💬｜general-chat

[clear brook]

Hello all you gentlemen and ladies, it’s evident everyone here is so so talented with stable diffusion, but would anybody be willing to BLESS me with the most advanced to date text to speech ai software, both normal voices and than celebrity/character deepfakes. Thanks in advance 🙏 this community is filled with helpful people

And I don’t want to believe the first ad that pops up on Google, so if anyone is aware, I would love to talk.

[untold flax]

.............................

I apologise for even starting a 'help me help me' message. Figured it out - cheers for the pointers

[untold flax]

bruh there are many videos, like..... loadssssssss of tutorials.

bless the ai path you walk with some time spent in the tube. Stable Diffusion is for pictures though so maybe shuffle this question to the right place and you'll get what youa re asking for. Literally loads of guides and videos.

You fuggin G - thank you

[pseudo bough]

anyone know why sd mov2mov does nothing in the command prompt when i start a process clicking generate? get no result the web ui just hnags with the interrupt button and skip button grey with 0/75 stuck on it

[hidden jungle]

helloooooooooo

[ornate flame]

No announcement here for StableCascade?

[rugged spoke]

as i understand no good solutions to use automatic on windows with amd now?

ive read some ppl happy with results on 7900 being near high end nvidia but i suppose thats a linux users? 😑

[ornate flame]

It's worth it to learn to install Arch Linux imo

[rugged spoke]

its just really makes mad, to have expensive card that performs badly -_-

[ornate flame]

Yeah AMD's software team seems lacking

[rugged spoke]

maybe ive didn tweak it to the top on windows, any attempts ending up error in memory mostly.

i do add same commands as ppl having also 7900 xt

but for me they dont seem to work for some reason

neither medvram or --no-half --precision full

seems to cause errors

[broken cave]

i have a 7600 on my desk for validating comfyui on windows. the problem is i don't have time

[rugged spoke]

only optimization engine that seem able to make highest resolution images is sub quadratic, although lil bit slower than others

i wanna use automatic 😦

is 25 steps 768x768 in 26 sec is a bad speed or no?

no turbo it can make somewhere1024-1024 in 12 sec but turbo so rare for now -_- , and that often memory errors

[broken cave]

it all comes down to people who can program also having the hardware you want to test, and you will discover that choosing bad hardware is misaligned with having the knowledge to program

[rugged spoke]

the most annoying problem is that 20 gig of vram seems to clog up very fast , and sometimes it tells me that even 1280 on 1536 is too much of a resolution

not sure if CPU matters much , but its somewhat weak for now -_-

[broken cave]

you are at the beginning of a long journey

it really depends what your goals are

[rugged spoke]

the most weird part is that this --no-half --precision full -medvram that work for most 7900 xt users doesnt work for me. Ending up alot of errors =\

Im not sure about the goals... to have fun. Have at least medium speed generation - not to have this often memory errors

[broken cave]

to have fun.
is IT fun?

there are UIs that make more IT decisions for you and should make things a little easier

otherwise bing image creator is a lot more capable for most people and use cases

not meaning to sap your excitement or anything

but when i help people get acquainted with this stuff i get them on bing image creator first because it actually works

[rugged spoke]

on the top of all it feels like randomish or stars position ones. maybe depends on model & promt. One day it can generate in img2img 1536x1536 with no problem, other it cannot generate higher than 1024. It can go with hires fix upscaler as high as 2500x3000 . But often generating very high resolution image once or two times ends up in vid memory full error

ppl seem to enjoy amd on linux, mostly) . But ive got a feeling setting it up there is frustrating -_-

[broken cave]

there isn't really an audience for super high resolution imagery so nothing is optimized for it

[rugged spoke]

Heard theres rumors rocms coming to windows soon, and that might end suffering of amd users)

[broken cave]

it sounds like you are mostly having trouble upscaling?

[rugged spoke]

y generating in high resolutions

and some random memory full errors

upscaling via res fix-scaler or extra tab seems to be more less fine-stable

having different promts with additional loras to generate in txt2img in high res( one is generating even 1280x1536 other cant go higher than 1024-768 without errors). Can 1-2 additional lora & styles affect vram clog up that much?

