💬｜general-chat

[jaunty wind]

I am hungry, can someone AI generate my food?

[wet zealot]

or a "release"

[flint eagle]

still not local release?

[wet zealot]

the deadman is already in place

[jaunty wind]

I want AI generated girlfriend, food, friends

and AI generated AI

[rare parrot]

That's so funny man

[astral goblet]

so long as these "ai assistants" are cloud based, i won't use them. That's a massive privacy concern. I don't trust these companies to protect my data

[lapis wigeon]

StabilityAI has fallen...

[rare parrot]

Wdym, they will always 100% guarantee your data will only be sold to whoever pays for it and used to train their AIs

[late brook]

Weights when.

[rare parrot]

They didn't so far

[wet zealot]

never

[flint eagle]

sd 3 in cloud not diffrent from midjourney or dalle-3 if you SD team want to diffrent any not bury by them needs to release SD3 local

[rare parrot]

When they're finished from what I know

[wet zealot]

stability is too financially fucked to be doing open releases

which is not a good thing for us

[late brook]

its only 10^25 floating point operations how long could that possibly take

[astral goblet]

lol business as usual. I'm more concerned about the hackers that can scrape their databases. A personal assistant would be something with intimate knowledge about me

[rare parrot]

"Give me currently paid thing for free right now, that's how you can save your business"

[astral goblet]

only a matter of time before one of these assitant systems causes billions of dollars of identity theft damage

[lapis wigeon]

Didn't you notice? This AI model is an scream for help. They are on the edge of bankruptcy

[cobalt turret]

for a split second I thought they are going to release it, but nooooo...#📣｜announcements message

[static cape]

Seriously at this point just release the weights and let us pay 100 bucks or "whatever you want" - I'd pay 100 bucks if that meant certainty and not waiting any longer.

[wet zealot]

there is news

but

[flint eagle]

why there is more opitions why going with sd?

[wet zealot]

its not news for here

[warm relic]

release the weights already, your researchers left 2 months ago. not paying for a bad api with worse images than midjourney.

[ripe merlin]

thought sd3 weights dropped 😦

[wet zealot]

wait patiently and you may receive

it may not be SAI's doing, should they die

[eternal ledge]

At this point, I'd be happy to pay $500 for a copy of the weights 😂

[astral goblet]

theres a couple other locked behind subscription models that are better too

[rare parrot]

Wait you're saying you'll buy a thing that costed several million dollars to train for $500? That's a crazy good deal for them tbh ngl

[flint eagle]

like I said if sd contunie like this midjourney and open ai got this

[wet zealot]

it will be freely available

[eternal ledge]

Even if I had to buy it per GPU or whatever

[wet zealot]

the matter of how is now up to fate

[rare parrot]

That's not how it works and whoever would've gotten the weights would leak them right away most likely

[static cape]

And how many more API-announcements until then? This was basically the THIRD.

[wet zealot]

it will be freely available by some means

whether stability or not

it will come out

[rare parrot]

It does work as far as law is concerned and that is what matters

[desert dagger]

any docs on how the new edit feature works?

[eternal ledge]

It can work however they want to license it (people will indeed pirate anyway and so on), MSSQL for instance licenses at a cost per core

[rare parrot]

If you don't get charged that's a free trial

[desert dagger]

they have been honest. the peopel that are not being honest are all the users insisting that SD 3 will never be open source - sort of like you - and causing trouble

[wet zealot]

I doubt it will be official means by which the model is released?

[rare parrot]

It's not deception tho

[astral goblet]

Kind of tangential. Theres actually evidence that ancient militaries long before newton had the capability of calculating trajectories and would've had some kind of theory going on. Newton might not have discovered the laws of motion. He might've just been the first to publish it widely

[rare parrot]

Most services are like that

[wet zealot]

there are things about this to be known

[rare parrot]

When they do free trial

lol

Like every single one

[desert dagger]

i'd worry about more important things in life, like the bombs dropping all over the planet, and the tornados that just wiped out a lot of people, and get a grip

[static cape]

2 Weeks TM ...

[desert dagger]

stop spreading misinformation then

[astral goblet]

tbf, those things aren't important to most people's lives.

[wet zealot]

if stability is bought out or goes under, i know SD3 will get out

how? up to interpretation.

[astral goblet]

Most distinguished men were a free mason or part of some society. Before the internet, social clubs were what you did

[desert dagger]

you posting stuff like SD 3 will only be paid, or SD 3 will never be open source - as you have no information to go on - is you spreading misinformation based on baseless assumptions

[wet zealot]

tbf its a pretty reasonable assumption given the direction other industry players have taken

[astral goblet]

the illuminati nonsense about free masonery is pretty modern. They were just one of many clubs men would join

[wet zealot]

mistral tried it, openai did it

also

FUD

[astral goblet]

FEAR IS THE MINDKILLER

[eternal ledge]

Time to go wander off until the next #📣｜announcements that probably won't be SD3 being released 🥱

[wet zealot]

this is a cope word -- it is perfectly reasonable to doubt.

so get out with that shit

[desert dagger]

it's a baseless assumption as there's no information about this specific thing. that's like saying "because everyone else grows apple trees, even though i said i'm growing pear trees, i'm obviously lieing and i'm growing apple trees

Stability.AI has SAID they will open source. untill such time as they SAY they've changed their mind, anyone that says they aren't going to is just tyring to cause trouble and doubt

[astral goblet]

I must not fear. Fear is the mind killer. Fear is the little death that brings total obliteration. I will face my fear. I will permit it to pass over me and through me. And when it has gone past i will turn the inner eye to see it's path. Where the fear has gone there will be nothing. Only I will remain

[wet zealot]

stability said it'll be open release
keeps pushing back the timetables
makes paid API

at a time when funding is critical

hell will freeze before they drop -- they are in no position to drop.

however, others are

[desert dagger]

they are still training the model. period.

[wet zealot]

SD3 will come

[desert dagger]

and hell will freeze over before you go away, too

[wet zealot]

not sure what gave you a wedgie but ok

[prisma horizon]

we got the dragons in war of thrones? we will get sd3

[wet zealot]

shilling helps nobody, though that'd be clear. SAI is no friend of yours.

cope

[prisma horizon]

well, considering that ai is a tool that can be abused, some form of identificantion is pretty reasonable

[narrow kernel]

Well if they don't have that kind of information, you can just create a bot to get unlimited trials... It's not rocket science

[wet zealot]

let's watermark airbrushes and photo editors too

while we're at it

[prisma horizon]

all the metadata and watermarks can be edited out

[wet zealot]

chat full of the same kind of people that shill cryptocurrency

[prisma horizon]

i am

i suggest the assumption of a daily dose of healthy glue.

[wet zealot]

"FUD" says it all

crushing reasonable suspicions based on hype

[prisma horizon]

what game?

[warm relic]

• devs exit scam and rugpull your shitcoin *
  trust the devs, it will bounce back, just have faith, to the moooooooon!!!

[prisma horizon]

ultima online?

ragnarok online?

[astral goblet]

Businesses won't leave money on the table and so long as SD3 is bringing in subscriptions, the managers in charge of the business will hesitate to pull that trigger. Not until new revenue streams are ready to go.

[wet zealot]

this

[astral goblet]

My guess is they'll release the lighter versions but keep the big SD3 model exclusive to the API

[wet zealot]

SAI is fiending for cashflow

and an open release would not give that to them

[astral goblet]

Well, there is the commercial use aspect. Once they release SD3 weights, services won't be able to deploy them until they pay. That is a huge revenue stream but it has to be ready to go

[prisma horizon]

yeah, sex too is fake, right?

[wet zealot]

their cashflow is absolutely not from commercial image gen

[soft sparrow]

where is SD 3?

[astral goblet]

Consider quantum immortality. The moment someone dies, they just begin observing a reality where they didn't die. You observed them dying in your reality, but their quantum state has moved to another. Hey? Yeah?

[prisma horizon]

the developers are monopolizing it to gen porn XD

[soft sparrow]

what the hell is the topic of discussion!?

[wet zealot]

worst interpretation of the many worlds theory out there

[astral goblet]

SD3 will be released with the stability license taht requires payment for commercial use. Like the lightning models have. ( i think lightning are stabilitys)

[wet zealot]

that shit is all philosophical and has no analytical basis

[astral goblet]

is it though? think about it. wave forms and particles bruh

[wet zealot]

come on...

[astral goblet]

double slits bruh

[prisma horizon]

lol

[astral goblet]

motion sickness from the chat scrolling too fast?

feel better!

[wet zealot]

this chat is slow wym

[knotty fog]

Question: Is the prudishness here "over the top" or just no hard core stuff. I ask because I don't see much in terms of women in a swimsuit, not, in any disgusting pose, but just as is. I always try to follow the rules on a group, so it is useful to understand if there is heavy handed admin'ing thinking that nearly everything is NSFW.

With knowledge I can operate appropriately.

[astral goblet]

it was about lead into gold too

[prisma horizon]

i have a solution, we can suggest the mods to slowdown the chat

[rare parrot]

Comfy Manager, ComfyI2I because it adds built-in image and mask editor, rembg-comfyui-node-better adds a bunch of handy features like running the pipeline to just a single node instead of all, ComfyUI-Custom-Scripts for even more handy features like being able to save pipelines as presets in UI instead of exclusively as files

[knotty fog]

I know. That was my point. I was browsing content and most seemed very tame except for some "scary" stuff.

