stability gives them favorable commercial terms doesn't communicate that well though. Trust that businesses are never honest. Especially ones that are still in the red and spending money to establish themselves.

[sage burrow]

https://pitchbook.com/profiles/company/530262-28#funding

[craggy crest]

not all businesses - just the ones that have a track record of their CEO ....ing over his friends and business partners

[desert garnet]

wasnt that what happened with facebook/napster

[torn wharf]

napster they got hit with the riaa. after they were acquired by microsoft, they actually did a cool spotify style service which was way ahead of the times

my bad wasn't microsoft that got them. roxio. but they did a streaming subscription thing it was handy dandy

at facebook i think when they restructured the company they fucked over joseph gordon levitt and he didn't play his cards right. cut throat but yeah thats how its played. schmiddy from Google fame, isn't exactly a 'nice' guy either.

[torn wharf]

wait mb. it was another guy who played that character in the social network. the amazing spider guy

[raven fern]

im not sure, but i think people are waiting for the improved version to come out before training

[bitter hearth]

2 weeks

[sterile pendant]

Oh god, not this again lol

[craggy crest]

just feed her catnip, she'll be fine

[bitter hearth]

there's too much water in this planet

[torn wharf]

death to the water! its time humanity take back the earth

[craggy crest]

@naive sparrow

[sacred jewel]

https://tenor.com/view/tiktok-dog-awkward-dog-confused-awkward-confused-dog-gif-2881577801438605193

[torn wharf]

https://tenor.com/view/i-dont-believe-you-ron-burgundy-will-ferrell-anchorman-youre-lying-gif-15177693

[bitter hearth]

might go back to SDXL for a bit to get the stronger version of CFG++

not sure if it will be possible to port it fully to SD3

at the moment only "alternate" mode works in the comfy node for SD3, which lets you half your CFG

but the original full version lets you set your CFG to like 0.6

[hollow apex]

guess that would make sense

lets hope its slightly less censored to shit

[bitter hearth]

hmm never mind the full CFG++ at 0.6 CFG wasn't that much better lol

[rain current]

[bitter hearth]

[verbal epoch]

Where do I find trained sd3 models

[vast condor]

The base model is on hugginface of course, if you mean finetuned by the community, civit is still not hosting yet to my knowledge, and the early ones weren't really that mind blowing anyway. Gotta be patient

[craggy crest]

you don't, they don't exist yet

[mortal mesa]

Failed to fetch response from Ollama.

[supple gulch]

How about sd3.1? What's schedule of realese?

[mild bramble]

Is that ai?

[mortal mesa]

ya was error message, it failed to load my LLM 😆

[edgy kelp]

Man, I lost so many Balls (TM) while I was offline

[edgy kelp]

Biggest BALL

[mortal mesa]

uh oh, SD3 just popped out topless with appropriate bits unprompted, what do i do, email it to trust and safety

[edgy kelp]

Oh man, you'll have to destroy that model for it has been disentangled by your PC

SAI will persecute you

Or prosecute, whatever

[mortal mesa]

maybe both

[bitter hearth]

2 weeks I think

[edgy kelp]

Man, in two weeks I'll be billionaire

[sage burrow]

What is something that is SFW that SD3 does NOT know, that Pony and/or SDXL knows? I'm testing my model merges.

[tough oriole]

Yall having luck with aesthetics on SD3? Its learning my subjects but it doesnt look good.

[supple gulch]

Hopefully it's a 4B model as emotional compensation

[mortal mesa]

the ole girl grass meme

[bitter hearth]

SD3 has easily the best aesthetics of any model out there
the problem is structure (and also subject knowledge)

the 16-channel VAE is currently unmatched

I think midjourney is still the best model in the world overall because it has good structure and aesthetics and also good subject knowledge (often training in a legal grey area on hollywood movies)

but SD3 can produce better quality because of VAE

[torn wharf]

unets feel so old and sloppy since stability dropped mmdit on us.

looking forward to the future

[mortal mesa]

i don't use them but from what ive seen on a different server Ideogram looks better

[bitter hearth]

Ideogram is also very strong

[torn wharf]

lots seem to think that if we just get a 16 channel vae into sdxl or sd15, then it'll all be good. but those are still unet models that barely understand text comprehension

[bitter hearth]

I actually think Kolors looks very good a lot of the time, but its not a generalist model so its not actually that useable

[torn wharf]

does ideogram run locally yet? i've all but forgotten about it

[bitter hearth]

no its like midjourney sadly

[torn wharf]

yeah it's entirely irrelevant to me if i cant use the weights. the architecture is meaningless

[bitter hearth]

yeah its sad

[mortal mesa]

just another diffusion model you cant really play with

[bitter hearth]

Lykon said that Unet won't even scale to 8B
so the big model wouldn't have been possible without the DiTs

[mortal mesa]

looks nice though

[torn wharf]

kolors is a unet . i like what they're doing there, but they're really late to the game to try to get a new unet model started 2 years after stable diffusion 1.

[bitter hearth]

the sad part is Auraflow didn't get the 16 channel VAE

[torn wharf]

auraflow is v0.1 and the current release is just to kickstart community engagement. i doubt they'll keep it to a 16gb requirement too. likely will release smaller versions and then versions with different finalization steps. vae or otherwise

i haven't even bothered to load auraflow because it's a 16gb file and i've only got 16gb vram

[sage burrow]

Comfyui model merging with SD3 plus anything else, just basically leave you wish SD3 only. There was a very very slight colour vibrancy increase when I merged it with Juggernaut, but that's about it.

Now I wonder if my SDXL plus SD1.5 merges actualy worked or not 😦

[junior peak]

I'm getting somewhat nice results using big tiger gemma 27B as a prompt enhancer

somehow it seems to make these pics a bit better

[coral sable]

Oh rly good, is that Ultra? wats the prompt

any ETA for SD3 2B refreshed version?

[edgy kelp]

Reportedly: "When donkeys will fly"

If it was actually meant to be a beta version, they would have kept on training that model even before the release

[mortal mesa]

i await the surprise 4B release

[dry wave]

this is so nonsense

[edgy kelp]

I await Stable Cascade 2.0 LOL

[dry wave]

I have the feeling I had this discussion here so many times and it's useless to repeat it over and over

but the unet architecture is not anyway worse than a dit architecture

just because it uses a unet does not mean it is not a transformer

same way all dit architectures use some convolutional operations under the hood, too

[dull star]

If only they didn't can it

[dry wave]

it's not a question about unet or dit or mdit

it's often rather the data and data annotation that makes a difference. And also the stronger vae that allows for more details

[bitter hearth]

SD3m with claude 3.5 prompter

[torn wharf]

not quite true. a DiT architecture is just better suited to the task. it was engineered for the purpose. Unet was just what was available back then and they prototyped on it.

[dry wave]

and a better text encoder for sure

[torn wharf]

if you're going to open with "this is nonsense" then we'll never have an honest discussion on the matter. i give up immediately

[cunning lintel]

lucky you! https://jingjingrenabc.github.io/ultrapixel/ thought think of it as csascade++

[torn wharf]

before you reply, just know that you win. pat yourself on the back

[dry wave]

dit is just a transformer architecture on image patches, what's the difference to the unet?

[torn wharf]

wurstchen 4.0

[edgy kelp]

Ah right, I forgot about this... wonder when they actually release the thing tho

[dry wave]

the only significant difference is that unets are not using positional embeddings. You could add them, though

[torn wharf]

you already won chill

[cunning lintel]

it's released already

[edgy kelp]

Woah 2 days ago

[dry wave]

there was even a paper comparing different architectures and found the sdxl unet architecture outperforming the pixart dit architecture on the same data trained

[torn wharf]

tbf pixart isn't that great and doesn't have the same structure as an mmdit

[dry wave]

which also doesn't necessarily mean unet is better.
unet is the more complicated architecture. PixArt with it's dit architecture showed you can reach same results with simpler architecture

[sterile pendant]

Did anyone else peer review it to verify the claim and ensure there wasn't cherrypicking or biased?

[dry wave]

peer review? in ml ? 😅

mmdit is something else, yes

[sterile pendant]

Exactly... Claims are cheap and it's super easy to bias and cherry pick

[torn wharf]

unet has been worked with longer. more complicated means less easy to improve upon. simpler is usually a better design in many cases. and it shows in colors wiht sd3

[dry wave]

I don't really believe that mmdit is parameter efficient. But that's my subjective opinion. There is no good evaluation yet

[edgy kelp]

In no way a pear reviewed this paper

[torn wharf]

there's lots of results coming out. auraflow for instance. there might not be any great finalized review yet, but there are MANY indications

[desert garnet]

we need a pear to review this math function to see if it works

[dry wave]

the paper evaluated several architectures, not their own one. So there is no reason to cherry pick

[torn wharf]

why am i bothering though. we've already agreed at the beginning that you won

[edgy kelp]

Pear reviewed fruit juice (made from pears, for pears, reviewed by pears)

[desert garnet]

but is the math real ?

