🆕｜sd3

[sage burrow]

Sai is adding some models to Amazon servers, how do I make that one sentence into a long article I can put 4 ads into? 😉

[errant dust]

I also found it interestng to read that SD3 Ultra is simply SD3 Large, but with a better workflow

I had actually said this a few weeks ago but was told no, I am wrong and know nothing. But here the CEO himself says exactly that.

[noble coyote]

So ClipDrop = SD3.Large?

[sage burrow]

I hope the next sd3 release is epic. Though I'd be happy with just awesome hands. They obviously hace it, considering flux.

[bitter hearth]

Emad, McMonkey and Lykon said Ultra is a workflow I believe

[errant dust]

I tried it a few weeks ago and commented here in the SD3 Discord, and was gunned down for my ignorance.

[bitter hearth]

I thought I was the one saying it was a workflow in that conversation 🤔

[sage burrow]

Which one us 8b? (My fave) or us that different?

[noble coyote]

ClipDrop and Glif - both run off SD3.Large?

[bitter hearth]

#🆕｜sd3 message

[sage burrow]

Has anyone created sd3 loras? Is it worth the $2?

[bitter hearth]

I was saying there that it was a workflow

sd3 isn't worth it now

[errant dust]

Your communication skills could do with improvig then

[bitter hearth]

even if you had a subject matter that sd3 could handle, its too fragile in structure

ye I'm not great at communicating on discord

[errant dust]

I honestly doubt this issue is exclusive to Discord. That is not a realistic belief. No offense.

[bitter hearth]

yeah people will have trends of how they communicate regardless of platform

[errant dust]

Not accusing, just noting.

I have no issue accepting that was what you meant

[bitter hearth]

its common in STEM fields
to have good math skills but terrible English skills lol

[errant dust]

Myeah. It is common, but there is little excuse for it IMHO, and I feel it is a common cop out for just not reading.

Not that this is you, just that is my belief

[sage burrow]

This one looks fun https://civitai.com/models/158194?modelVersionId=807336

[bitter hearth]

you're right people use it as a cop out

[bitter hearth]

ye Flux is better at soft styles already

was not expecting it to get this good this fast

[errant dust]

Today people will point fingers at the generation of internet, YouTube and so on as the culprit, but truth be told, this is nothing new. Before the blame was TV, and so on. Frankly it is the way reading is taught that is the real blame.

I am a reading nutter, but my sister was the exact opposite. For her reading was a punishment, a horrible tasting medicine that had to be taken on occasion. It was very frustrating to me, and she knew it could not be so with my example, but could not share it. Then when she was 12, I gave her one of the funniest books I had read, easy, and hilarious, pictureless, and told her I'd stop pestering her to try if she read only 3 chapters. If they left her cold, I'd never bring it up. It changed her life.

[bitter hearth]

ye if people connect with a book when they are young it makes a big difference

[errant dust]

it showed her that the books forced in school might be boring to her for the rest of her life, but it was those books, not the reading experience per se

I summarize it simply as: when you watch a movie or series, you are watching the story. When you read it, you are living it.

[bitter hearth]

they do choose odd books for education, we had Frankenstein

[sage burrow]

We had lord of the flies

Grade 5 I think

[errant dust]

Well, I was raised in France through my adolescence, and their reading list is... all French lit as you can imagine. Most countries are undoubtedly the same. In France it was Balzac, Maupassant, Pagnol, etc. Here in Brazil it is Machado de Assis, and so on. But none of this helps develop the love of reading, which brings so many benefits. It is why I am all in favor of Harry Potter and so on, even if I have ZERO interest in them. If it changes reading to something magic and fun, I will support it day and night.

As to me, well, I was a weird weird kid, who was reading non stop as of 5

[sage burrow]

We also had animal farm

[errant dust]

maybe earlier, but I know for sure then. Burned through all the Oz books by Baum at that age

[noble coyote]

I read Sven Hassel at school 😄

... and later did a Bachelors in English Lit

[errant dust]

People think of Oz only in reference to the Wizard of Oz, but Frank L. Baum wrote well over a dozen Oz books, and it spawned a massive following and business in which dozens of Oz books were then taken up by other authors

[sage burrow]

I didn't find the in school books interesting until I hit college.

[noble coyote]

Yesterday I learnt how Wizard of Oz got its name ...

[sage burrow]

Nor the teachings, yawn lol

[noble coyote]

The guy who wrote it had two filing cabinets - one was for all entries A to M - and the other for all entries O to Z ...

[errant dust]

Even Mary Poppins is but one of several Mary Poppins books. Cool reading too at the right age.

[noble coyote]

See what he did there?!?!?!?!?!?

🙂

[errant dust]

I am guessing he was a radical Anti-N guy

😄

[noble coyote]

Was his name Frank N Baum at all?

[errant dust]

It was Frank L. Baum

[noble coyote]

L OK

[sage burrow]

I read regular normal books until about age 12 (Nancy Drew etc), but then I found MAD magazines 😄

[errant dust]

Yeah, I read them too, and the Hardy Boys and... bah. When I was 7 I'd fake a head ache, go to the nurse, be told to lie down, and read

[sage burrow]

SAI isn't very good a mimicking spy vs spy 😦

[errant dust]

I only stopped after i had read the dozen or so books she had on her shelf

lol

[noble coyote]

I have 1,500 books in boxes - love them dearly - but they're taking up space and collecting dust (all BA English Curriculum)
They need a good home!

[errant dust]

Yeah, I starved myself to read. My love is my ereader

[noble coyote]

After all, I have the same number of books on my Kindle

[sage burrow]

You've only ever moved a couple of times at most in the past 10 years? 😄

[errant dust]

https://medium.com/@Albert_Silver/kobo-aura-one-a-near-perfect-ereader-67a7a2b4b57e

[noble coyote]

My latest paper read is Mary Queen of Scots, by John Guy

[sage burrow]

EReaders are awesome, you can red then in the sun 🙂

[errant dust]

Yes, and if you read my article above, you will also discover the ideal size

[noble coyote]

Been here 38 years come Easter ...