[astral goblet]

There are better places to post AI porn. Just go there to share your spank bank material. Lots more people that enjoy mutual ------tion on those servers

[fringe karma]

Does anyone know what is the best way to create images with SD and dreambooth in the cloud?? (My PC is very bad)

[prisma horizon]

alchemy was in swallowing down mercury to get high

[astral goblet]

i love the old proto sciences. Alchemy was chemistry, but without models and understanding. They were just going for it

Astrology was studying stars, but blended with story telling. They still took great measurements and mapping. It's kind of sad what astrology is today. It doesn't even study the actual stars anymore.

[prisma horizon]

depend from your necessities

extensions can be dead weight, a a gold sand, depending from what you are looking for

[knotty fog]

Ah, the old rude response approach to people. I clearly said NOT porn. Not beat off material. I DID wonder where the line is. But I just saw a post in #1019361238234443776 by @low moon with some "side butt" which wasn't deleted so full body camouflage isn't needed. 🙂

[astral goblet]

Why look for the line? "Porn" has always been one of those "if you have to ask..." situations.

[rare parrot]

masquerade-nodes-comfyui and ComfyMath if you want to automate stuff, for example I have a thing that downscales images to be not more than 1280 in each direction for workflows where I feed it an image, like img2img and inpaint

[knotty fog]

I understand. But when I read the rule in question it came across in a way that gave me pause. I'm ok with this now that I see more content here.

[prisma horizon]

that's a good idea

[rare parrot]

I wonder which LLM powers this fella

What does it do?

[astral goblet]

we're all just holographic interferences on an infinite manifold

[rare parrot]

Looking at their repo, a LOT of those things already exist lol

[astral goblet]

dude just deleted their whole account or what?

he didn't feel very LLM'ish. i don't even know anymore

[rare parrot]

I think got banned

It did feel very high temperature LLMish

[astral goblet]

does banning often delete all their posts? i can't even review now

[fervent thunder]

yes

[rare parrot]

Like CR Seed, you can literally just right click on KSampler, convert widget to input, then make a Primitive node type and it has that functionality

Not sure, what were you using it for?

[astral goblet]

Just seems like deleting content that isn't spam is weird

[fervent thunder]

hey yall, quick question. how specfic do you have to get for a good image? i went on citivai and there are some amazing images that im trying to replciate and yet i can never do it

[eternal ledge]

I do like rgthree seed though, it's instead of a primitive node and has nice controls for changing to "random every time" vs "new fixed seed" or "whatever last got run"

Haven't tried CR seed, but I assume it's not too different

[astral goblet]

Prompting is king. Sometimes it's not even specifics but it's embeddings (textual inversions) and flavor tags (greg rutkowski)

[rare parrot]

KSampler has that already tho right?

[fervent thunder]

flavor tags. ill look into that. thank you

never heard of those haha. i have easynegative which is good

[astral goblet]

basically anything that is meaningless to your prompt but helps stylize the image. like artist names, types of photography, styles, etc

[fervent thunder]

interesting

[eternal ledge]

It has "after generate", not convenient buttons, you can think of "new fixed seed" as being a combination of setting the widget to fixed and setting a new random seed. And KSampler's widget doesn't have "use my last seed", you have to load from history and update to fixed. More a matter of convenience than possibility

[rare parrot]

I see

[astral goblet]

i just pack them onto prompts randomly using an extension. "Dynamic Prompts" has s thing where it loads a gpt2 LLM to expand the prompt with them

[rare parrot]

Oh yea

rgthree stuff is kino

[fervent thunder]

how do y'all know all this stuff

i had no idea how complicated it was

[astral goblet]

reading and experimenting and being in this scene for 2 years

[rare parrot]

Reading thru the readme of extensions in ComfyUI Manager

And talking with ppl

Looking at their pipelines

Etc

[fervent thunder]

i really want to know the best sampling methods and schedule types. but like really know what they do. like what are the differences between them and stuff. it fascinates me

seems like it would

im looking into more embeddings

[eternal ledge]

I have an unfair advantage as one of those people with a compulsion to consume information 😂 Also comfy is shaped in a way that comes very natural to my coding-since-I-was-six mind

[astral goblet]

the differences are the math they do on the noise sampling

math. ugh

[fervent thunder]

math lok

gotta love it

[rare parrot]

dpmpp 2s ancestral is my favorite currently, it's a really good one tho slower than euler a. Also check this out: https://stable-diffusion-art.com/samplers/

[astral goblet]

talking with chatgpt about all of this stuff is awesome sauce

just verify what it tells you

[fervent thunder]

okay. so why is that your favorite? does it tailor to what you want to make better?

[eternal ledge]

Also helping people out in places where other experts are, so you can be corrected by those who are more experienced and learn even faster

[astral goblet]

jinx you owe me a discord premium!

[eternal ledge]

the best way to get an answer on the internet isn't to ask a question, but to give the wrong answer
-- I forgot

[rare parrot]

It is less likely to have issues with stuff like hands compared to other samplers from my experience

[fervent thunder]

you know whats funny. my eyes are the hardest parts to adjust. my hands are usually perfectly fine

[fervent thunder]

ah i see

[astral goblet]

um. seems about as valuable as discord nitro. sure

[fervent thunder]

whats that?

[rare parrot]

Like if I give the same prompt and seed to 10 different samplers and then do that test like 10 times, on average dpmpp 2s ancestral had the least errors

[fervent thunder]

lol i am a youngin

okay

[rare parrot]

That's a VN bro 💀

visual novel

[eternal ledge]

Reminds me, have you found something better than shoving everything through the Impact nodes' Detailer for getting full-res renders into arbitrary inpainting mask-segments? Like say I want to turn the gumdrop buttons into soccer balls on a gingerbread man: all the ways I've tried either end up with inaccurate soccer balls due to taking little space in the image or a circle of grass around each ball because the model isn't aware of the surrounding visual context well enough

I want something similar to the A1111 "masked only" inpainting, but in Comfy 😂

[fervent thunder]

i really appreciate this

its so fun to learn

ive never tried impainting but i know what it is. im trying right now for the first time

is there a way to previously edit an image you already made? i know impainting is that technically, but what if i want specfic areas to do specific things?

[proper karma]

lol they removed the bad emojis from the announcement posts like 😐

[eternal ledge]

That's pretty much exactly what inpainting is for

[fervent thunder]

hahah okay i will

thanks yall

yeah

🫡

[fringe karma]

Hey... which one do u think is best for a bad computer, webui, comfyui or any colab?

ye, the webui automatic 1111

[astral goblet]

LOL i wouldn't recommend comfy to an absolute beginner. Then they have to wrestle with learning and understanding node graphs before they can get anywhere. That's worthy on it's own but it's a whole other thing.

Most people already know how to use a website with fields. Recommend fooocus or A111 to absolute newbs

[fringe karma]

probably, i'm using a good pc from my university to do it. But i would like to make some test at home

[astral goblet]

You can't effectively use comfyui without undestanding a LOT of technical details

stable swarm is good stuff too

node graphs are visual coding essentially. comfyui is a little bit of a hobbled node graph system though, since it can't do loops

node graphs are all over. many production tools use them. editors of all kinds. video, 3d, shader, music, etc

[fringe karma]

good. I'm doing research on using synthetic data to optimize an artwork classification algorithm

and im using drambooth with the a111 to do it

[astral goblet]

rexeeting *

[fringe karma]

I'll try those things you said

[astral goblet]

stable Matrix is a good package management app. lets you install many UI's easily. sharing the checkpoints between them

also lets you browse civit an neat things like that

https://new.reddit.com/r/StableDiffusion/comments/1d1zw74/mobius_the_debiased_diffusion_model/ cool research here. Looking forward to seeing the resulting model

I've been using oit for a while and i like it a lot better thank pinokio. i think that one requires virtual machines and it causes all sorts of gears on my windows 11 to crap up

yeah thats what i've seen. virtual environments would easily break

[onyx hedge]

eatch time u use SD, i can see a commit try to send offline, and done online ^^ that is normal ?

is it normal ? ♥

[sage reef]

looks cool i guess, idk will see when it releases

[eternal ledge]

I got so confused and thought reddit was broken 😂 I managed to have forgotten new reddit exists for several months

[fervent thunder]

is clip skip the same as denoising strength

[astral goblet]

lol sorry i sometimes use the new subdomain so i can embed photos in posts. old doesn't allow it. Funny enough, new.reddit isn't even the newest reddit design. its the old new reddit design

[eternal ledge]

I hope the new new design is new.new.reddit

[astral goblet]

no. it has to do with unet layers. this is probably the part of unet architecture i understand the least, how clip skip works

[astral goblet]

old.new.reddit

[onyx hedge]

time base is up.. all is the same ?

[fervent thunder]

i just got a message saying i was out of storage. do my images get automatically saved somewhere in my computer?

[onyx hedge]

you can check that in settings =)

[fervent thunder]

i see

[onyx hedge]

normaly somes was on %applocal% as usualy for logs and elses

[fervent thunder]

okay

yeah its defintiely my c drive

its talking about my GPU

[onyx hedge]

you ask was about your memory of your gpu stack ?

