💬｜general-chat

[solid kindle]

is it sufficient ?

i mean the VRAM

[silk condor]

Well I am running an 2080ti atm with 11 GB and been testing to run local GPT and works fine.

I do use stable diffusion 1111 everyday on it, but I am not able to play and do it at the same time.. I CAN DO.. but there is frame drops and "lags" here and there.

I want to run my own Local GPT and stable without compromission my gaming PC 🙂

[inner plaza]

hi guys, can i reccomend paid services here?

finally found a good background removal API that is cheap

[balmy hemlock]

Hi there, anyone comfortable with ComfyUI and mask nodes here 🙂 ?

[wild girder]

anyone comfortable with diffusion 111? If there is someone,

[balmy hemlock]

Hello there! Aynone comfortable with ComfyUI / Mask and inpainting nodes ? Thanks

[low moon]

No. Use comfyui.

Comfyui or get out.

🤣

[fervent thunder]

Aint nobody got time for that 😂

[fervent thunder]

it was honestly easier for me to learn comfy

than work out what Fooocus was doing

[bleak matrix]

Good morning, everyone! 🌤️ How are we all this beautiful day?

[fervent thunder]

hello

[solid kindle]

does GPU speed matters for image generation assuming both GPU have a same VRAM ?

[junior urchin]

When were we expecting sd 3.1 to come out 👀

[hallow nacelle]

Is there an extension that shows the filename of the image used for img2img

[alpine hound]

How to create a.i. Images using stable diffusion on discord ?

[junior urchin]

yes, true

Read about 3.1 "coming in the coming weeks" a few weeks ago, but then it went eerily silent

It was the questionable "contacted by a sai member on email" and "I have a contact in sai, who told me this"

Had hoped it had flavor this time

[trail lion]

Yup won't be able to blame the community this time, but perhaps the fear of falling behind now that other alternatives are surfacing. Reminds me of marvel trying to blame the fans for making crappy movies that flop, like Madame Web

[junior urchin]

We did that the last time

then gaslighted by them too

[cedar salmon]

you were saying things like that

ok, ya i was refering to the gaslighting

[junior urchin]

yeah, sai gaslighting us, yes

saying that it worked and everything

[cedar salmon]

rewmember you trying to put all the weight of their terrible performance on the comunity

LMAO

go support disney

[junior urchin]

While on the topic of 2B; Didn't they say they already had 4B, 6B and 8B ready?

At least functionally

not available, but ready

(or cancelled/deleted)

[cedar salmon]

i still blame Ella

trust and saftyed themselves out of the market

Comfy had a 4b in development, verified, not anymore though

i dunno your weird, no ide of that connection or what your trying to say

ide blame capitalism though

mmm ya i do see much of that

they are still on track to be the first to fail

of the big boys

[cedar salmon]

LOL Ella Erwin twitter/linkedin profile says AI safety @ Meta, guess she got the boot from stability hahaah

idjots

[unreal sonnet]

hello :)

[quartz night]

is there any update from stabilityai on sd3.1?

We havent had any news from them in 2 weeks.

[cedar salmon]

please hold restructuring

[sand trellis]

Hi,

A bit lost with the new flux models and all new extensions on comfyui. What is the best workflow actually ? Is there one which is stable and have all the up to date functions like ip-adapter, etc... ?

[abstract quarry]

I think it's too early yet for ip-adapter or controlnets in Flux

there are early versions of them, but they are still undertrained and won't work as well as we are used to it in SDXL

[sleek otter]

Here is a Flux img2img workflow which includes Florence2 (w/f embedded in PNG) #🆕｜sd3 message

[sand trellis]

Alright thanks

Is there a place where there is frequent updates on everything related to flux ? like a discord or else

[tender nebula]

Yeah

[sleek otter]

https://github.com/black-forest-labs/flux

[scenic robin]

does anyone know a discord bot that uses stable diffusion that i can add to my server and mess with?

idk if its against the license for that to be allowed

id assume not

stable artisan looks good

[unreal sonnet]

im not sure but i think that's ilegal

[crystal echo]

A photorealistic image of a sturdy dolly made of steel, with an adjustable platform that securely holds 17 and 27-gallon totes. The dolly has a retractable handle that nests within the frame, vertical posts for stacking support, and four swivel casters with locking mechanisms. The platform has a non-slip surface to prevent totes from sliding.

[desert dagger]

@shell tendon https://arxiv.org/abs/2409.00587

[shell tendon]

whoa that is way cool

[desert dagger]

yeah. thought you'd be interested

[shell tendon]

definitely, thx for sharing

5090 better have like 200k cuda cores lol

[desert dagger]

heh. not for training on a home machine

[hard swift]

yo

[iron vigil]

Nice try

[orchid ember]

Greetings to all AI and SD specialists.
I came across the following question. In SD is it realistic to make ONE image from different positions ? For example: Girl: Front, back, side view. At the same proportions, appearance that would not change, I want to create a 2d game, for this I need the same sprites from different positions

[winter shoal]

you can use controlnet with certain loras to help you to achieve what you wish. You want to build an openpose skeleton for the character and then use that in the controlnet, that will most likely build what you wish for with some clever prompting.

[orchid ember]

Is it possible to create a ||naked ||character and then on his body to make different clothes ?

[worn thicket]

what does score_9, score_8_up, score_7_up, score_6_up

in ai generation

sorry

wrong chat

[winter shoal]

propably doable. its just about you figuring out the workflow. i tested it earlier. i worked on some pixel style characters a year ago but just as a proof of concept at that point.

[orchid ember]

Can I write to you in private messages?

[winter shoal]

sure. this whas my reference back then. https://www.reddit.com/r/StableDiffusion/comments/1180fls/sharing_my_openpose_template_for_character/

[abstract quarry]

you don't need controlnet for that. You can achieve that simply by prompting or by training a lora on such example sprites

SD is already trained on such spritesets, so it has a rough understanding what that is and can create it

Flux is probably much more powerful for that

[orchid ember]

Flux doesn't have models that copy looks from one image to another.

[abstract quarry]

you don't need that

[orchid ember]

I was told to use 1.5 SD, I'll try to do something like that there, the question remains open about clothing, but it already seems to me need to discuss with experts in the field of gamemadev

[abstract quarry]

clothing on off is also possible. Best to use loras for that

make a few example images of characters next to each other with clothes on and off, then train a lora on it. Afterwards you can apply that on your images

just copy&paste your character on the left, inpaint on the right

SD and Flux both will make the character on the right exactly same looking as on the left if you prompt for it

just as an example, look at civitai and search for models with "on/off". You will see what I mean

[orchid ember]

@abstract quarry Alas, I tried once to teach LORa I failed, probably because of location (I am from Russia) I did it both Russian and English language guides, but I got an error, but the idea is good, I will have to take it into account, thank you

[bleak matrix]

Good morning, everyone! How are we all this fine, charming day?

[remote trout]

Cześć wszystkim. Jestem tu nowy. Czy jest ktoś kto mógłby mi pomóc w pierwszych krokach w programie ?

[low moon]

It's not Friday but it's okay.

[scarlet smelt]

Hey i need some help. I try to install Automatic1111. When launching web ui user . bat, i encounter this error : "could not find a version that satisfies the requirement torch : 2.1.2 no matching distribution for torch 2.1.2

[scarlet smelt]

i should have read the entire message in the command box. It appears that i just have the wrong python version.

i will try again by donwgrading it

[vernal swan]

If you find a solution to this please tell me i have the same problem

[scarlet smelt]

i had a wrong version of python

i try to reinstall it with the recommanded version

[vernal swan]

i actually have the recommended version but still the same problem

DDL load failed while importing cv2 ERROR

[bleak matrix]

At least you can look forward to it, eh!?

[warm junco]

Come to #🤝｜tech-support with a full cmd log

[warm junco]

Same for you, in #🤝｜tech-support we can help

[feral pike]

prompt gen recommendations for automatic1111 ?

[feral pike]

@warm junco any ideas?

[low moon]

SD15 = very elaborate and loads of negaitves. SDXL = less negatives still elaborate. Flux = elaborate, short novel - no negatives.

[feral pike]

what?

i am looking for SDXL

some extension that can help with prompts

[low moon]

there are so many

promt enhancers

[feral pike]

yes thats why i ask

which one is good

[low moon]

Florence 2

[feral pike]

https://huggingface.co/microsoft/Florence-2-large

this?

[low moon]

yeap

[feral pike]

how do i use it in automatic1111

[low moon]

Try to convert to comfyui

[feral pike]

💀

[low moon]

I know

i waited 3 months too and used A1111

eventually

i gave in

[feral pike]

i will install it now

[warm junco]

There is dynamic prompt extension and wildcards extension is pretty nice

[feral pike]

what is the name of extension?

[warm junco]

Or use any other local LLM model and make your own prompt gen

[warm junco]

Dynamic prompts
And the other one
Wildcards

Wildcards is not really for prompt gen but for getting many different images pretty fast

[feral pike]

oooo

let me try , thanks

this automatic webui has soo many options , its so cool!

https://huggingface.co/spaces/gokaygokay/FLUX-Prompt-Generator

this looks cool for prompt gen

it says its for flux

but i guess should work for anything?

[low moon]

Also if you really like A111.. why not try forge?

Its like A111 but better

[feral pike]

i thought a1111 better

what u mean

[main junco]

What is the best way to make similarly looking face - like a brother.
ip adapter?

[feral pike]

how do u use loras in webui?

[pliant iris]

What does Hires.fix and refiner do?

[quartz siren]

Hires fix upscales image basically and refiner basically refines the image. I don't really recommend the refiner, as it usually makes things more smoother which might remove some detail. Hires fix is good though.

