🆕｜sd3

[errant dust]

Look, Christian was no big actor, but it did not help he was paired to some of the world's worst movie lines ever, including the most moronic lovelorn scenes I wish i could forget

[hallow lion]

but before they wer eruthless

thats why George horhey Lukacs said fk this

im selling it go to hell

I agree he is a decent actor

"sand is coiarse" tho

veyr coarse

[bitter hearth]

I loved the prequels at the time

[hallow lion]

I hate sand

fk sand

it gets everywhere

[errant dust]

I'm ok with Lucas selling the franchise. he isn't going to continue work on it, and would be a bigger disservice to fans locking away any future to it for all time

[bitter hearth]

they don't need to be perfect movies

[hallow lion]

i can attest to this

it sucks

sand sucks

[uncut river]

yeah, beach is nice but sand is bad

[mortal mesa]

my buddy has the original trilogy on laser disc, its the highest quality you can get of the non edited originals

[hallow lion]

mmmm

lasedisc

[uncut river]

or at least there is never a good switch between footwear and barefoot, with sand

[hallow lion]

the forgotten awesome media that never had a chance

[uncut river]

maybe I should bring boots to the beach

[hallow lion]

lol

[errant dust]

I beg to differ. Want the best, absolute best theatrical reproductions? Seek out the Despecialized versions

[bitter hearth]

also the sand line isn't bad
this line got memed into oblivion

[mortal mesa]

non edited

[uncut river]

(not me, its sd3)

[hallow lion]

espcially the 4k ones

they take up 50 gigs

but huly crap

u get the 1977 experience

if you have a projector

[errant dust]

Yes, exactly. The despecialized version is a reproduction of the original films that came out

[hallow lion]

mmmmmm

[uncut river]

digital needed a long time to beat video

[errant dust]

as in the original theaters

[hallow lion]

it barely works

[errant dust]

Laser Disc isn't actually digital in the same sense as DVDs or BluRay are

[hallow lion]

like my player complained but somehow it does work

omg

its like the original movie

[uncut river]

2024 - 1977 = 47

What would be the state of generative AI in about 47 years? small chip included in your watch, with the power of Flux now? Nah, that's not a real jump

[errant dust]

This is interesting if you didn't know this already: https://www.youtube.com/watch?v=GFMyMxMYDNk

[hallow lion]

yes

Horhey Lukacs is a hack

Or is he

nah he is not

but he did get alot of help for those early films

he was tremendous and excellent in a new hope as he had rela passion for the 50s scifi tv specials and childhoos stuff

he wa sin pain like all of us yearnign to get back to our childhoods

so it wa shonest

in empire he was still ok and he knew when to hend over the reigns to other so make sur eits a good movie

in the return of the jedi he lost it a bit with the ewoks and toys because now he had to pay taxes and run a company

[errant dust]

That has nothing to do with why he did not direct it, and he had final approval over it all

[hallow lion]

and its ben downhill ever sicne

XD

A new hope comes from a good place

George almost died in a car crash

[errant dust]

Although I shared a SFX documentary about Empire Strikes Back a few days ago, narrated by Mark Hamill and that was televised at the time, there is another great documentary about the actual making of the movie, and it has tons of astounding footage. Not too long and really interesting to see. https://www.youtube.com/watch?v=YeB-uXGGaEU

[hallow lion]

The Cantina ad Jabba's pallae scared the crap out of me as a kid.

Little girl!

Please

thes emovies were not for kids :))

they were so scary

The Cantina was a horroshow

I wa scared for days about tentacles coming out from under the sofa after I saw it

[errant dust]

I was 10 when it came out (cough), and despite waiting in line for well over an hour, only got seats on.... the aisle stairs! but once the movie was over, I hopped to the best seat possible and watched it a second straight time. Ahhhhh the good old days. 🙂

I was not alone on those aisle stairs mind you. lol. They were packed.....

[hallow lion]

It was so good wasn't it? 🙂

[errant dust]

A different time for sure. You'd organize with friends for such movies, as I did for Return of the Jedi, so someone would come real early to get a good spot in the line

[hallow lion]

It came out years befor eI wa sborn and still when I saw it on those black and white KOMMUNISTA Tvs in eastern europe I was hooked

[errant dust]

I was the usual go-to for the early liner since I was a known reading maniac and would be there with my paperback of the day oblivious of the time spent

[errant dust]

Anyhow, the documentary above is super cool IMHO. You see extensive fencing training by Hamill for his famous duels (including injuries!) as well as the HUUUUUGE hangar set they built from scratch for the entire swamp world of Yoda. Just amazing.

[hallow lion]

Yeha it is.

Horhey Lukacs also financed a lot of Empire coz the bank said no.

He had a vision and wanted it seen thru

He hated hollywood to

He wa sa rebel

and

[mortal mesa]

i remember watching the 2 TV movies, i was too young for the theaters, Ewoks: The Battle for Endor was the good one

[hallow lion]

befor ehe became the empire he handed it off

The designs made those movies

and

[errant dust]

I don't see any value in calling people names. He made them happen, and overcame challenges and obstacles you and I could never begin to fathom

[hallow lion]

the believability

it looke das if that world could exist

and the designs were industrial - again the ships looked as if they coudl exists

while other schifi shows hade weird shit that looked too weird for it to look believable

Ralph was a industrial designer

when he heard those concepts he thought as a real world engineer

how would these tie fighters look like

[errant dust]

They had a half dozen models of the Falcon for filming a variety of scenes, from the size of a coin, to a 1:1 model that was left on ground (the famous one in the Hoth bunker)

[hallow lion]

yep

the amount of hate he got for the prequels was probably soul crushing

[errant dust]

Cameron was a lot like this too with the Titanic. They had a variety of models, some of which were just enormous

[hallow lion]

I woudl have walked too

long befor ehe did

[sage burrow]

SD3

[errant dust]

I like the idea, but the head seems a tad oversized

[sage burrow]

does NF4 have any benefits over regular flux, except for speed?

[errant dust]

no

[errant dust]

NF4 is essentially the ultimate compromise allowing you to run the closest possible model to Dev at a fraction of the memory. But it is a compromise, so the pure unadulterated Dev will be better.

[sage burrow]

I'm so screwed when glif becomes paid ROFL

[compact phoenix]

How can I create text to image

[sacred jewel]

https://discord.com/channels/1002292111942635562/1237461679286128730

[sage burrow]

I think that perhaps flux and GPT4 didn't quite understnad the concept...

[errant dust]

What was the prompt for these two signs?

[sage burrow]

Lame a$$ ai (all programs) give the males proper fur, but the ladies only cat ears, or only the face of said animal grumble

[sacred jewel]

Beautiful octane render illustration Text "Still Loving ART" steampunk style

[sage burrow]

that armour doesn't look comfortable lol

[sage burrow]

Picasso? Flux, really?

The flux pro devs forgot to put Picasso into the dataset 😦

[bitter hearth]

Hello again

[frail shoal]

i have a rtx 2060 mobile. Flux works very well with fp8. But when trying to gen with nf4 with the new comfy node, it get's stuck at the ksampler step, just keeps loading and does not generate an image. Has this happened to anyone else ?

[torn wharf]

this might be my favorite so far because i wasn't expecting anything from a pormpt as dumb as "horse with rockets for hooves" but here we are. it nailed it in form and function.

[bitter hearth]

People were talking about that

Saying it's hit or miss if your card supports

[sage burrow]

Even better if you didn't prompt the speech bubble! 😄

[bitter hearth]

On the 20 series

Ctrl f you might find it

[torn wharf]

lol the prompt is more detailed than just that, but that theme of it was championed so well

i added the go go gadget horse because i thought it was more realistic that way and as you know, i always shoot for realism

the image has to be believable

[sage burrow]

Now SD3 on the other hand, has likely heard of Picasso

[real terrace]

Hi, I was wondering, how much difference in speed is the clip fp16 vs the fp8 in speed and quality, for Flux Schell?

[torn wharf]

speed differences would be huge if you cant fit fp16 weights into your vram all at the same time

[bitter hearth]

What I'm having right now

[torn wharf]

did you put the lime in the coconut? doctor said

[real terrace]

I have 12 GB of VRAM

[torn wharf]

did you say "DOCTOR?!" an then he's like "you drink it all up"

[torn wharf]

you'll want the fp8 weights. i have 16gb and can't fit the 16bit weights. but i use that file and just load it in fp8 mode. i'm not sure there are any benefits from using a prebaked fp8 file, other than file size

[torn wharf]

i guess loading the model would be faster

[frail shoal]

it's a shame, was really excited about it. I don't get an error or anything, it's just maxing my ram and gpu and keep loading

[real terrace]

I'll try it then, thanks!