[astral goblet]

Ahh storage woes. The solution is obviously buy a new 3tb M2

[fervent thunder]

it says OutOfMemoryError: CUDA out of memory. Tried to allocate 4.05 GiB. GPU 0 has a total capacty of 12.00 GiB of which 1.11 GiB is free. Of the allocated memory 8.53 GiB is allocated by PyTorch, and 961.34 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[astral goblet]

thats about vram memory. do you have xformers enabled? I've noticed a lot of new users don't enable this

[fervent thunder]

i think whenever i boot it up and i saw my run console it says its not enabled

i havent ever had this problem

[onyx hedge]

you may be be full from a other service in back end services

[astral goblet]

first time ! wowwweee. youll remember your first oom error forever! (you won't)

[glad sail]

Where could I find the latest guide on how to make videos with first and last frame images?

[astral goblet]

just buy a new storage device and call it a day. (be me with 6 devices)

[fervent thunder]

well its not my storage

[astral goblet]

seen a recent post on reddit about motion steering for animatediff. it's really good i'l go grab it for you

[onyx hedge]

sli gpu memory ?

[astral goblet]

https://www.reddit.com/r/StableDiffusion/comments/1d1fkt3/5_new_steerable_motion_workflows_for_travelling/

https://github.com/banodoco/steerable-motion

[onyx hedge]

old methode dont work with simple skip clip+ 1 ?

[glad sail]

That was fast! Thank you! 🙌🏽🙌🏽

[astral goblet]

6 storage devices not 6 vram devices haha. vram unfortunately can't share between cards last i looked into it. nvlink would be what you want for that too. sli is obsolete

[astral goblet]

had a tab open for it already. hurray! i plan to sort thoruhg it later. new comfy workflows are like learning an entirely new ui

[cunning ice]

If I end up with 4 pictures that each have 25% of the ideal finished product, is there a tool for Frankensteining them together with SD or another program?

[astral goblet]

Photoshop?

[opal hedge]

See announcement Icon > Real Shit?!
Read Announcement > Sleep

[cunning ice]

Yeah that's fair. xD

[astral goblet]

yeah any cloud based assitants are a big no go zone for me. I won't even use the old generation of phone personal assitants. it's all data mining for the purposes of marketing and manipulation

[jade bronze]

What is a workflow and how do I use it on SD Forge? I've seen them for a while on Civit but never known what they do.

I tried googling it and looking for guides but couldn't find any. I downloaded a Workflow .json file but I'm not sure what to do with it.

[astral goblet]

workflows are configuration files for comfyui. every different task you want to do in comfyui needs a specific workflow.

you don't use them in forge

[jade bronze]

Ah, I see

Thanks

Does ComfyUI have benefits over Forge?

I moved to Forge from a standard A111 installation and find it a lot better, expecially for SDXL.

[astral goblet]

forge is much easier to go from task to task on. the benefits of comfyui is that you can specify the technical parts of the generation very closely. there are custom nodes that allow for a lot of neat things.

you could look at stable swarm. it's a UI up front with a lot of creature comforts. has comfyui on the back end. i'm not sure if it does face swaps by default though and you hae to bring in a custom workflow and get it working in the UI to do that

do you have an older gpu? forge's optimizations would benefit you a lot there. they'll be ported to base A1111 eventually.

[jade bronze]

I just got a 3060

[astral goblet]

low vram. it may help you there. i'm not sure what else about forge is different. the unet patcher but that's all back end and end users don't see it

[jade bronze]

Yeah I'm fairly inexperienced and I've just been doing things the easy way, I think I'll keep it simple since most of that went over my head.

[balmy rune]

🥱 still no weights

[rapid ore]

where do I put an upscaler pth file I put it in the gfpgan folder but it does not show up

[strange dagger]

Disini ada orang Indo gak?
Aku masih pemula.

[fervent thunder]

yo

vc??

ladies and gentleman i have installed xformers

[opal forge]

not cool

it's not free 👺

[fervent thunder]

oh fr? i just went to a video tutorial. i didnt get it from a website or anything i just edited my batch file

[stray pond]

helo I need an expert in doing stablediffusion ControlNET. There are certain poses I have seen and would like to create I would like someone teach me or just make a video on the process of doing it. I will pay money for your time. Thank you and have a nice day!

[strange dagger]

ada banyak mau ku tanyain, tapi aku gtw mulai dari mana.

[fervent thunder]

hey noob trader. idk what you are trying to make but if its for anime, i would get the extension that lets you use danbooru tags. they have a good amount of positions on there. also look into different loras that use those positions

https://civitai.com/

this is my preffered way of getting loras

[stray pond]

I have tried that but there's a specific way to do with ControlNET that you can replace two models with characters of your choice with that specific pose. LorAs are limiting in that regard

[solid whale]

Quick question, I do not have any experience whatsover with training Lora's. On average how much active time (me actively working) and how much time in total (active time + the time the computer needs to compute all the information), would it take to create a Lora?

[wraith cedar]

how do i orchestrate different prompts for my subjects again? like for example 2boy, boy1 is smoking and boy2 is drinking

if i have more than 1 subject i cant tell either what to do without all of htem doing it, i forget the proper way to do this

[fallow veldt]

There's not enough control to do that in a consistent way. Best practice I think is to try to produce an image that can be use as a base and inpainted until you get there.

[wraith cedar]

ah okay, thanks calcu!

[fallow veldt]

no problem, I mean you can try, or test Regional prompter, but you can considerer yourself lucky if you get the two characters in the correct pose or position, style etc. Then inpaint until they are doing what you want.

[wraith cedar]

yeah it seems easier to just tell it how many subjects and then suggest smoking and drinking is something happening in the picture

[fervent thunder]

how do people make such great images with low sampling steps

is it just the quality of an upscaler and loras and checkpoint

[astral goblet]

turbo models, hyper models, lightning models, lcm loras, there are a few ways

[fervent thunder]

turbo and hyper models.. ill look into that

i am learning a lot lol. you are a big help flow!

[astral goblet]

glad to. as long as i just have to point in the right direction

[fervent thunder]

yeah

im currently trying to replicate an art style using a lora

ill look at some images poeple make and they look incredibly good. i have the same lora and checkpoint but somehow theirs looks leagues better so im trying to learn a lot

[cunning aspen]

Stailityai is looking more and more like Openai... Sad...

[fervent thunder]

shoot

[low moon]

We will never join the dark side

Emad will bring balance to AI

He just has to fix some hands and borrow money from Musk

If i had to choose bwteen sd3 weights or a tool that gives perfect consistency i pick the later

im talking 100% not 90% not 80% not 98% no tricks!

[little lava]

gm

[silent dome]

Anyone ever have a problem where their positive and negative prompts become invisible? Like their still there and can produce the appropriate images but the text itself doesn't show?

[sage creek]

Hey, Can anyone help me with a logo? I need to get the text correct

[fervent thunder]

hi

[narrow kernel]

There are quite a few ways to pull it off, but if you pick up some basic Photoshop skills, you can do it in a fraction of the time. Generate template images with the two characters in whatever kind of poses you want(one image for each character) and one for the scene. Don't worry a ton about their quality, it's just for drafting. In Photoshop, just manually kitbash the three images together using the select tool to cut them out and then you can resample the composite image to blend it all back together. Canny/depth controlnets help as well, and use a pretty high denoise like .6-.8.

[fervent thunder]

ddddddddddddddddddddddddddd

[elder raptor]

Can someone explain to me how to create images by discord?

[trail lion]

You can check this link, they just announced it recently. https://stability.ai/stable-assistant

[marsh hound]

Guys one question, After how many times getting removed from Creativity Program you can't reapply no more?

[solid whale]

When training a Lora for Stable Diffusion, is it important that the images used to train all have the same resolution or aspect ratio?

[teal iris]

No, it is not necessary in kohya-ss

[narrow kernel]

Yes and no, depends on what your training set is. If they aren't all the same, you're at the mercy of random bucket cropping, which can cause some issues.

Like if you're making a style lora, you'll be fine. If you're training one for portraits of some specific person, it depends on the images and where the faces are in the images

[teal iris]

I think you are wrong. Images with different aspect ratios can be trained without cropping unless you choose to crop them. However, images can be trained with different aspect ratios, but you will only need to train them for a little longer in order for the results to be good.

[frail sonnet]

That auto queue button is dangerous LOL

[oak latch]

do u guys use any online hosting to get images faster and better?

i tried shadow pc and airgpu and they suck balls

[solid whale]

Great, thanks for the answers guys 🙂

[fervent thunder]

Huh

[sage reef]

[fervent thunder]

guys do u think i should buy 4080 super now or wait for 5080 for SD performance

also how much time it takes to generate 5-10 images at 1920x1080p resolution?

[teal iris]

kaggle or colab free

[teal iris]

in colab 16g vram 1 img tike 1.30m or 2m

or 1024*1024 tike 35s

[tulip yarrow]

Alright so how many steps is good to train an SDXL LoRA

Of an anime character

ping me when you answer please

[low moon]

are there koyha stormtroopers if there is a koyha ss?

[narrow kernel]

They will get cropped to whatever resolutions you set in the training parameters. Most sdxl loras you find are trained on one size in those parameters, like 1024², sometimes two sizes. Each size you add to it will massively inflate your training time.

[hardy lava]

How do i use SD3 via the discord?

[tulip yarrow]

How DO you even train an SDXL LoRA model properly?

[low moon]

same as sd15

[tulip yarrow]

How many steps are needed?
I put my trigger word as "A girl has medium length teal hair with twintails and parted bangs. She has brown eyes"

[low moon]

just takes longer

[tulip yarrow]

I mean like how many steps are recommended

[low moon]

2-3 epochs

[tulip yarrow]

I mean like about 1000?