[pliant iris]

Why is hires. fix only in txt2img. Is there way to do from img2img?

[feral pike]

more steps or more words , which one is harder for the model?

[desert dagger]

more words

[feral pike]

what happens if i do 150 steps

do i get a better image

whats the sweet spot?

for SDXL models

[main junco]

Guys how does fooocus work? Is it soemthing more that model?
Does it use LLM to expand your promt?

[fervent thunder]

ye it does

[main junco]

something like GPT2?

[fervent thunder]

yeah something like that

I personally use GPT 4o for all prompts

or the mini is fine even

not yet
is this the new llama or a different meta service?

[main junco]

Hvae anyone tried to use ip adapters to make similar faces? I want to try implement face obfuscation.
Become invisible from face recognition software

\

[fervent thunder]

yeah it works to a certain extent

[main junco]

What is a best approach to it? Ip adapter - and swap face?

Or something else?

[fervent thunder]

its controversial because people have different viewpoints
my opinion is that if you really want good results you should do fine tuning for this
either lora dora locon loha lokr dylora
or a full checkpoint if budget is huge

you can still combine this with ip adapter, face swap and control nets anyway

[main junco]

But how to train a LORA to make some changes in appereance

Like I want to get faces similar to mine

but not quite

Also does flux support fooocus?

And does flux have ip adapters or similar face swap capabilities?

[fervent thunder]

most people seem happy with the pre-made lora makers like the replicate one
it is likely that it would be fine for you

I'm working on a personal pytorch framework for fine tuning, but my goals are different

[main junco]

Are there any that you can run locally?

Without paying )))

[fervent thunder]

koyha, SimpleTuner and OneTrainer
are probably what you want

there is also Diffusers

[main junco]

Even for LORA?

3080 12gb )

Or much more hefty?

[fervent thunder]

its very situational

[feral pike]

hyper models are cool!

especially with hires.fix

[fresh ruin]

which one is better for SD, 3090 or 4080?

[umbral solar]

So what's the best current model for logo generation?

I've tried most of them, but not sure if the tech isn't there yet or it's just my prompting

[neat cypress]

guys i need your careful analysis and your opinions which check point looks best

trained style images as a single grid : https://huggingface.co/MonsterMMORPG/3D-Cartoon-Style-FLUX/resolve/main/Trained_Style_Images.jpg
trained checkpoints grid comparison : https://huggingface.co/MonsterMMORPG/3D-Cartoon-Style-FLUX/resolve/main/Trained_Checkpoints_Massive_Grid.jpg

[quartz siren]

Depends on what you mean by SD but usually 3090 is better. You can usually run bigger models and 24gb vram is a very nice spot.

[quartz siren]

With text? Flux is definitely the best, without it, I guess logo loras for sdxl.

[umbral solar]

Any hosted sites you recommend for logo loras?

[quartz siren]

sadly no, I am sure there probably are but I just run most things locally.

[silk condor]

I am currently running an ESXI that currenty hosting an home file server, game servers, image generator auto1111 .. But I have trouble to get local GPT to run with an UI due to the issue that you can run GPU and VM together..

Would it probably be better to run a Windows Server that runs Auto1111 and Local GPT ( need docker ) and then run virtual box or something to run the other stuff?

Would love to hear some ideas

[fervent thunder]

GPU passthrough possibly

[silk condor]

I have looked and most of the threads etc people say it works if you got like 2 slots of GPU and bla bla etc.. Not sure if it's worth the hassle :/

So Im looking maybe into Windows server or should I just go with a Windows 10 or 11 and virtual box from there?

[fervent thunder]

are you sure you have to split across two operating systems?
which of each thing is on each operating system?

[silk condor]

What do you mean?

[fervent thunder]

why do you need a virtual machine?

[silk condor]

Well I wanna do some stuff at home and try and expriements but currently running media server at home, game server for friends etc, home assistance, file server to store files etc and then stable diffusion and local GPT is next

[fervent thunder]

could you do all of that on windows? so you wouldn't need a virtual machine?

[silk condor]

If I do it on 1 machine and I need to maybe restart something.. I would have to restart it all.

[fervent thunder]

in that case running virtualbox would work okay

[fresh ruin]

Ok ty. Anyone knows if intel or amd CPU makes any difference in SD performance?

[quartz siren]

No the cpu doesn’t really have a large impact. The gpu does all the main computation. Don’t get a very horrible cpu but get a decent cpu. It doesn’t matter which company.

[fresh ruin]

yeah, ty

someone recommended pairing R5 7600x for example with it

[fluid charm]

hi, my webui bat file is stuck at this when i launch it, can anyone help?

https://i.gyazo.com/8c3171912d9701d911bd6b1fa91d4146.png

[kind cosmos]

hello, may i use the free credits in dreamstudio to test stable diffusion for an article review I am writing about Stable Diffusion? I am planning to publish the image results together with the article review, but I'm not sure if Im allowed to do that

[quiet mason]

Hey guys

I installed stable diffuser webui for nvidia automatic1111
But its been giving me a connection time out error

I tried clearing the cache, changing browsers doesn't seem to work

I have a rtx 3060ti ryzen 9 laptop

Can someone help me out

Yeah amd

Aight 🙏

[swift quest]

yo where can i ask question on how to use this

?

like questions on how to use this thing like how do you use more than 1 resourse i see people use different loras but not characters more like actions or faces and stuff on 1 chosen character

[quiet mason]

Wait but I install SD webui for nvidia

So i shd read the nvidia guide right

[swift quest]

bet thanks

[quiet mason]

What my laptop has is nvidia gpu + amd cpu

Lit thanks a lot

[undone barn]

um, hi, i want to download stable diffusion 1.5 checkpoint. where do i get the best version from? civitai?

[trim magnet]

yes civitai

[undone barn]

oh, I see it got removed from huggingface

[trim magnet]

the base yes,the finetunes are still there

[undone barn]

is this what i want? https://civitai.com/models/155256/stable-diffusion-v15-bf16fp16-no-emaema-only-no-vae-safetensors-checkpoint

[fervent thunder]

I am not sure if I could ask a question here, but I would like to generate realistic character for a game. However, I need to generate the body parts separately. The reason is so that I could move the body parts independently. Also I need the dismembered version of the limbs. Let say I want to display the arm is chopped off and the character is still shooting with the other arm. I have tried with SD1.5 and SDXL but can't really generate what I want. Is that even possible? Do I need to train myself?

[warm junco]

No 150 is to much. You get nice images around 30 steps

[undone barn]

lasso tool in software like clip paint studio or paintstorm studio or krita, friend

the issue is not using AI here but knowing enough to "paint in" the missing pieces and separate them out

[cerulean birch]

hello, i used this last year and i wanted to try it again, is there a latest installation guide with nvidia gpu on windows locally? thank you in advance!

[fervent thunder]

sad that I still have to do manual editing as the work is probably not tidy and need to be done per character

but thanks for the reply, would be appreciated if there is an existing lora/dreambooth/etc

[warm junco]

Hey, yes you can find it in the pinned messages of #🤝｜tech-support

[undone barn]

If you have the dough, i could do your editing for you- i'd animate it too but spine is really expensive

There are some open source software sort of like spine but probably not as robust

[fervent thunder]

well that's the thing I am trying to save cost by doing it all myself haha, I am just developing indie game for a hobby

[fervent thunder]

it is not about the animation itself, more like the sprite I am looking for

[undone barn]

Oh?

Why else do you need separate parts then?

[fervent thunder]

so I can animate them independently, not like a whole different character sprite per action (like jump, attack, etc)

something like this https://toppng.com/uploads/preview/body-parts-2d-character-body-parts-115630068843ern7dx6j2.png

but I want it to be generated so I can do multiple characters at once

[undone barn]

That sounds like a use case for a lora, custom

I happen to have made a character set up like that, but pixelart

I was planning to use pixelover, its just a bit ... You know how it with new software, its not always obvious right away how to set things up

[fervent thunder]

yeah it always happens aye

[undone barn]

So far the best answers i have are dragonbones and blender

[balmy hemlock]

Hello there, any expert with comfyUI ( especially mask / inpainting nodes ) ?

[undone barn]

For like rigging such a sprite like the above, assuming you will want joint bending

[fervent thunder]

right ok thanks

[cerulean birch]

thank you!

[fervent thunder]

have you heard about spine2D?

[cerulean birch]

whats the best model for img2img nsfw? or nsfw in general? im using A1111

[fervent thunder]

a LOT, juggernaut, majicmix, can search in civitai

depends on your style preference

[undone barn]

$369

[fervent thunder]

lol I got it for free, but I dont recommend my method to get it to anyone

[cerulean birch]

thanks, i did check there but no idea which one to choose xD

[main junco]

How do I use roop or other face swap alternatives like (reactor or face swap lab) on existing image

I just have an image

[frank fern]

So SAI goes AWS, huh?

[cosmic light]

hello

[main junco]

Also does reactor or face swap lab uses model itsel?

So is there a difference for roop or reactor if i am using flux or sd 1.5

[warm junco]

You can use Reactor with any model

[fervent thunder]

hello

[sand trellis]

Is it possible to outpaint using forge and fulx ?

[grave rampart]

hey fellas

[quiet mason]

yo

in the nvidia a1111 guide

it says edit the webui-user.bat

but i dont have a file with the ".bat"

i do have a webui-user as a windows batch file

and a webui-user as sh file

and normal just webui file

which one is the guide reffering to?

oh ig its the windows batch file

[sudden ruin]

The batch file

[quiet mason]

launch.py: error: unrecognized arguments: --no-half-vae--medvram

its showin this error

ah a space

oopsie

[undone barn]

like paint tool SAI?