[sage burrow]

My thousands of books dwindled whenever i moved lol. My lapidary collection also

[errant dust]

As a teen I had to give away boxes every few months. It was torture

What was I to do? My shelves had three rows of books one in front of the other, and then I was relegated to making piles on the floor next to my book case

[sage burrow]

As long as they were replaced with new ones 😄

[errant dust]

I kept favorites, and ones I wanted to revisit or reread

[sage burrow]

That's one thing I've never enjoyed is rereading fiction, or re watching a movie

[errant dust]

It depends on the book. For revisiting, it might be a passage to try to illicit the same strong excitement or emotion I had when I first read it. For a reread, well, this really only came about when I was 17 when i had an unexpected and powerful experience.

I had read 1984 when I was 14. It was school imposed but I had enjoyed it. It ended up not being given away in the many boxed books, and when I was 17 I found myself perusing it. I remembered the story just fine, so why would I reread it?

[sage burrow]

Because it's non-fiction? 😉

[sage burrow]

You had an interesting school!

[errant dust]

The thing is, this time, as an older, more knowledgeable, and more mature version of my self, things in the story were impacting me differently. the story itself was being reexamined in a very different light, and things hidden before were now

The love story between Winston and Julia hit me differently, so I reread it and left with a profoundly new experience of a book i thought I knew

[sage burrow]

We at least were taught about ancient Egyptian archeology in grade school 🙂

[errant dust]

curious, i repeated this experience a few years later. Now all the political ramificatins and metaphors struck me like a hammer

which I knew, but had never really 'understood'. Not this way. I then realized that rereading this book, this literary onion, was really teaching me about myself and how I as a person had changed and evolved. And this book, and rereading was helping me to see this in a sharper way than i could have otherwise.

[noble coyote]

I d/load free books from Project Gutenberg - and then make them into audio files - which I listen to in the car!
I think I'm on my 3rd re-listen of Moby Dick and Don Quixote both!

A colleague of mine once lived in the same apartment that George Orwell used to live in!
It was where he wrote Keep the Aspidistra Flying!

[sage burrow]

Project Gutenberg is one of the best online resources ever!

[noble coyote]

1984 was written on an island in Scotland (Jura)

[errant dust]

For me, Moby Dick is really two books in one. You have the enthralling and beautifully written tale of Ishmael, and you have that really weird 100-page interlude about whaling techniques. As if he wrote an essay and forgot to remove it. It is not uninteresting but is by no means essential or even important to the story itself. You could easily skip it and not miss a beat.

[sage burrow]

And one of the few that doesn't require current university enrollment!

[noble coyote]

It is sheer poetry to my ears. I love the part where all the Whalers meet up at random in the middle of the Ocean ...

[errant dust]

Yes, the book and its writing are truly breathtaking. It is not unlike Hemingway at his best, where they are able to write amazing prose without ever being about the prose or dragging down the tale or pace at its expense

[noble coyote]

The boy who "swims" in the whale-oil must have stunk - why, they all must have reeked the whole time!!!!!

I love Hemingway! Fiesta, A Farewell To Arms, The Old Man And The Sea ...

[bitter hearth]

ye here in london some of the rly old perfume shops still use whale flavour

[noble coyote]

The "whale flavour" is called ambergris

[sage burrow]

Here it's all fake now 😦

[noble coyote]

If you find some on a beach anywhere, it is worth its weight in gold

[errant dust]

Eau de Orca

[hexed dirge]

https://civitai.com/models/722757/vintage-movie

[sage burrow]

Never found any ever! Had to buy it

Hmm a 1984 lora 😉

[errant dust]

Visually, per the book, for a 1984 LoRA you'd need a cross between Soviet brutalist architecture of the 50s, and the Fallout 3-4 game tech (in the game) of some alternate universe with high tech set in a 50s retro backdrop

[bitter hearth]

that really was an amazing book

[errant dust]

The modern brutalist architecture of North Korea would work too

[sage burrow]

BTW you guts should checkout EEBO if you haven't already 😉

[errant dust]

?

What's EEBO?

[sage burrow]

Sounds Hella fun 🙂

[sage burrow]

Early English Books Online 📚
For once Google lads right there too

[dull star]

https://tenor.com/view/hell-yeah-breaking-bad-jesse-breaking-bad-jesse-pinkman-walter-white-gif-24546693

YESS

You are strictly prohibited from using this Lora for the generation of pornographic content.
oh wow, I've never seen something like this

[errant dust]

Are you talking Early English works like Beowulf? (good luck if it is not translated) or just older works in modern English like Shakespeare and up?

[sage burrow]

Those as well, but also more interesting academia than you find elsewhere

That's one thing I love about the internet, so many translations!

[noble coyote]

You alo need the Senate House Building of the University of London aka The Ministry of Truth in 1984

https://en.wikipedia.org/wiki/Senate_House,_London

[sage burrow]

Digs out manhwa loras...

[errant dust]

I'll take a look, but I need to be honest in saying I am by no means an academic in English Lit, even if I no doubt have read most if not all the major works. Interestingly some older 'fun' books can be a major let down. One that comes to mind is The Three Musketeers

[sage burrow]

The most interesting books do happen amongst the dry ones too. English was my worst uni subject I think. I just didn't get their metaphors (arg)

[errant dust]

No, really. Some of the key plotlines that were not filmed are the reason why it is held in such high regard. Example in case, for those not in the know. The key villain is a manipulative shrew who is imprisoned about 2/3 into the book. She escapes and the three heroes are forced to hunt her down and kill her. But... how does she escape? Oh this is rich. She is seen praying day and night (she cares nothing for prayer per se) by the jailor through her door. She asks him to pray with her and so moved by her purity of spirit (this drivel goes on for DOZENS OF AGONIZING PAGES) he joins her to kneel in her cell. Finally he is so moved by the beauty of her... spirit, that he helps her escape. UGH.

[sage burrow]

So in other news, has everyone seen the NaNoWriMo vs ai debates lately? Death threats, really? Some internet people need to get a life or something.