[low moon]

short trigger words work better

or something

how many training images u hav?

25 at least i hope

so u have to calculate it base don that

[tulip yarrow]

I have 14

For the other one, I have 27

Oh wait... wait oops wrong number

For the other one, 18.

I usually just divide the amount of files in the dataset file by two, but turns out there were also npz files

[opal hedge]

It's time for my daily 'wen sd3' comment

[grave stag]

Last time I tried I tried and failed to replicate something that looks like oh dam that's really the anime character I tried to replicate and I especially didn't manage to replicate the example image I tried to replicate.

Now I have a better CPU and 12gb of ram with my 4070 instead of my 3060 ti, I had before.

So I want to really sit down and learn this program so I can make OCs and anime fan art for my RPs and stop paying people for it if I can?

Anyone able to really help teach me how to use this?

[static schooner]

I heard ai can now generate 3d models is that true and do you think it would work with 3d printing, lol?

would be epic to build anything in a instant with the perfect dimensions with ai

[inner geyser]

Which command should I use here to create an image?

[frail sonnet]

No, it doesn't at all.

Yet

[frail sonnet]

None here, you have to go to the artisman area (read the faq first). then you use slash dream

[oak latch]

is there any paid alternative that is really good

[carmine radish]

becky

HEY

[hollow light]

hello

[median kernel]

hi

[silent briar]

hello

[median kernel]

@silent briar Nice to meet you

[silent briar]

I need increase the rate limit

who I can chat with?

[limber nest]

hi

[jade mason]

When sd3 weights?

[low moon]

The way AI works now it's called a crapshoot. You do things and sometimes you get amazing stuff, sometimes garbage. Either way there's nothing you can do with it. You admire it and move on. No stories will be told.

[urban heron]

Hey guys, on a 3090 in forge I get about 3.4 to 3.7 it/s in ponysdxl 25 steps sde ++ karras 2m. I run xformers, cuda stream e.t.c. I am running on an undervolt of 893mv( around 330 watt) at 1860mhz and a +250 boost on memory.(9705) Any way I can speed this up more? My GPU is zotac amp extreme holo

[fallow veldt]

How to add an extra path for models, aside from the Auto1111, in Comfy UI extra model paths, so it loads from both paths?

[stray wolf]

https://cdn.discordapp.com/attachments/1235271104512397395/1245230502990970932/the.gif

[worn nimbus]

If I get professional membership on stability.ai can I download sd3 model?

[hearty jetty]

hi,good afternoon

[limber geode]

hello

[verbal osprey]

was sd3 cancelled?

Thought it would have been out by now

[vale steeple]

base finetuning take time

clearly it is not cancelled because you can still see the activities of partners

[dusty dawn]

Good mornging everybody. I am training a lora. My dataset only has images of 512x432 resolution. Should I still set the Max Resolution to 512x512 for training using Kohya SS?

[pearl ocean]

Who knows, as far as we know, they never started it

[little lava]

gm

[gilded folio]

in 2 weeks

[low moon]

Soon.

[craggy musk]

yo

how can I use the image generator?

[static cape]

Does anyone know if we'll be able to watch the live sessions later? I'll probably miss half of them at least, but I think I'd like to join in.

https://www.studios.thehug.xyz/lab

[tall berry]

Any peeps online good with ComfyUi? I need some help with stuff haha

[raw burrow]

Hello! I'm on the HUG team that's collabing with the Stability AI team for this course.

Yes - the live sessions are conducted over Zoom and recorded. The recording and accompanying materials are then emailed to everyone who is registered a couple afters after the live session ends so you can tune in on your own time. There's a gated Discord so you can ask questions at any time as well.

[static cape]

That does sound good. I really hope SD3 will finally have released it's weights until then.

Is there a catch / condition to get the deposit back?

[raw burrow]

@static cape To get the refundable deposit back, you'll just need to finish the course and complete a simple Feedback Form that covers some points from the different sessions.

[low moon]

Onl 2 more weeks.

[ocean mulch]

How do I generate images. I am new pless help

[low moon]

You press "make image".

[ocean mulch]

Can you send a screen shot

[surreal jay]

wow, ten dollars!

[hot relic]

That's crazy

[surreal jay]

if 11 people send $0.955 in 3 days

[tall berry]

MLM on the rise over there

[hot relic]

https://tenor.com/view/yas-hyped-lit-feeling-it-party-gif-14473619

A dollar is all I need

[surreal jay]

lol

[karmic brook]

thanks for chiming in on this ❤️

[river spade]

Needing help from someone with Kohya_ss I'm running into a Error I cannot fix

@warm junco can you help me

[shell stump]

hey i have a question regarding running SD on colab. is there a way to use my google drive disk space instead of of the runtime space?

[grave stag]

Anyone able to help me buckle down and really learn how to master using this?

[desert dagger]

i dunno about you, but I first spend sometime learning how the AI i'm using thinks, and then when i craft a prompt, it understands me and gives me exactly what i'm asking it for.

[desert dagger]

look at the top of the page you're on, you'll find a way to copy it to your google drive. do that, and run it from there.

[desert dagger]

start by using single word prompts. lock the seed, generate, change the seed a few numbers, genereate, get a feel for what the AI is going to think about by default when it sees the word. go from there.

[grave stag]

That's both simple and more complicated at the same time.

If I give you what kinda OC I want to make can you help me with what I'd have to do to learn how to make said OC or even just I'm practicing on said anime character lora how come I can't get it to look like them help?

[desert dagger]

i'm aware, but that's where you start if you really want to learn how to talk to stable.

[surreal jay]

interesting, never thought to do it this way. Smort!

[sly eagle]

hey do you know a model that makes good anime backgrounds

or basic anime backgrounds

and what is a pony diffusion?

[desert dagger]

just use the phrase "anime background" as your prompt

[ornate flame]

it's a great sdxl model for anime

[lucid cedar]

edit: moved my question to #🤝｜tech-support

[devout oxide]

In this server can I generate images through text?

[lucid cedar]

I think it's #1237459938901491852 channels but i'm not sure, haven't used them

[devout oxide]

okok

[sly eagle]

in XL?

[desert dagger]

in any version of stable

[sly eagle]

oh ok thanks!

[elfin turret]

Hello guys Im searching for a job of anything md me if you are interested

[tulip yarrow]

Alright what's the difference between setting a bunch of epochs for the lora model than just using repeats

[sage reef]

im no expert with training, but i think repeats is "per image" and epochs is per dataset, so epochs is amount to go go over the full dataset and repeats is how many times to go over each picture, but im not sure

[tulip yarrow]

And which one is the better choice

[sage reef]

i personally would say repeats cause it goes over each picture to learn as many details as possible, but ultimately, if your gpu can take it, increase both values eventually

[static cape]

Does anyone know if there is a similar Node for ComyUI? controlling LoRA's during inference like this is extremely powerful.
https://github.com/cheald/sd-webui-loractl

[sage reef]

https://github.com/cheald/sd-webui-loractl/issues/30

nothing currently

[low moon]

controlnets and canny and prompts arent as precise as pen and paper or 3d modelling programs yet

AI is like a box of chocolates. you never know what you gonna get

And even when you do get something u like u cant reproduce it or iterate on it

not with good enough accuracy

and of course: h a n d s

[lunar raft]

Hey guys, im new to ai image creation. I want to create realistic looking fictional planes. I did some research and after trying dreamshaper a few tries, I was very very impressed, but I knew there was more, that being model training. When I searched that up I have gotten a bunch of different ways to do it. What is your guys opinion for the best way to cutom train stable diffusion without costing a ton of money?

[low moon]

koyha SS

[astral goblet]

a ton of money is a very subjective term

[lunar raft]

20+ dollars a month

[astral goblet]

on a 100month schedule you should just buy a pc

if you intend to do this stuff that long

[lunar raft]

the thing is I dont know how long I want to do stuff, let alone how to do stuff

I want to get in and see how good of images I can create

[astral goblet]

rundiffusion is a good service to rent training servers on

i think you can use google collab to learn still, if your intention is legitimately learning. i think they only block the most popular scritps to stop people who have very specific use cases

https://github.com/hollowstrawberry/kohya-colab this seems to be up to date

[fervent thunder]

what do you think the best way to get only one subject in a generation? i am only using a 768x768 canvas and im utilizing 1girl and solo girl as my prompt and then multiple people and etc for the negatives, and uet im getting either multiple girls or hands reaching from out of frame'

because i know most models dont go past 768

and that can generate weird deviations

[astral goblet]

if you're using an SD15 model, its because of resolution attention. start at 512-640 level resolutions and then hires fix it up to a larger size

[quiet imp]

@lunar raft better to start with leonardo.ai on the web. It produces very good images. I'm new here and I have no idea where to begin and I've been doing prompt to image/video for awhile.

[lunar raft]

are they pretty realistic?

[quiet imp]

@lunar raft yes. And free mostly.

[lunar raft]

alright lemme try that

thanks for the tip

[quiet imp]

@lunar raft no prob.

[teal pagoda]

SD3 never.

[fervent thunder]

gotcha. tahnks

thanks

[lunar raft]

I tried downloading Kohya but got stuck when I had to download homebrew

[low moon]

hmmm supir kinda fux up eyes

[lunar raft]

what is the best model to use/train for a single subject?

also I dont need to train it, just make good images

[astral goblet]

if you don't need to train, what is this message all about?