[dry wave]

yeah, future will show. The architecture of mmdit is unnecessary complicated in my opinion. You are right that simple is usually better. So I would go for the simple dit architecture of PixArt and improve on that

[torn wharf]

we've seen that recently. tyson publicly peer reviewing 1x1 = 2 or square root of 2 is 1. was it necessary? maybe not widely. terrance needed to hear it tho

[edgy kelp]

Math was done by pears, so hell yea brother

Reportedly 1 pear + 1 pear equals 2 pears (but this is not 100% pear reviewed yet)

[torn wharf]

auraflow simplified stability's implementation of mmdit. it has a lot of potential and i'm sure it has caught stabilty's attention

[mortal mesa]

you can destroy numbers by taking your numbers individually and multiplying (1 x 1 = 1) till you are out of numbers

[torn wharf]

depends on how you define the unit of a pear. is a 100gram pear worth 1 ? while a 110gram pear worth 1.1 ?

further. the spelling of pear always fucks with me. peer? pear? pare?

[edgy kelp]

That's why to some degree 1 pear + 1 apple might or might not equal to 2 pears

[torn wharf]

need to pull out the calculus to get statistical propabilitys of that apple possibly being a pear

[edgy kelp]

I'd call it Pearbability

[desert garnet]

schrodinger pear

[torn wharf]

sounds naughty

[edgy kelp]

Sorry brothers, I'm impeared

Also, let's talk about the fact that you can impeach a president and not impear them

[torn wharf]

i saw that debate and there was some impearment

[edgy kelp]

I can't stand this, it's unpearable

[torn wharf]

well i saw the highlights. i'm canadian so i don't really pear

[edgy kelp]

I'm italian, I pear even less

[torn wharf]

yeah pearing near the border of the pearnited states has me pearing a little bit

[edgy kelp]

appearently

[torn wharf]

i pear for the economy sometimes. it's already all peared up. now our largest pearing partners are peared

imagine if over on your side of the pear, pearis started the revolution up again. one might pear that could mess your pearconomy up too

[edgy kelp]

Y'all gotta fuel them cars with spear pear juice bought from Pearmany in the Pearope

I reckon Pearmany is a leading expearter of pears

[torn wharf]

theres always pearaguay but that might be less available in the eu

[edgy kelp]

I think we should stop if we don't want to be peared... uh banned

[torn wharf]

i'm running out of bad pear puns anyways

[edgy kelp]

Hahahah

[torn wharf]

winners imo are impeared, appearently, pearis and pearaguay. thats my pear review

[dry wave]

new models always look better. I doubt, though, that it has much to do with the model architecture itself. SD 1.5 was trained on LAION and when you look at the data it's no surprise the model looks so bad. The most efficient way to train models nowadays is to train on high quality artificial Midjourney, Dall-E or Ideogram data. Also auraflow was trained in ideogram data as crazy. The availability of more and more high quality synthetic data is the driver for better open models

I would wish there are more evaluations where people really test different architectures on exactly the same data to see which method works best

but such evaluations are rare. Ni surprise, they are basically burnt money

I think we can agree that you need something more powerful than clip to get better prompt understanding. But do you really need mmdit? Did ever someone compared mmdit against cross validation and found it superior?

[bitter hearth]

pear pressure

[pseudo owl]

dit models are pretty nice but whats more important is probably better training data, better text encoders, and better vaes

[torn wharf]

i'm not sure why that would be more objective of a measure. dufferent architectures might prefer different data. captioning may work differently on models with t5 instead of clip vit . there are plenty of considerations. ultimatley i think testing the best product of one against the best product of another is how to contend them. and even then, objectivity will be hard to maintain

[torn wharf]

it's kind of a whole system. the architecture it's all built upon certainly matters. like an ICE built around carberateurs, fuel injection, rotary, diesal or whatever. the steal it's all built with matters too. the driver is especially important. many factors working together.

Kolors is a good showcase of SDXL with better data. there's an improvement but nothing that breaks it out of it's mold

soon we'll get better vaes adapted to work with sd15 and sdxl. i don't think we'll break the mold there either. it'll still mostly be the same

[pseudo owl]

mmdit seems better for prompt following but worse for image quality i guess?
next-dit(from lumina) seems to do be a middle ground
plain dit seems to be somewhat a middle ground too but slightly worse img quality
unet seems to be also a middle ground but slightly worse prompt following

[torn wharf]

Ella adapters certainly haven't taken off because they actually don't offer much improvement

[dry wave]

I don't think it has much to do with the architecture. prompt following comes from a good text encoder and a good captioning if your data

[torn wharf]

maybe well need to wait for some models to show up refined with ella adapters

[pseudo owl]

only sd15 has 1 ella adapter and that one was kinda ok but nothing too good

it definitely was a lot better then sd15's original prompt following but not much better then sdxl's

[dry wave]

there is no much technical reason why a dit should be better in prompt following..From what? I mean, nobody really understands at all how these transformers work but I wouldn't overinterpret anything here

ella adapters are just adapters after all. They fix the text encoder, but if your model itself does not have great understanding it's hard to fix that afterwards

newer models are all trained on synthetic captions. That together with better text encoder gives them the better prompt understanding

[pseudo owl]

yeah i guess i was kinda wrong on that, its probably more with data but auraflow is really good at prompt following(maybe best?) and text rendering. However, it only has a pretty small text encoder and only 1(t5 xl). Maybe its just better dataset

[dry wave]

train sdxl on synthetic captions and t5 and you will probably get a model as good

you only need one text encoder

I don't know if mmdit has better text understanding. Could be.

[pseudo owl]

Oh yeah a image i got from a nice lora called 'better lora' from sdxl, improves text rendering a lot

so good text rendering might not be architecture specific

[dry wave]

but shrek is a known word

if you use words that are not known you will get in trouble with clip

I mean it works to a certain extent

but t5 seems to be really the better choice when you want text and prompt understanding

[torn wharf]

you're not considering the MM in MMDiT. multi modal. it has blocks exclusively for text comprehension

[dry wave]

anyways, I just think we shouldn't overemphasize some of the architectural differences. dit or unet - the difference between both is not that huge

the multimodal in mmdit is a joke

there is not much multimodal in there

[torn wharf]

That's a big technical reason why the architecture is more suited towards prompting images

[pseudo owl]

not too bad
prompt: a image of a man holding a sign saying 'xfnk',

[torn wharf]

oh.... so the technical paper is just kidding. got it

[dry wave]

they project text and image patched into the same latent space and apply self attention on them.
Before that they projected them into different latent spaces and connected them via cross attention

[dry wave]

it's a company. They want to advertise. Don't take everything they claim serious

I would say the biggest difference between the dit and mmdit architecture is that the text is processed and interveiled with the timesteps, too, while in the dit and unet architecture the text was frozen

[torn wharf]

seems like ad hominem and not a real reason. You said there's nothing different that would imply better text comprehension. now when i point out direct evidence for it, you insist that they just made that up in an academic research paper, for marketing purposes

[dry wave]

this makes the model much more computational expensive, but it could allow the model to adapt/connect the text understanding to the image it processes

[torn wharf]

https://arxiv.org/pdf/2403.03206 more people need to read the paper before they talk

[dry wave]

🤦‍♂️

they call the model multimodal but it's just a model with two domains: text and image

SD1.5 is also using two domains, text and image

so why is sd 1.5 not a multimodal model?

[torn wharf]

no one said it isn't...

[dry wave]

because they named it so. It's a marketing thing. You need to name your new architecture somehow

[flint sapphire]

Someone speaks Spanish

[torn wharf]

it's named such because the transformer blocks themselves include two networks for different modalities.

it doesn't mean no other text to image model is multimodal

[dry wave]

calling it multimodal dit sounds for sure better than calling it Würstchen

[torn wharf]

wurstchen is a unet

[dry wave]

😬

[devout schooner]

I've been getting great results with Florence-2 captions and SDXL recently for Loras. I find they're far more "human aligned" than the ridiculous thesaurusmaxxed GPT-4 descriptions of images.

[dry wave]

yeah, and it's super fast

[torn wharf]

you've had a track record of calling factual information "a joke" or "wrong" when it doesn't suit your understanding, so i'm going to name this face Mr Kreuger.