[fair forge]

hey everyone

i had being out of the loop when it come to stable diffusion for quite a while i (at least a year if not more so) i dont know how many changes have being made, and i can use some help in installing it, and before anyone saids anything, yes i was looking for toturial onlines, each one of them shows something bit diffrent and im not sure which one is the best to follow

[minor star]

Hey everyone
i want to train a LLM SDXL fine tune model with about 100k images i have trained 84k images model till now but haven't gotten any better results till now, can anyone tell me how to start it?

[nimble moth]

Hey, started today I started experimenting with a comfy rewrite from scratch:
https://x.com/JulienBlanchon/status/1831719118434709868

The idea would be to ship a very minimalist ui + very minimalist server. Without any opinion on the usecase, no deps (not even torch). But super-extensible, so people can easily extend the user interface and server to meet their own needs. If some of you want to join, contact me 😉

[warm junco]

Hey, in the pinned messages of #🤝｜tech-support you find the right install guides

[low prawn]

Just so its known, my body is ready for Flux Auto (and lineart) 💚

[iron swallow]

stick it up ur ass

[quiet mason]

guys

do i need to download controlnet models

cuz i seem to have the preprocessor

but no model

[copper crystal]

how else would you expect your gpu to load them?

[quiet mason]

o;

i was running them on normal model checkpoints

makes sense aite

[copper crystal]

preprocessors just convert images into something that a controlnet model can interpret

[quiet mason]

ahhh

makes sense thankss

[copper crystal]

sometimes, like if you start with an open pose image, you don't need to convert it, so you'd select "none". simple thing that is often is overlooked when people are first getting their controlnet legs.

[quiet mason]

is there a website for them i was looking thru civitai does that have these as well?

[copper crystal]

"model zoos" on hugging face are good. i use git clone on those sometimes.

[quiet mason]

ooh okay ahah yeah my day 1 on sd

[copper crystal]

https://huggingface.co/lllyasviel/sd_control_collection/tree/main these are a bunch of controlnets for XL.

https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main these are all the 1.5 models

maybe not all. might be some others out there

https://huggingface.co/h94/IP-Adapter there's also these. ip adapter models, not technicall controlnet but work similar

[copper crystal]

welcome to the circus!

[quiet mason]

tysmm

[quiet mason]

i need to place the .pth files for the control net models in models-control net, correct?

[copper crystal]

\Models\ControlNet mine are all in here. Sometimes a node or an extension will want a different directory. I hate that so i just create a symbolic link in that case.

[quiet mason]

lit lit

thats where i put them as well

[quiet mason]

woah it really is

been trying to get a proper coloured animated img2img

w good eyes

its p fun tho haha

[split nebula]

Good for you, want a cookie?

[rain mortar]

Hello, i was recently looking to use Stable diffusion, so i joined i was actually looking to see is there a place in specific where i can download pretrained models?

[copper crystal]

it's a spammer. don't engage.

[split nebula]

I haven’t seen crypto scammers in discord before

This is new

[copper crystal]

lol. what?

[split nebula]

Come on why do people keep pulling this shit

[copper crystal]

its not new lol.

[split nebula]

Who even falls for these

[copper crystal]

1/1000

[rain mortar]

old people i would assume

[split nebula]

I haven’t come across crypto scammers in discord until just now

[rain mortar]

like the prince story from zimbawe or what ever

[copper crystal]

there's new suckers born everyday

[rain mortar]

give me 5000 usd and ill give you in exchange a couple 100K when i can take out the money

[split nebula]

I mean discord assumes the person knows a little about the web

I think

[copper crystal]

they've been around. crypto scamming and NFT shit has been lucrative for a long while now

[split nebula]

I’ve seen them in like youtube comments and such but this is the first time i’ve come across one in discord

I’ve seen tons of the steam link scammers on discord though

The nitro stuff

[copper crystal]

everybody thinks they're invincible to scamming till they're scammed. once at a job. clerk at a pizza shop. i was left alone for the night. hot girl came in with credit cards but they wouldn't swipe so we punched in the numbers manually. (her idea). yeah it was a scam. thought i was bullet proof until then but she had a nice pair so i wasn't thinking

i've grown empathy for people who take the bait since

[split nebula]

Irl scams can get quite convincing in some cases yeah

Social engineering is a real thing

These nonsense random emails/comments are impossible to fall for if you’re more familiar with the internet though

But you can have more complex scams that combine existing info on you or a company and such to build a very convincing story and therefore more dangerous

[copper crystal]

Often it can happen when someone's in a vulnerable moment too. even if they are experienced. Bunch of bills stacking up. Girlfriend wants that trip she been talking about. You see a guy promising financial success online as you stressing about it. Lets go!! (happened to a friend who worked in IT around the time Kanye Coin was coming out.)

[rain mortar]

So is there a place where i can download models or Loras?

[copper crystal]

oh boy is there. Lets just send you to the new version of the site though before you discover the rest on your own. @rain mortar

https://civitai.green

[main snow]

The diffusion is indeed stable

Lemme know when flux works on webui like a normal checkpoint

[fervent thunder]

I don't rly know webui

I might go learn it some time
so I can understand what webui users are saying

[split nebula]

As bad of a state as you might be financially, there’s simply no such thing as easy simple financial success. It’s just not real. You can fall for it if you act impulsively and without thinking straight, but that’s much more difficult to happen in an online situation

In a real-time example you can get pressured into making bad decisions through social engineering relatively well. They even try to imitate this sense of urgency in certain scam emails by adding in deadlines and such, but so long as you don’t act impulsively you’d be fine unless it’s a more elaborate ruse

[fervent thunder]

sense of urgency is classic social engineering yeah

the best / coolest social engineering trick is a hard hat and a clip board

[split nebula]

There are many interesting ones yeah

[copper crystal]

i agree that a good rule is "there's no free lunch". When you are saying "you can do this or that" you're talking about yourself as you roleplay the situation in your head with full cognitive pause. What any other of the billions of people on planet earth do is something else, especially in "the moment" with nothing but the immediate information in your working memory.

It's just not hard to understand how scammers operate at all is what i'm trying to say. A little bit of empathy and being honest with yourself is key to bridging that understanding. It's not always "old people" or "inexperienced people". It can happen to you or anyone you know.

some of my fave social engineering is low life low tech https://youtu.be/MIlwXXoJQxw

that lora would work well with sdxl models. not pony versions. juggernaut, hello world, mobeus, proteus, cyberreal xl. it should.

[main bronze]

Can someone help me my controlnet preprocessor on comfyui isn't working

[split nebula]

Yeah, I agree completely, just something that would be impossible in the context of an online scam given you can avoid thinking “in the moment”, unlike an irl scam like the one you ended up falling for

[copper crystal]

there are lots of ways to apply guidance with controlnet as well. you can combine loras with contorlnet canny or the qr code ones.

[split nebula]

Because you can easily stop and think before taking a decision when it’s not playing out in real-time

[copper crystal]

i'm more of a green smoker. dabs make my tolerance go sky high and they get boring to me lol. i've long grown out of my heavy smoke days i think.

beer drinker vs liquor shots lol

[split nebula]

Thinking stuff through in real-time is much easier said than done though, yeah, especially if you’re tired/under pressure

[main bronze]

You know how to use controlnet on comfyui?

[copper crystal]

i do but it's difficult to explain. its something you just need to figure out. node graphs are fun like that. scott detweiler has a good playlist for comfyui and how to approach any workflow idea you want to do

[fervent thunder]

for the most part control net only gives you two levers to pull anyway
strength and schedule
so just play with those two levers

[main bronze]

Bro my preprocessors aren't popping up i just see the tile preprocessor

[fervent thunder]

99% of control net uses I see outside of tiled upscale
are canny and depth
sometimes canny is soft edge or something like that
very rarely depth is normal map or marigold

[copper crystal]

i live in PNW where the plants have some the highest terpene levels on the planet. (hydroponics not included). high grade stuff is in crazy abundance always. not a lot of reason to distil

[fervent thunder]

oh I forgot openpose
my bad openpose is huge

[copper crystal]

i think the highest was the indian alps? a youtube channel i like talked about it "Strain Hunters"

Pacific Northwest. BC Canada, but also washington and oregon count i feel.

[normal heath]

what model do yall recommend for generating graphic logos

[fervent thunder]

the latest realvis or jugger SDXL lightning

amazing speed

SDXL lightning is a model type

4-6 steps

[copper crystal]

checkout "fooocus" ui. it makes generating pretty easy and fun to play with. Works best with XL models

[fervent thunder]

hyper is pretty big too its a lightly different tradeoff

notably there is flux dev hyper

I don't know much about hardware requirements cos I use cloud

[copper crystal]

my theory too is that all the draft dodging hippies from california came up this way and established the cultivation culture here. People started hybridizing strains here hard and its where modern concentrations really first started. That might be a bigger part of it than the climate, but the climate is still goldilocks for the plant.

[fervent thunder]

I use Vast.ai and tend to rent L40, A40, A5000, A100 or H100

[quartz siren]

It's pretty slow but it will use the same amount of vram as sdxl if you use the nf4v2 version.

[split nebula]

I use it on a gtx 1060

Same

[fervent thunder]

if you are not upscaling flux is pretty short wait

when I went from my 8k SDXL upscale
to flux 1024x1024
flux felt way faster lol

[copper crystal]

i run it with quantization. my 16gb fits it into 12.5. i think you can go smaller too. 8gb would be slow with some swapping

[split nebula]

8k upscale is wild

[copper crystal]

i'm happy to be called a tech nerd. happy to gab too

[fervent thunder]

ye I love high res like 8k

[split nebula]

I can’t upscale sdxl at all with 6gb vram

It just runs out

[fervent thunder]

you can with tiling

[copper crystal]

you could use tiled diffusion or ultimate upscaler though couldn't you?