If you're training loras you train on the base model so it has wider compatibility. But you don't want to train so i don't even know how to talk to you now. Getting jerked around here.

[sage reef]

maybe if you dont want to train, use ipadapter?

[lunar raft]

I am not saying I dont want to train, I just want the best images

if I need to train it I will

[sage reef]

but wait... you want the best images for training or how to generate the best images?

[icy ridge]

Hi everyone. I'm new here so not sure where is the proper place to mention this. Please do tell me.

I've just open sourced a Ruby SDK for Stability AI API Image Generation. It currently supports core and sd3 generation, and I plan to add the rest of the endpoints asap. Check it out here: https://github.com/OlympiaAI/stability

[lunar raft]

how to generate the best images

[icy ridge]

heads up @still glacier and @warm junco

[sage reef]

ok but then that has nothing to do with training at all, that is then just a combination of a model and settings that will output good pics

[lunar raft]

Okay there we go

[lunar raft]

Sorry for jerking you around, I am very very new to ai generation as stated in the message above, so my research was, and still is very scant, so I didnt know what gets the best images, and I thought that I have to train a model to truly get the best images, I guess I dont need to do that.

[astral goblet]

@sage reef "How else can I offend you today?" troll profile. he's stringing you along now that i've dropped it.

[lunar raft]

So then my question would be what is the best model to generate images?

[astral goblet]

sd 2. 1

[lunar raft]

whats the difference from 2.1 and SDXL?

Is SDXL for training models?

[astral goblet]

Pronouns: USA

[lunar raft]

whats wrong with that

[astral goblet]

🙄

[sage reef]

well the thing is the models out there, depending on what you want to generate, might not have that subject in the dataset, so generating that specific thing might not be good on any model, which technically means that yes it would have to be trained i guess. now assuming it is in the dataset, it's hard to suggest cause i dont know what you want to generate, but go with either high end models with general stuff like Juggernaut or haveall, etc or if it's something specific, like some anime stuff, go with animagine xl, etc, again i dont know what you want to generate to begin with. and also you would have to adjust the resolution and settings anyway.

[lunar raft]

I want to generate images of planes so maybe photorealism?

[sage reef]

well planes are common so i guess the general models should have that, so try juggernaut

[astral goblet]

fooocus comes with it by default last i looked at it

[sage reef]

i mean heck even base sdxl could be enough for you idk

if you have enough disk space, it doesnt hurt to download a couple models and just switch and try

[lunar raft]

Alright I will try juggernaut and the standard 2.1 and see what happens I guess

[lunar raft]

I use the web version does it matter?

[sage reef]

ah...

well go ahead

[lunar raft]

what?

[sage reef]

just keep in mind 2.1 is lower resolution than sdxl models

[lunar raft]

oh okay

[sage reef]

also, if you are using web stuff and if some money is not an issue, you can technically even use sd3 idk

[astral goblet]

ideogram

https://ideogram.ai/ probably the best paid model right now

[low moon]

lol

ideogram

[sage reef]

i mean... sure it might be lil weird profile, but to be honest i saw worse LOL

[low moon]

its ugly

[lunar raft]

bro I can train a model on the web

[low moon]

but does EVERYTHING

[astral goblet]

skill issue

[vale steeple]

Dont talk about paid models here especially some people who can be...

Well

Have some opinions on paywalling models

[astral goblet]

sd3 is a paid model. what do now?

[vale steeple]

Either leaks or just late-release

There's nothing much we can do in short time

[sage reef]

inb4 some random tech company releases something very similar to sd3 tomorrow out of nowhere 🙂

[astral goblet]

thats what pixart sigma is

[sage reef]

i mean pixart has some potential

[vale steeple]

Pixart Sigma diffusers you mean

This community have gone far to "We should unite ourselves and fight against the greedy corporation who closed source our models"

[astral goblet]

why is that a correction? pixart sigma is a DiT model

[vale steeple]

... Probably just those who had loud speaker on their hands

[astral goblet]

https://github.com/PixArt-alpha/PixArt-sigma

[vale steeple]

Wdym

[astral goblet]

i'm asking you. you added "diffusers" to correct me

[sage reef]

i always have to laugh a lil bit at their repo name and account name, it's like PixArt alpha presents PixArt sigma... instead of them having an account name that is not related to their repo names, but anyway lol it's like sd1.5 presents sdxl

[vale steeple]

Didnt the dev of PixArt have a name

[astral goblet]

Sheldon Cooper presents Sheldon Cooper & Amy Fowler's Fun with flags

[sage reef]

lol

[vale steeple]

Uhh, oh welp guess I just mistaken

[sage reef]

i mean diffusers is not part of the title anyway

[vale steeple]

Oh well they have none

Just Dalian-Hong Kong-Huawei Noah crossed develop

[astral goblet]

There's a lot of people rallying on Pixart Sigma since it's an openrail licence

expect some community models soon

[sage reef]

did they release training codes?

[astral goblet]

yeah and theres a beta branch of one trainer

[sage reef]

nice

[vale steeple]

While me dont even understand which allowed commercial and stuffs

lol

OneTrainer exclusively waiting for diffusers to compatible to train PixArt Sigma

So yeah

[astral goblet]

Doesn't matter who made it. They're released on a permissive license and they're pretty effective weights for a base model

SD.Next has support for it too i noticed

[vale steeple]

It uses int8 version of T5 though

( although you can easily change the entire folder to fp16 or bf16)

[astral goblet]

i don't think that matters for inference

[low moon]

whats the best universal VAE to use?

[astral goblet]

there isn't one. it's always use whats best for that case

[vale steeple]

Yep, just that my GPU running on int8 are like

Slow

[astral goblet]

int8 is quantization I think. thats a lot different.

needs a lot different instruction sets for integer math

[fervent thunder]

should the img2img sampler match with your original sampler?

[fervent thunder]

also, when im impainting, should i get rid of all my prompts and only add the things i want fixed, or should I add the things i want fixed to my exisisting prompt

im gonna assume add to existing

[fallow veldt]

it depends, it is something I wonder myself too. But it depends on the case.

For whole picture and trying to fix something, for laziness I usually use the same prompt. Sometimes in that case as you say I add the thing I want to change in the second place of the prompt after the style.

For inpainting "only masked" , or when trying to change something, I would try to add only the style and medium prompts, and prompt the specific thing. As the whole prompt wont make sense to SD with the context presented, and it will try to do the whole prompt in that space.

The context is really important, so for example you try to add another character, and you do "whole picture", fill (in Automatic1111 I'm talking), SD will think you are talking about the subject shown in the context, and wont create anything there, usually (or when I tried that sort of thing)

[fervent thunder]

hm okay. this is really awesome. thank you so much

in the past few days my brain has grown tremendously lol'

[lavish ravine]

hi

[nocturne perch]

hi

[astral goblet]

no it doesn't matter. the only thing that img2img uses is the starting pixels colors

[fervent thunder]

roger

[astral goblet]

inpainting is i2i on just the marked parts of the image

[fervent thunder]

okay cool. thank you

one more question for the night

i just downloaded a samler and followed the instructions on how. i dont see it. is there a setting i need to check off or something?

[astral goblet]

you don't just install new samplers. they have to be in an update. they're part of the core app usually. what did you download?

[fervent thunder]

ah

https://civitai.com/models/35966/dpm-2m-alt-karras-sampler

[astral goblet]

lol it's a hard patch to automatic. nifty. i'm not a developer so much but i think if i had made that i would've just done it as a pull request. the instructions are all there

oh it's only for 1.6 though. you've probably got 1.9

you probably don't need it

[fervent thunder]

ah i gotcha

thanks

[normal birch]

Hey all, I'm building a tool that lets you train a digital double. I'm using ComfyUI. While I'm happy to play around with it, I'm also happy to hire someone to help with me it 🙂 - Who here is an expert at Lora, Dreambooth, ComfyUI who wants to make a couple thousand $? 🙂

[woeful galleon]

I wants to earn but don't know anything about these things 😢😢

But I want to learn these

[regal rain]

i have a 3 part logs of sanoma I installed stable diffusion web ui

[dreamy dust]

hi

[regal rain]

hi guys i installed sd web ui in sanoma after following chatgbt for 24 hours 😄

https://drive.google.com/file/d/1IBgfLqReWwhhWNXvnSCJH1gtQscgWPTV/view?usp=sharing

https://imgur.com/rYfd6lA

https://imgur.com/aMqdy4z

[pearl ocean]

To think, SD3 might never exist

[little lava]

gm

[calm plover]

🎨 🤦‍♂️ 🤕 🔨 🤖

[limber grail]

hi guys
I am a teacher and I am trying to teach SD to children at a very basic level, im using comfyui
first of all i want teach text2img but I don't want them to produce explicit photos. what should i do ?

[late surge]

666

[teal pagoda]

Simple, just don't teach the kids. This is a too powerful tool for some children.

Teach them to manually draw for now.

[vale steeple]

for what? though

like what they learn after they prompt and click generate

I think it is better to convert children's drawing into real image using SD like some of us did with the help of ControlNet

[charred mesa]

guys its the end of may, im pretty sure it will come out 100%, very real and legit

my uncle's friend who doesn't work at stability said so

[teal pagoda]

sure thing

[opal hedge]

sd3 wen

[teal pagoda]

never

[opal hedge]

been feeling that way

[valid beacon]

sd3 about to come

[opal hedge]

Is there evidence from SAI for this or is this hopium

[potent spire]

did they even find the much needed investors?