[dry wave]

oh yeah, I'm so sry that I called it a joke when I should rather wrote "it's a name they used for marketing reasons, because obviously all other models are also multimodal"

good that I added an explanation afterwards

[torn wharf]

i dont see it on marketing material though. it just seems like what they named the architecture. it's a bad hot take. you also shouldn't appologize where you don't actually mean anything by it. it dilutes the potential for sincerity

https://stability.ai/news/stability-ai-secures-significant-new-investment their latest marketing video for sd3 is here. nothing about mmdit or multi modal. i can't find it anywhere in their marketing material. it's only regarding the individual transformer blocks of the architecture. Sort of like Unet is named that way because it's sort of shaped like a U.

the marketing reasoning just seems made up and fraught with bias

[dry wave]

right, I'm not sorry. I think I just cannot discuss with you because you seem to just react on single trigger words without trying to understand what I wrote or even put them into context. Same way you seem to just post trigger words yourself. Like you write mmdit is multimodal, so that is a proof that it's better in text understanding. No, this is not a proof. It's just a word (and you seem to put just way too much meaning into single words). I talked about what in my opinion is the difference between mmdit and dit and why that could mean it has a better text understanding but there is lack of any evaluation on it.

As said, I try to add context and explain my opinion as good as possible but it seems that we both just communicate on very different ways

so let's just stop that, it leads nowhere

[torn wharf]

" Like you write mmdit is multimodal, so that is a proof that it's better in text understanding." I believe the reasoning was elaborated much differently than that

hyperbole is better suited to poetry and literature. i never said anything of proof. i said words like "evidence", "seems to imply", and "indicate". Where i questioned you was your insistent that there is zero technical reasoning for thinking it could be better.

#💬｜general-chat message here's the last time i used the word proof on this server and i was wrong then. so i am reserved with such language now. being wrong is a great opportunity to learn.

[sterile pendant]

Damn, you guys are still going hard at it

[torn wharf]

i wouldn't say going hard. i'm not sure what his goal with all the misinformation is though. legitimate research and a legitimate architecture being boiled down to a marketing stunt , honestly has me confounded. Im' not sure how anyone seeking the truth could arrive at that conclusion without having ulterior motives.

there's a reason why the researchers behind auraflow are using the mmdit approach as well. It's a strong architecture. Calling it a marketing stunt is like calling master view controller programming a marketing stunt.

[bitter hearth]

I found florence2 as an amazing automatic prompter (needs to be given an image)

gemma 27B is probably amazing for this yeah

its gets tiring writing image prompts, I quite like LLMs for this task

ironically you could get a second T5 to write the prompt for the SD3 T5

[craggy crest]

what makes you think he has a goal beyond just talking?

[dry wave]

actually that's a big disadvantage of many models:/ they are so trained on synthetic captions that they cannot really deal anymore with short human captions. It's a bit weird that you need to feed your prompt into an llm to get better outcome 😬

[bitter hearth]

something that didn't get mentioned that I am pretty excited about is the rectifying flow part of SD3

it makes the paths more straight

when the paths get wiggly we need to either use the slow solvers or lose accuracy

and in the SD3 paper it said that the bigger the SD3 models have even straighter paths

if you look at how bad euler images are compared to DPM++ 2m images
that difference is entirely down to wiggly paths

if the paths were perfectly straight then an euler image and a DPM++ 2m image would look the same

it would also only take one step!

we will never get to that level but rectifying flow is a huge step in that direction, because it directly trains on straight path objective

[dry wave]

I heard from other people that rectified flow is more vulnerable to overfitting, though. It memoizes images faster. Dunno how much this is true, though

sd3 despite all it's failures sometimes produces extremely sharp and detailed images. I would give the better vae the credits, but could also be the used diffusion method, who knows

[cunning lintel]

i thought rectified flow lowers cfg effect and thus negative prompting hardly has effect

[bitter hearth]

I would also give the vae more credit than rectifying flow, but I think it also helps

[mortal mesa]

i dunno what all this mumbo mumbo is but i generated a cat with neg prompt only

[torn wharf]

I've heard this a few times but I get good results when using negative prompts. Tested with colors.

I think the meta negative habits like "bad hands" don't work as they do in refined models, so people are reverse reasoning their way around that

[pseudo owl]

not a good idea, t5 models are way worse then modern llms and multimodels

unless you train it a LOT

[bitter hearth]

yeah BERT-likes and T5 have to be fine-tuned for each specific task, I agree

they are very parameter efficient once you have a good fine tune locked in, that is not too underfit or overfit

but I wouldn't want to do zero shot or few shot on them

[torn wharf]

There's long clip that I want to see some research done with. Same old clip models with larger context iirc

[bitter hearth]

would be cool to see more done about improving clip models yeah

if I remember rightly long clip only benchmarked slightly better on most tasks

but on one task it was much better

[torn wharf]

This guy https://github.com/beichenzbc/Long-CLIP

[bitter hearth]

yeah I think I saw it on reddit

I might be getting it confused with a different clip

I had a go with SAG, PAG and Free-u today in SDXL

but it doesn't help that much for my type of image I think

because I do very high steps but very low CFG

Align Your Steps or Karras scheduler also doesn't help as much if your steps are very high

its a bit personal taste though, I could definitely see PAG/SAG making changes I just wasn't sure if they were an improvement or just different

SAG did definitely improve concept bleeding

[torn wharf]

Yeah the concept bleeding. So much of it on a unet since the cross attention is so limited

[bitter hearth]

its so bad yeah

[torn wharf]

I'll have to play with sag

[bitter hearth]

going back to SDXL after SD3

concept bleeding everywhere

[torn wharf]

Even cascade handles it better than sdxl but its there

[bitter hearth]

I like to prompt colours a lot and it keeps adding the wall colour to their skin

[torn wharf]

Using an actual color word is very heavy. I almost rely on regional prompting when I want to direct color

[bitter hearth]

I never liked cascade
my opinion is a bit unpopular there
but I don't like the side effects of the compressed latent

it affects the structure

regional prompting sounds good I haven't tried that yet

[torn wharf]

Yeah. I mean ... I don't use it often.. but its there to consider. There's a few problems with it but I cheer anyone finding use from it on

[bitter hearth]

clownshark gets the best images out of anyone I have seen and he uses it
so I gotta respect that

[torn wharf]

Have you toyed with omost? A phi-3 directed regional prompting system for sdxl

[bitter hearth]

not yet, aside from the demo, but that looks amazing yeah

[torn wharf]

Thought it used an LLM but from what I learned I guess phi3 isn't an llm

[bitter hearth]

I was saying on the Rundiffusion server that attention maps is an area that hasn't been explored enough

there was a paper about attention map injections where they found a lot of potential improvements to explore

its hard because you lose image quality from doing this

a bit like how we lose image quality with CFG

[bitter hearth]

wow really good text

it got the fancy font

[torn wharf]

Worrcerestffs sauce. Spelling checks out

[mortal mesa]

whats this here sauce!

[bitter hearth]

my new project is to forget about fancy solvers and just work on clownshark-style noise injection

[pseudo owl]

it is

[bitter hearth]

and just solve with DPM++ 2M because its good enough mostly

this noise injection video is amazing its a huge detail boost
https://www.youtube.com/watch?v=59-3RZknRgk

you do lose some convergence speed though

[torn wharf]

all things i've read about it is that it is just a regular language model and isn't trained on a large language set

[bitter hearth]

some people call Phi 3 a SLM
for small language model

but I think its fine to call it LLM

I call BERT an LLM sometimes

[pseudo owl]

llm is just a name, it means large language model but even models sub 1 billion parameters are called llms. Anything that basically generates text in an autoregressive manner can be called an llm honestly.

phi3 can be very much called an llm since its trained on a very large dataset and can be considered "large"

[bitter hearth]

I think people just wanted to differentiate it from like 70B+ models

but it gets quite unclear

[torn wharf]

yeah LLM seems to be the generic term now. the vernacular landscape is still being developed. microsoft calls it an slm

[sterile pendant]

Oh I agree, mmdit is the better way forward. It will scale better as models get larger. Unet is still good, but it's a very old architecture by comparison. Wanna say it was originally created for the biomedical world for aiding in CT or MRI scans. In another year or two, wouldn't surprise me to see mamba based diffusers or something along the lines. The field changes a lot, but doesn't immediately make prior tech obsolete, they just typically end up being less efficient. Mmdit technically still has a lot of wiggle room for efficiency as well, as the aura creator brings up with his model

[bitter hearth]

there's a prototype mamba diffusion model paper already out, its really cool 🙂

they didn't replace all the attention layers, just some of them though

[edgy kelp]

When pear-based diffusion models?

[torn wharf]

i just think unet models are dusty and are not the way forward. while still fun to play with too, its just i can feel how dated they are now.