& also why something that work for others might not work for me? In webui commands. I might have missed something to install ? It can depend that much on how different every people setup & brand of a hardware?

why other ppl 7900 work with medvram and mine doesnt o.0

well if remember right it gives error somewhere near at 93%. or after generated once. More likely 93%

[sturdy crown]

我找会中文还是应文的web工程师。Dear,

I'm excited to help enhance your online presence.

After reviewing your website and social media, I'm confident we can elevate your brand with:

Modern, user-friendly website development
Daily content creation and strategic social media management
Logo redesign for contemporary appeal
Sign up now for our comprehensive services and enjoy the first month of social media management for free. Let's discuss further during a brief call or meeting at your convenience.

As a special offer, we're flexible and willing to discuss pricing that fits within your budget
Looking forward to collaborating!

Demo Website Pdf Availabe

Best Regards,

[nova zodiac]

Zluda is coming…

[rugged spoke]

Is there some magic way to not have theese vram errors & problems with clogging up on AMD now?

[gray wyvern]

https://openai.com/sora what is this ???

[lusty beacon]

Check this out https://openai.com/sora

[winter pike]

those gens looks super real, i never thought that level of quality video generation would be achieved that soon

[lusty beacon]

same

[sinful ruin]

you and me both....I am absolutely gob smacked it's here. I knew this would be possible someday....I was thinking years though

I was just fucking around with creating motion thru the caveman methods available right now.......

and then this drops

out of nowhere

[gray wyvern]

imagine this realtime ?!?!

[sinful ruin]

this is almost too much to process

mentally

the implications

[gray wyvern]

tech paper out later today

wondering of you could use motion LoRAs to have ID specific microexpressions, animations

[crystal creek]

Hello everyone. Give me some advice, please. I want to train LoRA on a large number of images (for example, 10k), I know that such a number is acceptable in the case of styles and complex objects/items. How many steps should there be for each image (i.e. what number should I put at the beginning of the folder name)? I realise that a lot depends on the train dataset and other settings, but I would like to know at least an approximate range.

[warm junco]

For any amd question message me xD

[warm junco]

bigger model = more vram usage, what was your gpu again? it shouldnt crash with 5gb

[iron imp]

@oblique ivy whats up? You sent me a friend request. You a bot?

[warm junco]

yes was spam

[lyric dirge]

Hello everyone, I installed automatic1111 via run.bat and finished it with update.bat, but I don't even know how to open it. Can anyone help?

[desert ember]

yea, I could really use some help as well, I followed the guide in the support channel and its still breaking during install

[fervent thunder]

hi guys i downloaded the ip-adapter-faceid-plusv2_sdxl_lora, for lora, but when i place it in the lora model file, and restart stable diffusion, nothing pops up

does anyone know a fix?

[silent dome]

Anyone here ever mess around with Silly Tavern?

[nova zodiac]

have you got a sdxl model loaded at the top?

[fervent thunder]

yea

[nova zodiac]

I don't think anyone is familiar with the one click install

[desert ember]

damnm

[rotund coyote]

where's the AI art bot

[nova zodiac]

#1047610792226340935

[rotund coyote]

All is lost

[nova zodiac]

well no - heaps of free generators around - civitai have one, happyaccidents.ai, aiscribbles.com

[rotund coyote]

hope they are good

need a book cover

[nova zodiac]

I mean they will be as good as the bot here was - if you are looking for something with more control, you could try https://tinybots.net/artbot/create

[rotund coyote]

thank u

[left gulch]

https://youtu.be/8f4ohKxypRM?si=xlQ-Wb-GSBoFydgU

[olive mauve]

Is there a guide or tutorial someone can point me to in order to install Beluga? I have no background in coding and a little lost once I am redirected to Hugging Face

[nova zodiac]

Got a link to beluga?

[olive mauve]

https://huggingface.co/stabilityai/StableBeluga2

[nova zodiac]

Try https://huggingface.co/spaces/Sidaddy/Beluga2ScriptGenerator

[olive mauve]

Im sorry I am not sure what to do with this. I only have experience with Midjourney and ChatGPT. I am a bit over my head trying to figure out how to install this. Not sure where to start

[nova zodiac]

That shouldnt be an install, thats a use it right there job

[olive mauve]

I see, ty

[atomic bane]

Error: [Errno 2] No such file or directory: 'C:\Users\richardkim\anaconda3\lib\venv\scripts\common\Q33RPXYW.DOCX, when install on window how can i solve?