[errant dust]

She is caught by the musketeers who stick a sword into her. Then, we are entreated to 30 pages of her whining and crying about her life, what she wanted, the injustice of life, death, the universe.... and you are left screaming at the book (like me), 'Will you just die already??"

[sage burrow]

For newer fiction, Lovecraft I'd one if my faves!

And I'll admit I did enjoy the Harry potter books mostly

[errant dust]

I listened to them all read by Stephen Fry, and he was brilliant.

Newer fiction.... lol. Lovecraft is over 100 years old

But yes, a singular style and fun goth horror, a genre he created

[sage burrow]

In uni my absolute fave books were the anthropology ones. Fortunately the good authors creates very thought provoking and society questioning articles. Which is essential for young college students imo.

[errant dust]

He was Bram Stoker on roids

I mean Bram Stoker was all about creating this immense and almost suffocating sense of foreboding, which is quite brilliant. Lovecraft was this in the spades.

Though Lovecraft was essentially a short story writer.

[noble coyote]

The Fencing Master is a great read - it is by a Portuguese author - Arturo Perez-Reverté

[noble coyote]

Gothic Romance/Horror - Melmoth the Wanderer; Memoirs of the Ettrick Shepherd; The Mysteries of Udolpho ...

[errant dust]

I'll look this one up since I obviously speak fluent portuguese. It'll be interesting since I don't know it.

[noble coyote]

El maestro de esgrima

[noble coyote]

https://en.wikipedia.org/wiki/Arturo_Pérez-Reverte

[sacred jewel]

I have a couple I made... meh, so so output and experience. Better spend the $2 on a Flux LoRA IMO

Ballz

[sage burrow]

Oooh 🙂

Hmmm, I wonder if it will be possible to make loras for it one day 😄

[hexed dirge]

prompt/lora ?

[sacred jewel]

Damn I am away from PC right now. Hit me up again later.

No LoRA iirc

[errant dust]

Venom and Alien had a love child?

[bitter hearth]

sometimes you can steal prompt with florence2

today I forgot a prompt for an old image but florence2 brought it back

did anyone else break the good news here?
https://x.com/RunDiffusion/status/1831729215747608962?t=BmWOe6ra38t5KqKf---utA&s=19

RunDiffusion.com (@RunDiffusion) on X

Exciting news! 🎉 Introducing RunDiffusion Photo [Flux]. Our latest model designed to elevate your photography game with stunning, AI-enhanced visuals trained on Flux We are testing this model now and enjoying the results.

#RunDiffusion #AIArt #aiart

Twitter · Sep 5, 2024, 4:20 PM

[bitter hearth]

https://tenor.com/view/flashbang-flash-cat-betterttv-gif-10690574645280585754

[sacred jewel]

🤪

[pseudo owl]

got this from reddit(not mine) but it's amazing. 16ch vae is great with flux quality.

[bitter hearth]

someone in this channel was doing those but sd3 a while ago

lmao

funny prompt

[pseudo owl]

yeah sd3 can do it too, but i definitely prefer flux much much more.

[dull star]

I wonder who

[bitter hearth]

wow look at the arms she's not orange

maybe flux realism is soon

[sacred jewel]

DMd you.

[bitter hearth]

https://media.discordapp.net/attachments/194772750957674497/1039982173618438164/1034723981825159228-1.gif

indeed

sus

[bitter hearth]

what game is that

[winter frigate]

happy buny

[sacred jewel]

...and evilish one 😛

Scary drone but also weird neighborhood

no thanks

[winter frigate]

!generate image of asuccessful neighborhood store in Bogotá, Colombia, promoting the sale of ice cubes. The scene takes place on a sunny day with warm lighting. The store is bustling with happy customers, and shelves are neatly stocked with groceries and food items. In the foreground, a young, friendly woman dressed in traditional Bogotá attire is smiling as she sells ice cubes next to a modern ice machine. The store feels lively and welcoming, with a sense of community and success

[sacred jewel]

That's a lot of cocaine 🤭

[sage burrow]

Flux?!

[dusky thistle]

yep

[coral sable]

why post flux in SD3?

[dusky thistle]

it's 2024 and biden is president, in case you got lost after climbing out of that cave you've been living in lol

[coral sable]

this defeats the purpose of channel, but if mods don't care who am I to say otherwise

[dusky thistle]

literally everything posted in every channel has been flux

i've at least been posting tons of cascade stuff and developing code for it and sharing it on github

i see you haven't been here for a few months so i guess you didn't realize

[coral sable]

I got email from Stability yesterday about new things, and it was full of stuff I knew months ago. No word about SD3 2B remake

[dusky thistle]

yep

[coral sable]

hopefully in few weeks means this year, but still better not to rush.

[dusky thistle]

yep that was the mistake last time

[coral sable]

I'm getting 16gb vram GPU tomorrow, what you have?

[dusky thistle]

that and hyping shit up too much for too long, that never works out

4090... been itching to get a whole rack of em but trying to hold out for the 5090

[coral sable]

how would you get a rack, with GPU risers, or it's done differently?

[dusky thistle]

well i meant it more as buy a pile of em

def gonna get some machines with dual 24gb something gpus soon-ish

[coral sable]

i got 4060ti just so i can run stuff, but maybe i return it will see tomorrow

speed is not importaant, but if 16 isn't enough I'm just trowing money

[dusky thistle]

you can run most things with that

[coral sable]

I wonder if it's possible to use AMD GPU with high VRAM in same system, and use it's VRAM instead of system RAM in caases 16GB isn't enough

[dusky thistle]

theoretically but it would be hell

[coral sable]

as of my understanding hyararchy is that when VRAM isn't enough, RAM is used, then SSD

what if i could put 2nd gpu on 2nd spot

[dusky thistle]

i have that shit disabled

if i run out of vram, it just throws an error and quits

its not worth it

it's deathly slow

hundreds of times slower

[coral sable]

yeas speed suffers

[dusky thistle]

it doesn't really slow down it basically comes to a crawl or stop

[coral sable]

i was able to run some stuff on 8Gb that I wouldn't otherwise, by utilizing system RAM with workable speeds (talking 2-3 minutes)

but not worth the nerves for sure

i hope 16 will be enough for some time

[coral sable]

that's you?