[quartz siren]

you can probably maybe fit it in 6gb vram? it would be really slow tho.

I honestly would reccomend runpod or something and if you want a free service something like colab or kaggle. They provide way way more vram. In kaggle, you can actually rune the full version of flux

[split nebula]

How do i do that

And is it absurdly slow

[fervent thunder]

in comfy ui there is both tiled ksampler and tiled VAE

[copper crystal]

sdxl has pretty good models. they are each kind of unique too. hello world is a fun one

[fervent thunder]

yeah its slow

[copper crystal]

forge has it built in and on webui it's called "multidiffusion" which combines both.

[split nebula]

I take like 2 minutes to generate a regular sdxl image 1024x1024

Slightly under

[fervent thunder]

training lora is what you want
rather than searching existing latent space images

[copper crystal]

i found very different results when i upscaled flux with tiled diffusion and when i used sd ultimate upscaler. both did well in different ways. i might prefer sd ultimate upscaler

[fervent thunder]

I'm switching to multi GPU

cos they are better value as less people want them

sadly I have to leave comfy for swarm to do this

[split nebula]

I have a laptop lol

[fervent thunder]

same

[split nebula]

Multi gpu?

[copper crystal]

i thought they were basically the same project now. swarm's spooling works in tandem with comfy perfectly

[fervent thunder]

on vast.ai

you always see these massive servers like 10xA40

[copper crystal]

if you have a comfyui install already set up i think you can point swarm at it too

[fervent thunder]

they have amazing price cos most people don't want the hassle of multi-gpu

[split nebula]

Ah, cloud gpus

[fervent thunder]

I joined the swarm discord today seems good

[split nebula]

I just use sd for fun, it’s not really worth paying for me

[copper crystal]

its probably more economical for them to serve up multigpu servers instead of one server one gpu. so lower price may be a cost leader to encourage people to deal with the hastle and tool around it. They got you hook line an sinker bud! (lol not really its a good deal)

[split nebula]

I’d rather have slightly longer wait times

Personally

[fervent thunder]

ye there is no rush

you can leave it generating overnight for example

[split nebula]

Yeah lol

[fervent thunder]

in other areas of tech, you have to wait for compile to finish, or regression to crunch

or blender render

its not that different

you end up with too many image anyway

if you use cloud a lot

most people don't have a need in their life for thousands of images

[split nebula]

Yeah

[fervent thunder]

flux dev is too strong

for image gen you just want flux now

yeah

I am not happy that it cant take negatives

but the image quality is a big step up

no but they used to work there

the main science team

now called Black Forest Labs

oh its a separate, rival, company

its quite dramatic what has happened

[copper crystal]

lol flux is pretty good. it's made by the original researchers that made the first version of stable diffusion before they left to RWML and then came back to Stability and then left Stability to form BFL after they developed the architecture and training code for SD3.

You can still use SDXL since there's so many tools out for it. It still has great longevity. That lora you found will work great with it. Flux might know rosin too though. It knows buds pretty good.

[copper crystal]

the weights are stored with a 16bit decimal number format. i turn on an option in the UI i use to load those weights into memory using 8bit decimal number format. This means a 23GB file fits into 12.5GB of space. there's more you can do to get it to work on lesser hardware as well.

[undone barn]

hi

just updated my clip paint studio from 1 to 3

[normal heath]

any specific prompts i should give or negative prompts for better looking text in images

[copper crystal]

forge ui. there's lots of tricks with it. it would use the ssd if the file is stored there. but that's just how it works anywhere really.

[covert kestrel]

I'm using Forge and seeing a lot of integrated extensions. Can anyone sum up what are those and what those does?

It currently has:

• Dynamic Thresholding (CFG-Fix)
• FreeU (newly integrated)
• Self Attention Guidance
• Perturbed Attention Guidance
• Kohya HRFix
• Latent Modifier
• Multi Diffusion
• Style Align
• NeverOOM

[stray pond]

helo guys is there a website/discord where I can find experts for SD? I want to use controlNet canny to change the face and body of from a complex image with two characters. I would like to know if there's a paid service out there that let me have tutors for this

[inland stream]

Hello everyone 🥳

[low moon]

Have an excellent Friday.

[fervent thunder]

I can tell you 7 out of 9 of those:
Dynamic thresholding rescales certain values of the latent each step to avoid CFG burn
Free U affects backbone and skip factors and can boost coherence, saturation and contrast
Self attention guidance pushes the model away from blurry images
Perturbed attention guidance pushes the subject away from degraded images
High res fix runs a second pass at a higher resolution, there are a few versions of this on different platforms
Latent modifier changes some of the values in the latent, there are many versions of this
Multi diffusion is a way to generate using tiles and has a way to overlap tiles to try to avoid seams, at the cost of sometimes hallucinating at the place the seam would be
Don't know what style align is
Don't know what neverOOM is

[sand trellis]

Hello there,
How would you suggest me to proceed for training SDXL, SD3 or FLUX on a set of different marbles? The idea is to create a model for each kind of marble or 1 model that can generate each one of them.

[fervent thunder]

in order of decreasing difficulty:

1. pytorch
2. diffusers
3. OneTrainer, Koyha, SimpleTuner
4. replicate, civit
5. paying a freelancer

I recommend doing the earliest one on the list that you can handle

[pine tide]

yo guys do you have any tipps o get not as much fucked up hands? literally 80% of my propts are waisted because of too many fingers 🥹

[low moon]

Use SAM selector or adteilar to fix those hands. This is our only weapon. We have nothing else.

[covert kestrel]

StyleAlign attempts to retain the style across multiple batches of image generation.
NeverOOM caps the VRAM usage at 1.5GB.

Source: CoPilot

[ashen rune]

https://www.youtube.com/watch?v=zo6w-Md7DUo

[fervent thunder]

thanks

I do this by using IP adapter and color match, with the first generation as the input

[feral pike]

is 943 seconds for a 10124x1024 image okay? using flux dev and 7800xt gpu

[fervent thunder]

is SDXL no longer available for free in discord?

[low moon]

Was it ever?

[forest trout]

It was for a little while, but it's been at least a year now since it was available.

[worthy quiver]

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?

[desert dagger]

sdxl would be fine. you'll need to train a lora on the logo though

but for text, you'll need flux

[worthy quiver]

thanks

I want sth avaiable on hugging face

[tulip mirage]

🤩

[abstract quarry]

SDXL sucks for logos

at least I could never make it to train properly on a logo that is unknown to it

with a full finetune maybe, but not with lora

Flux, in contrast, works really well for that

it is also capable of generalizing the logo into 3D

[quartz siren]

I honestly recommend flux over sdxl, you don’t even need anything, you just need to prompt. It’s great at logos.

[feral pike]

how do u move a model to gpu ?

'GPU is available in the environment, but no device argument is passed to the Pipeline object. Model will be on CPU.'
i get this

[quartz siren]

what are you using? If diffusers, just very easily do pipe.to(“cuda”)

[warm junco]

When?

[feral pike]

its this thing https://huggingface.co/p1atdev/wd-swinv2-tagger-v3-hf

[warm junco]

Where do you use this? In auto1111

[feral pike]

nah trying to make a tagger python script

[warm junco]

Ah ok

[trail lion]

comfy is now my tool of choice for tagging

[fervent thunder]

yeah you can use comfy as if it was Langflow or Flowise

it might have more features at this point TBH

[trail lion]

I havent used those, but I was able to create a very specific workflow using some LLM models that does exactly what I need

[fervent thunder]

its similar to comfy but for LLM

[trail lion]

there was a checkpoint on civit that made some kind of comment like "you can train men and women separately for flux but not together" , paraphrasing heavily, but I'm putting that to the test

made it sound like attributes tend to get copied between them or something

[fervent thunder]

its all unknown yet

need to work out if it was progressive or adversarial distillation, for example

[abstract quarry]

the answer "never" is missing. Very optimistic 😁

[graceful sphinx]

Can somebody help me with Forge? I'm going crazy... Thanks

[pulsar venture]

Hi, how is call the extension for using controlnet on comfyui?

[desert dagger]

purz has a very cool video out showing how to use stable audio inside comfy

[trail lion]

gotta be careful with those kind of community polls, we dont want to get blamed for the quality of the next release

[fresh ruin]

yo, guys, is the difference between 32gb and 64gb of ram significan in SD?

[oak latch]

when training for my lora how should i remove tattoos and shit like that froim the pictures

[hasty hornet]

What is SD 3.1? Was something new announced ?

[heavy blaze]

Hello every one,
My laptop Spec is 16 GB RAM with GTX 1650 is this okay? for "Stable Diffusion & ComfyUI"

[desert dagger]

not really. Vram is what you care about, not ram. you want your gens to be on your GPU, not your CPU

[fresh ruin]

idk, seen lot of folks on reddit say it consumes a lot, specially sdxl and flux

[desert dagger]

i'm going to guess they have a machine that either doesn't ahve enough vram, so the CPU is being used too - or they have something set up incorrectly

[fresh ruin]

mm

[midnight flume]

Hello! I am trying to train stable diffusion to create "minecraft skins" (which are essentially 2d images that could be mapped to 3d)

So far, I am shooting for SD 2 for this task, and rendering the image into a 768x768. I'll be running it on one A100 and very low dataset count (50) with danbooru-style tagging (gender, clothing, accessories, etc.) in a comma-separated fashion

I am aware it's a weird territory, but would this be too low of a count? And would using SDXL benefit me better for this?