[valid beacon]

sadly not yet,only ai companies that get free billions from the government are the ones that make ai models for drones to use them in wars

[potent spire]

they dont need billions from the government

that wouldnt happen

but private investors, individuals or companies

but with that policy they werent going to get there

and some of their competitors have now taken that place already and have companies and investors on their side

[valid beacon]

yes should have made drone models instead,money from gov is better because it never runs out,you always get more and more and no risk of getting regulated or banned as long as you stay under their boot

[opal hedge]

What's lil bro ranting about

[valid beacon]

gib money

[potent spire]

well regulation doesnt stop growth

and Midjourney managed to generate $200 milion without investment by investors

Stability AI apparently barely got to $5 million

[gilded folio]

that's so nice. teaching children about new technologies is what we all should do. But as the others mentioned, it'd better to let them draw on their own and turn these drawings into realistic images in front of these kids. That would make them love these techs.

[wet grotto]

hi guys o/

I'm architect, I'm doing a course of sd to architecture where will be presented at the architecture council of my country, brazil

this is still a novelty for architects, I've been working with it for a year now and I have a youtube channel all about sd and architecture

I came asking if anyone have some cool material, extension or anything interest to be apply in architecture and can share with me will be wonderful

can I post the link of mey channel here? idk

if anyone wants the link to show my yt channel just ask me 😉

[balmy abyss]

Hello
How are you?

[wet grotto]

fine thanks

[keen shard]

Thank you, everything is fine, how are you?)

[balmy abyss]

Also good!
What do you do?

[charred mesa]

sd3 will come out, just not soon 😔

[wet grotto]

next can run controlnets of sdx models?

[charred mesa]

we'll see what controlnet models SD3 will get

[wet grotto]

i dont try next yet, dont know much about it

will be cool 😄

when sd3 will come out?

@charred mesa

[charred mesa]

no official news

we are all just guessing

[wet grotto]

hum ok

hey

there are colab to next already?

[soft iron]

yeaaah

you finally understood what I wrote

[dusk scaffold]

chatgpt now can read images. how can i read them locally? which local llama can do it?

[wet grotto]

you need VL models

deepseek

you can found the space them on hf

https://huggingface.co/spaces/deepseek-ai/DeepSeek-VL-7B

@dusk scaffold

[dusk scaffold]

thanks very much

[narrow kernel]

Deepseek, llava 1.5 or 1.6, moondream2, there are a bunch of vision models.

[dusk scaffold]

There's a lot of model-00001-of-00015.safetensors, from 1 to 15. I need download all them?

[charred mesa]

just look for gguf, exl or awq versions or whatever

[wet grotto]

u r welcome ^^

[dusk scaffold]

yep. i'm downloading them using lm studio

[wet grotto]

lm studio is a backend?

@dusk scaffold

I use koboldai as backend, is opensource and free

[dusk scaffold]

LM studio to find and download the models

I'm testing the mode llava-llama-3-8b-v1_1-f16.gguf but it's not working properly.. I'll try other models

[dusk scaffold]

I enjoy kobold+sillytavern for some fun

[bold dirge]

gm guys, may I ask why I can't find lora training tab in Automatic1111 even I've installed dreambooth extension?

[dusk scaffold]

I'm wanting to use this method to see if the AI ​​creates a UI similar to this (https://images.examples.com/wp-content/uploads/2017/04/Desktop-Application-UI-Design.jpg). It doesn't seem to be working.. I'm using kobold + the models mentioned, but it's not working.

[raw burrow]

awh of course! feel free to tag me with any questions! i'll try to keep a look out 👀

[dusk scaffold]

It's working now. I forgot to download the mmoproj

[viscid tiger]

Is this possible to use IP adapter to change floor texture?

[wet grotto]

yes

[tall berry]

Boyyyyssssss and girls ! Be aware that there’s a website out there taking credit for other people’s loras saying it’s theirs 💀💀💀

I came across it when I was looking for more images to retrain my Lora and they said it was their ai 💀💀

[astral goblet]

it just stops small players from entering the field and competing with meta, alphabet, or openai

[potent spire]

tbh they cant compete anyway imo

or in rare cases

[astral goblet]

stability and midjourney wouldn't have happened in a regulated market

or wurstchen

or controlnet

[potent spire]

depending on how regulated, yeah

[astral goblet]

The automobile market was controlled the same way during it's most innovative era

[potent spire]

i dont like how they used the term "research" as a bypass to regulations

and now they are for profit

had never anything to do with research but with bypassing laws for profit

looking at you MJ

[astral goblet]

i dont think MJ claims research

[potent spire]

they call themselves "research lab"

i wonder if they wil lget in trouble as well btw

although they have apparently a revenue of $200 million on an annual basis

[astral goblet]

yeah i see that on their x. i've never heard midjourney called a research only model though. It's been subscription based since the start and the guy first marketed it in published interviews as a tool for rapid prototyping that artists can use

[potent spire]

possible, and now its a bit late for that anyway. They have established themselves at this point

[astral goblet]

you've decided they're a research model at this point you mean?

[potent spire]

compare those $200 million to like $5 million by SAI

[astral goblet]

back to the point, they wouldn't have ever gotten a foot on the market if there were regulations

[potent spire]

wdym?

[potent spire]

depending on how it would have been regulated yes

[astral goblet]

you were the one who says they're publishing research models and deceptively commercializing them. wdym?

ok, if the regulation was "you have to be nice" then yeah they could've played still. But we're talking about the kind of regulations that are being planned atm

[potent spire]

no, i mean that they were just selling themselves as a "independent research lab"

[astral goblet]

good luck arguing with yourself. this has become circular

[potent spire]

well in that case they wouldnt come far, indeed

neither them, nor Stability AI

possibly not even OpenAI with their stuff?

who would even remain besides of Adobe?

[astral goblet]

OpenAI is Microsoft backed. they're fine.

[potent spire]

yeah but i mean their initial stage etc.

but thats all what if ofc

[astral goblet]

bigger players than adobe. They're staying to ai solutions for images and video. they're more likely to buy tech from bigger players like MS, Meta, Alphabet

[potent spire]

they can afford it, especially due to their TOS and if a regulation allows those companies to use TOS the way they use it

so Meta and co. can legally train on your stuff and get away with it and so on

[astral goblet]

Copyright should never become so strict that a person can't give permission to someone to host their content. and if you want meta to host your content, it must have rights to it. TOS are just those terms. If you upload your stuff, we get to use it.

Education is needed here not regulation. People need to realize if they dont' want a company using their content, don't give that company content

Best privacy feature is the option to not publish

[potent spire]

its hard to avoid those tbh

like i really want to see some of the anti-AI people boycotting anything and everything they mark as pro AI

and avoid getting trained on

[low moon]

Why don;t I get the same result sin comfyui as forge is u se the smae image siz epromt and sampler settign and seed?

[loud solar]

Some settings in Comfy are different ...

[low moon]

thats not fair!

[loud solar]

Forge is kinda balancing some of the users mistakes out ...

[low moon]

sigh

i thought seeds and settings and those at leats give some extra consistency

please tech mature!

i hate beeing on the bleeding edge

its fun but i bleed

i noticed this with fooocus at firts. foocus realyl makes it less refined looking even with high steps

the eye sare smuddgy in foocus

[loud solar]

That's one reason why Youtube tutorials not always work ... they use comfy ... you wanna try on forge ... and fail ...

Sometimes you can see it cause the scale is different ...

[low moon]

i even added loras and perturbed attention guidance and everythign - literarily i copied every setting from forge

different image...

[loud solar]

One difference can also be: Forge using XFormers and Comfy maybe not ...

[low moon]

yeah

i hate to say it but forge does better with eyes

and in general you get much better results in forge with just a basic promt

comfy is better coz u can experiment mor etho but take smore work and patience

comfy eyes are also a but deformed and smudgy like fooocus

and dont ge tme started on upscalers....

omg

[loud solar]

I like the fast way generating some pictures with forge, take the best and work more on them ...

[low moon]

even Supir fux up eyes

[loud solar]

Comfy fells more like ... you've had a lot of work and this is your result now. Don't try again!#

[low moon]

ud thing a couple of circles ar ethe easiest thign to get right

lol

there are things u simply cant do in forge tho 😦

llike use 2 samplers or 5 samplers or start with cascade and end with sdxl

or start with ella sd15 go to cascade go to sdxl and supir and much more

[loud solar]

you can do in forge but not in 1 step ...

And sure some things you can't do ...

[low moon]

What upscaler yall use?

one that adds some detail but doesnt change too many things and keeps eyes as circles not turns them into blobs

[loud solar]

I try to do IMG2IMG and if needed SUPIR

[median temple]

has anyone tried https://www.reddit.com/r/StableDiffusion/comments/1d4cwi9/pcm_phased_consistency_model/

looks like a successor to lcm/lightning from animatelcm author? good 2 step results

[wet grotto]

@low moon is normal used one model togenerate a image thehn other to use controlnet?

for example, SDXL to generate a imagem adn 1.5 to use ipadapter?