I was driving a car from 1986 until a couple years ago. carburetor go burrrrrr but i tell you, it was a piece of shit car. not a good car no no no. its a piece of sh carrr 🎶

sold it to a guy who used it in a total destruction back yard derby for $100

had it's uses but, old and dusty and not modern

[sterile pendant]

funny you brought that up, i actually almost used that as an example lol

[torn wharf]

old cars are fun af. you get a model that doesn't have anti lock breaks ? sooo much fun in snow

[bitter hearth]

appearently not going to be a thing sadly

[torn wharf]

or just the rain or whatever. hard to get all sideways with ABS

[edgy kelp]

I hope at least for a balls-based diffusion model

[torn wharf]

https://tenor.com/view/this-week-got-me-like-struggle-swim-ball-pit-gif-15956297

[craggy crest]

have never, ever, had that problem - with any models.

[sterile pendant]

i used to be a dumbass that was into drifting... yeah, i know it all too well lol... but yeah, the whole field is a giant min/max game. reminds me of a rover i had to design ages ago where i ran into the engineering hell loop of min/maxing. it started with a goal of a rough frame size, which had to be made out of welded steel, but then based on that, it has a weight, so you need motors that can move the weight, which means you need a battery of a certain size to operate for X amount of time. so then you find that the frame isn't sturdy enough for the 25lb battery, so you need a sturdier frame, which is heavier, which means you need bigger motors, which means you need a bigger battery and so on lol. models are a lot like that as well where you're juggling vram size, parameters, training time, inference time, perplexity and so on

[edgy kelp]

The question is: when will Stable Diffusion models learn to output face expression more complex than 'smiling' or 'sad'? Is it too advanced maths right now? Do we have the technology?

[sterile pendant]

make better datasets

they only comprehend what they are captioned with

[edgy kelp]

I think SAI developers were themselves not trained with a dataset that lets them understand this

[pseudo owl]

its an issue with sd3 and auraflow(if you dont add any detail).
its more towards humans

[edgy kelp]

Current SAI datasets 'the picture shows a person, but now let's talk about how many leaves are there in that particular tree over here'

[bitter hearth]

vision models are not that great yet

[sterile pendant]

high quality captioners are slow, they rely on things like cogvlm that are fast as hell, but definitely not perfect

florence2 is a reaaaaaaaaaaaaaaaaaly good one though

[bitter hearth]

florence2 prompts better than I do LOL

[craggy crest]

i don't think the model is the reason you can't get other expressions

[bitter hearth]

ok but at lower CFG

[sterile pendant]

florence2 large chugs out "more detailed captions" in like 1 second per image on my pc with an old 2080 in it and it's accurate a good 95% of the time

[edgy kelp]

'angry' is a basic face expression that was handled even by SD1.4

[craggy crest]

okay so sad, angry, smiling, what other expression are you wanting? resting 'b' face?

[pseudo owl]

cogvlm is not fast at all, its like 17b parameters. Unless they built some optimized library which they could have. Cogvlm produces very detailed and high quality outputs but its a bit too much unneccesary detail.

Current sota multimodel is internvl, thats comparable to gpt4o in vision.

[edgy kelp]

try winking, smirking, pensive, disgusted... to say some

[bitter hearth]

there are some concepts where the model requires more CFG than others

expressions tend to be further along the scale where they need a decent CFG injection

[sterile pendant]

it is fast compared to vlms like llava 1.6, which are miles better at captioning, but exponentially slower in inference time per image

and 17b parameters is nothing, they caption hundreds of thousands of images per day on their servers with thousands of a100ss

[pseudo owl]

cogvlm or florence2? and again cogvlm is a lot better then llava 1.6, and other llava models.
florence2 is also better at image captioning then llava models but florence can't do vqa.

[craggy crest]

find me some images for those on google that shows me what you have in mind

shocked

[sterile pendant]

florence2 is like going from llama2 to llama3. llama3 8b > llama2 70b.

[edgy kelp]

That's a decent one, but why do you need image references for common use words?

[craggy crest]

laughing hysterically

[sterile pendant]

it's not completely perfect for say a company captioning 100mil images when they have 20million dollar datacenters, but for small scale use, it's a beast and the fact that it's so tiny is mind blowing

[pseudo owl]

cogvlm(and cogvlm2) is still better then florence2 and most llava models, its just vram intensive and slow for common gpus.
florence2 is great since its very small and fast and takes little vram but its not state of the art

[edgy kelp]

That's laughing but definitely not hysterically

[craggy crest]

cause those are very generic terms and can mean a range of facial expressions. and i cant' read your mind to know if what YOU consider a smirk is what I consider a smirk

[bitter hearth]

yeah only the bigger llava models like the Yi 34B one can match cogvlm 1

[craggy crest]

it is in MY opinion

[edgy kelp]

https://tenor.com/view/sensible-chuckle-gif-3532650

That's the sensible chuckle dude

[bitter hearth]

it looks like a big laugh to me

[craggy crest]

and? is that what you think laughing hysterically looks like?

[bitter hearth]

I think crystalwizard is right with this one

that's much more than a chuckle

[pseudo owl]

llava models are meh since they use a much smaller clip model then cogvlms, the new ones are a bit interesting tho, they probably are similar to cogvlm level.

[bitter hearth]

yeah llava method is a bit old now they just keep strapping bigger LLMs on

which will probably work for a bit longer

[edgy kelp]

https://www.google.com/search?sca_esv=1a233ca9bb5af21c&sxsrf=ADLYWIIZ5jybpgjk0DHCAeDNc-aFeAe5tA:1721169551235&q=hysterical+laughter&udm=2&fbs=AEQNm0Be9hsxO5zOUoY5v2srYNPRIvTz_02aG-_CVE5t-hWDE-o4NknGVSyJJ2o-ODmr6AzzAs2HXn22T5ul0WKsZVMmJpVctjfuP_7P6Q0vzDgg0UR2OBy8kQja4_o4EYJ19-Z4htN5rihC5e5nWN0tX3m-NShysbeSfMs-yIqyBGYs3SN_Nt8&sa=X&sqi=2&ved=2ahUKEwiij8yd0KyHAxX_0QIHHbFfAYcQtKgLegQIDhAB&biw=1517&bih=674&dpr=1.1#vhid=uDo4fihP6KRBrM&vssid=mosaic

https://www.google.com/url?sa=i&url=https%3A%2F%2Fstock.adobe.com%2Fsearch%3Fk%3Dhysterical%2Blaughing&psig=AOvVaw1hS6uo2qN-noQ4Xp8inC1v&ust=1721255955211000&source=images&cd=vfe&opi=89978449&ved=0CBEQjRxqFwoTCNidgKDQrIcDFQAAAAAdAAAAABAE

https://t4.ftcdn.net/jpg/00/45/64/63/360_F_45646344_AtXGQ7lxvVMhpO9NBANcujDHQpCrfWRg.jpg

[sterile pendant]

i use llama3-llava-next-8b a lot and it's pretty good most of the time, but since florence2 came out, F wasting the inference time

[edgy kelp]

Shouldn't the mouth be slightly more open?
https://thumbs.dreamstime.com/z/hysterical-laughter-businessman-hysterical-laughter-businessman-man-business-suit-laughing-raised-hands-over-white-127027304.jpg

[craggy crest]

i notice on that page that there are a whole lot of diffrent ways people laugh hysterically. one person has their head thrown back mouth open, another has their face buried in their arm, etc - the problem isn't that the model doesn't know how to create the expresison - its that you're not telling it the specifics of what you want the expression to look like. you're expecting it to read your mind

[pseudo owl]

the best multimodel is definitely internvl2, quality is actually comparable to gpt4o and gpt4v.
internvl can even do grounding, video, multi-image. new images should 100% be captioned using internvl2

[craggy crest]

if you want him laughing with his mouth wide open, say so

[bitter hearth]

ah yeah I haven't tried internvl2 yet but I saw it on open_vlm_leaderboard

[sterile pendant]

yeah, but at what efficiency in terms of "caption quality per gb vram per unit of time"

of course big ass 2000000b models are going to be really good, but lets talk consumer level gpu captioners

[pseudo owl]

ok florence2 and moondream2 are the best then(the good part about moondream is that you can vqa as well), both are roughly same speed but moondream2 is slightly bigger and similar vram usage

[edgy kelp]

Alright dude, if you say so

[craggy crest]

painful grimace

[edgy kelp]

That's a good one

[sterile pendant]

here's your average PC based on steam hardware survery(probably at least 100m samples)

so yeah, i like to talk more about the models we can actually run at home without $10k pcs

[pseudo owl]

i just use kaggle since it gives 2 t4(combined 28gb vram) for free lol, my home gpu has 4gb vram

[willow void]

Just put in my 3090

[bitter hearth]

this one is really really good

has less CFG burn also

[craggy crest]

a man laughing hard, head thrown back, mouth open, hands in the air, excited (using SD3 2B for all of these)

[pseudo owl]

3090 is a bit faster tho, t4 is pretty old but runs still fine enough

[willow void]

His hands about tree fiddy

[sterile pendant]

he was trained on the aliens from total recall

[craggy crest]

i didn't run them through the hand refiner

(we were discussing expresions)

[willow void]

its better this way

[sterile pendant]

QUAAAAAID START THE REACTOR

(screenshot from the actual movie, not a generation)

[craggy crest]

intense concentration

[torn wharf]

Open your miiiind

[bitter hearth]

[craggy crest]

@bitter hearth workflow i used for that

[bitter hearth]

thanks but 4gb vram

[craggy crest]

ouch 😦

there's always the online comfyUI websites

[tough oriole]

Working on a SD3 cartoon concept. (Picked a random one from my sdxl dataset) learned the style well enough but it still cant do good poses and I'm not sure how to fix it.