[torn wharf]

you'll still not fit all of the weights into vram, and will be using shared memory. but less of it so it should be a lot faster

[bitter hearth]

Try the other channels, ask if people got it working on your same card or check GitHub

[frail shoal]

yeah i asked, but nobody seemed to be having my card for the moment

[real terrace]

oh yes I never understood how I could work with a model that is larger in size that my whole VRAM

[torn wharf]

shared system memory is really helpful. it's swapping it out over your pcie slot to your system ram, then swapping it back when its needed. this just adds a lot of time to the operations.

some people don't have enough system memory to do it with, and their page file, a storage device file that acts as cache for the system memory starts getting used which adds even MORE time to the operations

[errant dust]

Thanks. Fed it to SD3 Large, first three tries it messed up the text. BUT it did respect the steampunk aspect more:

DE3 got the text wrong in 3 of four samples, but also good on Steampunk and one ws spot on:

[bitter hearth]

Furry

[bitter hearth]

I monkey mode opened the coconut

is good

[errant dust]

Quite liked Ideogram's output which nailed the text with no surprise. Here are two of the samples it fed me:

[torn wharf]

one of my favorite vacation memories is a man hacking coconuts open with ease and precision using a machete. poped a straw in it and handed it to me. mmm fknyeah

kinda wish i could just machete a coconut into a perfect little drink cup

step1: live somewhere there are coconuts

[bitter hearth]

Who the hell let this dog in

Take him out

[sage burrow]

I created a new Flux pro glif 😉 https://glif.app/@LadyLalita/glifs/clzu7qzsv000viusphhjghcyn

[torn wharf]

https://tenor.com/view/lady-and-the-tramp-spaghetti-gif-23646201

he's mirrin

[sage burrow]

Chinatown has coconut cracking hatchets 😄

[errant dust]

I hear ya. I made one for friends here too: https://glif.app/@AmericanCarioca/glifs/clzoyqqv00000jjj58x96kqd3

[torn wharf]

i'm not sure the chinatown near me would. though i do love that place. 2nd oldest one in north america. only a few blocks, but it is such a distinct culture shift. Especially for the kind of stuff you can buy. when i was a kid that's where we'd get firecrackers.

[errant dust]

I may edit it so a generic entry like X will produce a random but creative prompt and image

for those short of ideas but want a 'demo

[sage burrow]

this particular server has way too many censorship filters!!!!

I didn't prompt c l e a v a g e. Flux u r a p e r v e r t

[runic tusk]

That dog looks like he's just coming to a realization that the cat is fuckin' right.

[sullen moss]

https://civitai.com/models/645943/fluxunchained-artful-nsfw-capable-fluxd-tuned-model-by-socalguitarist?modelVersionId=722620

[rain current]

I think he is scared, very scared

[real terrace]

Is there a way to use even more RAM? I have 64 GB of RAM, and this is the usage right now

[runic tusk]

It uses what it needs.

[bitter hearth]

I love the dog face

[bitter hearth]

[rain current]

And the cat: 'Seriously? I'm going to slap you...'

[torn wharf]

using more ram won't make it work faster. there are multiple demands on the system. one is it needs to have enough memory to store the weights while processing them. more than enough memory doesn't offer much to improve generation times. once you have enough memory, speeds are improved with less steps or faster processors.

[torn wharf]

https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/67e8efd5-c67f-4b02-b4b5-e1dee24d7b13/width=450/0001_2024-08-14-1723651415.jpeg bruh her elbows are at her belly button. i dont think this author has a good eye for proportion. that does not showcase how he finetuned flux

[sage burrow]

FLUXING, it was supposed to say FLUXING lol

[bitter hearth]

[torn wharf]

whose the cat? is erich schmidt the cat?

that guy has a TON of investment headaches lately

[bitter hearth]

its the panik

cat version

[torn wharf]

have you ever watched a cat during a solar eclipse? you can tell they're panicking

[bitter hearth]

when she says she likes cats a bit

[torn wharf]

those are all accurate cats

[bitter hearth]

too smol resolution but still cats in their ai heart

[torn wharf]

the pedermobile

[sacred jewel]

Can I run Ideogram locally via ComfyUI? Otherwise, I am out 😛

[errant dust]

You are out

You can use it 10x a day for free though, with four samples each

Same for Dall-E 3, which you can use 15 times

(more, but then in 'slow mode')

[sage burrow]

When ideogram first came out I tried it and it was horrid. So I haven't tried it since. Has it improved?

[errant dust]

It is easily one of the very best

[sage burrow]

!

[errant dust]

they are in their second model and a new one is coming out soon

[errant dust]

supposedly even better, which is impressive

https://ideogram.ai/t/explore

[sage burrow]

I find Meta pretty awesome

[errant dust]

you can click the Top to see the most upvoted images

[torn wharf]

meta.ai is 3.1 llama in canada now

[errant dust]

of the day, week, or month

[torn wharf]

but they are stunting the image generation. it's supposed to have support for 512 tokens but it won't do them

[sullen moss]

Well, this is just the beginning. When SDXL first came out, I downloaded everything that was available for testing, and later, I started selecting what I liked the most and what suited my workflow.

[torn wharf]

the beginning is the base model. maybe i'm just old fashioned but i expect soemthing called a fine tune, to be tuned.

[sullen moss]

By the way, I have a feeling that through FLUX, Musk might be compensating for his grudge against OpenAI, and I think they might be turning FLUX into a full-fledged competitor to DALL-E 3. With computing power, datasets, and overall investments, this could be a strong move. I'm not saying I'm a fan of Musk in general, but in this case, it could be a big plus for regular users in terms of competition.

[real terrace]

great, just for the record at least the RAM see more use when doing img2img with dev

[real terrace]

I tried fp8 instead of the fp16. With dev I think it was a minute less, from 8:30 minutes to 7:30, I think

[mortal mesa]

if you switched to CPU only rendering youll see full ram usage

[real terrace]

oh I don't want that... I remember when I started using SD 1.5 I run it at the same time on CPU and GPU (!?)

Both were long, and glithcy

for some reason I could only run Automatic1111 on CPU, and to use GPU there was some shady other GUI or installation (because AMD things)

Black generations from time to time, great stuff

[rain current]

In Grok 2, Flux.1 is already integrated. In theory, it's a partnership, but if Musk buys BFL, we won't see future versions of Flux on our PCs. However, if it remains just a partnership with economic participation or investment, that would be good news. But I'm afraid Musk will acquire BFL

[mortal mesa]

what would he be buying really, his boffins could spin up a lookalike model next month

[sterile pendant]

Might have something to do with why there hasn't been a paper published for flux yet?

[torn wharf]

shaolin soccer nice nice

[sullen moss]

https://www.theverge.com/2024/8/14/24220173/xai-grok-image-generator-misinformation-offensive-imges

[craggy crest]

you need to post this out to twitter 🙂

[mortal mesa]

cant tell if the article is supposed to be positive or negative

[cunning lintel]

eh, it's clearly negative.

[errant dust]

I always knew he was Chinese

[bitter hearth]

Drama this drama that

All the same

[bitter hearth]

Burn the witch!

[sage burrow]

Twitter lets me post nsfw, going to test out their new AI image creator 😉

[bitter hearth]

[dull star]

https://www.reddit.com/r/StableDiffusion/comments/1es91bu/major_bug_affecting_all_flux_training_and_causing/

Boreal Flux will be even better

[sullen moss]

More tests with analog

[torn wharf]

This seems like the project to follow. I'm feeling the simple-tuner is kind of rushing their bad code out to be "first". I haven't seen anythign quality from their tooling yet. ostris is sponsored by glif is nice to see. I love when these coompanies that build off of the open ecosystem give back to it in big ways

[dull star]

I understand.

[sacred jewel]

What's "analog" in this context?

[torn wharf]

mutant kneecaps

[bitter hearth]

Booba

👍

[sullen moss]

https://civitai.com/models/646517/gustave-dore-woodcut-style-flux-d?modelVersionId=723271

[torn wharf]

x has free flux pro now

oh wait not free. part of the premium x

[mortal mesa]

https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1079

[raven fern]

haha some of the pics on that page

[torn wharf]

https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union-alpha

https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050 also new ui features in forge for loading the specific t5

[sage burrow]

😦

[mortal mesa]

ya they seem to be having issues

[sage burrow]

back to glif lol

[raven fern]

are the pony people still waiting for a new version? i still see v6

[sage burrow]

I was wondering how we nearly went an entire 24 hours without a new model releasse 😄

[torn wharf]

Last I heard he's dataset building. Should just publish the dataset already imo

That's the true open model

[raven fern]

is he working on a flux version?

[torn wharf]

flux kinda knows the beastie boys. thats' sorta like adrock on the left. kinda sorta like mike d on the right. the middle dont look anything like MCA /hat over heart/ god rest his soul

[sage burrow]

He didn't sound interested in sd3 nor fux (moreso a different model), but also said will see how things turn out. Just search his name in this forum to see the convo.