[nova zodiac]

Youve gone astray somewhere if the venv is in the anaconda3 folder - what are you trying to install?

[haughty cypress]

Guys where can I generate text to image here, I can't find a way!

[prime swan]

which chat room should I go to if I'm looking for LoRa or checkpoint suggestions for a project I'm trying to get?

[halcyon willow]

Sora is amazing

How even is that possible?

Shame ClosedAI won't share any technical details

[crude garden]

Does it bot will catch with the Sora

Sora is very good when i saw it just now

[clear brook]

I have a video that I think was made with stable diffusion, I have $20 for anybody who can recreate it for me and teach me or just tell me the softwares needed to recreate, it’s pretty simple. DMs open

[static cape]

The compute they need for that behemoth can probably not even run like people are running Dalle3 at the moment. I bet it needs huge amounts of Vram AND since it's from ClosedAI it'll bever be available Opensource / without a subscription and behind closed doors.

Our only hope is that they a) Share at least some technical details so it can help the OpenSource development of alternative Models or b) that sometime in the future something leaks (similar to how the NAI leak skyrocketed SD v1-5).

[fierce raptor]

Definitely will not run locally - not a chance in hell considering the compute it doing when generating these. The research paper is interesting (even if very light on detail) with the concept of diffusion 'Latent Patches' being equivalent of LLM text tokens. I really, really, really want SD to give us a video equivalent we can run via services like Mage that have a lighter touch policy on censorship as compared to OpenAI. Heck, I would pay for a service providing this quality from SD themselves (up to say £35 a month), but they are under heavy scrutiny as it stands with how people are miss-using the tech.

[jolly karma]

Hey everyone!
I'm working on a photorealistic character who's pregnant using Automatic 1111 with Realistic Vison V-5.0, but the stomach keeps coming out horrendously disfigured and honestly terrifying. Does anyone know what the best method would be to get a natural looking pregnant belly? Mabe using inpainting, or any +/- prompt suggestions?
Any help is greatly appreciated!

[agile tiger]

maybe use open pose if there is one

ipadapter

then add some embedding if someone has made

or give more power to the negative prompt

[warm junco]

there is a lora for this

[jolly karma]

@agile tiger Sweet I'll give this a try. Thanks!

[clear brook]

Sora?

Is that legit

[agile tiger]

yes

jobs bye bye

no more actors

[clear brook]

Why would anybody use stable diffusion

Is it out now?

Does OpenAI have a text to speech software?

[agile tiger]

they have it all but no enterprize would want to get into a closed system

[clear brook]

Wdym?

[jolly karma]

@warm junco Do you happen to know where i would find it? SD is pretty new to me, so I'm still learnin.

[warm junco]

on Civitai.com you find everything, its the largest model database.
the lora you need is flagged NSFW, so you need an account to see it

[warm junco]

nope

[warm junco]

SD is OpenSource and Free, also no censoring and it runs localy on our machines instead of a cloud from a corp

[clear brook]

@warm junco are you aware of the most modern text to speech software? With ai voices and deepfakes and all

[jolly karma]

@warm junco Awesome. I appreciate the help

[warm junco]

eleven labs is good

idk any free ones so far

[oblique willow]

I hope stable audio become overpowered like diffusion

[drowsy robin]

Quite a few, all of which are of lower quality however.

[oblique willow]

At this point everything is copyrighted. And audio is just a pattern of sounds mostly repeated. Ai should replace it asap.

[drowsy robin]

TortoiseTTS, Whisper, OpenVoice, and Bark.

Some others that I don't know on the top of my head.

[warm junco]

nice thanks! didnt looked much into tts

[oblique willow]

I don't remember the name but there is ai which lets u generate audio in popular people's voices.

[drowsy robin]

Well, RVC.

But that isn't TTS.

That's VTV.

[oblique willow]

Hm. I guess VTV would be more realistic than TTSm

[drowsy robin]

Well, you have a real voice to base it on.

So ye.

[oblique willow]

Ye if they can speak..

[drowsy robin]

A TTS can be fed into RVC however

[oblique willow]

Is it totally free though? It needs credit igm

[drowsy robin]

RVC is open-source.

[oblique willow]

I dont remember much it needed something like credits which you can get by letting it use your gpu for few hour.

I am not sure if that is diffrent ai

[drowsy robin]

Idk what you're talking about.