[craggy crest]

it's just a style, right? maybe use controlnet openpose along with it?

[grim pivot]

Heii

[tough oriole]

Oh yah forgot about they added support ight will try

[bitter hearth]

What about openpose with balls gens

[craggy crest]

how do spheres... pose?

wouldn't you need an openBalls controlnet?

[bitter hearth]

Open balls lmao

[sullen moss]

Are there any updates on the 8B model?

[mild bramble]

Is that from only sd3?? Or some upscales?

[craggy crest]

when there is, you'll see annoucements posted all over the internet

[edgy kelp]

Hilarious: I'm trying SD3 on Huggingface spaces and I wrote long prompts about some random celebrities, the best results in resemblance were obtained when I included anagraphical data like zodiac sign and whatnot similar BS, that must be the T5 I guess

[bitter hearth]

feel like its was just a case of a small sample size

don't think zodiac sign would be in the captions much

maybe T5 is interpreting it in a funny way yeah

and zodiac sign is close to other terms in the internal embedding inside the T5

[radiant ledge]

or maybe that celebrity just happened to look like a bull or something

[edgy kelp]

An aries celebrity would look like a goat demon, big time

[hazy hatch]

Skibidi toilet

[lucid swift]

Balls

[edgy kelp]

Balls(TM) but also Pears(TM)

Notice the insane difference between the two images, neither have the best resemblance but some supposedly filler words did... something.

image on the left features the extended prompt with "useless" details:
"a highly detailed close-up of the famous taurus actress christina hendricks, born may the third of 1975"

image on the right features the shorter prompt, more concise:
"a highly detailed close-up of the famous actress christina hendricks"

Both images used the same seed and same everything, except for the slightly different prompt... I subjectively think that the extended prompt has somewhat better resemblance

[sterile pendant]

No it wouldn't, if anything, it would just look like a ram. You gotta remember that the concept-space mappings for it would mostly map to constellations and the sky. That shits light-years away from a human celebrity. You'd likely need to prompt better to get a human-ram hybrid.

[edgy kelp]

You took me seriously for real or you are continuing the joke? I can't tell properly

Did the same with Ryan Gosling, it improved the face symmetry but ruined the picture, I wonder what happens if I remove the negative prompt

Without negative prompt he looks like a middle-eastern

[strange grotto]

RuntimeError: expected mat1 and mat2 to have the same dtype, but got: float != struct c10::Half

met this error

1

[limpid drum]

Hey folks, I've been way for a bit (almost had to go on bereavement leave) . No that I am back, am reading about flroence2 and how it's great for tagging and performs well despite the smaller file size. Question: Can it be used for prompt generation or is it strictly for image description?

[bitter hearth]

yes I use it for prompt generation

it is very good for it

its better than I am (as a human)

[craggy crest]

Odd, I've run it around 50 times with no errors. Is that all that's showing in the console log?

[bitter hearth]

I found 0.5 shift could get a bit rough in terms of structure

when it works, it does look nice though

[craggy crest]

for SD3, what Comfy said is the lower the number, the more the details. The higher the number, the more the shape/structure. so yeah, at 0.5 the details are going to be the main focus, not the struture

[bitter hearth]

yeah my experience was similar

it can be a tricky trade-off at low steps

at high steps you can get both but it takes long

I didn't expect it to be like this because the older models do better with karras schedule

[craggy crest]

shift 1 as opposed to shift 8

[bitter hearth]

these look same to me

[craggy crest]

sigh. it didn't copy the second one to clipboard so they were the same image

here's with shift at 8

[mortal mesa]

with and without glasses

[bitter hearth]

Shift 0.1 is better

thanks wow the difference is huge

[craggy crest]

yeah. i will go down to .5 at times, but i usually stick between 1 and 2

[bitter hearth]

that's a good range yes I did a lot at 1.5

I use lower CFG than you do, in general, so it gets a bit more squiffy sometimes

but running a batch of 20-30 will mostly still yield one good one

[mortal mesa]

shift 3 detail

there would of been bugs at lower shift

[bitter hearth]

that looks nice

I think shift is similar to CFG
lower is higher quality but you lose the control and the structure

[mortal mesa]

cant really put my finger on it, its not quite that, seems more about fine details vs coherence, you can still get detail just not the pimples on her face or the bug is the jungle

[bitter hearth]

its hard to explain yeah

I like this look that super low CFG gives, I would describe it as "wispy"

Would be easier to test if you guys prompt for shining balls

Much much better than humans

[craggy crest]

yeah, i don't usually like what i get if cfg is lower than 4

shift specifically changes the time_step value

[bitter hearth]

most people don't like low CFG that much its just a personal thing

it kinda goes low contrast and hazy, with pastel colours

[craggy crest]

from the Ksampler docs: cfg
The classifier free guidance(cfg) scale determines how aggressive the sampler should be in realizing the content of the prompts in the final image. Higher scales force the image to better represent the prompt, but a scale that is set too high will negatively impact the quality of the image.

other places call it CLIP guidance

shift is dealing with the actual time_step value

[bitter hearth]

yeah the Ksampler I use doesn't actually take in prompts

instead it has an input called guidance

and then I use separate nodes to add the positive and negative

[craggy crest]

right, just pointing out that shift and cfg are nothing at all the same. one is determining how much effect the CLIP neural network will have as far as data gathered from your prompt, the other is effecting the sampling math itself.

[bitter hearth]

oh yeah you are right

I just meant they are similar in some ways

in the sense that higher values can lower quality

[mortal mesa]

oh hi you were using the adams family stuff, ya? i tried and i get either one step or 4 steps, anything easy i was missing ya think?

[bitter hearth]

4th order implicit adams yeah

[limpid drum]

Thanks and apologies for the late reply, working on some major enterprise software agreements here and I get to fall on the sword in two weeks if they're not done and we lose $ 2 billion in tax benefits. Which Comfy node do you use for prompt generation with Florence?

[bitter hearth]

I'm afraid I always forget the names of things
but I just searched "florence2" in comfy manager

I didn't need to get a custom node from github there was one built in

[bitter hearth]

someone else on this server couldn't get the adams solver to work either

but I am not sure why

[mortal mesa]

k, ill play again sometimes, thanks

[bitter hearth]

if I remember rightly this one

was right

you can take workflow from there

also this one, but the image quality is degraded by noise injection as an "experiment"

can't remember if the cat was SD3 or SDXL though

[mortal mesa]

right off the bat, Value 300 bigger than max of 150: max_steps

[bitter hearth]

oh no is the max 150 steps?

[craggy crest]

here - https://huggingface.co/gokaygokay - just go to his huggingface page, click on the Auraflow space - it allows you to use florance OR his long captioner

his long captioner does a better job in my opinion

[bitter hearth]

ah thanks I love captioners as I don't like writing prompts

[craggy crest]

he's got 6 extremely nice spaces to play around with

[bitter hearth]

I need to look into max steps if the max is 150 then
I have been doing 300 steps for no reason lol

[craggy crest]

good grief why?

[bitter hearth]

I mostly do stable diffusion to do weird experiments rather than actually make images to use

[mortal mesa]

uh i dont know, i couldnt do 300 if i wanted to, it errors and stops and puts that message in console

[craggy crest]

(watches you make the ai retrace the same lines over and over, putting smaller and smaller dots in eactly the same spot)

[bitter hearth]

I didn't have errors

it might be a VRAM issue

as I rent datacenter GPUs so VRAM is high

[craggy crest]

there's a point of diminising return. stick with around 32 steps

at a certain point, you're just racking up GPU hours that even a microscope couldn't tell the difference in

[bitter hearth]

yeah in the DPM++ paper he compares against a powerful 4th order Runge-Kutta solver and he says its not worth the time

[craggy crest]

it's not. pull an image up into photoshop, zoom to the pixel level and - that's as finely detailed as you get

the AI is going to do one full redraw of the entire image for each step. at a certain point, you've got what you're going to be able to detect and you're just wasting compute time

[bitter hearth]

I'm not sure yet what my conclusion is

sometimes I quite liked the changes I got from the expensive sampling

[craggy crest]

if you're doing something like medical imaging, or some sort of scientific application, that's different but for stable diffusion? stick with around 32 steps

[craggy crest]

it's your money

[bitter hearth]

yeah I'm just wasting my own money so its ok

[limpid drum]

Thanks again, appreciate the help

[bitter hearth]

no problem

I didn't realise the default Ksampler tops out at 150 steps though

that's good to know

[mortal mesa]

non converging samplers don't necessarily add detail at high steps, it just keeps changing it

i think my issue with adams is i have a crap card and have issue with flash attention

[bitter hearth]

are you sure the max is 150 steps?

maybe it got changed?