[trail lion]

@desert dagger is right, I upgraded from 32 to 64GB ram, not much difference fwiw. I saw a massive difference though going from an amd card with 16G vram to an nvidia with 24G

[desert dagger]

skins for objects in minecraft?

[midnight flume]

No objects, people! With what's called "transparent layers" (which are additional details that appear around the base layer)

[desert dagger]

right. but you do realize those are unfolded squares, right?

[midnight flume]

I do, that's what the 768x768 is (scaled up map of the 2d unfolded squares)

[desert dagger]

by SD 2, are you referring to 2.0 or 2.1?

[midnight flume]

I am thinking of stabilityai/stable-diffusion-2, would that be 2.0?

[desert dagger]

yeah. i wouldn't suggest you use that. mincraft image files are low res, pixelated - even the high-res reality mods are. you can use sd1.5 for that jsut fine but SDXL is probably what you want to use becasue that's what most of the trainers out there like

[midnight flume]

Ah, I am okay with any! I am using my own cloud server with accelerate, or is this more on that I should use fine-tuned ones from other trainers?

[desert dagger]

and i jus use seamless image textures unless i'm doing something like a flower - you can't create transparent backgrounds with stable diffusion, you're still stuck with photoshop

[midnight flume]

Right, would it make any sense to fill in what's transparent with (255, 255, 255)?

Or an uncommon color that would otherwise not appear in the dataset

[desert dagger]

no. because then you can't have white in any of the image. use something like neon green or drop out blue

[midnight flume]

That makes sense, okay, I'll use something like neon green for this to make removal easier. I think I can then remove it via PIL (or any image libraries) later?

[desert dagger]

can it target one RGB value?

[midnight flume]

I think so! I can convert it to an array and replace any matching 3-tuple values with an alpha value of 0

[desert dagger]

then yes, you can use that to set alpha channel based on that RGB value

[midnight flume]

This is really insightful, would you suggest anything I use with SD1.5?

[desert dagger]

i'm not understanding that question

[midnight flume]

I know there's a lot of tricks out there, LoRA to have less parameters to finetune, there's dreambooth which changes the theme with little sample images

[desert dagger]

sure - but this is really simple photoshop work, are you sure you're not overthinking this? how many skins are you gonna make?

[midnight flume]

Truth be told, I have a free weekend and I figured, why not have this be my first foray into learning about diffusers!

I don't doubt it's not time or energy efficient, but previous works I've found of people from this was really unsatisfactory to me, and I'd like to see if I can do any better

[desert dagger]

okay, it'll work for that. i'm not sure what your question about anything to use with sd1.5 is asking throuh

here's a good seamless texture prompt for SD 1.5: green random seamless Textures for 3D, graphic design and Photoshop gold trim

[midnight flume]

Right, I have to fine-tune SD1.5, I am wondering if I can just plug it directly into accelerate or if there's any options that would make fine-tuning faster

[desert dagger]

😉 well ... you said you wanted to learn this and, if you experiment, you'll know the answer to that in detail

[frail sonnet]

Does anyone know how I'd fix this? "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\custom_nodes\fluxgym>env/Scripts/activate
'env' is not recognized as an internal or external command,
operable program or batch file.

C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\custom_nodes\fluxgym>

The fluxgym install says to: "Now activate a venv from the root fluxgym folder:

If you're on Windows:

python -m venv env
env/Scripts/activate
"

[midnight flume]

On that topic actually, I've tagged my data like man, tanned, white shoes, black hair, chains, ...
Would this be bad?

Since a lot of datasets out there contain what seem to be full sentences, but I also see models that are trained with this style as well
I imagine CLIP would be able to adapt, but it's hard for me to tell

[desert dagger]

go watch this, https://www.youtube.com/live/XbPCu-fGFcA?si=Z5CHWdIIhjwCwb_3 pay close attention to when he starts talking about python venvs

[midnight flume]

😭 That's true, and SD1.5 is probably not too large anyway (<1b is my happy range)

[desert dagger]

remember - your labels need to tell the AI what the thing in the image is. you're going to use those as keywords when you prompt. if you show it pictures of a banana, and labled those as white shoes, when you prompt "a guy wearing white shoes' it's going to draw you a guy with bananas

[midnight flume]

That makes sense, I try to include all the details I think would be nice to generate, and leave out stuff like "white shirt (that has stains, emoticons, faces, ...)"

and let SD have variations in many kind of white shirt, so that on inference I can pump in high batch size and have many different shirts

[desert dagger]

for 1.5 - be specific, concise, don't ramble. tell it exactly what you want. you'll see a LOT of prompts out there that are packed full of modifiers that are basically just random dice rolls for data. but someone copied a prompt someone else used, and others copied form them, and... and not one of them knows why they are actually using those prompts

[midnight flume]

I also did the beautiful, ugly tag to indicate what kind of skins I like and don't like

And that's true! A lot of prompts I see are.. paragraphs long, many

[desert dagger]

here https://docs.google.com/spreadsheets/d/1bdidA4w5pB2BQMyxhkFu710Nu5bzKM1E-Wh_38ZZlT0/edit?usp=sharing my prompt spreadsheet. you're welcome to make a copy for yourself.

it'll help, i think. most of the tabs are SD.15 stuff unless labeled otherwise.

[midnight flume]

This is all really nice! I didn't realize google sheets API allow for images, but thank god it does!

[desert dagger]

i insert all of those, one at a time, by hand 😉

[midnight flume]

I am impressed by your dedication to this 😭 Kudos, I can't even annotate >50 samples by hand

[desert dagger]

i've been building that sheet since 1.4 released. hopefully it'll help

i've got around 3500 hours in on studying how SD works by this point

[midnight flume]

It definitely does! The idea is that I can pull from these to get an idea of what a corresponding "minecraft" looking skin would look like

At this point, you seem like you reached enlightenment. I have around ~1000 hours in NLP and I think I still have lots more to learn

[desert dagger]

it's a very deep pit, the rabbit hole

[midnight flume]

It seems like you made it out to the other side!

[desert dagger]

naw, but i got distracted by flux for the moment.

[midnight flume]

Out of curiosity, another approach I have found someone do is like this: https://docs.monadical.com/uploads/3adad129-34ba-40a5-af14-da7cb5fa2624.png

Their idea was "to guide SD into learning about how the skin parts map to the actual 3d part" (sorry for embeds not loading 😭)

[desert dagger]

that sounds much more useful

[midnight flume]

My thoughts was "since images are abstract anyway, SD definitely doesn't learn or view things like humans do, so maybe a 2d map might make sense to it?"

[desert dagger]

and you might look into the text to 3d that's out now. Stability just released their text to 3d as well

[desert dagger]

i think you need to learn how to do matrix multiplication

[midnight flume]

Ooh, the text-to-3d model looks really good

[desert dagger]

yeah. and probalby much more what your project wants

[midnight flume]

I have a.. weak but somewhat okay enough understanding of linear algebra

I figured having the 3d model wouldn't be ideal since what I have looks like 2.5d, or am I misunderstanding? And that it also needs to learn to generate 2d

[desert dagger]

okay, first order of business - binge watch this guy's channel https://www.youtube.com/@Green-Code

[desert dagger]

minecraft uses 3d squares and rectangles with textures mapped to them.

[midnight flume]

😭 I promise I am not that bad, I can read most arXiv papers that come out

[desert dagger]

yeah, but he's fun

[midnight flume]

That's more or less in-line with this, but would the 3d model not also have to learn the 2d map with it?

I don't think I am against the idea of using Zero123, my only qualms is that in my head, I am thinking that this task seems simple for SD to learn

[desert dagger]

i don't think so, the code is going to create the map for it after it generates.

however i don't know. you'll need to read the information for it to see how they handle that. i don't think it's making anything but the model - and then you map it in something like blender

[midnight flume]

Oh, you mean as in to create the 3d map. That's okay! I already have code to make an isometric view of the minecraft skin

[desert dagger]

just use umapper 😉

[midnight flume]

😭 mentions of umapper just drags me to a github library with a few stars

But, in any case, in summary of all this. I can do a training set with SD1.5 and probably still get good results, and it'd be pointless to train with an SDXL?

[desert dagger]

it's a uvmapper for 3d models. just a second

[desert dagger]

here https://www.uvmapper.com/

[midnight flume]

That is definitely beyond my needs on this 😭 Although I like the name uv, linear algebra base changes and all

[desert dagger]

it's much more simple than that. see - we use X, Y, Z for 3d cooridnates - we needed 2 more for up/down and right/left for maps, so we just use U and V

[midnight flume]

That's pretty interesting, I can't really visualize it, but I can appreciate the added details!

[frail sonnet]

Does this mean.... something similar to loras for 3d eventually? 😄 https://stability.ai/news/introducing-stable-fast-3d?utm_medium=email&_hsenc=p2ANqtz-8eWlqHd4HC0UUG-kEsNVAq5IrP2_6Xm3LOYT9VZTuYDsaoA-1m4F7pdvXJAzs9lbOOF3Epg5DcEdg1gFn0z4vdKAmx3w&_hsmi=94321401&utm_content=94321306&utm_source=hs_email

[desert dagger]

just means they've finally released their text to 3d model

[fervent thunder]

you could make a lora for it in pytorch right now
but if you mean universal tools like Koyha then yeah that might take some time

the use case for 3D images like this is fairly slim at the moment, at current quality levels

[desert dagger]

waiting for the text to 3d to be useful in blender and 3d printing

[fervent thunder]

yeah that would be cool

[desert dagger]

right now, the demos all look good, but what I get out of things like meshy is... no good

[fervent thunder]

I started making depth maps in blender for control net but it turns out SDXL doesn't follow depth maps that well, its fairly loose

[desert dagger]

that's what controlnet is for

[fervent thunder]

yeah I don't think 3D AI gen is really there yet

[desert dagger]

it's not. it's coming along though

so far, everything that's put out is a lump - you couldn't rig the thing if you tried

and it's sure not generating with rigging in place

[fervent thunder]

lighting is gonna be another issue

[fervent thunder]

yeah what I found was that I was making changes to the depth map that looked meaningful in the depth map image but SDXL made the same image each time

only a big change in the depth map did something

[desert dagger]

olivio put out a video where he used SD inside krita with depth maps.