[torn wharf]

pony is 90% data. once the new data set is made he could train it on any model available if the community has enough support for that. they're working towards auraflow but things change all the time. once you have the dataset its a no brainer to point it at a new model with new tools.

personally i think they should just publish the dataset on a torrent. it would ice all the CSAM concerns (or vindicate them)

[sage burrow]

I'm pretty sure they also doing it as a for profit business, so that sort of thing would defeat that purpose. Though putting out the data publiclly, would probably actually lead to a deficate, and not even being able to break even after training costs.

[torn wharf]

they offer generation services as a for profit business and they still ahve to tune the model with their settings and skilled supervision. other people won't be so easily capable of just poppin another out nor get the rep they have for producing smething with the same dataset

[torn wharf]

yolo

[torn wharf]

the only thing i can't figure out with the new forge is how to prompt each of the encoders separately

[torn wharf]

ahhh yes. xlabs has put out new custom nodes and new controlnets

[torn wharf]

this one is dope like the one my fam had when i was a wee ice ball

yeh i found a perfect pic of the one i had growing up. 79 brown special . fusion really good at knocking out those og fords

fusion i mean flux

[bitter hearth]

Now make it a ball

[torn wharf]

https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-Q4_0.gguf 7GB flux anyone?

[bitter hearth]

oh that's the guy who did the NN latent upscale

citty96

he makes cool stuff

[torn wharf]

I also find it funny that nobody knows about the xlabs comfyui nodes because its being hella supressed on reddit. their accounts are all spam banned. people posting about it are getting removed. i dont know whats going on. mod team is just ripping contributors to shreds over there.

people think you still can't lora with nf4 quants

[bitter hearth]

as far as I am aware their location may be in a sanctioned country

I don't want to speculate incorrectly though

[torn wharf]

really cool that the bar for flux has been lowered to 8GB gpus now

[oblique bridge]

https://tenor.com/uSmM.gif

[vital crag]

How does the quality compare?

[torn wharf]

what i found https://www.reddit.com/r/StableDiffusion/comments/1eso216/comparison_all_quants_we_have_so_far/

[bitter hearth]

I wonder why people didn't do quants with SDXL or SD3

[vital crag]

seems to lose a little prompt adherence, the dreadlocks are missing on the lower models. quality still remains extremely good though

[bitter hearth]

really can't judge quality on cartoon

I wish people used photos for testing

[radiant ledge]

Apple did quantization for SDXL, but the arch is not very tolerant to quantization and lost a lot of quality. SD3 should be better.

[sterile pendant]

there you go

[bitter hearth]

I don't like either because of Flux's style for realism
but they both look the same image quality so the quant should be good 👍

[sterile pendant]

oh i prompted for that

[bitter hearth]

oh no

[sterile pendant]

"A professional photo of" and threw in other hipster shit like "rustic decorations"

(this is my new go-to prompt for trolling because they always say the proof is in the pudding)

[real terrace]

I love the good details in a straight generation

that's flux dev, long generation time but the most quality it seems

you can zoom in and there is good detail

hands are perfect MAN

[bitter hearth]

Perfect man

I decided instead of R2D2 I am just gonna start prompting stuff that makes no sense
like Rococo furniture in desert cave

[sterile pendant]

At any rate, I knew it was only a matter of time before they started gguf'ing these dit models. In the llm world, q4 really is a local machine standard and is usually within 5% or so of the q8 version and the q8 version is usually within a fraction of a percent of the q16. They could also do similar stuff to ggufs where you have _k_m or s where some important blocks stay full sized.

Makes me wonder if the next sd3 model will take advantage of this.

[bitter hearth]

I guess this is a big advantage of DiT over Unet

cos the smallest Flux Dev quants are now smaller than Kolors Unet

for example

[bitter hearth]

Fancier than most houses I see

ye its got fish

the seed where it put the fishtanks into the cave wall is way better

[sterile pendant]

Unets are quite different though. They have some operations that are transformer based, but the rest of the model isn't a regular transformer. They don't quantize very well due to operations that rely on the whole U shape of the network

[cobalt moon]

hm

lets try quantized Flux on 2GB VRAM

I guess it gonna be a ticker time bomb

[bitter hearth]

I'd rather see a quantised model that targets a 3090/4090 24GB
but started out way larger
it would be the best possible way for consumer GPUs

[bitter hearth]

512 next to run on my phone

there are phones with 24GB ram

they could run not only the quants but the original flux dev

[cobalt moon]

like you can literally run LLM on it lol

[bitter hearth]

yeah I use LLM on my phone constantly

[sterile pendant]

yeah that's possible, like running 70b llms that are quantized down to fit in 24gb vram

but i can't even begin to imagine the compute time for making an image model that size...

but also keep in mind that the average consumer PC has 8gb vram

[sterile pendant]

barely even 5% of pcs have 16gb or more vram

so for now, targeting 8 or 12gb vram is idea and will likely stay the ideal for at least another year or two

[bitter hearth]

open source isn't really necessarily about targeting the average person
even though open source does often get interpreted that way

[sterile pendant]

you're in an echo-chamber of the same couple hundred people though, the bulk of users are the silent majority

that never make a post here or on reddit or twitter, etc etc

[dull star]

idk man they should target H100

🔥

doesn't everyone have an H100 at home

[sterile pendant]

when the have models that size, they keep them to themselves for api revenue. 99.9% of local users will never donate even a single penny to the companies that spend millions training them and paying researchers

[bitter hearth]

well I don't think donations are the only way, or the main way, open source is funded

a lot of open source comes from companies that make their revenue elsewhere

[sterile pendant]

no, not really anymore. now it's big companies that back them like nvidia or now elon backing flux. once that happens, yeah the quality goes up, but then comes the censoring that everyone complains about (i want censored models). then you end up with an sd3 and everyone cries about not being able to make their fetish stuff.

though elon might be down with the degeneracy

so flux won't likely be forced into being censored

(it's already censored)

[dull star]

I want censored models too if people can train stuff back in (if not then even SFW models don't come out, because no porn = "don't train its bad")

I'm only interested in like 2 loras right now and both are SFW

and anatomy isn't utter rubbish

[bitter hearth]

I missed the Elon news
Which flux model is X using?
and is the censorship/guardrails the same as the open source one?

[sterile pendant]

probably dev for now, but maybe their pro version, we don't know

[bitter hearth]

the media is complaining that it is uncensored
but its surely just the same as the open weights. right?

I guess the media didn't know about it until now

the mainstream

[dry wave]

how do you know that it is Flux?

haven't read that anywhere except here

[sterile pendant]

one hour on that planet is 7 years on earth.

[sterile pendant]

it's been stated

[rain current]

yeah! In a scene of Interstellar a woman with lots of tattoos and long blond hair wearing a tight black shirt and tight black pants, the shirt is Nike's, it's on the water, and the spaceship in the background along with some galactic astronauts

[bitter hearth]

Flux has a look that cannot be mistaken

[sterile pendant]

just google for blackforestlabs and x

it's been publicly announced already

[dry wave]

yeah, found it now

[bitter hearth]