Voice.AI?

That thing kinda sucks.

[oblique willow]

A ai which converts your voice into voice of famous people for example donald trump.

Very realistically

[drowsy robin]

RVC is real-time.

So.

Lol

Go look up "Okada's RVC"

[oblique willow]

I think its not rvc then. I dont remember name.

[drowsy robin]

Being honest, what I really want is 2D to 3D conversion.

The only thing I know of that does this well is MIDAS 3.1.

[clear brook]

@drowsy robin you seem to know a lot, where do you get your info

[drowsy robin]

By researching.

Lmao

[oblique willow]

Its Voicemod net.

[clear brook]

And money isn’t a obstacle, what is the best tts out rn

[drowsy robin]

Elevenlabs.

[clear brook]

Right lmao but all due respect the top 20 searches are advertisements

[drowsy robin]

Nothing beats Elevenlabs.

[clear brook]

Love to hear it

Was about to drop a bag on the tenterprise for lovo.ai

[winter pike]

rvc is the best for transfering voice to voice, weights(dot)gg is the biggest rvc model catalog, its the civitai of rvc models

[clear brook]

Glad u saved me

[drowsy robin]

My guy, I have been a part of the AI space since SD1.0. I just keep up with Reddit/AI stuff.

DallE 1.0

[oblique willow]

I was talking about Voicemod . Net

[clear brook]

What’s your time zone

[drowsy robin]

Why do you want to know my timezone?

[clear brook]

Because I’m abt to dm you for contact info id like to speak to you more

If you’d have the time

[drowsy robin]

Let's just say I'm close to EST. Also, I'd like to not do that.

[clear brook]

Fair enough

[drowsy robin]

I have this discord's DMs disabled for good reason.

[clear brook]

You know you know your stuff and everyone wants to ask you

[drowsy robin]

Well what do you want to know?

I am in this chat right now.

[clear brook]

Just wanted to dm you so I could ask you more technical questions as I get more involved

No big deal though I understand need for privacy

[oblique willow]

I dont think anyone in 2024 really uses dms.

[drowsy robin]

Too many scam DMs.

Too many untrustworthy people.

[oblique willow]

Ye

[clear brook]

I sure don’t accept them

But like we all agree once sora is accessible to the public it’ll be #1

[marble steppe]

I picked up a Prime B550M Mainboard, 16GB RAM, Ryzen 5 5500 and have an RX 6800. Going to be installing Arch Manjaro on a 128GB NVME.

What do I need to know to install stable diffusion, and/or is there a recommended guide to follow?

[drowsy robin]

Unsure. https://www.reddit.com/r/StableDiffusion/s/Xubpk5sLKg

SORA is good, but closed source.

I will wait until Emad and stability one-up SVD.

[oblique willow]

Text to voice is dull though. The moment people hear ai voice. They close the video..

[clear brook]

Svd?

[drowsy robin]

Stable Video Diffusion.

[clear brook]

What 😭

[drowsy robin]

#▶｜stable-video-diffusion

[oblique willow]

I never watch any ai voice video now matter how good it is..

[clear brook]

Where in the world did you make that assumption

That’s you man

Rule #1 never assume you are the average person

All due respect

[drowsy robin]

Text to voice (elevenlabs) is still the highest in emotional, quality output.

It is good enough for audiobooks. And speaking.

[oblique willow]

Okay lets be honest you tryna create youtube shorts/long with an Ai voice over it. I would say average people would not like it unless its a 3 minute tutorial. I am only saying it becausee, in case you expecting it will go crazy and work as good as real human. It would not. Everyone can recognise Text to speech. If u are putting effort in it. Then remembwr audio is 60% of the video. People can watch video with audio only. But best vudeo with dull audio is garbage.

Ofcourse no one means to demotivate you or anything

Just a friendly advice before u waste too much time

If you trying to do this*

[clear brook]

Well I’m not but I’m sure sny1 who would wanna do that appreciated your advice 💯

[drowsy robin]

Lmao I'm not gonna take your advice on that and I apologize.

[clear brook]

^^

[oblique willow]

Advices depend on people and situations. No one is forced to follow/ignore.

[drowsy robin]

Thank you, Captain Obvious.

I simply do not agree with the statement you gave based on the multitude of repositories and projects that focus on that very thing.