[fervent thunder]

ah okay thanks

I realised just now that I am forgetting normal maps

and those coloured depth maps

(I used black and white)

can probably get better results adding other types

[desert dagger]

you need this https://shadermap.com/home/

[fervent thunder]

oh thanks this is great

I didn't actually know what a normal map is until reading about it in control net stuff

[desert dagger]

🙂 welcome to the world of 3D graphics - this talks about blender, but it's mostly just an explination of what normals are/do in 3d rendering

https://youtu.be/cn5BC3Vzcsc?si=qLmqCYPKfTlph5P9

[frail sonnet]

I mean eventually, I wonder if it will be possible. I'm hoping 😄

[cedar salmon]

indubitably

[frail sonnet]

Very awesome! Unfortunately my skill levels are more along the Kohya lines. One day though, that would be quite awesome 🙂

[fervent thunder]

thanks this is great

[fervent thunder]

you'd be suprised by how nice pytorch code can look

if you ever look at Comfy node code you are often looking at a bit of pytorch

[desert dagger]

welcome 🙂

[fervent thunder]

on reddit they were saying normal maps affect the image a lot more than depth maps

should be good

[desert dagger]

go watch the video, then you'll understand what normals are.

and then go experiment. maybe they do affect the image more, maybe they dont'.

[midnight flume]

Update on my weird minecraft skin thing, so far, not bad! https://imgur.com/a/vmFARe4 (3000 steps, ~1hrs in on an A100, SD2.1)
And of course, isometric render: https://imgur.com/a/izXOC9P
It seems like SD can sort of comprehend the anatomy of a minecraft skin.. ish

ps: Absolutely hate how checkpoint loading in diffusers is, but managed to get it figured out. It's unfortunate that we don't have the trainer API

[desert dagger]

that's not bad at all. not steve, but sorta steve the creeper

keep going with it 🙂

[midnight flume]

Thank you! Really happy with the resulting stuff, you suggesting SD2.1 and using neon green definitely saves me so much trouble! 🎉

[desert dagger]

🙂 it's looking really good so far.

[midnight flume]

Really appreciate the kind words! And money well spent! https://imgur.com/a/xLuCmIA (for 7000 steps!)

[pine tide]

dumb question, is SDXL and Pony the same?

[trail lion]

SDXL is a model, Pony would be a finetune of that base model, but Pony is not really like other finetunes in the sense that it changed a very significant portion of the weights

[pine tide]

alright thats why sdxl loras also work on pony and pony loras work on xl right?

[silk flax]

hello

[abstract quarry]

they often don't. As Pine said, Pony is deviated from SDXL so strongly that most loras won't work in both

at least not in good quality

[pine tide]

alright thanks

[main bronze]

Is there any ai discord community that isn't dead?

[trail lion]

yes, but I cant reveal its identity...but man it's a really great discord

[main bronze]

Bruh just send invite

[mossy edge]

hey guys, I`d love to just understand how to keep the same face on my realistic AI models ? Do all of you edit the photos after with faceswap of some kind or ?

[fervent thunder]

ye its possible that redditor was wrong about that, will have to see

SDXL, SD 1.5, Cascade, Lumina, Kolors, Pixart, Auraflow

are some others

if you are willing to use a pytorch workflow there are many more, but without a nice inference GUI like comfy UI

Poisson flow for example

or autoregressive

[main snow]

The diffusion is stable

[fervent thunder]

they can do img2img

img2img is actually a much, much easier task for the model, than "regular" text to image

cos it already has something to start with

instead of a cloud of random noise

[delicate trellis]

Like what i offer job I'm well certified

[nimble lichen]

hello !

[remote steppe]

Hello I need to make a stable difusion model
should be an image blender 2 images, and fine tune it with lora or dreambooth, than deploy it on replicate

if you are interested on this job and want to help me
DM ME

[nimble lichen]

Hello can we generate some picture free?

[quartz siren]

all flux models can do img2img, they can do inpainting as well. Infact, they are better then sd3 medium in those tasks I believe.

[quartz siren]

What's your image strength, you might need to decrease it then.

[copper crystal]

I saw invoke Ai people saying that on the first day. "Flux can't inpaint"

I don't know why they lied, but those lies persisted and spread far. It's not uncommon for people to make this huge mistake about flux now.

I think when it landed, a lot of business managers panicked and started spreading rabid misinformation. Such as "Flux can't inpaint" which is why we see it everywhere

[quartz siren]

Yeah their's an inpainting space now, you can try it as well. https://huggingface.co/spaces/Gradio-Community/Text-guided-Flux-Inpainting

[copper crystal]

i use forge ui and just use the built in feature. but before forge updated, i was using comfyui and didn't bother because there were people saying it couldn't be done

[fervent thunder]

I can somewhat see the logic of their concern

as far as I can tell it would be possible to distill a model in such a way that inpainting would not work well

[copper crystal]

i wouldn't call it logic. reactive reasoning maybe.

lieing is never logical

[fervent thunder]

was their original statement worded quite so strongly though?

[copper crystal]

i forget. i saw it as a screenshot. essentially, "You can't inpaint with flux"

i think it came from their server

[fervent thunder]

what I was thinking was if a model, in theory, was distilled such that it was jumping to the endpoints of sub-trajectories
and they passed over the sigma range that would have been ideal for that inpainting task
that could be an issue

[copper crystal]

yeah they just said can't i think.

https://www.reddit.com/r/StableDiffusion/comments/1ej8sb3/invoke_staff_insisting_that_inpainting_with_flux/ it's decimated in there though. that was the post i saw

[fervent thunder]

we don't have the paper for the new model so we don't know for sure but
generally diffusion models are trained for 1000 steps for the forward pass, and then we run it in reverse for less steps
so the model weights have "learnt" on the basis of having seen 1,000 seperate sigmas
then for distillation the model just gets taught trajectories, that will start on one sigma and end on another sigma
in theory this might be an issue if your task heavily relies on a sigma which the trajectory flew over

[copper crystal]

i dont think he did any qualitive research. it was posted on the 1st if i recall right

[fervent thunder]

yeah I remember then making that call really early

lightning was done like this:
128 steps -> 32 steps -> 8 steps -> 4 steps -> 2 steps -> 1 step

[quartz siren]

I would recommend nf4v2 instead. You can't fit fp8 in 8gb vram, It will use shared ram which will make it much much slower. Nf4v2 can fit with cpu offloading I believe and it will be far faster.

[mossy edge]

Would love if someone can please tell me how to set up and which tool to use to set up a proper AI model creator ?

[quartz siren]

Yes but it's a pretty considerable change in speed. fp8 is similar quality to nf4v2. q8 gguf is slightly better then nf4v2 and fp8 but will be also much slower the nf4v2 since it can't fit in 8gb vram.

[fervent thunder]

I found the original quote on the invoke discord
he just said it and then never explained lol

at least, not the step-distilled version.```

to be fair this means he tried Schnell

and then he assumed it was the same for Dev

at least to my understanding Schnell would suffer heavily from the issue I said above
it only saw the trajectories for 4 step generations
if your inpainting needs to start on a sigma that fell in between the sigmas for one of the 4 steps
then it hasn't directly seen your task before

[copper crystal]

I went to the server to look at the context. non commercial license was mentioned in the message he replied to. schnell doens't have a non commercial license. guy just responded to him with "can't"

regardless, the rumor still persists to this day. it sucks. misinformation is never cool.

[fervent thunder]

I think it was a slight misunderstanding cos his comment was ambiguous
cos the context of the previous comment does make it seem its about dev
but this line at least, not the step-distilled version.
makes me think he meant schnell

[copper crystal]

i've also just found spaces for inpainting with schnell that appear to work great

[fervent thunder]

oh that's good to know

[copper crystal]

in my experience with people, they lie. I think it's more likely that invoke staff were in a period of panicked lying instead of them trying to be accurate but misunderstanding.

Flux freaked people out

[fervent thunder]

not sure, I personally don't take part in the Stable Diffusion Drama
and there is always a drama of some kind

[copper crystal]

"drama" is a broad term. feels lame to lump criticisms around misinformation into the same pile as "lykon should be fired" style childish drama

[fervent thunder]

yeah I agree your claim was much more legitimate

from my reading he essentially said "at least not Schnell", implying he had tried Schnell only

but this is quite a tentative prediction I am making

[copper crystal]

I really believe that business schools are teaching their protege's that we live in a "post truth society" and when i see upstarts like invoke behaving as if truth doesn't matter, it really works me up.

I kind of have a bone to pick when it comes to entrepreneurs and lying.

[fervent thunder]

when you refer to them as an upstart
what's the reason to view them that way?

[copper crystal]

because they're an entrepreneur startup. they sell a product.