    @classmethod
    def INPUT_TYPES(s):
        return {"required":
                    {"model": ("MODEL",),
                     "scheduler": (comfy.samplers.SCHEDULER_NAMES, ),
                     "steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
                     "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
                      }
               }
    RETURN_TYPES = ("SIGMAS",)
    CATEGORY = "sampling/custom_sampling/schedulers"

    FUNCTION = "get_sigmas"```

[mortal mesa]

not sure at all, console spit it out

[bitter hearth]

which sampling node do you use?

[craggy crest]

he said he has a "crap" card...

[mortal mesa]

your workflow from the cat

[craggy crest]

how much vram do you have?

[bitter hearth]

flash attention might have been an issue

yeah

comfy is complex and sparsely documented so there may well be a 150 step limit somewhere

gonna try to take a look through

[mortal mesa]

2080 ti 11gb, ya i think i have to downgrade pytorch to get flash attention on my card :/

[craggy crest]

ummmmmm you might be better off renting GPU time from AWS or even just using google colab

[bitter hearth]

join cloud gang yes

[mortal mesa]

im good thank you, you can see images i post and videos and audio

[craggy crest]

i'm actually really surprised you're generating on that, at all.

[mortal mesa]

now your being silly

[craggy crest]

i'm not. you've got a very small system. and a lot of this stuff takes a lot larger system than you've got.

[bitter hearth]

I don't actually know what the vram minimums are for stuff

[mortal mesa]

4gb

[bitter hearth]

ah ok

[mortal mesa]

never heard of small/large my computer is quite large

[bitter hearth]

sadly machine learning suddenly started using massive amounts of VRAM

it didn't used to

Alexnet was trained on RTX 580s lol

[craggy crest]

out of curiosity, can you run Da Vinci resolve on that? or the latest version of Blender?

[mortal mesa]

im actually waiting for 50 series pricing, when i get up off the floor ill probably buy a used 3090

[mortal mesa]

no clue, no desire

[craggy crest]

are you generating videos some other way than editing video clips together then?

[bitter hearth]

generally there are tricks to reduce vram for most of this stuff

it mostly involves a lot of tiling

that's why people disagree a lot about how fast upscaling is

[mortal mesa]

adiff and modelscope hotshop all kinds of things

ya sure it takes forever for length but its possible

3072 x 1728 24fps on my puny card

deforum type Adiff with some audio sync, udio song

see the little card that could

[bitter hearth]

those are really cool yeah

[mortal mesa]

some plain animateLCM

this small computer is winning ya see

its not the size of the card its how you use it (this also isn't really true)

[edgy kelp]

Prompt used: "typography text saying: "i smell earwigs in here", flowery background, motivational, detailed exact text rendering"
Negative prompt: "missing negative prompt"

[bitter hearth]

SpongeBob jumpscare

[sage burrow]

Has anyone ever been able to get a drow elf with black skin, using SD3? I keep getting pale human ladies. SDXL on the otherhand makes amazing drow!

[mortal mesa]

lol ya thats what i was going for hahah

[torn wharf]

but we've got the biggest balls of them all!!

[bitter hearth]

Try harder 👍

[bitter hearth]

I don't think sd3 is very good at pointy ears stuff

[sage burrow]

Have you managed any?

[sage burrow]

The black skin is the most difficult part 😦

[bitter hearth]

No, I don't try elves anymore cause sd3 is really bad at it lmao

The anatomy of the ears are ugly

[sage burrow]

I made an SDXL drow lora, wonder if I can trick SD3 into using that 😉

[bitter hearth]

You can just generate half with xl then sd3

[edgy kelp]

I swear I used to make the very same bad quality memes with paint a few years ago using impact font

[mortal mesa]

now with more lazy

its the AI way

[edgy kelp]

Hell yeah brother

[mortal mesa]

must be true the AI said it, thats commng

[torn wharf]

https://tenor.com/view/terminator-fence-terminatorfence-nuclear-bomba-gif-21219334

[edgy kelp]

Sad matters

[mortal mesa]

i live under a rock, i dont know what Drow is and neither does the AI i guess

[edgy kelp]

That looks like a Night Elf from Warcraft franchise

[torn wharf]

dark elf in DnD. While DnD is popular, the art is quite niche and not found to far outside of DnD contexts.

[mortal mesa]

i play some ancient game, they got dark elf and high elf

[edgy kelp]

I think drows exist in some european folklore/mythology though

[torn wharf]

Thats right. DnD draws largely from folklore and mythologies. But you'll find that art for the traditional myth Drows are not the same as the DnD stylizations.

WoW is largely a DnD rip off the same way that Overwatch is largely a TF2 ripoff. i use that phrase loosely here. "Heavily inspired by" might be more fitting

[sterile pendant]

You'd probably need to use every other word to describe them other than elf or elven. Pretty much all of the concept mappings are going to associate an elf with pale or fair skin

[mortal mesa]

dark elf popped out some horror type with black skin for me, i didnt save em

[sterile pendant]

You might also try just using a controlnet to get the form and use terms for dark skinned

[sage burrow]

but, but, charcoal black obsidian skin, should be universalt right? 😄

[sterile pendant]

Without mentioning elves

Always ask yourself "what would images like this be captioned as in a dataset or by a non-knowledgeable person looking at the image"

[sage burrow]

PSA gliff sucks, bikini armour isn't allowed grumble

[torn wharf]

idealy, the base model captioning should be so robust that the knowledge is generalized and it can zero shot the image. but if you prompt a hulk elf, it's going to be either mostly the hulk or mostly an elf.

some concepts seem so over fit that you can't blend them into other concepts so well. like doing car models as a ball is hard to prompt for, since cars are so over fit and it can generate specific models so well

[bitter hearth]

[sage burrow]

My furry glif got the skin colour correct ROFL

[bitter hearth]

Always furries

[sage burrow]

Apparently SD3 wasn't the problem, it was just my bad prompting! Claude helped!

@bitter hearth apparently SD3 can do pointy ears just fine... with the help of Claude 😄

my SD3 drow failures

[sterile pendant]

yeah i was just testing it out myself and was having no problem getting elves with dark skin.

[sage burrow]

black skin is always difficult though, even prior to SD3 :(. Niiice ears 😄

[sterile pendant]

yeah just throwing the first one it made out here

[bitter hearth]

looks ugly

[sterile pendant]

that ear is a little wonky though

[sage burrow]

How so? What's wrong with the image?
Or do you mean the ones where I didn't use Claude?

[torn wharf]

finally finished my coffee and got around to it

[sage burrow]

Btown skin though 😦 Awesome ears :>

[sterile pendant]

this one turned out alright as well. had to remove fur from "primative fur garments" because it kept giving them gorilla forearms

like this

see the huge forearm parts lol shit had me rolling

[sage burrow]

What happens several years after a furry and a drow get frisky....

[sterile pendant]

lol

[sage burrow]

it even got the white/silver hair without me asking!!