journalists are going wild
```Musk seems intent on letting misinformation like this pervade the platform. By allowing users to post Grok’s AI images, which seem to lack any watermarks, directly on the platform, he’s essentially opened a firehose of misinformation pointed at everyone’s X newsfeed.````

[dry wave]

weird that mainstream media all don't mention it

[bitter hearth]

they don't seem to know that SD3 2B is even more realistic

mainstream media has never come even 1% close to understanding tech though

[dry wave]

it's all just Grok's AI images

[dry wave]

that's unfortunately right

[sterile pendant]

well keep digging, there was some official post that linked the two. you're free to keep wasting your time speculating if you want though.

[bitter hearth]

to be fair the Google search I did, did not come up with great sources in the first few results

[sterile pendant]

here you go https://x.ai/blog/grok-2

"In collaboration with Black Forest Labs, we are experimenting with their FLUX.1 model to expand Grok’s capabilities on 𝕏. If you are a Premium or Premium+ subscriber, make sure to update to the latest version of the 𝕏 app in order to beta test Grok-2."

there's the actual hard proof straight from the horse's mouth

[sterile pendant]

thought so, fucking love that movie

[bitter hearth]

Grok-2 is strong too apparently

[sterile pendant]

idk how it does, i know grok was hot garbage in the llm world. like a total waste of parameter size apparently vs models a fraction of the size. maybe they've improved it. i don't really touch anything that weirdo has his hands on. i didn't have a twitter account before and i sure as hell don't have an x account now lol

[bitter hearth]

its important because he is the most likely to open source it with a good license

[sterile pendant]

maybe, that's one good thing he has going for him. but knowing him, he'll find some dumb way to bias the models for the lulz or be a puppet for those who pay him to do so. he's autistic, people know that and plenty of people take advantage of the fact he's easily manipulated. you know i don't even have to name a recent one...

[sage burrow]

BC there's no reason anyone would pay extra to use X under normal circumstances

Or imo even use X lol

[bitter hearth]

once the weights are open I don't really see the risk

cos you should benchmark a model a lot on your task anyway
before actually deploying it

[sterile pendant]

My point is, if someone has their hand in Elon's pocket, and Elon has his hand in black forest labs' pocket, any future model coming out will be biased with some kind of BS

[bitter hearth]

I see what you are saying yeah

[frail shoal]

neither gguf nor nf4 works for my pc. They max out vram and gpu. However fp8 does not max my vram neither my gpu utilisation and works fine. I'm thinking the custom nodes for those are not working with lowvram option on comfyui. They are trying to do everything in gpu. Is there a way to fix this ?

[sterile pendant]

try this node and set it to cpu

https://github.com/city96/ComfyUI_ExtraModels

i have to use it like this to work on this 8gb gpu and it works fine for me

and maybe add --disable-smart-memory to your comfy launch flags

[frail shoal]

i will try this

[sterile pendant]

it will force every model to unload from the vram after they complete their task

[frail shoal]

--disable-smart-memory to your comfy launch flags ?

[sterile pendant]

yeah, the same place you added --lowvram

[frail shoal]

this will ? I though it is automatic --lowvram. I don't place it in comfy, but it does that by itself. Will try both then

[sterile pendant]

find your run_nvidia_gpu.bat and right click, open with notepad

and add --lowvram --disable-smart-memory to it

[frail shoal]

yeah i know how to do that thanks. I made my own bat

[sterile pendant]

no problem, that's where you add those kinds of flags. but that node i showed helps a ton.

it will be slightly slower for handling the prompt encoding step though, depending on your cpu, but for mine, it's like two seconds (13600kf)

i'm also still finding that i have to periodically click unload models from the comfy manager though. sometimes my it/s will be double the norm, so i have to cancel the current job, click unload models, and do it again. it will probably take comfy and the addons a little bit of time to hammer out the memory management aspects of it all, but for now this works

[rain current]

+loras

[icy drift]

Using Flux Dev's native character consistency across game asset types. A detailed description may render IP Adapter obsolete. (Alternatively IP adapter may prove vastly more powerful for Flux than it has for any prior model.)

[bitter hearth]

sadly character consistency in a base models is seen as a bad thing in machine learning

it means the sampler has low recall

[icy drift]

This is consistent adherence to a detailed description.
woman mage with pale skin, long dark hair, a red cloak and hood, with a wooden staff that has a red diamond on top. Under her red cloak she has a long-sleeved full-body suit black shirt and trousers that are skin-tight and show off her muscles, and a gold belt with leather pouches.
You are thinking of a model's inability to generate diverse characters from simple prompts. This is unrelated.

[frail shoal]

did not change anything : ggml_sd_loader:
GGMLQuantizationType.F16 476
GGMLQuantizationType.Q4_0 304

model weight dtype torch.bfloat16, manual cast: torch.float16. Does this mean it's trying to load it in fp16 ?

[bitter hearth]

yeah I see what you mean

if the prompt is simple then low diversity (recall) would be bad

but if its a very detailed prompt then its not as bad

I would still rather have more diversity than this though, as far as I understand its a side effect of distillation

[sacred jewel]

I have 4 👀

[sterile pendant]

how much vram do you have?

because you could just be getting screwed by cpu or system ram speed if it's having to do most of the work on cpu

and make sure you update your comfyui and that you aren't trying to use loras with the gguf or nf4 versions of the model, the loras only work with the fp8/fp16 versions of the model(for now)

[frail shoal]

but why does it work with fp8 ?

[frail shoal]

i'm not using loras

[sterile pendant]

because comfy has native support for it and it likely meshes well with the memory management system used by comfy

[frail shoal]

most of the things are being done in cpu with the fp8 model. I'm not having any problems. It seems like int4 gguf is requiring a ton more resources than fp8. It does not make any sense

[sterile pendant]

to run the quants, you need addons

those addons might not mesh well with comfy's memory management

[frail shoal]

yeah i'm guessing comfy needs to support it

[sterile pendant]

what card are you using?

[frail shoal]

rtx 2060

[sterile pendant]

yeah that's probably why

it's probably not fully compatible with some of the data types

or at least not full speed compatible. i know some models of cards had issues with half precision(fp16) around that generation

[frail shoal]

wondering if forge supports it well

[hollow swift]

where can i find force/clip set device node?

[bitter hearth]

extramodels for comfy ui

[hollow swift]

ok thanks!

[severe phoenix]

please is there anyway to sorta merge sdxl and flux models?

[bitter hearth]

definitely no

too different

[sage burrow]

Use a workflow that starts with one then ends with the other. Not merged but same idea kinda.

[frail shoal]

have you guys installed the new front end for comfyui ? How can i do that ?

[severe phoenix]

i just heard of sthng called block merging thats why i asked. thanks guys

[bitter hearth]

--front-end-version 'Comfy-Org/ComfyUI_frontend@latest'

you are in the comfy discord too lol

[frail shoal]

well they didn't respond

[bitter hearth]

but did the flag work >

?

[frail shoal]

have not tested yet, will try now

[sage burrow]

If you search merging in this section, some folks who know a lot about such things explain why it's not really a merge.

My experience with merging 2 different models has produced results that are always exactly like the first model selected.

Really easy to tell when I "merged" sd3 and pony.

[hollow swift]

i cant run the gguf model it says this, any help?

[frail shoal]

same

[dry wave]

you just cannot merge different models. It's not possible.
The problem is that:

• many tools do not output error messages but just do nothing and so the user don't know that the merge failed
• sometimes part of the models match like for example SDXL and SD 1.5 both use CLIP-L text encoder, so when you trained CLIP-L then this part if the model is merged while the remaining part is not merged

[sage burrow]

@severe phoenix

[sage burrow]

Flux pro via glif LOL
I made a flux pro img2img with prompt, glif btw. Unfortunately glif only offers 20 free per day currently though. https://glif.app/@LadyLalita/glifs/clzvb1932000bs5zl0fjerf3r

Flux Pro vs Flux Dev img2img via glif

[dull star]

I don't blame them

I have been hogging free usage of SD3 8B for quite some time until 2B came out

[sage burrow]

Wait, Flux has 8b? I thought it was large/4b? Think I need to search glif more!

[sage burrow]

Though you can get many more daily credits if you apply for them as a glif app creator via their discord... though I think you might have to repply daily

[severe phoenix]

thank you so much. please what does clip models do? there's apparently this new clip model for flux that can work for also sd3 and sdxl

[sage burrow]

I just tagged you so you could see that person's reply, you probably want to ask them (I don't know nearly enough about such things, I just look at image results 😄 )

[zenith hemlock]

Just made a benchmark prompt 😄

Band poster: bass guitarist on the far left, drummer beside him, guitarist in the center, another guitarist next, and female vocalist in a wedding dress holding a microphone on the far right. Background features a circus with fairies riding elephants, creating a lively carnival atmosphere

flux dev

[zenith hemlock]

it couldn't follow the prompt completely but did something anyway

😄

and DALL-E

in chatgpt*

[muted dove]

Flux 😛

[zenith hemlock]

flux won this

[severe phoenix]