[low moon]

its not normal

but it can be done

sometimes i think sd15 is more refined with textures and shapes but it sucks at prompt following and anatomy

so u can use sdxl to start up and do the last 15 steps with sd15

lol

hax

u cna combine whatever u want btw cascade sdxl sd15 sigma...

and in two weeks u can with SD3 too

but if u use controlnets u have to keep it in mind

controlnets only work with soem models

so

if u use controlnet or ip with sdxl then when u refine with soem othe rmodel the effects of those plugins wont work

you have to use the contorlnets/ipadapters on the first/main model in your generation, the rest of the phases are for aesthetics only

[wet grotto]

@low moon ok I get it, I'm doing some teste now, if necessary I can ask you something?

2 weeks, wow!

will be wonderful

sigma is a model? never used it before

I was a bit out of sd some time

was in LLMs

[pearl ocean]

Is it actually coming out?
Ppl have been saying it will come out in 6 weeks every week

[low moon]

its a meme at this point

[polar valve]

I just learned about Forge's existance. From what I've found about it, this is a godsend for my 🥔 pc. And I won't have to relearn the entire UI since it's practically auto. Are there any hidden holes or is Forge just a straight upgrade?

[keen falcon]

Does anyone got videos specifically explaining about sampling methods and all these weird parameters pls ? its kinda hard finding good resources on all these... Most tips I even get are "f*ck around and find out"

[river spade]

Is there a way to train like a LORA but for videos XD

besides using a SD Model and animating it

[trail lion]

Yeah it's better, I Guess the possible downside is it will lag from a development features perspective

[wet grotto]

ip adapter dont works to sdxl?

[trail lion]

It works

I use it all the time, what a great addition

[wet grotto]

I will test again

yesss ip adapter is very usefull I use a lot too 🙂

[tulip yarrow]

Why are there no SD2 LoRA trainer?

[trail lion]

Sd2? Because nobody uses it, it's a dead abandoned model

[tulip yarrow]

oh... that's why

but why do more people use sd1.5 then

since it's (persumably) an older and worse model

[trail lion]

Well tons of content, easy to train on

[polar valve]

And is way faster to run on a lower end rig

In comparison to xl

[trail lion]

It's particularly popular with the waifu crowd

[tulip yarrow]

I wonder what SD3 loras would be like... and when SD3 would ever release to the general public...

[wet grotto]

pls can you told me the name of preprocessor and model op adapter's controlnet that you use?

i can't make it run maybe using the wrong controlnet

[forest turret]

hi everyone 🙂 does anyone know if SD can generate sprite sheets for video game characters?

[viscid tiger]

I’m doing some research but didn't find any example or solution
Could you send me some references?

[fervent thunder]

guys is SD more easy with mac like faster with mac coz of NPU than desktop pc?

like faster than 4080

like macbook air or macmini m2 pro

[brittle coral]

Can you install both comfyUi and automatic1111?

[fervent thunder]

with NPU

[brittle coral]

I seen it done on Civitai.

[brittle coral]

I am about to find out. I got a M2 16gb air, M1 Pro 64gb and a windows box with 64gb ram and a 12gb 3060.

[shell tendon]

Yes

Very easy

[supple garden]

hello lovely peeps!

how can i start from zero creating images on stable diffusion?

what are the easiest ways to start for someone who hasn't done anything yet?

[loud solar]

Easiest way to use an A.I. without paying and for getting a first idea is to use: https://www.craiyon.com/

If you are getting the idea you are on the right way you can care for using other webservices or install Stable Diffusion local ...

[raven agate]

Text to Spritesheet maker now lol... https://civitai.com/models/448101/sprite-sheet-maker

I also have a comfyUI voice assistant workflow as well now lol.

[wanton igloo]

yo

[pseudo fulcrum]

yo

[teal pagoda]

Forge is better, maybe only some extensions won't work as intended.

[pearl ocean]

I got thr M3 Pro.
But I use my gaming pc for SD

[wet grotto]

sure, I have a tb channel focus on architecture, I will send you the video using ipadapter

well, I don't record yet the specific video changing texture but I have this one using ipadapter and you can see other videos on the channel

https://www.youtube.com/watch?v=mxin2EjzvQM&list=PLRvM4LTxFnNVZxqIJlkB0ff3V6drhUD-r&index=2

any questions you can ask me

guys

is hard traine a model?

for example I'm architect, I wanna a model about a specifically architect and I have a lot of photos, how can I traine the model? any video to recomend me?

[fervent thunder]

thx man eagerly waiting excited dm me with results wheneever ur free

[inner heath]

yo

[fervent thunder]

yo

[surreal jay]

I got a 1TB SSD yesterday specifically to store SD and all the models. Loads pretty fast compared to NAS!

[gritty walrus]

@wet grotto you can train a Lora using OneTrainer or kohya_ss if you have a good GPU (not awesome, just good). If not, you can try to do it on collab or one of the online services, like civitai.com that has one. I dont have any video at hand but just search 'lora training' on youtube and pick one

[blazing wolf]

https://tenor.com/view/dancing-cat-dance-cat-cat-meme-chinese-cat-gif-12629347036627000898

[hazy bramble]

Hello

Can someone help me style transfer my vacation photos

I can pay for credits

And toss some change for more credits left over for the help

I used neuralstyle.art but results are varied, im not sure I'm too happy with the results

I only have DALLE with gpt4 rn

[wet grotto]

thanks I will check it!

[little lava]

gm

[dawn mulch]

Anyone know how ComfyUI downloads files from huggingface?

So im using windows, and a VPN that gives me free internet with xfinity wifi. but the speed is capped at 100 kB/s, its slow but fast enough to watch youtube videos at 480p. But i recently saw that when using the CMD and certain thinhs, such as ComfyUI, i can downloads things at 3000 kB/s, sort of circumventing the speed limit. When i try to downlaod these files through a browser then I am restricted to 100 kB/s. But again using COmfyUI. and curl for CMD, I get 3000 kB/s. I dont know how this is happening but i assume it has something to do with ports, prxoies, etc. It doesnt seem to be unique to COmfyUI it seems. Does anyone have any idea?

Note that this is a recently new thing because previously it was not like this because everything was downloading at 100 kB/s.

With the speed limit it can take 3 days to download a 7GB file. But with this interesting option, it takes less than 1 hour to download a 7GB file. This is quite amazing and I want to understand how it works so I can try to ultilize it for other situations.

[teal pagoda]

Guess I can help if you answer

[hazy bramble]

Hi

[hazy bramble]

Here

DM or whichever channel is fine

[gritty walrus]

Huggingface has an API. all projects just use the lib and consequently that API. I tried diving deeper into it because that fella was filling up my C drive but gave up and made a hard link pointing to another drive instead. Too troublesome

[elfin turret]

Hi

I have a problem with stable diffusion

It gives me an error when I try to install

[wet grotto]

how are you tryng to install?

[tulip yarrow]

Alright so are there any other LoRA trainers other than SD1.5 and SDXL

[proper karma]

stable diffusion is not something you can install, you'll need to be more specific about the exact software that you're trying to run, what kind of computer you're trying to run it on, etc...

[tulip yarrow]

yeah it requires GPU which i don't have 😭 wish it was just an installable instead of a gpu

[elfin turret]

ERROR: Could not find a version that satisfies the requirement torch==2.1.2 (f

I have a good grafics card

[proper karma]

LoRA is a technique for adapting models using a small subset of params that tries to approximate "what's the minimum we need to change to be able to do this", it's not a piece of software that's specific to stabilityAI's model releases.

[proper karma]

you should probably stick to maybe using invokeAI (like a single-install-and-it-works kind of thing), or just use ChatGPT or bing chat to generate images bud

[elfin turret]

Invoke ai allows changing the model?

[oblique elk]

i wish there is a website you could enter words like invoke ai and get these results:
Other features

Support for both ckpt and diffusers models
SD1.5, SD2.0, and SDXL support
Upscaling Tools
Embedding Manager & Support
Model Manager & Support
Workflow creation & management
Node-Based Architecture

[hollow flame]

i'm wondering if a very specific extension exists. i use an LLM with instructions I've written for it to generate great SD prompts. it's kind of a pain copying and pasting each new prompt into SD, so I want to generate one prompt for X amount of images, switch over to the new prompt automatically then generate for another X amount of images, etc etc. is this possible? I'm on Forge

[elfin turret]

Thanks

@oblique elk

[lusty beacon]

SD3 next week? 🤔

[pearl ocean]

They say that every week

[supple garden]

Thanks!!!

[loud solar]

Have an inspiring time

[lusty beacon]

One Youtuber said that

[loud solar]

I'm a Youtuber, too ^^

[teal pagoda]

So May is out of the game now

[lapis urchin]

anyone have experience with dynamic prompts?

[teal pagoda]

SD3 will never be released

[lapis urchin]

lets say i did this {X1|X2|X3|X4} is there a way to make it so the first image gets X1, second X2 etc?

[brittle coral]

Has anyone tried the Krita Stable Diffusion Plugin?

[royal slate]

is it possible to add preview images for models, etc. that don't have them? Editing them has a 'replace preview' button but says "no image in gallery".

[boreal dew]

in a1111?

[royal slate]

yes 🙂

[boreal dew]

there's an extension called CivitAI Helper

[boreal dew]

CivitAI Helper, lemme go get a screenshot of what it does

https://github.com/butaixianran/Stable-Diffusion-Webui-Civitai-Helper

ahhh i can't post images in this chat, lmao. woops

that should tell you what it does though

[royal slate]

just paste that url into extensions tab?

[boreal dew]

I believe so, yeah

from 'install from url'

Extensions > Install from URL > Paste that link in the first box

[royal slate]

can I also install controlnet the same way or is that different?