[fervent thunder]

I think the main thing I like about Invoke is they are node-based
like Comfyui or Chainner

[copper crystal]

someone's account got hacked and is botting now

dont use the same password everywhere people

[mossy edge]

can someone bless me with an answer to my question ? xD

Would love if someone can please tell me how to set up and which tool to use to set up a proper AI model creator ?

[fervent thunder]

do you have a graphics card?

[copper crystal]

you want to do face swap? start with fooocus for learning basics like that

if installing stuff like fooocus is hard, start with stability matrix to help with basics like that

[mossy edge]

yeah, I do.

I got comfyai with sd 3medium

photos are pretty generic in a way,

[mossy edge]

ai model, as in a female model, to grow and instagram channel with.

[mossy edge]

as realistic as possible.

[warm kestrel]

hello fam,
idk where to ask or who so im asking it here. Is there a prompt for SD/SDXL animal fusion, character fusion like MJ has its own. I was playing with it a lot of time but cant get what i want

[mossy edge]

yes.

[mossy edge]

what is controlnet?

just started watching and can say its insane x)

[sand trellis]

How would you suggest me to proceed for training SDXL, SD3 or FLUX on a set of different marbles? The idea is to create a model for each kind of marble or 1 model that can generate each one of them.
I have a 3090.

[desert dagger]

plan to use online trainers, and create a lora for each different type of base model

[sand trellis]

Do you know a good tutorial I could follow to learn about it ?

[desert dagger]

a tutorial for creating loras? or a tutorial for one of the online trainers?

[sand trellis]

So apparently there are several ways for doing so, LoRA as you said, then Embeddings, Textual Inversion, Hypernetworks, Dreambooth and Fine-tuning. Is LoRA the best solution ?

[fervent thunder]

Embeddings, Textual Inversion are same
Hypernetworks is old
Dreambooth is a particular fine tune method
all of the above count as fine tuning

lora is probably what you are looking for

[sand trellis]

Alright and for the online trainer, what does it have that I couldn't have locally ?

[desert dagger]

you've got a slow machine... online trainers are going to give you access to faster gpus and a lot more vram

[sand trellis]

Ok thanks for your advices. If you have a link for an online trainer and on how training a lora, it would be much appreciated.

I searched on internet and seems like there are a lot of differentes solutions

[desert dagger]

to start with, your lora has to be trained for a specific base model. a lora trained for SD1.5 will only work with SD1.5 or checkpoints/fine tuned sd 1.5 models

so first you need to decide what base model you want to create the lora for

[sand trellis]

In my case it would be for a texture. A texture of a specific marble type. Then I could apply it to any marble object in every generation

[desert dagger]

all you need is a prompt for that. however, that still doesn't really matter. you still have to decide what model you're training it for. do you want to use sd.15 to create with, or sdxl, or flux, or... you can think of the models as dictionaries. they all have words, but they don't all have the same words. And even the words that are the same in all of them are not on the same pages. when you use a Lora, you basically tell the AI 'pull up your dictionary, go to this page, find this word, and change the definition of it to this" - but if you have a changed definition for the sd 1.5 dictionary, and you try to give it to sdxl - the word on the page you specified... wont' exist - so it can't use your update

[sand trellis]

I would go for Flux

[desert dagger]

then you need to use a flux trainer - the one I would recommend is currently not working. I know people are using trainers for it, i just don't know what they are using.

[sand trellis]

Alright no worries, now I know where to go

[copper crystal]

i've been using the flux branch of bmaltais project

[trail lion]

same here, it works really well, it's kind of for advanced users though at the moment, no gui integration yet

at least the version I pulled last

no sample images, etc, it's sort of an early release or something you could say

[neat cypress]

published my first public model - contains huge amount of research info : https://civitai.com/models/731347

[warm garnet]

pip install Skynet

[desert dagger]

did you deliberately make a model that created what looks like rubber toys?

[oak latch]

i made my first flux lora but i got a question i did not provide nude pictures. can i generate nude pics?

[neat cypress]

yes it was a test of a style

[desert dagger]

okay. seems cool

[fervent thunder]

do you think A6000 is a good choice for flux lora, out of the data center GPUs?

[oak latch]

thats what im doing rn

good ash

i just dk how tf to get nude loras working wit no nude photos

[fervent thunder]

I've been using L40Ss a fair bit

same VRAM though

[abstract quarry]

boah, dude, just shut up.

[oak latch]

why?

[abstract quarry]

this is not a discord for such sick questions. go somewhere else and annoy people there

[oak latch]

how is it sick

how old are u lmao

12?

[fervent thunder]

I have to switch to multi gpu though
at least on vast.ai the value for money is so much better

very luckily, for inference Stable Swarm makes it super easy. Its not easy on comfy

[oak latch]

vast is way cheaper but runpod is so much easier to setup

[abstract quarry]

yes, I assume that's the age of people asking question how they can combine nude loras with other peoples' face lora

[oak latch]

its just there and works

[fervent thunder]

for the most part vast just works

[oak latch]

robot ass response

its not that serious relax little man

[oak latch]

u use a1111?

or

[fervent thunder]

I use a mixture of comfy, diffusers and pytorch workflows

[oak latch]

got u

im learning comfy rn

[fervent thunder]

but I am switching comfy to swarm for their multi gpu setup

[oak latch]

so confusing

[fervent thunder]

yeah comfy was very confusing at first and it has no docs

[oak latch]

what u like most

swarm?

i use flux and im pretty sure i cant use it on a1111

[fervent thunder]

swarm actually has full comfy inside it

so I will still be using it the same way

[oak latch]

ah shit

i didnt know that

cool

[fervent thunder]

I'm very interested in invoke as well

cos they are doing something a bit different

they are sort of making the photoshop of stable diffusion

their new canvas is coming really soon

[abstract quarry]

invokeai is really good

it's just always behind in terms of features

but as soon as they are implemented, they are working out really nicely

in theory they also have a node-based system like comfyui, but I guess, not many people are using it

[fervent thunder]

I think A1111 and Comfy have strong brand name effect at this point

really starting to realise how good starting your image out in blender can be

its a lot of effort though

[pulsar venture]

Hi, can someone help me with controlnet openpose. What is scale stick for xinsr cn?

[grave forge]

hi

[slender rain]

Guys quick question

What is the recommended training steps for creating a lora?

I'm creating poses with 12 images only

[dusky sierra]

hi

[vale karma]

Hi

[exotic pasture]

Hallo,everyone

[hidden sparrow]

Is there an AI or a way to use an existing AI to edit photos? By editing, I mean tasks like color grading, lighting, etc., without changing what is actually seen in the photo.

[errant moat]

where can i download the weights for sd3?

[rancid ore]

hi where is the place to hire ai-artist for picture concept like this kind of artwork https://www.artstation.com/amirzand ? i want to make artwork story like manga/manhwa, for my game-lore system, we need to generate many based on my concepts and some also repetitive place/characters

[desert dagger]

sure, there's a #1092446741984444416 forum for those sorts of posts

[ashen rune]

https://www.youtube.com/watch?v=9DbRJDitVhA

[trail lion]

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

[golden valley]

hi everyone! Who wants to work as a moder or developer in my project?

[copper crystal]

who want's to be my underling but not get paid!?

how far off am i?

[lament pewter]

Hey

[grave forge]

hru

[lament pewter]

fine. wbu

[grave forge]

good how is your day going

[lament pewter]

well. How is your weekend?

[fervent thunder]

Hello, I'm looking for recommendations on what libraries for low level development

I don't care if I need to code what has been done already, I only want to never have to deal with python again

ideally C/C++ libraries that would enable me to implement the basic building blocks

[fervent thunder]

that's no problem at all, there is a C++ API for pytorch
https://pytorch.org/tutorials/advanced/cpp_frontend.html

[fringe shoal]

I am looking to create an image that never before has been uploaded to civitai. I want to recreate a underwater base in the deepest parts of the ocean where lights hardly can be seen. Im trying to recreate a horror setting, like the game SOMA. I have looked around to find a model for this setting without much success. The closest I have found is this lora which im trying out right now: https://civitai.com/models/622686/underwatermovielora

I would like someone to help me to find a good model for SD or SD XL and /or some lora that fits this description 😅
Problem is that most loras/models often generates the surface or sunlight through the water, which destroys the creepiness. I want it to be deep, deep into the open ocean. The game Subnautica is also a good inspiration source.

[fervent thunder]

I want to stay off anything python related

[fringe shoal]

Inspiration: https://wallpapercave.com/wp/wp4396591.jpg

[fervent thunder]

yeah there is no python involved

The C++ frontend exposes a pure C++11 API that extends this underlying C++ codebase with tools required for machine learning training and inference

:/

well... I understand

thanks, really... but after months losing hair over dealing with pip and python coders... I'd rather avoid anything even relating to python

https://github.com/leejet/stable-diffusion.cpp

looks alright for example

This one is lower level and I was eyeing it too: https://github.com/tensorflow/tensorflow/

[grave forge]

good

[fervent thunder]

tensorflow and pytorch both have the same situation
the underlying code base is C++, and then you can either use python front end or C++ front end

[lament pewter]

Are you a developer?

[fervent thunder]

yes, that's why I still haven't started using tensorflow either ☹️

[grave forge]

huh no I'm not just seeing how everyone's doing

[fervent thunder]

are you aware that you can get them working with zero python in your codebase?

they both offer ways to do that

[lament pewter]

What do you mean?

[fervent thunder]

that's true, my avoiding python at all costs is more like I see that things coming from people that do python first are overall just lower quality.

I'm more inclined to clone and test a repo with zero python in its codebase, that's just my educated-prejudice though.

“educated-prejudice” pls, lol

[clever musk]

any recommanded model for digital art?