lool ok thanks

[sage burrow]

😄

[sage burrow]

What do you mean your computer won't run Flux Dev, skill issue! 😄

[dull star]

https://huggingface.co/fal/AuraFlow-v0.3

[tough viper]

Can soneone point me to a flux getting started guide?

[zenith hemlock]

fat city

[sacred jewel]

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

[bitter hearth]

Do they sell humans in stores instead

[zenith hemlock]

😮

let me check

[bitter hearth]

You'll become a product

[zenith hemlock]

@bitter hearth

😰

looks like they are..

[bitter hearth]

[zenith hemlock]

A detailed butcher shop entirely sculpted from smooth, golden butter. The walls, counters, and display cases are all meticulously crafted from butter, with a lifelike buttery butcher standing behind the counter. Inside the freezers, there are real red human meats, also made from butter, giving the appearance of frosty, cold storage. The shop is warmly lit, causing the buttery surfaces to glisten slightly, creating a surreal yet captivating atmosphere. Every detail, from the signage to the hanging hooks, is sculpted from butter, blending realism with the rich, creamy texture of the material

[bitter hearth]

The lamp is a piece of cheese

Lmao

[zenith hemlock]

it really looks like a cheese ahah

can u give the prompt

[bitter hearth]

I did not use much brain power before hitting generate

"Cheese people behind a counter"

[zenith hemlock]

no way 😭

😄

anyway lets try on flux ahah

[bitter hearth]

Id eat this whole restaurant

[zenith hemlock]

tinges beady

[bitter hearth]

There

[tough viper]

is dev better than schnell?

[sage burrow]

MUCH! (pro is even better than that!)

[tough viper]

pro?

Is there way to run pro non locally?

[sage burrow]

The only way I can think of maybe doing so is to use comfy, and that run api node. Though I'm pretty sure you would have to pay for the api tokens with this method. At least that's how it worked with SD3 when I used my comfy SD3 api node.

[errant dust]

It is worse than that. You need to consider the very reduced number of users who have the technical knowhow, interest, and motivation to install and run models like these.

[bitter hearth]

its a tiny amount of people I agree

but Llama 3 405b FP16 weights are over 800GB and there was still value in that being released

[errant dust]

Sure, but Llama also has smaller ones

And Meta's value in this likely has nothing to do with the motivation Flux or SD had, even if the result is a publicly shared relase

[sterile pendant]

Yep, talking percents of percents of percents. So out of the millions of people using genAI, you're left with a few thousand people in the communities that are actually active in even vocalizing anything. Of that vocally loud minority, most of them are weirdos obsessed with making NSFW content and showing off their fap factories and waifus

[errant dust]

I know, it is beyond ridiculous, but whatever. Look at it this way, I'm sure they do the same thing with LLMs. 🙂

[bitter hearth]

but Booba

[errant dust]

Anyhow, it doesn't matter. The fact is some or many image AI generators have really established a not unimpressive audience. MJ showed that with their millions of accounts early on. Last year when I asked they had 16 million active accounts, and I cannot imagine it shrank since then.

[sterile pendant]

They do, it's all about weirdos doing ERP. Been in the LLM community longer than the diffusion community

[errant dust]

And unlike DE3 or even Ideogram, there is no free images a day option

[bitter hearth]

I actually don't rly think the best diffusion model within the next few years will be SAI or BFL (flux)
I think at some point Meta, X or Google will drop an open source diffusion model

[zenith hemlock]

only through api for now

[errant dust]

Google? Cmon. I have not seen any sign or hint they have any interest in joining the free for all. Meta might be. They did a lot of this even before, such as bankrolling a free super Go model to compete with the AlphaGo. Since, while Deep Mind may have shared how to build a super Go model, they ddi not share the one they trained. So Meta, still FB then I think, stepped up

Not a lot of people know about that project of theirs

[bitter hearth]

not sure if you missed Google Gemma 2

Gemma 2 is one of the best open source ones in its weight class

[sterile pendant]

gemma 2 is actually really good

but mistral nemo stole the thunder of both llama3 and gemma2

[errant dust]

Then I did indeed miss it. Saw the name, but did not pursue it

[zenith hemlock]

Meta is finally doing something useful for humanity by releasing open-source models, but it feels like there's some shady shit behind it 😬

[bitter hearth]

ye Gemma 2 turned out to be good

[zenith hemlock]

It looks like early 2023 stable diffusion 1.5 semi-realistic anime outputs 😄 cool output

[bitter hearth]

TBH people are suspicious and critical of every large company no matter what they do, so they may as well just do what they want at this point

[errant dust]

Unlike now?

[bitter hearth]

Meta doesn't rly make their money directly from the public anyway, since they are an advertising firm, so it doesn't matter too much to them what the public perception is

[zenith hemlock]

because the only goal of these companies is to make more money, people don't buy into their fake sincerity

[sage burrow]

cuase the nsfw ones suck currently lololol

[dull star]

no because idc about nsfw

that's the point

[sterile pendant]

a lot of the diehards in the genAI communities are on the spectrum, so things like autistic paranoia and paranoid schizophrenia are common.

[bitter hearth]

that's what I'm saying, there probably isn't anything they can do to make you like them, so they may as well not try and just do whatever they want, for better or for worse

[errant dust]

Nonsense. You need peopple using your services to advertise to them. But that is a digression. The fact is they have tons of projects and whatnot the public won't know anything about

[sage burrow]

if censored I'd just use MJ prob

[bitter hearth]

its that if they get a bad enough reputation then they will lose enough users yeah but that mostly just happened to X

[sterile pendant]

the reality is that you're never going to get an uncensored quality base model again. those days are gone

[zenith hemlock]

When I first encountered Stable Diffusion back in 2022, I used to really enjoy creating NSFW content, but thankfully, I've outgrown that teenage phase 😭

[bitter hearth]

most social media companies just slowly fizzle out
rather than going down in a blaze of fireworks like X

[errant dust]

The fact is, X's use of Flux is an amazing coup for image generators in general

Not just for Flux

[bitter hearth]

I'm really happy about the X and Flux news yeah

[zenith hemlock]

However, trying out something with FLUX wouldn't be a bad idea

[errant dust]

Flux has a lot of amazing strengths, and depending on your focus, can be a pure godsend

[sterile pendant]

oh and one thing i forgot to mention earlier was that a huge chunk of image gen users in general are from countries where pron is illegal. places like india and china

so they tend to throw the largest hissy fits over models being censored

[dull star]

paid + not offline

[errant dust]

As I really enjoy diagram and illustration art, using styles and more, it falls flat a lot in some. But like mall f these players, it is aboout knowing who does what best

[errant dust]

Offline means little to me. Sorry. My machines are online permanently, and there is no real bandwidth consumption in sending a prompt or downloading the image

[zenith hemlock]

Actually, it's more about your control ability than whether it's online or not. Midjourney is an incredibly limited service in terms of control ability

[errant dust]

would be inteesting to hear what that means. Control ability.

[mortal mesa]

i never realized how censored this discord it, just hit a filter twice in a row

[errant dust]

There are tradeoffs, like anything

[sterile pendant]

they have a ton of layers to their system though, llms parsing prompts for nsfw content, llms expanding caveman prompting into useful prompts, the image gen, then layers of nsfw filtering for the outputs

[bitter hearth]

midjourney needs a new model now

they are basically behind at this point

[errant dust]

I take it you don't use MJ

[bitter hearth]

no I don't, you are right 😄

[errant dust]

Or you'd never even dream of such a comment

[zenith hemlock]

With local models, you can train them however you want, create LORAs, fine-tune them, and edit different visuals to your preference without having to answer to anyone or deal with censorship limits. You can also benefit from features like ControlNet. In other words, the possibilities are endless with local models

[sterile pendant]

MJ is really solid. my sister uses it a lot. i don't really mess with it much

[errant dust]

They released a new model two weeks ago. v6.1. It is still quite iffy in text, so what else is new, but anything else it is... quite amazing

[sterile pendant]

https://artificialanalysis.ai/text-to-image/arena

[zenith hemlock]

For people who just want to generate and use a single image without worrying about the technical aspects, Midjourney is a perfect fit. However, for those whose goals are more technical and experimental, local models are ideal. So, determining which is best really depends on your usage and needs

[sterile pendant]

yeah 6.1 is a little lower ranged than 6

[errant dust]

and you have pages of parameters, flags, you can add to modify the behavior on an image

save styles, and so on

[dry wave]

I wanted to write the same

in arena Midjourney is clearly beaten by Flux

[errant dust]

Arena?

[sterile pendant]

thats new though, flux just came out

[dry wave]

MJ 6.1., too

so what's the difference?

[sterile pendant]

not many samples yet for it

but flux is definitely lightyears ahead overall

[dry wave]

and MJ options to control and modify images are a joke compared to the huge ecosystem of SAI

[bitter hearth]

need to check their code at some point
because its very easy to get ELO systems wrong
ELO is quite tricky get right in practice

[dry wave]