[lapis urchin]

adetailer seems to be messing up constantly

[boreal dew]

it's the same way I believe

Sorry, just saw this

[proud niche]

You mean like: https://github.com/adieyal/sd-dynamic-prompts?tab=readme-ov-file#basic-usage ?

[lapis urchin]

those are examples

[blissful relic]

There was a very happy and joyful expression on his face.

[proud niche]

That extension should do what you want. Tutorial here: https://www.youtube.com/watch?v=s-1L6MCVh-E

[opal hedge]

It's officially june in UK no SD3 yet

[royal slate]

should I be renaming .pt to .pth?

[boreal dew]

no

[whole vigil]

Hey does someone have experience with swapping faces with stable diffusion I used control net and reactor but in both cases the faces don’t look like the original they look partly like it but not enough. Is there a way to do make the swapped face look more like the original?

[little lava]

gm

[narrow kernel]

Get it like 80% of the way there in terms of likeness->Photoshop to get it to 90-95%->resample it with a fairly low denoise.

I don't really ever do face swapping, but the above still applies to pretty much every workflow where you're trying to make an image a very specific way.

[fallow veldt]

something like @narrow kernel said, but use Face-ID while inpainting for the face swap. It can be used for generating, and it is good as it will try to respect the subject characteristics (except you prompt against it) but it is kind of cumbersome as it will always place the face in the same place and change the style a bit, better to have a generation and then change the face via inpainting.

Also depends on the reference image, multiple reference (all at once or separate) might help, but it may take several inpainting steps at various denoising levels, and some judgement from you about the likeness.

[junior crescent]

Hi

[novel shard]

Hii

[steep wave]

create an african woman realistic image showing high end frequency separation, with pure black and gold color

[brittle coral]

Not including gender specific variations there are 134,719,200,000,000 individual "looks" (?) for humans.

Feel free to look over my spreadsheet:
https://docs.google.com/spreadsheets/d/1IpyQ4TyLrQukqpWisenW_GwzNayYXVIEjJuVnhT9maU/edit?usp=sharing

Part of a prompt generator project. Comments welcomed

there is a related discussion in #🧣｜comfy-ui I don't know how to put the thread is this reply.

[wet grotto]

hi guys

somenone can explain me or recomend me a video about the using of "(", "{", "|", "BREAK" and ":1,3" in prompts?

hey

SDXL yet need more then 1024 to generate a image or can be any resolution?

[pure kelp]

Whenever Stable Diffusion generates what I would call "messed up complicated objects" (like bad hands, faces, eyes), which are quite hard to fix with normal inpainting, what I do is
-make a copy of the generated image
-crop it down to just the weird part
-go to img2img, put the cropped part there
-set the resolution to 3x higher than what the cropped image has
-generate the image
-photoshop it into the old one and inpaint any seams
I've had relative success with this method, but it is quite the hassle to do it every time. Is there a tool for maybe automating this? Correct me if I'm wrong, but doesn't ADetailer do that? I've tried using it, and it works very well for faces (not so much for anything else).

[real karma]

What do you guys suggest for getting the most realistic lighting? (I use fooocus AI) Do you guys have any best to use loras etc?

[drowsy robin]

I've not done image generation in a while, and I've recently gotten back to having installed ComfyUI for Intel Arc.

I can't seem to get IPAdapter to work, so for now I just want to see if I can do controlnet instead.

[wet grotto]

guys, what are the best models for achitecture?

I use realisticivision and juggernaultXL, someone else to recomende me?

[arctic sedge]

@sleek otter Hey man. Do you know what model of SD Omost is running with? I can't seem to fiure out which one. I assume it's a custom model that can only be used with Omost?

[trail lion]

the models posted on civit all have descriptions. the majority of the trained checkpoints are going to be trained on people , so look for 'building' or 'architecture'. otherwise, go with the base. the 2 you listed are going to specialize in people

[wet grotto]

Really? I understand... This models are so good to architecture, so the buildings models are better? It's great!

[loud solar]

https://civitai.com/models?tag=buildings could be a start

[wet grotto]

Yes I would check there 🙂

[trail lion]

civit is a pain to navigate when you're looking for something, it should be way easier than it is, but it seems to get slightly better over time

but it's free so beggers cant be choosers

[loud solar]

Hope you'll have a lot of fun exploring A.I.!

[wet grotto]

Thanks!!!

[trail lion]

interesting, I've never had an issue with inpainting, but you do have to know how to nudge it in the right direction, which may require openpose or inpaint-sketch (with an import from gimp/photoshop)

[loud solar]

PineAmbassador is more a pro than me ... just wanted to give you a quick link. But ... it can also be a big fun to check your prompts with other models ... so better keep the good ones. SD Forge and A1111 have a Text-Script to run multiple prompts one after the other ...

[umbral fog]

Hola, tengo una duda, al momento que hago el cambio de rostro, si un objeto esta encima de la cara por ejemplo la lengua está la distorciona y tengo que modificarlo en modo paint. Hay alguna forma de que reactor pueda hacer las modificaciones sin hacerlo en inpaint

[pine tide]

yo guys the stable diffusion resolution of pictures max is soo low 512x512 till 756x756 how do people get so crispy pictures out of it? 😮

[loud solar]

Using SDXL ....

[pine tide]

are there different methods or is this one the standard everyone is using?

[umbral fog]

Hello, I have a question, when I change the face, if an object is on top of the face, for example the tongue is there, it distorts it and I have to modify it in paint mode. Is there a way that reactor can make the modifications without doing it in inpaint

[loud solar]

SDXL is trained for 1024*1024 ... depending on the User Interface you use you might need a different method....

[pine tide]

how to get this up to 2k/4k like every other normal picture? Is using an upscaler a must after the picture prompting? (im very new sory for the newbie questions)

[loud solar]

Well you can upper one aspect ratio and lowe the other one ... and than ... depending on the User Interface you use send it to IMG2IMG and double the size before you use an upscaler

[pine tide]

thanks this clears some stuff

[loud solar]

For example you can use with SDXL 1280*720, double it in IMG2IMG and than use an upscaler ...

[umbral fog]

can someone advise me??? I have a question, when I change the face, if an object is on top of the face, for example the tongue is there, it distorts it and I have to modify it in paint mode. Is there a way that reactor can make the modifications without doing it in inpaint

[pine tide]

nice im trying it already out, thanks G

[loud solar]

Have fun!

[lapis urchin]

is there a way to use animediff or something similar without an input video?

[low moon]

does ELLA work with controlnets?

😄

[wet grotto]

What is ELLA?

[low moon]

https://www.reddit.com/r/StableDiffusion/comments/1bbxr7h/ella_equip_diffusion_models_with_llm_for_enhanced/

[wet grotto]

checking

one technical question

diffusers is for generate images like LLM are the term for text generate models?

make sense? XD

[dull yarrow]

hello, um has anyone checked tooncrafter?

[full lark]

It looks cool af, too bad I get an error when trying to install it :C

[valid charm]

america is a 3rd world country

[cold carbon]

where do we write our dream commands

[zinc bane]

Hey, I was wondering if it is possible to train a lora to a person( full body, close ups, etc.) and if so, how to go about it and what checkpoint is better at realistic photography for this job.

[lime gust]

i downloaded it on my pc and the promts im putting in are fairly normal, it looks really realistic but alot of people look like fucked up morphing nightmares

any reccomendations for settings/config ?

[bronze spear]

Have you started looking at models yet?

[lime gust]

no how do i do this

can i send nsfw example

i put it on the basic model and now it looks mostly better but the faces dont look as good

like default settings

[shut umbra]

Does anyone know a good lora for motorcycle racing suits?

[little lava]

gm

[gilded folio]

Guys I havent had the chance to try omost. Is it as good as they say?

[tender cove]

@echo marsh lol

[echo marsh]

oh hey

[jade sentinel]

Hi,I 've tested the Faceswap with Fooocus.
Is it possible to generate video with Fooocus?
Users upload their selfi photos and generate their video animation?

[wet grotto]

IC Light exist for A1111?

normal map is better then depth maps?

[teal osprey]

its been 3 months since i last used SD. were there any updates or anything i should do before going into it again?

[wet grotto]

ic light
in apint upload is a workflow that I learn now and it's very helpfull

ELLA

[teal osprey]

sorry i dont understand anything can you elaborate please

[oak latch]

could i pay someone to help me gen very realistic images i can offer my 4090 to use

[fluid atlas]

hi guys

is anyone here?

[cedar salmon]

sorta

[vestal tundra]

not talkative much... people doin their stuff... more images I mean

[unborn heart]

does anyone have any tips for getting solid white backgrounds through prompting?

[hoary fog]

why does it refuse to listen to me when i try to make it a night setting

[wet grotto]

ic light is new, ELLA is new, in paint ulpoad is usefull

[fervent thunder]

IMPORTANT QUESTION:

When training a character (in my case, celebrity) LoRA, it is said that one must not include too many pictures of one particular attendance of an event (Golden Globes, BAFTAs) as this will result in bias within the LoRA. However, if I included 10 pictures of each event, for every event, wouldn't that get rid of the bias as it is now evenly distributed? I'm asking because, there is sometimes great variety in camera angles, composition and facial expressions in photographs of one event, but I am not "allowed" to use all of them, because the maximum is one for each event, greatly reducing the potential of the LoRA.

[rancid venture]

k