[oblique willow]

Its okay. I am not suprised that you do not understand.
You aren't creator or someone who does this.

[drowsy robin]

You assume wrong.

[clear brook]

Catcher cmon man you injected yourself in our convo nobody asked for your negativity

[drowsy robin]

I'm not wasting any more time on this conversation. You clearly seem to think you're in the right.

[clear brook]

@drowsy robin SD can add motion to photos? Do you have a link to a good tutorial on that

[drowsy robin]

SVD can handle video motion.

It's a different model is it not?

But animdiffusion does exist.

#▶｜stable-video-diffusion

[oblique willow]

Again. This is a public chat. And i can say whatever i like. I am neither toxic nor I am saying anything which you need to care about. All I did.. writing something I know and have experienced so that "someone"'s time or efforts can get new direction if they are reading. Agaij, nothijg is focussed on only you.

[drowsy robin]

Yeah, no. No thanks.

[clear brook]

So elevenlabs.io is top notch for tts

I assume you are a big SD fan

[drowsy robin]

I am a fan of open-source.

[clear brook]

But in any other fields are you an advocate of ai softwares for?

Like for example anything to do with organization etc it’s a shot in the dark

[drowsy robin]

Organization?

I mean, I consistently mess with LLMs.

[clear brook]

Basically asking if you use any ai for anything else in your life that you’d like to share

Shot in the dark

ChatGPT 😇

[drowsy robin]

Nah, open-sourced.

LLama2, LLava, Mistral.

[warm junco]

GPT4All

[clear brook]

And what are their purposes

[drowsy robin]

https://os-copilot.github.io/

LLMS are language models. They simply process a token context input and then output a response suitable to said input.

A guessing of tokens that seemingly is rather coherent.

OS-copilot, along with other such softwares (Open-Interpreter, Autogen Studio and Self-Operating Computer) take this to their advantage.

[clear brook]

I have a lot of sauce on GitHub but my skill set isn’t aligned with code too well

Anything with a UI?

[drowsy robin]

I cannot code.

Lol

[clear brook]

I can’t even understand it

One upped you

So how do you use GitHub?

[topaz parcel]

Not sure if right channel. to ask this - but does anyone know the process (using A1111) to create the individual images at the various stages, from non-villain to full villain? How do i create the separate stages/steps - https://civitai.com/images/4936746. (once i have 5 images from beginning to final, I ca then use a video editor to create the animation - but how do i get each image needed)?

[drowsy robin]

^ anim diffusion does this.

[topaz parcel]

is that an extension?

[drowsy robin]

It should already be part of A1111, I thought.

[topaz parcel]

oh ok - i'll google it, to see if any tutorials online

[drowsy robin]

Well I mainly just clone and test out repositories based on the instructions given for said repository.

Which are usually just in the readme.md

[clear brook]

And once you successfully do that what is the end result

[drowsy robin]

That... depends on the repo. Lol

[clear brook]

Like it still doesn’t have a friendly ui

[drowsy robin]

If it isn't designed for one, no.

[clear brook]

Honestly I gave up after spending hours trying to effectively get GitHub settled on my Mac

[drowsy robin]

Huh? Can't you just install git on Mac?

[clear brook]

Whatever I was doing was not working

I had to go through a terminal

[drowsy robin]

https://git-scm.com/download/mac

[clear brook]

I dmed any chance you’d accept

[warm junco]

nope its an extension

[drowsy robin]

In SD.Next it is native.

That'd probably be why I thought that.

[topaz parcel]

So - I am relatively new to SD (although not to AI art generation). And I sometimes get a bit thrown by some of the inclusions I see in prompts that I experiment with from civit.ai.

For example - I see this in prompts. ~~aesthetic~~. and seems to be random, without meaning? But it must mean or do something.

And then also I see prompts that have certain words of phrases in single, double and triple brackets (like this) ((or this)) (((or even this))). what do they mean / do?

Is there anywhere that has a good guide that breaks down, in simple terms, all of these little nuances and features that make up a prompt?