so I think with Flux MJ is behind. Not just because image quality is worse than Flux, but also because Flux is open source and will develop faster neer features than MJ

that's said it might be possible that MJ 7 then will strongly outperform flux. We will see. I won't say the battle is over, just that MJ is behind now and have to catch up

[errant dust]

Myeah... the thing about these so called tests is they are pretty biased in what they test IMHO. Let me give an example. Here is Flux asked to make an impressionist oil painting of two young me playing chess in a park.

[sterile pendant]

no, it's really not. everyone uses it and it's a decades old system stemming from the chess world. you're thinking of videogame elo where it's you with four other idiots on a team and a bunch of chaos. when it's just 1v1 elo, it's hard to screw up

[mortal mesa]

MJ has the userbase to levrage

[dry wave]

this is correct and usually is a huge advantage for MJ

[zenith hemlock]

With the release of flux fine-tunes like RealVis, MJ will fall further behind

[hallow lion]

How did MJ become so popular?

[dry wave]

because MJ is optimized to death to please the user

[errant dust]

It was the very first open to the publilc

[sterile pendant]

it was high quality, way ahead of the rest of the competition and you didn't have to run it locally

[errant dust]

of all lthe image generators

[mortal mesa]

type some junk into a webpage and it just works

[zenith hemlock]

easy to use, artistic look, and it was free at first 😄

[hallow lion]

Ah, the right place at the right time.

[bitter hearth]

its more complex than this

[errant dust]

It came out before Dall-E and before SD

[zenith hemlock]

Actually dalle came out first

but yeah it came out before SD

[errant dust]

No

[dry wave]

Dall-E was first

[zenith hemlock]

isnt first dalle released in 2020?

[hallow lion]

Who's behind MJ?

[errant dust]

not to the public

a few journos allowed ot test it don't coount

[bitter hearth]

MJ may have been first yeah

[bitter hearth]

I can't remember fully the order things came out

[zenith hemlock]

ooohh the test version

yeah I remember that time

[errant dust]

I can. I wrote articles about it in late 2022

[zenith hemlock]

youre right

[sterile pendant]

dalle was first, but the whole ai fever didn't really hit the limelight until MJ

[errant dust]

Dall-E 2 was the first model opened to the public and it came after MJ was opened to the public

byt about a month to be fair

[hallow lion]

I never made an image with MJ. Generating on Discord is just more confusing than Comfyui for me.

[dry wave]

anyways. MJ was always ahead in terms of aesthetics. Not much in prompt understanding, but it got aesthetics better than its competitors for years

[hallow lion]

And the results are meh.

[dry wave]

and currently MJ is really nervous after flux was released

so they themself see the problem

[hallow lion]

The company that cracks consistency next will be the next MJ.

[sterile pendant]

even their current versions are basically still SDXL level prompt understanding. they're probably working on a dit version now to compete with sd3 and flux

[mortal mesa]

Puffy Pope jacket thing got MJ the eyes

[bitter hearth]

I definitely think MJ was ahead at first

[hallow lion]

Real consistency not cobbled togetehr with IP adapters and SAM masking and all that jazz.

[bitter hearth]

but not as much these days

[errant dust]

I tried both at the time. Honestly, DE2 was better at adherence even then, but its results were realy weird. It often looked like a cut and paste of imagery. MJ looked like art, but it too had massive weird issues. Like it was ful of image noise it could not get rid of. This persisted into MJ v3. MJ4 was a MASSIVE upgrade and game changer

[bitter hearth]

MJ has a threat from the 16 channel VAE models like SD3 and Flux
but also SDXL fine tunes have still been steadily improving over the last 9 months
even if they can't always win on fine details, the more recent SDXL model versions can do very good compositions

[hallow lion]

Dall-E is like eating paper, Mid Journey is like French Cuisine with snails and stuff. Stable Diffusion is like a normal meal.

[sterile pendant]

MJ uses a bunch of that under the hood if you're using img2img, but the end user doesn't have to deal with the chaos. the majority of genAI users are not tech savy and are ignorant/undereducated about it all.

[errant dust]

Not anymore. DE3 is king for a bunch of things today. Far from all of course, but in some areas it is unrivaled still

[dry wave]

naaah, Dall-E 3 is still one of the best models

[hallow lion]

lol

[dry wave]

it's still better in prompt understanding than Flux, although worse than Ideogram

[errant dust]

it really is

[dry wave]

it's really bad in photography.

[bitter hearth]

I agree Dalle 3 is the best or one of the best

[dry wave]

but on art stuff it's totally competitive

[errant dust]

yes, strengths and weaknesses

[hallow lion]

I tried some of my promts from ideogram with flux and all the flux ones looked way better

[dry wave]

I wonder if that is not even on purpose. Like maybe they don't want DE3 to make realistic images to avoid all these problems like fake images and so on

[sterile pendant]

dalle went a more artistic cartoony route with their model

it was originally made for art related things in the first place

[dry wave]

Ideogram is horrible in aesthetics. But it gets even the most complicated prompts right, even if they look like photoshopped afterwards

[bitter hearth]

it has been confirmed that it is censorship
because there are jailbreaks out there

[sage burrow]

and a new pricing structure, to keep up with the competition 😉

[bitter hearth]

I don't mind their pricing structure, I think we need a mixture of open and closed source

[mortal mesa]

there are no actual problems with fake images, interesting that, i feel protected

[bitter hearth]

I just want at least one open datacenter sized model
we have it for video (opensora 1.2 is 67GB VRAM) but not for images

[dull star]

I remember DALLE3 making discord images or something, it was hilarious

is GPT4(?) captioning still the best

[sterile pendant]

well that's why big API models like dalle3 do so well, they have all the NSFW shit trained into them, but mostly keep you from circumventing it. that way, you get good creations with good anatomy(usually) and they don't have news sites and people rioting with pitchforks over NSFW content generation every other week

[bitter hearth]

InternVL2-Llama3-76B
maybe

[mortal mesa]

i did a cost analysis, no it isnt

[dull star]

and what about knowledge

[sterile pendant]

if they publicly released the models, people would jailbreak it within a day and start chugging out pron

[dull star]

fr

[bitter hearth]

yeah that happened to flux dev already

[hallow lion]

PonyJourney

SmallhorseDalle

[mortal mesa]

PornHubAI you know its comming

[hallow lion]

for sure

[sage burrow]

MJ tried to impersonate the whole lora thing with their "tunes". The tunes sort of changed the theme a bit, but, still nothing like loras.

[sterile pendant]

well like i've said before, there are tons of laws in the works for pretty much every major country that are going to seriously kneecap NSFW genAI content

[errant dust]

I showed this the last couple of days, but it illustrates one of its real strengths. Both are takes on the same theme:

1. create a cup of cappuccino in which the milk foam and coffee swirls form a map of the world
2. create a large pizza in which the toppings form a map of the world.

Here are two DE3 results, and the next are by Flux (though others aren't a lot better):

[hallow lion]

ud think theyd be ok with ai porn

[errant dust]

This plays clearly to its strengths but is a decent example

[hallow lion]

its like regula rporn but no one gets hurt

[dull star]

gpt4o just gets heihachi mishima (from tekken) right first try

[bitter hearth]

was this the distilled flux?

I think distillation had bad effects on understanding and latent space

[sage burrow]

My personal fave is Claude

[bitter hearth]

the other distilled models from SDXL that released earlier in the year had this effect that flux has where they lock on to certain concepts and they cannot interpolate smoothly

instead they jump from one concept to anotehr

[hallow lion]

Im startign to get the same feeling when I sit doiwn to write a promt as I did before when I drew on paper. Has AI become the new pencil?

[errant dust]

Here is Flux Pro:

[bitter hearth]

with a distinct representation of each concept but no smooth interpolation in between

[sterile pendant]

like here in the US, they have the DEFIANCE act that's just the first of many to come. this act is mostly centered around deepfaking, but guess what it also includes for the first time: ANY computer aided manipulation, which also means photoshop. so if people create some kind of deepfake of a celeb, if the likeness is even within 50% of the person, a jury will likely agree it's a deepfake of them and that's that. so basically, every other creation will potentially look like someone and they can then potentially file a lawsuit over using their likeness for pron.

[sage burrow]

There's a reason my niche is characters which cannot be mistaken for real people. (I planned ahead)

[bitter hearth]

ah ok that's no better

[errant dust]

I can show the reverse where Flux knocks it out of the park and DE3 fumbles badly.

[dull star]

OpenGPT 4o tried

[errant dust]

The point is that these generic tests claming one is obviously better at adherence to another, even that simple an idea, is really a big "depends"

[hallow lion]

I wodner if there is a pron film out there titles "Mortal Cumbath"

[errant dust]

There are things it can't do at all, not to save its digital life, so if not tested...

[sterile pendant]

90% of it is that people are terrible at prompting, even if they think they aren't. good ol' dunning-kruger effect

[bitter hearth]

there are sites where you can compares hundreds of examples instead of just a few
that's why I don't really rate midjourney that much
I saw it lose to other models too many times for it to be supreme

[mortal mesa]

see we have laws to protect us, it would behoove the model makes to not pretend to be law enforcement