🤭

[dusky thistle]

lol

you can do fp8 precision for most models now

so if you're just doing inference, 16gb will fly for a while

[coral sable]

yeap

and for training will rent if needed

[dusky thistle]

i mean honestly, my advice has generally been to sell the passenger seat of your car and buy a 4090 with the dough

or a door or two if necessary

saran wrap and duct tape don't cost much

[coral sable]

also waiting to see what 5000 series offers, I doubt will go for 5090 but if it's worth i might as well

[dusky thistle]

yeah i think that makes sense

ppl will sell off a bunch of old cards no matter what

that's when i'm gonna buy a ton of shit next

[coral sable]

bruh just for inference i'd rather buy gold then have compute I don't need😅

[dusky thistle]

oh, you need it

[coral sable]

on mining craze i made 3x on hardware. We might be heading torwards something like that again with AI, but I think not yet...

[sterile pendant]

flux makes pretty good 720x1600 wallpapers for my phone. hadn't really tried any extreme aspect ratios yet

[bitter hearth]

not yet... almost every industry is using ai

maybe soon they will be crazy about it

I should start posting sdxl gens here too

[coral sable]

sd3 base model vibe reminds me on 1.5 days, just has some magic

[dusky thistle]

no sorcery quite like the "body parts all over the lawn" prompt spell

[bitter hearth]

haha anatomy bad indeed

[sage burrow]

We've all been sort of doing that since the moment flux was released 😄 oops lol

[bitter hearth]

steven is here

[desert garnet]

yea the server is dead so nothing good to post

[bitter hearth]

I miss the gold old days of #🆕｜sd3 with 1 minute cooldown on chat

[coral sable]

I get it, would be weird that Stability makes flux channel

[bitter hearth]

[desert garnet]

not anime

[fleet meteor]

yaay forge updated so now loras dont have to be patched

[coral sable]

sd3

[fleet meteor]

Nice!

[coral sable]

ty

[bitter hearth]

Sd3 good

[coral sable]

good!

[bitter hearth]

Bow and... Dark elf stuff! @sage burrow

[fleet meteor]

sd3 still has one of the best textures 🤗

[coral sable]

SD3 magic, also skin looks like skin

[bitter hearth]

But most importantly, balls look like balls

[sacred jewel]

You can go back to the day Flux was released and read the WHY from there. It's a loooooong story lol

[bitter hearth]

[fleet meteor]

what is thiiiiiissss

XDDD

[bitter hearth]

A masterpiece

[sage burrow]

And obsolete in 6 months.
My next computer will be in a cloud 😉

[fleet meteor]

I tried prompting a 10 words prompt to gemma 2 9b and I use the output for flux and it works

[sacred jewel]

Trying out the Florence2img workflow from @noble coyote

[bitter hearth]

You'll be like Jesus sitting on a cloud using a computer

[novel anchor]

#🆕｜sd3 create a ballon

[dusky thistle]

Here is the image you requested.

[cyan flower]

Retrato hiperrealista de una mujer mayor con arrugas sonriendo en una cafetería iluminada por la luz natural

[noble coyote]

Hyperrealistic portrait of a senior woman with wrinkles smiling in a cafe illuminated by natural light

[sage burrow]

AI and cartwheels don't mix lol
It didn't show ads for me at least so should be fine https://youtube.com/shorts/0qKn9YETeuU?si=5-sQpMzCqPx3Q7Jf

[bitter hearth]

flux totally made me give up pony nudity altogether

the flux realism and quality continues to fascinate me

also using ollama for prompt gen based on my own initial ideas is pretty neat

[mortal mesa]

so i havent been using loras, have swapped to gguf checkpoints, is there a summery on using loras with it. Do i use model only, or model+clip loader. and i remember people mentioning it acts different with different quaints, maybe it was adjusting wieghts /shrug

[bitter hearth]

the gguf model works with loras, those gguf are unet models btw

means you need to also load up dual clip

[mortal mesa]

one question was basically which to use

[bitter hearth]

always one with clip

more options

[mortal mesa]

ty

[bitter hearth]

also for every lora
try it once with strength clip set to 1
and on same seed with strength clip set to 0
then you know what effect it has on text encoding

[noble coyote]

I never see any effect changing strength clip ... or is that just only me?! 😄

[mortal mesa]

the dorito rat

[bitter hearth]

your average civit fine tune doesn't touch the text encoder

[dry wave]

because nobody trains the text encoder - it's just not necessary for Flux

in Flux (as in SD 3) the text prompts are part of the attention mechanism (in contrast to SD 1,2,XL who use cross attention)

so when you train Flux, you implicitly also train the text

therefore, you don't have to train the text encoder itself

[craggy crest]

there are a couple of new clip_l encoders out you might want to try

[dry wave]

not really

All these models were trained on frozen text encoders - better to not touch them.

[craggy crest]

it works well, it enhances it's text abilities

[dry wave]

also training CLIP is super hard as to do it properly, you need an insane batch size (and, therefore, large vram). The way we trained clip was always just a workaround cause it was just super difficult to train sdxl/sd15 on new keywords

[dry wave]

I know - I still think its nonsense

Flux is only using the pooled embed of CLIP

so CLIP is the last thing you want to finetune to get better text

[craggy crest]

https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/tree/main new flux encoders

[minor hull]

hi
Let's say I want to make an 3d ad image generator.
I would need to combine company's slogan and image.
which model is most suitable?

[dry wave]

I know this one. I just say I don't believe in that kind of training.

[craggy crest]

it works well

[dry wave]

just train Flux and it will work well, too

the pooled embeds cannot really encode longer text prompts anyways

[craggy crest]

faster to adjust the encoder's understanding. this is specifically for text

[mortal mesa]

i was using long clip l, it wasn't very impactful

[sacred jewel]

I just swapped out the model and clip to the UNET versions...