[loud solar]

Flux

[clever musk]

@loud solarThanks!

[loud solar]

You will need good hardware ... but choice of the moment ...

[clever musk]

my hardware is rather shit, rtx2070

[loud solar]

You can try with a few credits here: https://fluxpro.art

20 Credits/6 hours ...

[clever musk]

@loud solar did you mean this one? https://civitai.com/models/618692/flux
Also, why try it paid if I can do it for free with stable?

or 2070 is too shitty to run?

[loud solar]

I'm using a notebook 4090 with the biggest model ... it works but not fast

You can use Schnell model, lower textmodels ... but that would lower the quality. The page I told you you can set generations to 1 and the model is following your prompt so good, it should be fine

[loud solar]

For local generations you could use SDXL, Cascade or SD3 finetunes, Haven't really tested the last option a lot.

It's all about VRAM ....

[trail lion]

the whole point of modular framewords is to avoid having to implement everything from scratch. certainly there's nothing wrong with someone wanting to rewrite something, but given you're just 1 guy, whatever you write will end up being some super niche thing most likely. if you think a specific library is crappy code, that's why forks exist.

[copper crystal]

i'm sure microsoft is working on direct ML with the goal of bringing it up to pytorch levels. opengl used to be better than direct x too

maybe it was direct x 6 that was the first good one? i was a baby when it happened.

[fervent thunder]

I developed two e-commerce solutions from scratch, faster than getting kohya_ss to work (I still can't get it to work while Comfy, both Automatic all work here), the overhead of all the problems in python projects is way too much. (mismatching and conflicting dependency versions, GBs of downloads that end up in error, errors that make no sense and both authors and users without a clue of why things fail or why they think they fixed them.)

I've seen the code of many extensions for Comfy and Automatic, their principles are fairly simple, but the end result when trying to deploy them are always disastrous.

[trail lion]

then there's the whole matter of platforms, if you write it in C, which platforms will you support?

[fervent thunder]

C is basically universal, so I think that I´d be limited to the platforms that also give me facilities to use GPU.

[trail lion]

it's universal, but you'd need to compile it for each platform, just saying, since it's not a cross platform compile on demand type of thing

[fervent thunder]

scripting languages depend on their interpreters that were in turn compiled for each platform too

...and those interpreters technically compile the scripts too

[trail lion]

ok, so it's an interesting project for you, not something that solves a problem for anyone else. That's totally fine.

[fervent thunder]

it already exists though

next step would be to make something like ComfyUI use that in the backend

imagine a pure static Javascript client talking to a C++ compiled server side

[copper crystal]

Did you develop any ecommerce projects from scratch back when the dotcom situation was 2 years fresh and not quite booming yet?

lets say juustt before paypal launches.

https://imgur.com/mXGqSkm back when elon hadn't sold X.com off yet

[fervent thunder]

My first e-commerce was back in 2007 🤔

[copper crystal]

give ml 5 more years to catch up to that point of the infrastructure solving. Everything is obvious in hindsight

you dont know till its proven though

[fervent thunder]

?¿

in 5 years time ml will have much more mature libraries yeah

I think that my main problem now though is that people have little care with what versions they use

and little understanding of semantic versioning

pytorch and diffusers both have semver

as does tensorflow and keras

scikit-learn also

pretty much the entire mainstream stack for python stats and ml has semver

I just checked like 20 of them lol

@fervent thunder they may have, but not the rest that depend on them... and specially people who blindly add exact versions to the requirements of their libs

[copper crystal]

people who act blind but aren't actually blind. that's layer 8

can't fix layer 8

you could have the perfect most ideal solution implemented in the world, and have perfect deployement on it, and you know what will still happen? layer 8

[fervent thunder]

yeah I actually 100% agree
aside from some fun experimentation in Comfyui I've basically quit using all of the downstream libs
I'm just using pure pytorch and some diffusers/transformers, along with the typical numpy/scipy/scikit learn packages

that's the only part of the ecosystem that I feel is reliable for a proper project

[copper crystal]

i'm a big fan of bitsandbytes. its always been the one that boosts me

[desert dagger]

matteo has a challenge for you this week then 😉

[fervent thunder]

also I've been looking into trying to deploy in C++
for the most part there seems to be 3 common ways:

1. convert from pytorch to onnx and then compile with tensorrt
2. use the torchscript intermediate representation to then deploy using libtorch
3. use a tensor library like ggml

[copper crystal]

i stay away from matteos stuff myself. always seems to be pushing his donation pages. always seems to be riling his audience up into an angry ferver. im not a fan of that sort of influencer culture.

[fervent thunder]

he needs that IP adapter training money

[copper crystal]

fan of free libre software but also not all that

[fervent thunder]

oh awesome an SD 1.5 challenge

yeah I still use SD 1.5 loads

[fervent thunder]

out of interest which drama was this reference to?

[copper crystal]

a few times. most recently was sd3 he made a video about it thumbnailed with an angry mob and hyping that anger click

[fervent thunder]

ah yeah I went and looked

[desert dagger]

well of course he is. anyone that makes content will do that. however his work is very good, and fairly important to a large section of the comfy users

and there aren't very many that get riled up from what i've seen. those guys usually just hang out here

[copper crystal]

ah. no. anyone who makes content wont do that. influencer culture is new and different from old internet culture

[fervent thunder]

youtubers sort of are forced to play the algorithm

[copper crystal]

jake paul. fine bros. mr beast. h3h3. pewwwwdiepie. they changed things.

[desert dagger]

find me someone that's making content, tutorials and otherwise, these days - and putting them on twitter, who doesn't have a paetron or other donation page, or isn't monetizing their vidoes through youtube

[copper crystal]

ez but i think thats a disingenuous challenge too

[desert dagger]

my point is - this is now, not then. and for a lot of them it's their actual job and how they pay the bills

[copper crystal]

https://www.youtube.com/@Dice_Ai_Development

[desert dagger]

but re: matteo, i haven't seen anyone on his discord act out, get mad, and throw the sort of temper trantrums that happen on this one

that's dice

and he's suspect

[copper crystal]

you said anyone

[desert dagger]

anyone reputable

[fervent thunder]

that's like saying people use Windows or IE/Chrome because most find it superior... most never bothered to try anything else

[desert dagger]

that's the strangest comparisons i think you could have come up with but, okay

[fervent thunder]

I feel like that style of marketing is normalised now anyway

[copper crystal]

i'm not so much the nsfw training guy type, but those experts aren't allowed on youtube. so you go where they are. you find the seedy underbelly guides. those people give it all away

and are smarter

[fervent thunder]

they aren't allowed? D:

[desert dagger]

@fervent thunder i think this conversation totally derailed

[copper crystal]

there was a huge marketing push to glaze the internet with influencer culture

it was big tech collusion

[desert dagger]

proof?

[copper crystal]

now it's just normal and people are actually fighting for the rights of influnecers

[desert dagger]

which are now all AIs

[fervent thunder]

proof that big corporations collude to increase profits?

[copper crystal]

/gestures at everything

[desert dagger]

proof of your accusations

[copper crystal]

/gestures at the entirety of the internet

[desert dagger]

that's not proof, that's your opinion

[fervent thunder]

Gen Z seem to genuinely like influencer content
its not for us

[desert dagger]

some of us actually don't care

some of us have had youtube channels since youtube first came online

[fervent thunder]

could we count every action against monopolistic corporations as proof that corporations tend to... corporate? :/

[copper crystal]

https://tenor.com/view/they-live-eat-trash-can-coub-glasses-gif-3495333

[desert dagger]

can't use animated gifs in this channel

[copper crystal]

WWRRD

do you know how the internet works? its marvelous. by linking one electronic document to another electronic document, we've created a hyperlink. something you can click that will immediately go to a new document! computers do this fascinating feat using a U R L. No not like the Duke. The internet uses a Universal Resource Locator to look up exactly where the document is. How hyper is that??

[fervent thunder]

that's actually the web

[desert dagger]

what on earth does that have to do with anything we were discussing, or the fact you cant' use animated gifs in this channel

[copper crystal]

trying to bring back early web vibes. might've over corrected

[desert dagger]

I think we need to find you a usenet feed

[copper crystal]

keyword : ||click||

[copper crystal]

you dont have a usenet account?

[fervent thunder]

some of us may want to return to usenet... since the eternal September ended up ending in the end.

[copper crystal]

meow

[desert dagger]

i said we need to find you a feed, not sure how that translates to whether i have an account or not

[desert dagger]

https://www.usenet.com/ list of usenet service providers

[fervent thunder]

in that case I set up mine again

I like to own my server

[desert dagger]

go for it, it's still up and running, and lots of people still using it

[fervent thunder]

on reddit there was a comment once saying they should make a collective storage for AI models on usenet

it was in a reply to someone saying they should use torrents

torrents sounds reasonable

nntp only for nntp users I guess

[copper crystal]

no. we need to go blockchain. its the only way to be sure!! ||bitconnneeeeeeccct||

[fervent thunder]

sure of what though

to ensure data validity we don't need blockchain... although it depends

a ledger with cryptographic keys could work 🤔

torrents would be the best idea I think yeah

[desert dagger]

anyone heard anything more about a possible flux video release?

[fervent thunder]

I saw some Flux videos videos on Civitai I think

[desert dagger]

i mean an official release from black forest

[fervent thunder]

ah no idea

not yet

[desert dagger]

well drat

[fervent thunder]

someone finally beat UniPC https://arxiv.org/abs/2409.03755

its ODE so this might help Flux generate faster

[copper crystal]

https://github.com/wl-zhao/DC-Solver

[fervent thunder]

wow thanks its out already

[copper crystal]

was in the paper