Thanks

[noble nimbus]

Basically (Something) does the same thing as Something:1.1

likewise ((Something)) = Something:1.2

it's for emphasis, the picture you get is a bit more Something

[austere robin]

So how can I generate images in stable ai

Can anyone explain

[pale latch]

the extension is stupid too. it might as well be standalone. it doesn't tie into auto1111's systems at all and is just it's own tab with it's own underlying code. like running comfyui in an auto1111 tab

[sterile torrent]

Guys someone know how improve eyes quality? I use ADetailer in comfy but that helps with face and is good but eyes still almost bad, flat

[gilded ermine]

the extension is stupid too. it might as well be standalone. it doesn't tie into auto1111's systems at all and is just it's own tab with it's own underlying code. like running comfyui in an auto1111 tab

[viral prairie]

Was wondering if anybody has been able to achieve a good workflow for coloring manga with control net ?

[warm junco]

yea i think so too

[pale latch]

i'm happy to be playing with cascade, but theres no reason it needs to be an extension with the capability it has. could've easily been standalone.

i guess it helps get it out there for the masses

[velvet fossil]

hello

[hallow plaza]

whats the best place to download the insightface\models ?

[magic geyser]

hello

[vivid kelp]

can we use stability ai API img2img inpainting with Controlnet?

[pale latch]

insight face is a tricky one to install as it requires building. but you can find prebuilt whl files for it. let me dig a thread thats helpful up

https://github.com/Gourieff/sd-webui-reactor?tab=readme-ov-file#viii-for-windows-users-if-you-still-cannot-build-insightface-for-some-reasons-or-just-dont-want-to-install-visual-studio-or-vs-c-build-tools---do-the-following @hallow plaza

[halcyon willow]

Fookin ClosedAI

Cant they just make the research public

[chilly vector]

💰

I'm sure that with enough time an open source alternative will pop out

[gaunt horizon]

Does anyone have a colab for stable diffusion video?

Looking to something like the fooocus colab

[nova zodiac]

Emad already replied to a thread somewhere they have something similar already

[chilly vector]

I barely lurk this server so I can't say I know about that. I'm just a bit miffed about how most people seem displeased with the OpenAI footage released recently. I just can't help but think about the endless possibilities really

[pale latch]

If you're a qualified expert in the field, they may provide access. I'm not sure how they're rolling this oen out but they done similar things in the past. They do share their research. Just not with the wide open internet.

[nova zodiac]

The problems are two fold - one is the fact that it democratises movie making and will put a lot of creatives out of a job; and another that it competely screws with the “seeing is believing” truth that we have currently, and that the world is definitely not ready for a world where you cant trust anything you arent seeing in real time…

[marble steppe]

would someone be able to link me an install guide for SD on manjaro linux, I can't seem to locate it in pinned messages anywhere.

for AN amd CARD

[tropic frost]

say im curious, do you guys generate images in a small format until you get something close to what you look for, then copy that ones seed and refine it for better results?

[nova zodiac]

Yeah, doing seed hunting without hires fix on, then fixing seed and then doing it with the hires fix

[forest trout]

Yeah high-res fix is awesome.

[chilly vector]

it's as simple as asking someone.

[bronze umbra]

Excuse me, I'm new in here, and I still don't know how can I generate images. I'm sorry for my ignorance. Anyone could help me, please? Thank you in advance.

[nova zodiac]

Easiest way is go to aiscribbles.com or use the generator on civitai.com

[vast berry]

ngl cascade isn't even that great

[nova zodiac]

If compared to base sdxl or base sd1.5 its pretty good - give it a month or two for finetunes to come out then youll see its true potential

I dont know if it will end up better than existing models but we will soon find out

[pale latch]

is easier to train too so the potential may catchup and suceed xl even sooner

[clear brook]

Where is stable diffusion is it a website oorrrr

[nova zodiac]

Stable diffusion is a technique for making images from text prompts…. If you wish to make images, start at aiscribbles.com

[covert beacon]

Can SD now generate multiple but different people?

[nova zodiac]

Yep - sdxl is better at it than 1.5 out of the box, but there is a tool called regional prompting which is good for splitting up concepts in an image

[dim thorn]

gm

[frigid sinew]

who can help me with stable diffusion

Basically currently I have stable diffusion web UI

that is different than sdxl?

Also what is the difference between sdxl and sd web ui

[wise crystal]

no one can

[nova zodiac]

Stable diffusion webui is a user interface - sdxl is a model for using in that interface (as is sd1.5, sd2.1, and all the checkpoints on civitai)

[frigid sinew]

how do I check which one I currently have? @nova zodiac

If I understand correctly sd1.5, sd2.1, and sdxl are all models I can use on stable diffusion webui?