[mortal mesa]

very nice, ill tell you a secret, the prompts came from a MoMA curator for the deaf and blind

[sacred jewel]

Great prompts... I already stole them 😛 😛 😛

[mortal mesa]

https://www.moma.org/audio/playlist/3

[fleet meteor]

Nice!

[noble coyote]

https://x.com/cocktailpeanut/status/1832084951115972653

cocktail peanut (@cocktailpeanut) on X

Fluxgym: Train FLUX LoRA LOCALLY with LOW VRAM

FLUX LoRAs are super exciting, but there has not been an easy way to train them LOCALLY using LOW VRAM machines.

Fluxgym is a Gradio app for training your own Flux LoRAs on your own 12G, 16G, 20G+ VRAM computer for free.

Twitter · Sep 6, 2024, 3:54 PM

Florence2/Flux img2img (Public Domain Originals)

[sacred jewel]

Awesome...

[sacred jewel]

That was all Florence that interpreted a real photo... it got it right almost 100%

"I told you NO! I will NEVER date robots!" 😛

[sacred jewel]

Hey now!

[sacred jewel]

hello there...

[noble coyote]

Searge SDXL img2img

[sacred jewel]

FLorence and Flux combined are SCARY good at copying an original... wow...

[young blade]

misto controlnets any better than the existing flux controlnets?

[noble coyote]

Yeeeesss @hexed dirge 🥳

Train Flux Locally With Only 8Gb VRAM

FluxGYM

[noble coyote]

Vintage Movie LoRA?

[sacred jewel]

Hahahahahaa... NOPE... It's Flux.1 Dev rendering a Florence vision caption 😛

[sage burrow]

What about 8gb? 😭

[sage burrow]

What size dataset? Definitely have to try this!

[runic tusk]

Flux ya:

[sage burrow]

it'd be cool if Pinokio worked ever

[pseudo owl]

Kinda crazy how flux can do basic "motion" and has very consistent characters and background(this is 1 single image and 1 prompt), hard time making the motion move a lot though.

[sage burrow]

anyone who's installed fluxgym on windows have any idea about htis?

[sacred jewel]

ya

Exploded Assembly LoRA

Fuzzy Ballz before night night

[sage burrow]

I'm not sure what Flux has been smoking but...

[fleet meteor]

😁

[bitter hearth]

Nice pair of balls

[clear summit]

please create an image that related to the definition of new media operation

[noble coyote]

Searge img2img

[fleet meteor]

😁

[noble coyote]

Is that a highly chiselled toilet seat at Vodafone HQ?!?!? 🥳

[sacred jewel]

Let's just say I wouldn't take a 💩 in a toilet that has no deep bottom 😛

[solar kraken]

/a black and white linocut of a drummer a bassist and a guitarist in the style of chagall

[sacred jewel]

NOT SD3 😛

[hallow lion]

Garbage Disposal Since 1986.

[sage burrow]

@bitter hearth I would donate some to the Discord anti fraud fund.

[signal shuttle]

Can't wait to play this game

[craggy crest]

the word 'Marbleized' is a good modifier to add to your prompt

[bitter hearth]

OMG I got the loops working in comfy

they are fabulous

can get amazing fine detail by refining 10 times

[pseudo owl]

loops?

[bitter hearth]

yeah it was this https://github.com/BadCafeCode/execution-inversion-demo-comfyui

[craggy crest]

want to share a workflow and tutorial?

[bitter hearth]

something went wrong

could only make degraded images

on other server VykosX just said to try his model unloader + garbage collector nodes

when that happens

[craggy crest]

try rebooting when that happens - yo might have corrupted memory

ran out of it and stepped on it

[bitter hearth]

something like that yeah

gonna remake but with VykosX's loop node pack

cos he released the docs a few hours ago

some loops have been interesting, if you run ersgan more than twice or so it gets very low quality

[craggy crest]

saw that, but if you're running out of memory, and thus overwriting what you shouldn't, you might want to increase your swap space

[bitter hearth]

yeah I will have to look into how to do that on linux servers

[toxic bone]

when i used linux, i think all i had to do to change the swap drive was format a partition identified as swap and the distro i used just found it

[bitter hearth]

ah thanks that should help

[toxic bone]

yeah coming from windows where it's just a setting you point at a drive, it's not intuitive that it would be a partition type. it's the proper way to do it though. we're just all conditioned to the MS way

which i think is intentionally obtuse from how unix systems figured it out years before

[sacred jewel]

Is this considered Deep Porn? 😛

[toxic bone]

there's so many loras for flux now since civit started selling training

"flux can't be trained"

[sacred jewel]

ROTFL... I know... I see like 50-100 per day now...

It must have been a SKILLZ issue

Micrografia LoRA
No prompt
Strength 1.00

and my test prompt at various strengths with trigger words:

[sacred jewel]

Line Design LoRA

[sacred jewel]

Future Urban Day and Night LoRAs

Hyperborea LoRA

Tech Wear LoRA

[toxic bone]

woah

loras alone seem to be enough to extend the model in any ways. i'd love to see someone do a no bullshit analysis of lora vs finetuning

[sacred jewel]

Yeah, I guess finetuning will be more ideal for generating directly a specific style or content but definitely the LoRAs seem to be covering lots of areas quickly.

[late seal]

[toxic bone]

i think that anything a finetuner could add, could just be extracted into a smaller 100mb file instead of a 23gb file

[craggy crest]

flux dev - no lora, no fine tuning

[toxic bone]

people only need one copy of the full 12B. adding diffs to the weights seems like the best approach. contingent on it out performing or meeting the level of fine tunes.

[cursive frigate]

Hi folks, I have been working with flux Dev lately. It seems really awesome but i really need a workflow that can do just about everything. I found one that SEEMS to be awesome and can do everything but I cannot get it working. I was wondering if anyone has used the SDXL FLUX ULTIMATE Workflow from CivitAI?

https://civitai.com/models/119528?modelVersionId=730025

If anyone has other workflows they can suggest that can do all the things please let me know.