[torn wharf]

only one prompt to all 3 tencs. bosh3 solver. A jet black dark skinned man, jet black skin like obsidian, he has elvish ears and is wearing elvish style robes. his eyes glow with the magic of valinor

[sterile pendant]

going to try running a few with kolors to see how it handles it

yeah it's being a pain in the ass about making them pale skinned elves

oh nvm, i got one

guess it didn't like the concept bleed test of "alabaster colored clothing"

[torn wharf]

its a unet so any "color" is going to take over the entire image

adjectives go hard

[sterile pendant]

yeah for sure, just wanted to test it out

[craggy crest]

can you fix that with FreeU, I wonder

[sterile pendant]

usually just by rolling the dice with seeds, honestly

[torn wharf]

i've never found attention span problems to be solved by freeu. that one has always just felt like it does nothing imo

[craggy crest]

it does quite a lot, actually

[torn wharf]

Yeah i remember all the hype. No one could ever explain it to me or what i was doing wrong. I just don't see it and then since it can't be explained, it feels like people are just propping it up

[sterile pendant]

PAG+automaticcfg and/or dynamic cfg always better me far better results than freeu did

Freeu just always gave things that fake aesthetic plastic look, even if you used the correct settings for sdxl

But iirc, they touched on some of the issues like that in the paper

The reality is that there's no magic wand that magically makes generations better. They all have pros/cons. So it's better to kind of grasp what all the tools can do and know when to use them

[craggy crest]

okay, well - it's used on mage.space so i did a tutorial video on it, which is here https://youtu.be/1FMIZNR25jA?si=eGM7-vQPK70w4hEk and created documentation if you want that - i got tired of people asking how it worked

[bitter hearth]

out of FreeU, SAG and PAG

I got my favourite results from SAG

[sterile pendant]

mine were with PAG advanced where i left scale at 3 and set rescale to 0.5-0.7. the ksampler's cfg needs to be N-3, so if you want ~5cfg, you'd set the ksampler to 2

[bitter hearth]

oh thanks that is a huge difference

will definitely use this

[craggy crest]

you're welcome. pass the tutorial around as well if you find people that don't understand what it does

[bitter hearth]

I never like the ears, always weird

[craggy crest]

drow don't have cat ears

[hallow lion]

Eve is that you. I wouldn;t trust that slimey bastard

[hallow lion]

ye thats the guy he did it, all of it its his fault

So you feel less bad about sd3's ladies in the grass, here is auraflows "photo of a fat dog eating bananas"

[craggy crest]

Big Brother

[chilly fog]

Вот в таком стиле нарисуй персонажа злобного коммуниста, на прозрачном фоне png файл

[sacred jewel]

Big Sistah

[low stone]

I couldn't get any results like that. I'm using dpmpp_2m, 30 steps, sgm uniform

[torn wharf]

the first minute opens with the usual same hype. It'll take your images to that next level and is a free lunch. its so ez. Not much hope for this video. i've seen a lot of them.

[craggy crest]

are you dissing my tutorial?

how about you watch the entire thing before you get negative about it

[torn wharf]

i got to the end and each different example as you're showing and explaining it, doesn't make sense to me since the visuals are just slightly different noise solutions and nothing about finer details. this is my usual experience with freeu. you gave me the tutorial as if i've never tried to understand it's use over the unet. but you're mistaken. i've just never understood the point of free u and it's basically just waving a magical wand and calling it good.

While you address that theres no "best" settings, theres a narrow range recommended for every parameter still because outside of that tight clamp the denoising solution just goes absolutely bonkers

[craggy crest]

i suggest you watch it again, pause it every step, and study it. yeah, some are slight differences, some are larger ones. it's a way to tweak, as I said, the actual U-Net network layers and skip connections

[torn wharf]

i've never made any image better with freeu. one thing your tutorial avoided mentioning is all the hype about speeding up generations too. how you could get better generations in less steps. no one mentions that anymore these days

[craggy crest]

and i'm not mistaken, it affects each of the layers, and the skip connections that some of the data routes from one side of the network to the other

my original question was i wondered if this could be a way to adjust how much the network grabs adjectives and prevent it going overboard with them.

matt3o goes more indepth with this https://youtu.be/0ChoeLHZ48M?si=Z9iRoqE9gZ92n0e5

[torn wharf]

prompting the unet block block seems like over kill. ther'es a clip cutoff extension on a1111 thats useful. regional prompting is the best solution to the unet's short attention spans

[craggy crest]

the point is to try to keep it from grabbing a color, mainly, and using it on everything

[torn wharf]

yeah. the attention problem. i know.

[craggy crest]

so if you prompt the color in the last block...

[sage burrow]

I may or may not have previously prompted a red tshirt on my horror zombies to trick some places into producing better results 😉
Not just SD apparently

[shut sable]

but attention is all you need

[torn wharf]

red tshirt made of ketchup

[bitter hearth]

deepshrink, scalecrafter and high-diffusion all have a little bit of block by block action

[sage burrow]

It's odd, I've tried FreeU a few times, sometimes the results are truly amazing, the other times, the results are extremely horrible. Works with some workflows and not others for me it seems.

Re the vid, I liked it, far less sensationalism than say, 99% of youtube vids lol

[bitter hearth]

yeah I had that from PAG, SAG and FreeU
they are great tools but not for every image

[sage burrow]

and red paint, and red viscous liquid 😄 Though that last one is really pushing it lolol

[bitter hearth]

sometimes these things sharpen too much and I like soft images

[torn wharf]

viscous. good one

[sage burrow]

I feel kinda dumb now, what is PAG and SAG? something something generator...

[sage burrow]

I stole it 😉

[bitter hearth]

Self Attention Guidance and Perturbed Attention Guidance

they are both trying to improve CFG

[sage burrow]

I did use the perturbed code (from Civit) on the SD3 model when it came out. No idea if that's the same thing. Only a tiny difference though, and could have just been bias.

[bitter hearth]

not sure
I really struggle with things being named different in different places

[torn wharf]

i wonder what kind of cfg improvements we'll see created for sd3. those work on the unet style network don't they? i know people started perturbing sd3 in the first week. i'm just saying, there'll probably be some interesting attention solvers to come

[sage burrow]

It was out day 2 😄

[torn wharf]

oh i looked at SAG an i'm wrong. it can work on DiT too. Does it already?

[bitter hearth]

not sure

I read like 30 papers on samplers in the last few days
turns out I was wrong, 4th order implicit adams isn't going to be the best

DPM++ 2M still seems to be very competitive

also UniPC seems to be underrated

at low steps its one of the best

[craggy crest]

are you trying to turn zombies into red-shirts?

[sour edge]

https://tenor.com/view/dog-nervous-eyes-innocent-scared-gif-11802894372842921173

[sterile pendant]

https://tenor.com/view/displeased-jean-luc-picard-star-trek-the-next-generation-not-happy-gif-23455400

Prepare the away team

[bitter hearth]

wtf

What did the cake do

[craggy crest]

it's what it didn't do. it didn't get away from the cat in time

[craggy crest]

The Rockin' Racoons?

[sage burrow]

They are supposed to be the Punk Skunks 😄

[rapid moon]

any idea when stable-diffusion-3-large be released to huggingface?

[bitter hearth]

no news yet

[rapid moon]

😥

[bitter hearth]

its probably gonna be a while

because the effect of people attacking SAI over the issues with SD3 2b is that they will delay the release heavily

to get it as good as they can

[rapid moon]

large is 8b right?

[bitter hearth]

yeah

[rapid moon]

any idea how much vram it will take?

[bitter hearth]

they said on reddit it will fit in 24gb

[rapid moon]

THATS F***in AWESOME!!!!

i can use 4090 instead of A100

[edgy kelp]

That sounds like they are training another from scratch though

[bitter hearth]

not neccesarily

there's two issues with the model and both are fixable with fine tunes

1. not enough subject knowledge
2. issues with structure

and really they are both just the same thing

enough subject knowledge will teach it structure

I suspect they may well just train from scratch though yes

[cobalt moon]

eh I am pretty sure SD3 will not getting into the same fate as SD2 and SD2.1

SD3 is literally their deadly hit or miss

[bitter hearth]

it will still be SD3

its already WIP

the new SD3 medium

but its a new training run

was my understanding

having said that IDK if it was 100% confirmed that it isn't a finetune rather than a new fresh run

but the point is it will keep the SD3 datasets probably

what I would say as well is that the 8B can do structure fine

so its just a case of scaling it down to 2B

I suspect its the self attention that is the issue

[cinder junco]

Just throwing out a quick question in case anyone has tried it. I'm having difficulty using SD3 with a custom node for MultiDiffusion/Mixture of Diffusers tiled upscaling alongside SD3 ControlNet tile. When set up a similar way as I had it using the SD Ultimate Upscaler node, which seemed to work, I get tiled ghosting of the original image (first image below). If I use an empty latent for the second stage and completely rely on the controlnet to guide the final result, you can see that is the cause of the tiling (second image) -- it applies the full input image to the controlnet, instead of breaking it up into tiles. There doesn't seem to be a way of connecting the controlnet side of things (which changes the text conditioning) to the MultiDiffusion side (which affects the model) to let MultiDiffusion directly affect the controlnet. I saw someone else having what looks like the same problem using SD1.5, but there doesn't seem to be a definitive solution. https://www.reddit.com/r/comfyui/comments/19amano/help_cant_get_tiled_diffusion_controlnet_tile/

This is the custom node I'm using: https://github.com/shiimizu/ComfyUI-TiledDiffusion

[woven bay]

Is stable diffusion 3 uncensored or not?
And also it will run on a 6gb vram gpu like sd 1.5 or not?