[errant dust]

By all means: impressionist oil painting of two young men playing chess in a park. They are each sitting on a bench and the board is on a stone table. If you think my prompt is the cause of the failure, rewrite it to achieve an impressionist oil painting from that. In Flux

I will say the prompt is pretty straightforward

[sterile pendant]

then you'll have no models because the people training them will be slapped with lawsuits into oblivion. the people training loras of ana joy taylor will be slapped with lawsuits (saw people spamming a flux lora of her) into oblivion. see where i'm going with this? it's just not going to be worth the risk for even dudes in their basements training their pron loras on celebs. it will also become illegal to even share or distribute these types of photos, so people sharing their creations will be slapped with lawsuits into oblivion

[torn wharf]

I found that with them too but it's more prominently in flux and I think it's because of the size

[bitter hearth]

to be honest
I said this the other say but I am happy to just take the benchmarks from the papers
they tend to calculate scores like FID and Clip Score on 30,000 images at 1,000 steps
I kinda doubt we will get a test that is better than the standard ones the papers use

[sterile pendant]

wasn't directly talking about you or anything, just saying in general about the topic

[errant dust]

Ah ok

[hallow lion]

But even in photoshop you cna copypaste a celebs face on a pron image lol

[sterile pendant]

and with the DEFIANCE act, that will also be illegal to share

[errant dust]

I was simply showing an obvious shortcoming of Flux, and how the test can talk endlessly about how perfect a model is, but if none of those shortcomings, in any variant, ever come up.... it is easy to ace the test

[hallow lion]

and with some better PS skills you can blend it pretty well. AI just does it better and faster

[torn wharf]

New laws are being made for new tech. Image diffusion isn't photoshop

Old laws still apply
Don't harass people

[hallow lion]

yeah I mean common sense... we all have it right?

[sterile pendant]

the DEFIANCE act includes any computer aided manipulation, which is a big umbrella term that also includes photoshop

[torn wharf]

I like to hope

[hallow lion]

sound slike those evil acts the RIAA was pushing against music sharing

[sterile pendant]

well this is different from you downloading your favorite nickelback song...

[bitter hearth]

but if a shortcoming is rare enough that it doesn't come up in any standard tests and also doesn't come up in any of the handful of image arenas
then this is likely rare enough that its kinda okay that the models can't do it?

like they don't have to be perfect

[errant dust]

Asking for an art style is not THAT rare

This flub is not restricted to impressionism or a specific artist

even its 'comic-style' output all looks exactly alike

[sterile pendant]

flux isn't very trained on art

[errant dust]

I tend to think it was deliberateloy shut down on it to try to sidestep the lawsuits appearing

[sterile pendant]

because almost all of the images of famous art you know are owned by some museum or another and they have to give permission for them to be used in datasets

[dry wave]

Art is a shortcomming in Flux for sure

but it's one you can easily solve via finetuning

something that is not even possible with MJ

[errant dust]

fintetuning you mean, not art

[dry wave]

yes

[hallow lion]

IF you were to estimate how big is the MJ model? and the DALLEE one? XD

[errant dust]

There is literally no better model for art than MJ. DE3 is excellent too for many though

even brillliant

[uncut river]

flux schnell

[errant dust]

ask it for a new vision of Mona Lisa by Picasso in the style of La Guernica and DE3 performs brilliantly

[uncut river]

weird abstract comic style of a fish-guy fantasy creature, lines out of place, amateur style, many lines in lines slowly decreasing in thickness like a fractal maze

[bitter hearth]

its very hard to tell exactly but there was a quote from the MJ team once saying something about how big it was

it was 10s of GB anyway

[hallow lion]

:0

[mortal mesa]

whi is alowed to bring complaints, cuz i went on a date and this girl was not the same person

[errant dust]

The CEO of MJ is on record saying it could handle and imitate over 4700 artists

[bitter hearth]

you can tell the MJ UNET is huge because its compositions are very strong, its like the opposite of deepshrink effect

[sage burrow]

That's why I want an SD3 and Flux workflow 😉

[bitter hearth]

that could be done with SDXL sized params, that's more about what you train it on

4700 is a super low number on the scale of these models

[dry wave]

Flux is either not trained on artists at all, or these images are badly captioned

[bitter hearth]

one thing about MJ is they are the only model out of any that was trained on a significant amount of hollywood movie stills

[dry wave]

if you try in Flux to prompt for artists "by [xyz]" you will see that only CLIP-L is reacting on that while T-5 does nothing

so either they haven't trained on artists because of copyright issues, or they have labeled their training data fully automated with some captioning tool that forgot to add artist and style information

[bitter hearth]

since its a good team I suspect it was a choice

to evade lawsuits

[errant dust]

yeah.... I don't think so. It is low compared to what they CAN do, on individual models maybe. But SD XL is still a 2b model, and you'd need a very specific LoRA to do what i said above. They can do this innately, with none of the very advanced legwork

[dry wave]

I found SD also really good with artists styles. In particular the old SD 1.5

[sage burrow]

Cascade via H.F lol

[dry wave]

if you compare SD 1.5 -> SD 2.1 -> SDXL you see that the models less and less react on artists. Guiess they are removed from their training data

[bitter hearth]

being in the base model is very different to being added in a lora after, being in the base model is much more efficient
SDLX has way more than 4700 concepts, this is not a lot

[torn wharf]

it turned the mona lisa into a generic christian depiction of mary

[sage burrow]

WOOHOOO I got approved for 200 glif tokens today! What should I make? 😄

[torn wharf]

obviously balls

[dry wave]

but in the end I'm happy that I can just train Flux on arbirtrary artists

[bitter hearth]

I agree with you that MJ is a large Unet but I don't think the number of artists in the training data is evidence of Unet size

[dry wave]

even the ones MJ does not know 😉

[torn wharf]

t5 is interesting becaues it wasn't trained on image pairs so you can't just invoke art styles with an artist name

[dry wave]

you could

if you train for it

[torn wharf]

yeah. same with sd15. lots of styles weren't trained into it but they were there because of clip-l

[dry wave]

the special thing on CLIP is that you do not have to train the unet/transformer and still get the artstyles for free (although in much lower quality)

[sage burrow]

Flux Dev via H.G. It has that flux blurred skin look 😦

[errant dust]

Ok, then tell you what. Here is the prompt. Exactly. "A mesmerizing painting by Pablo Picasso, reimagining the iconic Mona Lisa. In this version, it is painted in the style of La Guernica, with its signature shapes and style" and here are two sample images I made with DE3 about two months ago.

https://www.bing.com/images/create/a-mesmerizing-painting-by-pablo-picasso2c-reimagini/1-666da9644c5547d098fb072a5db3df27?id=Ya4w9XbrKrkHs09JQKUeeA%3D%3D&view=detailv2&idpp=genimg&thid=OIG2.I8mNm8PZOEJTvhS4CXAZ&form=GCRIDP&ajaxhist=0&ajaxserp=0

[dry wave]

it might be that Flux is not trained on much arts due to copyright. But I also have the feeling it just wasn't their priority

like they wanted to show that their model can do anatomy right and good photorealism

[bitter hearth]

it could be a priority thing also yeah

[sage burrow]

Flux Dev via H.F.

[torn wharf]

yeah im' guessing flux 2 is coming sooner than later. these guys made sd 1.4 then followed it up

[mortal mesa]

Uh Flux does Picasso, not sure thats what your saying, maby words i didnt read

[errant dust]

I'm not arguing that, since it is obvioous, regardless of their reasons. I was simply pointing out that these ranked tests declaring one model to be the overhwleming best, is a big 'depends' according to your choices. Some things it cannot handele at all, and is not a case of one accidental RNG failure

[sage burrow]

When I use Claude to enhance my prompts, it does make Flux create better "Dali"

[torn wharf]

i've had good dali prompts with long elaborate LLM descriptions of his paintings

[bitter hearth]

yeah that's fine I agree with you there
the original thing I was saying was that Midjourney was not the supreme best

because there are situations where it is not

but this applies to the others too

[errant dust]

I never said it was. It is supreme best in some things, but nowhere near all

same for DE3 and no doubt for Flux too

and Ideogram

[bitter hearth]

yeah I don't think we disagree really

the only thing I was saying about the artists is that
it doesn't necessarily imply a larger Unet

[sage burrow]

Is Meta not in the running?

nm

[errant dust]

I haven't tested it, but I am not ruling out anyone. I was simply pointing out that these charts declaring one as the obvious best are to be taken with a major grain of salt

[bitter hearth]

I do think MJ has a larger Unet, due to image quality
but I don't think the artist knowledge necessarily requires that

[mortal mesa]

meanwhile in china no one is using thoes

[sage burrow]

I used to love to put Picasso's different, lesser known styles into MJ to enhance my horror 😄

[bitter hearth]

what I think is that these models have a really large number of concepts
possibly over 100,000

but also the Llama 3 paper showed that parameter count and training tokens are way more disconnected than we thought

Llama 3 paper kinda proved over 100 papers wrong
because they gave Llama 3 8B 15T tokens and it did ok

[errant dust]

The fact is that I originally asked DE3 to redo Mona Lisa in PIcasso's style, but it was not what I wanted. I then told it to do it imitating the style of La Guernica, and I was shocked at how well it did