[craggy crest]

you should probably check some of Matteo's workflows on his L2 discord

[noble coyote]

Cannot find EVALUATE FLOATS nor EVALUATE INTEGERS - not in Manager

[sacred jewel]

Very nice

[bitter hearth]

downloading massive universal workflows tends to not work well

its better to build it up slowly over time

I took a look through that workflow, its just control net, ultimate upscale and impact pack detailer
they made it look much bigger than it needed to be

[warm totem]

where can i download the weights for sd3?

[bitter hearth]

hugging

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

[sage burrow]

Flux pro is fun 😁

[bitter hearth]

the split image thing is cool

[warm totem]

there's a lot of files what do i download?

[bitter hearth]

biggest .safetensors

[warm totem]

thanks

[sacred jewel]

The baby multi-heads are sooo cute 🤭

[frail shoal]

can somebody send the link for the new flux clip L text encoders ? I remember seeing this some time ago but forgot now where to donwload the different versions for that

[sacred jewel]

https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/tree/main

These are the ones for Flux specifically

[sacred jewel]

The gitHub repo

https://github.com/zer0int/CLIP-fine-tune

Sadly, Flux has no idea who Admiral Ackbar is 😦

[carmine harness]

Excuse me, when using the AI toolkit to train LORA, will using a higher resolution training set (such as 2048 or higher resolution) result in better performance?

[bitter hearth]

you want your training images to match the images that you want to generate

if you want to use it to make 1024 images, you would want 1024 images in the data

if you want 1024 then using 2048 would not be the best idea

however if you did want 2048 then that would be good

generally it takes a large dataset and a full checkpoint to significantly change a model's abilities at different resolutions

[carmine harness]

The original resolution of the training set is 2560 and the details are sufficient. If I need to generate a resolution of 2048, does that mean I need to compress the resolution to 2048? Or is it that when the resolution exceeds 1024, no additional benefits can be obtained?

Understood, try to keep the resolution of training and use consistent as much as possible. Does Flux have a decay value, where details can no longer be supplemented after a certain resolution?

[noble coyote]

"Match of the Day!" 😄

[toxic bone]

i been trianing flux with 496 resolution images and the 1024 tests are mint. unet expectations may need to have a little bit of shattering.

[bitter hearth]

but have you compared it to training with similar images at 1024?

[toxic bone]

no i compare it to flux loras on civit that look grainy and sd15 style. also they look better than the same datasets i trained on sdxl at 1024

[sacred jewel]

My Magritte LoRA was trained with about 50-100 images (I did a few versions with different quantities of images) all between 512px wide and over 4K wide on the longest side. All different A:Rs

Not sure if the results I was happy with would be considered successful or not. But I did not see anything that would make me think the dataset did not have the proper dimensions ... I always generate at or between 1MP and 2MP as referenced for the FLux model.

[toxic bone]

i would need to get a cloud gpu to try flux at 1024. pretty sure its not necessary depending on the data.

i would've never tried training sdxl with anything lower than 768. and that had noticeably worse results

[toxic bone]

got a link? i'd like to study a bit. not to criticize. just literally to feel out how these all feel.

[carmine harness]

In the training instructions of simplerturner, 512 and 1024 are trained together, allowing 512 to accelerate the convergence of 1024. I didn't see similar instructions in the ai toolkit. Based on my own experience, I collected some science fiction-themed materials, such as cosmic materials, and trained the LORA, because this type of material often has dust, glow, and atmospheric effects. The files trained with 2048 materials resulted in a lot of fog effects in the LORA. Perhaps I should try 512 or 1024 again later.

[sacred jewel]

Feel free to criticize 😄 I love love love feedback... good or bad all of it is always feedback 😉

[sacred jewel]

https://civitai.com/user/gonzalu/models?sort=Highest Rated

Sorry, direct link here:

https://civitai.com/models/708199

[toxic bone]

ahh have to be logged in to download that one. maybe its about time i make an account. maybe not.

[sacred jewel]

I think I have it on my Hugg Facial thingy.. stay tuned

[toxic bone]

sweet!

[sacred jewel]

https://huggingface.co/gonzalu/Style-of-Rene-Magritte-Flux.1-Dev/tree/main

You may be wondering why 2 files? Well, Somehow CivitAI modified my file where the SHA256 hash is different. So I downloaded [it] and also have it just in case. Not sure why or how they're mucking around with it...

The -1.2 version is my original. the _1.2 version is the CivitAI hashed one.

Anatomica LoRA

[sage burrow]

I'm trying to install fluxgym, would anyone here happen to know how to get past this error? "C:\Users\oipte>cd C:\oipte\pinokio\api\fluxgym.git

C:\Users\oipte\pinokio\api\fluxgym.git>python app.py
Traceback (most recent call last):
File "C:\Users\oipte\pinokio\api\fluxgym.git\app.py", line 9, in <module>
import gradio as gr
ModuleNotFoundError: No module named 'gradio'

C:\Users\oipte\pinokio\api\fluxgym.git>"

[sacred jewel]

SciFi Room LoRA

[toxic bone]

activate the venv and pip install gradio? It might be indicative of a larger probelm though. This is installed by pinokio which should handle that stuff

i've not used much of pinok but i've seen it. it's a virtual machine thing isn't it?

[sage burrow]

I tried to install it via git before that, and didn't even get as far as the install starting... I'm hoping that I can figure out one version or another.

Has anyone else installed and used fluxgym just fine?

[sacred jewel]

TecExoGear LoRA

[toxic bone]

looks like it's a frontend for aitoolkit which is pretty exclusive to linux. the dev doesn't seem to want to give windows any attention.

almost like he updates it specifically to give windows users issue

just use bmaltais' kohya gui

[sacred jewel]

WALL-E Character LoRA

[cursive frigate]

@bitter hearth do you have a good flux workflow you could share?

[carmine harness]

Regarding the front-end of the ai toolkit, I was also unable to successfully install it. So I just gave up and ran it in the terminal using python instead.

[sage burrow]

Kohya doesn't work on only 8gb vram I don't think though 😦 However, the bmaltais version you say? Does it? 😄

[sage burrow]

How do you do that? I tried, but it wouldn't start??? 😦 I probably put in the wrong run command or something though.