[bitter hearth]

its censored

by the normal definitions of that

for legal reasons probably all base models from everyone will be mostly censored going forward

and then fine tuners will do whatever they want

there was a crowdfunded program to fund an uncensored model and the creators just ran away with the money

lol

it was crazy

I still don't understand why no one took them to court

[sage burrow]

Define censored? lol You can get naked people, but not || nipples || nor 🥒

[edgy kelp]

It's not censored, it just does not like people lying on grass

[sage burrow]

I've seen several NSFW models/apps and they all suck! (and not in the way that people are hoping)

[bitter hearth]

I don't want to say its a skill issue
but I have managed to do the woman lying on grass test

[edgy kelp]

Notice that this prompt worked ONLY at 768x768, SD3 is not censored, so far it seems only VERY undertrained

[bitter hearth]

if your prompt has a difficult structure (such as the grass one)
you can do a lot to help
dynamic thresholding, CFG scheduling, block level IP adapter, attention map injection, control net scheduling
and most importantly
one of the stupidly long solvers

[edgy kelp]

I think there is severe inconsistence between seeds... and I don't mean that every seed generates totally different images (for some prompt it generates about the identical image), I mean that some prompts generate good images in only a few seeds. But I don't have the optimizations though since I used SD3 only on the huggingface space of Stability AI

Well, that's true for the ""beta"" SD3 medium, dunno about what's next

[bitter hearth]

I agree about seeds yeah

I get some seeds that are so bad

and then some that are amazing

[edgy kelp]

Of course I can't write good prompts, but if the seed number 1 generates almost the perfect image with my bad prompt, yet any random seed generates garbage there was evident issues with the training

Almost every prompt works at least for 99% at the first seed, I usually try prompts at the first seed in any model for to test the prompt adherence

[noble coyote]

Embeddings made using AUTOMATIC1111 can get around such difficulties!

[bitter hearth]

trying lots of seeds is a good idea

[edgy kelp]

Yesterday the dudes were trying to generate drows or dark elves or whatever, I did these two images with the same prompt but the image on the left was seed 1 while the other is the first random seed that came out.

Prompt: "a high quality hyperdetailed close-up of a dark elf drow from DnD"

I used the dumbest and easiest prompt, the elf on the left is not 100% lore accurate but has pretty good composition and anatomy IMHO

[bitter hearth]

the first elf looks like a royal

and the second like a rogue

[edgy kelp]

But the ears are objectively better in the first one

Look at this, used the same random seed as the one on the right but lowered the resolution to 768x768

[bitter hearth]

IDK if the second ears are bad

just different

[edgy kelp]

The second ears are doubled

Or... I guess those are horns?

[bitter hearth]

might be a popper collar

[edgy kelp]

Either way it's a confusing composition, which is generally thought of a bad image

I know the rest is pretty good though

[bitter hearth]

yeah I know what you mean

[edgy kelp]

Is the instrument Saxophone a no-no for SD3 or there is a more "natural language" friendly CogVLM way in place of "playing a brass saxophone"/"lips on a saxophone"?

[cinder junco]

Musical instruments with complicated features are difficult for AI. Tubas are a good example for SDXL.

[bitter hearth]

SDXL also can't do sax

mostly

[edgy kelp]

Nice, looks like SD3 does decent saxophones if the prompt is simple and short, that's a funny one

"a man is playing a brass saxophone", yesterday I tried a long convoluted prompt fixed with LLMs and it generated only body horror

768x768, one handed sax playing (also best saxophone I generated with any open source model, bad luck for the... weird technique)

[stark ridge]

!generate A serene mountain landscape at sunrise, with snow-capped peaks and a clear blue sky, painted in a realistic style.

[urban arch]

I needed some Drow for my video. I simply described "Drow elves with blue skin and white hair."

[edgy kelp]

Did you use that exact prompt?

Because that exact prompt, using SD3 medium without any fancy optimization, gives me this

[bitter hearth]

gonna try it

[edgy kelp]

I'll have to fix the resolution though

I guess he has some other token in the prompt, it just outputs images that look from bad SD1.5 fine-tunes this way

I'm using the model on the official huggingface space by SAI

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

[craggy crest]

@sage burrow he's calling your name

[urban arch]

Yeah, I realized after I posted that I was using SDXL in Fooocus, not SD3. I also have some D&D specific loras I downloaded from CivitAI, but that is the exact prompt.

[sage burrow]

Try the taesd huggingface 🙂

[bitter hearth]

this is my attempt with that prompt

or this variation

[edgy kelp]

I tried it a few days ago with other prompts but it seemed even worse

[urban arch]

I also got some mixed results

[sage burrow]

I'm surprised, it's one if my faves! Also glif

[edgy kelp]

Let me try with my previous prompt, the one that actually got me decent outputs

[edgy kelp]

This one I mean

[edgy kelp]

Left one is normal SD3 (I used a negative prompt), Right one is TAESD3 that from what I see can't use negative prompts

Same seed by the way

Prompt was "a high quality hyperdetailed close-up of a dark elf drow from DnD"

[sage burrow]

Taesd knows how to fo dark skin apparently

Here's sdxl

[edgy kelp]

Same prompt/seed with SD3 but without the negative prompt

[sage burrow]

My sd3 from yesterday

For fun try dalle!!!! Best and most detailed drow ever!!!

[edgy kelp]

I use Dall-E 3 very often... I reckon on Glif you should have the 8B, right?

[sage burrow]

Lykons said glif is sd3 large most likely

[edgy kelp]

All prompts I did with SD3 medium on that huggingface space are more or less improved if I lower the resolution to 768x768

[edgy kelp]

Ah, false memory then, I was sure he said it was 8B

[sage burrow]

If that's the case, I impatiently await sd3 large 😄
Though on glif Claude helps?also

[edgy kelp]

I guess prompt expansion does big things, however I tried writing short novels in the prompt box of that huggingface space and it seems not to like my natural language

I guess I'm not natural enough...

🤖

[young blade]

random thought of the day
cascade is still the best model (just sayin')

[edgy kelp]

100% true

[craggy crest]

some of us are still using 1.5

[young blade]

still go back myself occasionally

[edgy kelp]

1.5 had the best pretraining (if you know what I mean), but being the first one was "technically" meh. SD2.0 could have been way bigger if trained properly

[sage burrow]

I think it's more about asking in specific ways

I use 1.5 daily! 🙂

[sage burrow]

Why are there no 2.0 loras and models?

[edgy kelp]

I tried doing this image with Dall-E 3 and it got wrong the text every single time, it worked only with SD3 at 768x768

[edgy kelp]

It was a massive letdown when it came out, also it was more difficult to train

SD2.0 was insanely undertrained, probably though in a more linear way than SD3 since it had somewhat better anatomy

[young blade]

think sdxl was announced to be in the works just after 2.0 dropped and people probably said they'd wait since 2 was giving some pretty bad results on base model

[sage burrow]

I haven't tried it yet, is it goid at nsfw? Do the 1.5 or sdxl loras work with it?

[edgy kelp]

Cascade can do some nudity, other loras are not compatible and is very likely more difficult to use than other SAI models

[young blade]

dunno honestly, was more referring to base model cascade, not sure what lora's are out there for it

[sage burrow]

Wait? This community can wait? --day 2 after sd3 repease-- where's all the loras, I'm going to make a lora, who cares if it's a finished model! 😄

[young blade]

i just use a styler prompt mostly with cascade, seems to do really well for what i need

[edgy kelp]

There are little-to-none loras out there for Cascade because it was not hyped at all being a surprise release

[sage burrow]

The file size of SC! 😭

Prompt some up! 😄

[edgy kelp]

To be frank SDXL was announced in July of 2023 while SD2.0 was released in November of 2022, also mind that there was a SD2.1 in between

[sage burrow]

I'll try some on my furry glif....

[edgy kelp]

Some furry balls?

Me when there are no BALLS in here

[edgy kelp]

Are those the BALLS of Ultron?

[sage burrow]

What's that?

[edgy kelp]

Some character from Marvel movies, I have no clue what pop culture references work at all in the present day

[edgy kelp]

Where's the cat with 0.5 GB vram?

[sage burrow]

Ball shaped furries! 😄

[spare bear]

Generated by facadeofrealityx

[sage burrow]

I think I want my glif credits back rofl

Not round darnit

[torn wharf]

those are cute i like them

[torn wharf]

reminds me of critters

[sage burrow]

Reminds me of dollar store toys lol

[torn wharf]

has some "boglins" vibs

[craggy crest]

2.0 was replaced by 2.1 within a couple weeks...

[torn wharf]

they were little hand puppets you could control the eyes

[edgy kelp]

Just like SD3.1! 2 more weeks!

[craggy crest]

nope

[edgy kelp]

2 more and 2 more!

[sage burrow]

SD3 has pietre Bruegel the Elder in it already, it's fine and done! ❤️

[torn wharf]

sd3 is a group of models though. so i'tll probably be sd3 2b/medium revised. they've sorta painted themselvs into a corner with the versioning pattern

[edgy kelp]

Ah I see, trompe l'oeil