[bitter hearth]

so you can really squeeze stuff into low parameter counts

[sage burrow]

OK new glif idea, I'm going to make a Picasso SD3 one (perhaps a flux one as well). But I'm going to try to include more than just his well known cubist style

[errant dust]

well, he is essentially defined as a cubist overall. and I think La Guernica really highlights that iconic look

The story behind that painting is quite fascinating, and he used to bring in interested parties to see its progress

and discuss his vision

[bitter hearth]

this sort of thing I think midjourney is amazing at https://old.reddit.com/r/midjourney/comments/z7fzn1/what_happens_when_you_mix_ancient_egypt_with/

or this one https://old.reddit.com/r/midjourney/comments/1c35tvs/ancientpunk_egypt/

[errant dust]

There are a lot of things. I saw one where they imagined ancient egyptian paintings but with modern model ideas. So that known stye and look, but with teens on cell phones. it was hilarious and brilliant

[mortal mesa]

color me not impressed they dont look real

[bitter hearth]

it loses on realism to most SAI models

that's just how MJ is

[mortal mesa]

my prefrence i suppose

[bitter hearth]

its for cinematic, art or painterly styles

[errant dust]

real? It wasn't trying to be real.

That has graphic novel look

[mortal mesa]

yup no personal use for that

[bitter hearth]

MJ doesn't have 16 ch VAE yet anyway
so they couldn't match SD3 or Flux on photographic until they get that

[torn wharf]

when i prompt for alternat comic styles they come out beautifully. just dont' do generic prompts. ez

[errant dust]

ummmmm

[bitter hearth]

for now VAE is the biggest divider

[errant dust]

https://www.tomsguide.com/ai/ai-image-video/i-tested-flux-vs-midjourney-to-see-which-ai-image-generator-is-best-heres-the-winner

[torn wharf]

tom's is a click bait rag. you've been baited

[errant dust]

So clearly they can match for photorealism

[bitter hearth]

this is compared to the wrong one
its SD3 2B that is good for photorealism not flux

[torn wharf]

i can see the washed out colors from the older vae on mj easily personally. I think i could blind test which is which , mj and flux, at a rate way higher than random guessing

[errant dust]

Says Flux in this comment, no?

Besides, how exactly would you know if they have 16 ch VAE or not?

[bitter hearth]

well Flux has a 16 ch VAE but its been overtrained on a non-photographic style

whereas SD3 hasn't

so SD3 is the better comparison

[torn wharf]

MJ would hype it soon as they got it

histograms would show for sure i'm betting

[bitter hearth]

there are some flux fine tunes on reddit that do better for photographic

[dry wave]

Flux is really difficult to prompt, or we just haven't figured it out yet

[bitter hearth]

yes you can likely test for it with a histogram

[dry wave]

I sometimes got images from Flux that were extremely realistic / couldn't be distinguished from real photographies

the images in the article by flux are all really bad

[errant dust]

None of this changes the fact that MJ is perfectly capable of producing outstanding photorealistic imagery

and there is nothing to suggest it is lagging behind

[mortal mesa]

the uniformity has gone to far, what was first good has become its Achilles heel

[errant dust]

I have no idea what that means

[mortal mesa]

cool

[torn wharf]

i'm impressed at how fast you can flip flop. the other day you were saying flux and sd3 photos were bad because they all have bokeh blur backgrounds. but here you are calling MJ impressive and link a collection of examples all with bokeh style photography

i guess blurred backgrounds are realistic enough photos when it suits your point

[bitter hearth]

I don't think MJ does compete with SD3 in photography
a comparison to flux isn't really going to change my view about SD3

[errant dust]

I think that for you to be able to make that comment, you'd have to be using MJ and making comparisons

[bitter hearth]

I've compared hundreds of images

[torn wharf]

it wasn't just a one off problem either. you were cutting into a few of my generations calling it that "god damn blur" or something. maybe exagerating but you were quite hung up on the blur

[bitter hearth]

there are sites that have big comparison tables

[errant dust]

so you have compared hundreds of images? Or these site have comparison tables?

[bitter hearth]

both? I don't see the contradiction

I compared them by looking at the comparison tables

[torn wharf]

ahh. here come the strawmen attacks . right on time. subtleone doesn't like to have debates they can't win, so it often comes to this. arguments from authority. slippery slopes. people got their bingo cards out? https://yourlogicalfallacyis.com/

[errant dust]

I have not seen anything special in SD3's human images, other than the mutant feet and legs and more

[bitter hearth]

its okay there isn't drama

the thing about SD3 is that it has the 16 channel VAE, which currently only flux also has, but crucially SD3 has a more general training so it hasn't been fine tuned out of a photorealistic style like flux has

[errant dust]

Yes, all this talk about how specail its VAE is has no bearing on what I said about its mutant humans

[bitter hearth]

but this won't be the case within a few months because Auraflow, the next one from Pixart, and the OMI model are all planning 16 ch VAE

[errant dust]

I also asked how you could claim knowledge of the VAE MJ uses or not

[bitter hearth]

cos what you said was just an opinion so I can't really respond to it
regarding the VAE, 4 ch and 16 ch VAEs look very different

[errant dust]

I'm still waiting for how you made the claim. What images in MJ show difinitively it is using a VAE of 4ch

Yeah, look I think it is exceptionally sketchy to make statements on the technical structure of MJ with zero information or data to back it up. Other than possibly wishful thinking to back up the statement it is worse than X, Y, or Z

[bitter hearth]

VAE is for every image its not for one image or another

[errant dust]

Each genration of MJ, 3, 4, 5, and 6, is trained and built from the ground up

and MJ6 is quite new

[bitter hearth]

it is backed up we already said before
4ch VAEs and 16ch VAEs don't look the same
and SD3 looks a lot better for photography
but since you don't agree with that, you aren't going to find that convincing

[errant dust]

You're right. I don't find it convincing at all.

[torn wharf]

Difference of night and day especially on photography

Some people's eyes won't be as sensitive to color detail as others

"the human eye can't see over 30fps" is another good one

[sterile pendant]

eyes aren't scanline sensors. while they do stochastically update individual "pixels" as photons raise their voltage threshold to the point to fire the signal, there's a cooldown to how frequently those ion channels can recharge. so yeah, technically, on a per "pixel" (cone/rod), there is an fps limit lol. but since they stochastically update in your head image, using all kinds of layered algorithms similar to genAI, your brain can fill in the blanks on things like using generative fill in photoshop. you have two very large holes in your FOV at all times where the vein/nerve bundles are, yet your brain fills in those holes

but in practice, anything over 100fps is a cointoss in a double-blind a/b test

[torn wharf]

Oh Yeh I love the neurology of optics. Some brains process the signal better too

[sterile pendant]

yep, but not much more

usually within some kind of tolerance range like +/- 15%

[torn wharf]

The one I always point at is ufo test. We can see lateral motion VERY well. On account of being apex predators

[sterile pendant]

yeah, but that's due more to it breaking that stochastic updating i was talking about. it's actuating more and more cones/rods and the brain is lazy and doesn't like having to update that much data at once

ever do the still stare thing where your peripheral vision starts to turn black? your brain is lazy lol

[torn wharf]

Yeah there's lots of physical stuff. Persistence of vision is literally the retina chemically retaining the image

Generative fill is fun to find too. There are lots of tools to find your Blindspot. That one always blows a few neurons out when I do it

[sterile pendant]

an engineering analogy is actually more in line with a capacitor. once it reaches a threshold voltage, it will trigger the signal, but until it does, it won't fire

but it decays rapidly

if not constantly being triggered by photonic energy

[torn wharf]

Yeah cool analogy. That's a gooder

I love rubbing my eyes in the morning, then seeing kaleidoscope. I call it degausing my eyes

[sterile pendant]

it's also why all eyes saccade very subtly, it's our brain's way of refreshing pixels

[torn wharf]

Also sub pixel rendering my theory

[sterile pendant]

the micro saccade needs computers to see, but there are macro level jiggles that happen that we can notice as well

[torn wharf]

It's a way for the simulation to hide the pixels

[sterile pendant]

in engineering classes, we talked a lot about the biologic analogs of common circuits and whatnot

[bitter hearth]

they are starting to make really good motion interpolation neural networks
which implies somewhat that its possible for the brain to do that too 🙂

[sterile pendant]

but anyways, our brains do some kind of similar shit to genAI image creation. you don't store pixels in your brain, the data is stored more like how shit works in a diffusion model, but in an obviously far more complex way, and then your subconscious supercomputer "renders" shit out for you on the fly

[bitter hearth]

our "latent space" yeah

[sterile pendant]

the one thing we don't have is back propagation, so we aren't a true analog to things like diffusion

[tough viper]

hi can someone help me figure out why i can't load a model in comfy?

[sterile pendant]

that folder should be the unet folder, not the checkpoints folder

[tough viper]

ohh

[sterile pendant]

move it and refresh comfy

[tough viper]

do I need to install clips?

I thought this one was u only need the checkpoint

https://civitai.com/models/628682/flux-1-checkpoint-easy-to-use