[toxic bone]

Kohya script should. Ai toolkit used their model splitting code to build theres. The reality is that all of these trainers borrow code from each other. They're just different UI's and if tey don't have something implemented yet, it's often just a paste job.

Things are different now because they all had to figure out flux.

the GUI has options to check for model splitting.

[sage burrow]

Well that's awesome news, ty 🙂

[toxic bone]

make sure tog et the flux dev branch

its not main yet

[sacred jewel]

Hyperborea LoRA v2.00

[sturdy dune]

Is this type of noise unavoidable in Upscaling?

The issue is that you can see the tiling as well,

[sacred jewel]

Ed Hardy Style LoRA

[sturdy dune]

I'm using sdxl 1

[sacred jewel]

1! A:R

[sturdy dune]

How do I know whatone is it from the model?

Sorry I Can't tell are you suggesting use something else?

[sturdy dune]

This is my initial prompt: https://civitai.com/models/415107/crystal-clear-onevs1

And my upscale https://civitai.com/models/139562/realvisxl-v50

Is that a good workflow?

[sturdy dune]

AFAIK Flux doesn't allow commercial.

This is the SD3 channel even so it's weird to seek sdxl advice in her I thought it was sdxl 😄

[sage burrow]

Oh!! Last time I installed it, I found it didn't support flux and though it didn't at all yet! Ty 🙂

[sturdy dune]

Okay I'm going to try out flux, a few questions I'm sure have been asked a million times:

1. it works out of the box for comfy?
2. any resources you would reccomend for getting up?

The best version 😄 (Hopefully it works with comfy)

[bitter hearth]

the big advantage of doing tiled upscale the noodle way
is you can re-run individual tiles if an individual tile had an issue

[sturdy dune]

Could you guys point me in the direction of the best way to get up and runnign with flux?

[bitter hearth]

there is a hybrid method
called Mcboaty for some reason

[sacred jewel]

colorful architecture LoRA

[bitter hearth]

https://github.com/MaraScott/ComfyUI_MaraScott_Nodes
mcboaty nodes here
give you per-tile prompt and denoise with just 3 nodes
its by far the best method I have seen if you don't want noodles

[sage burrow]

It works wonderfully with Comfy 🙂
Flux Dev is my fave. Some people like Schnell as well.

Just try out the workflows that come with it, at first. They actually work well.

There's dozens of loras out for it now :). In most cases you don't even need a lora with Flux!

[toxic bone]

i'm not sure if this guide is perfect, but it's short and coudl give you insight to your goal. sebastian i think is one the better ones out there. https://youtu.be/j-9m2hOcyiU

[sturdy dune]

Have you read the licence?

“Non-Commercial Purpose” means any of the following uses, but only so far as you do not receive any direct or indirect payment arising from the use of the model or its output: (i) personal use for research, experiment, and testing for the benefit of public knowledge, personal study, private entertainment, hobby projects, or otherwise not directly or indirectly connected to any commercial activities, business operations, or employment responsibilities; (ii) use by commercial or for-profit entities for testing, evaluation, or non-commercial research and development in a non-production environment, (iii) use by any charitable organization for charitable purposes, or for testing or evaluation. For clarity, use for revenue-generating activity or direct interactions with or impacts on end users, or use to train, fine tune or distill other models for commercial use is not a Non-Commercial purpose.
```https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md

That's pretty tight in stating it's for non profit.

hmmmm That's conflicting terms.

but only so far as you do not receive any direct or indirect payment arising from the use of the model or its output: I'm a little confused, "Any output" means any content. 😄

Makes sense, just heard its a pretty invasive licence, thanks mate.

😄

[sacred jewel]

Cyber Mono Style LoRA

[sturdy dune]

This I think is more understandable:

You can use the images (outputs) generated by FLUX.1 [dev] commercially. However, you cannot use the model itself for commercial purposes without obtaining an additional license.```

[sage burrow]

You should see the various ai chats coding advice! 😄 LOL

[sturdy dune]

To me that's what the licence says

[toxic bone]

yeah i'm not doing this argument lol. i got baited and i feel so dumb now

[sturdy dune]

There's no bait, you said there's no issue with commercial outputs, their licence clearly says "Non commercial".

[toxic bone]

you're still going. oh well

[sturdy dune]

What are you talking about... I said I don't use flux because of their licence, you said there was no licence restrictions, then I showed you them...

[toxic bone]

i'm long past that. dont use flux. use it. dont care.

i thought i was helping you upscale an image but , it was bait. "I thought this was #sdxl " no you didn't. you knew what you were doing.

[sturdy dune]

I'm not taking legal advice from a dude off discord, I said I didn't use flux Because of the licence, meaning I already knew the restrictions you said there was no licence, I checked again to make sure and the licence CLEARLY states Non commerical

When I asked you if you read the actual license you started to argue and now you're talking about bait.

[bitter hearth]

flux dev is sadly non commercial yeah

its possible that schnell would still be of use

schnell is not great compared to flux but it can be better than sdxl in some ways, and in other ways sdxl is better

[sage burrow]

Do you happen to have a url? Apparently the 2 I tried, still don't seem to recognize any of my flux model folders, nor list flux

[sturdy dune]

Sorry am I looking at the wrong model then?

I just need to find out if I can generate non commercial images with any flux 😄

[bitter hearth]

yeah there is a non commercial one called schnell

https://huggingface.co/black-forest-labs/FLUX.1-schnell

but it is like a turbo model

[tardy sonnet]

Is there any way to use Flux without limited quota

I cant run it locally

[toxic bone]

the way i do is git clone the project, then i git checkout ~branchname~ in the project folder

[tardy sonnet]

Do i need a pc?

[toxic bone]

think glif has ways

civitai too but you need buzz, which you can earn or beg for

[tardy sonnet]

To use it on my phone?

[toxic bone]

or buy

[tardy sonnet]

Dude i don’t wanna pay for quota

Can I do it on my phone or do i need a pc

[toxic bone]

i mean, its just a website...

can do websites can't it?

[tardy sonnet]

It can

But is it free?

[toxic bone]

"my phone" might be v though

[toxic bone]

i'm being baited again. waste of time.

[toxic bone]

you can use flux images commercially. people an their misinfo. smh.

if you chose not to out of an abundance of caution, thats fine. others will and it'll make their life easier

[tardy sonnet]

Yeah I was trying to make you pay for it

Unfortunately u found everything out too soon

😭

Fuck off

[pseudo owl]

There will obviously be no unlimited quota but you can try prodia, I don’t see a real limit on hf.

https://huggingface.co/spaces/prodia/FLUX.1-dev

[tardy sonnet]

Wdym no real limit?

The limit is like 7 or 8 photos

I’ve tried it before

Oh no

That was another space

So this one has a large limit?

[toxic bone]

hugging face spaces only rate limit you, not quantity

[pseudo owl]

Yes, this uses prodia’s api. It doesn’t seem to have a limit.

[toxic bone]

thats how we know you're just making things up and baiting people for "debate"

they can also cut you off if you're doing certain "content"

[tardy sonnet]

Haven’t u ever got the “you have run out of quota, sign in to get more or blah blah” on hf?

[toxic bone]

why would i sign out?

[pseudo owl]

That’s for zero gpu, not cpu. The space I sent uses prodia’s api.

[tardy sonnet]

Dude even if you sign in there’s a limit

Why don’t you try it out

On the official flux space

I mean not official but the most famous one whose name I’ve forgotten

[tardy sonnet]

Ah fine then thanks a lot

[toxic bone]

i could say anythign and it'd be contradicted. weird.

[tardy sonnet]

So why would u say there’s no unlimited quota

What’s the problem with this space

[tardy sonnet]

Dude believe me I’m not trying to “bate” u into anything

The other guy helped me out and I’m happy

So just leave it alone fine?

[toxic bone]

ew. i said bait. and i said rate limited too.

you are intentionally shifting my language. i know why. that's your M.O.

[tardy sonnet]

I didn’t “shift” anything intentionally. I’m not a native english speaker and that may be the reason I didn’t get what u were saying

[sturdy dune]

Mate, are you okay? No one's baiting you.

When you state something and someone responds with conflicting information that's not bait, that's either clarifying or a discussion on what you mean.

Like when you say "There's no licence" and I find a licence, I wasn't arguing I was just saying you may not be right.

[toxic bone]

i didn't even say "there's no license"

[tardy sonnet]

Anyway

Even if i explain myself looks lime I’m trying to bait u into something

[tardy sonnet]

Hey

Just out of curiosity

If you think you’re being baited

[sturdy dune]

Now you're arguing semantics.

[tardy sonnet]

Why do you even continue this conversation?

[toxic bone]

that'snot semantics

[tardy sonnet]

And with ANYBODY who says a word?

[tardy sonnet]

Why not just get to your business from which people are trying so hard to distract you?

[sturdy dune]

You literally said there's no non commercial licence. Meaning you said there's no restrictions on generations. I called you out and now you're arguing semantics, are you a bot?

[toxic bone]

• obvious concern trolling is obvious concern trolling.

i'm not sure 'm having a conversation. I'm just countering lies about me rn

[toxic bone]

that's incorrect

[sturdy dune]

Now I know this lads a troll.

[toxic bone]

not how that works

[tardy sonnet]

🤣🤣🤣🤣

[toxic bone]

Did someone roll some dice in here or some?

[sturdy dune]

I'm being baited

Can't even form a complete argument.

[tardy sonnet]

If I were you I would just delete my account on discord

Get back to your business sir

[toxic bone]

feels real seasoned all a sudden

[sturdy dune]

Salty enought to delete all your dumb messages?

[tardy sonnet]

Enemies are trying day and night to distract you and bait you into fake debates which keeps you from reaching your wonderful goals

[toxic bone]

There's a development in the air

[tardy sonnet]

Bro doesn’t even stop responding 🤣

[bitter hearth]

Nothing in my comment was misinformation. I didn't even mention flux images.
What I said was that the model, i.e. the .safetensors file, has a non-commercial license, which is the case

[desert garnet]

u were replyin to some guy who cant even use local and other guy who cant even use sdxl properly that was your first mistake

[toxic bone]

(i think it was roleplay)

[toxic bone]

you were confirming for the guy that the images can't be used commercially without realising it. or he misconstrued it intentionally.

[bitter hearth]

ah okay yeah I didn't realise that's what I was confirming sorry

the license is rly short I suggest people just read the license

[pseudo owl]

It’s basically unlimited, but you do get queued.

[bitter hearth]

ye rate-limiting

[toxic bone]

They know i said rate limited.

[sturdy dune]

“Non-Commercial Purpose” means any of the following uses, but only so far as you do not receive any direct or indirect payment arising from the use of the model or its output: (i) personal use for research, experiment, and testing for the benefit of public knowledge, personal study, private entertainment, hobby projects, or otherwise not directly or indirectly connected to any commercial activities, business operations, or employment responsibilities; (ii) use by commercial or for-profit entities for testing, evaluation, or non-commercial research and development in a non-production environment, (iii) use by any charitable organization for charitable purposes, or for testing or evaluation. For clarity, use for revenue-generating activity or direct interactions with or impacts on end users, or use to train, fine tune or distill other models for commercial use is not a Non-Commercial purpose.

[toxic bone]

theres more than that

but you already know that

[sage burrow]

Glif, huggingface, or Mage for free. Paid is Mage with more features. They can all be used from your cell phone 🙂

[tardy sonnet]

This one apparently lacks the quality from the original flux dev space. I think it’s because of the speed. Are there any slightly slower options with better quality?

[bitter hearth]

I agree in the sense that it is a non-commercial license yeah, and that's what I said originally
but what the other guy is saying is that Flux devs say they won't take ownership of the images

[tardy sonnet]

I’ll try mage out too. Thx

[toxic bone]

They say it in no uncertain terms really

[pseudo owl]

You can simply increase steps, change guidance scale. Seed matters too. From my testing, it does seem quantized.

[tardy sonnet]

I’ve set the number of steps to 50

And still it’s not so realistic