How do u recommend I use sdxl

[frigid falcon]

Hello everyone. New to this server 👋

[nova zodiac]

Download sdxl and put it in the stable-diffusion-webui/models/StableDiffusion folder

Then fire up the webui and generate away 🙂

[primal shuttle]

SD 2.1 is not for using. It's for taking up space on your hard-drive because you keep forgetting to delete it.

[nova zodiac]

It has a niche with the StableSR upscaler

[sleek otter]

I am using PINOKIO. It installs very comfortably. A1111 installation and InstantID installation are 6 x slower than my already installed A1111?!?!?!?
Only speedy installation is IPAdapter+FaceID

Now installing Kandinsky-3 (not via PINOKIO)

[oblique dome]

I want to make my own LoRA for decapitated heads, i'm using Kohya_ss to do so but i have a question.

my instance prompt is "Decapitated head"
my class prompt is "Human"

what should my regularization images be??

They are optional so i guess i could try without any but idk if it'll turn out well

[nova zodiac]

Dont generally need regularization images

Especially for single concept loras

[oblique dome]

alright, good to know

thnx

[warm glen]

Hi everyone. (if I'm writing in the wrong place, please correct me). I want to train my model, but not on people, but on furry art of one artist. Tell me, maybe someone was engaged in training a model not only on portraits of people? I have a couple questions to ask.

[nova zodiac]

#🔧｜finetune

[warm glen]

Ok

[low forum]

where i write command? help me!

[nova zodiac]

If you are trying to make an image here, you need to read #1047610792226340935 then head to one of the sites mentioned at the bottom

[brittle geyser]

Question, anyone have any good NSFW discords? Got questions that probably aren't applicable or appropriate for the general

And I mean stable diffusion discords

[nova zodiac]

I mean you can talk about creation of a single concept without actually describing the concept…

[brittle geyser]

Seriously? Troll?

[nova zodiac]

Yeah, second one in two days…

[brittle geyser]

I would but downside is I also want a place to look for ideas and references. Kind of good ideas of what's actually capable.

Just some kids sitting in his parents' basement wanting attention.

Ignore them and they go away.

[nova zodiac]

@sudden ruin @bleak matrix you about??

[sudden ruin]

[brittle geyser]

https://tenor.com/view/bane-no-banned-and-you-are-explode-gif-16047504

[nova zodiac]

Thanks chief 🙂

[brittle geyser]

Outstanding response time there mods

[nova zodiac]

Yeah they had one going for over 90 mins a few days ago - that one was rough

[brittle geyser]

I self pruned me trolling the troll.

No need to leave it hanging around now that the neck beard has been booted

[sudden ruin]

Im trying my best, but I also got a life and gotta sleep 😅

[brittle geyser]

Dude I got three kids, run a business and run a gaming server in my off time. I feel that 100%. IRL comes first but we do what we can.

[sudden ruin]

Feel free to ping and/or use the report function anytime

[brittle geyser]

What really bugs me is these war thunder and Arma experts commenting on the conflict and whatnot in Ukraine.

I did 10 years active duty army, 36 months combat time in Iraq. Al-Asad, tal-afar and Baghdad. I was actually in Baghdad for the 2010 Iraq election. That shit was rough.

These little "kids" are lucky they don't have the slightest fucking idea what they're talking about when they try to troll or make any sort of commentary on combat. 🙄

[sudden ruin]

Please move that to #🌶｜off-topic if you wanna keep discussing

[brittle geyser]

Just wanted to comment on the combat thing. I actually take a solid stance on refusing to ever discuss politics lol.

Politics, religion in music always galvanize and end badly.

[nova zodiac]

Me - im just excited x-adapter code has been released 🙂

[chilly vector]

I personally don't think it would be that bad. Fake news are easy to fight really

[nova zodiac]

It is now, but given how far we have come from Will Smith spaghetti abominations less than 12 months ago, it will quickly get to the point where the average human has to question everything they see online…. That sure as shit aint gonna happen

[potent spire]

how is this supposed to put a lot of creatives out of job?

[nova zodiac]

Just look at how many people thought the pope was wearing balenciaga

[potent spire]

thats when people are not looking

but i dont wonder with the Tiktok generation anymore

[nova zodiac]

Thats my point, people arent looking and wont look, especially if it reinforces their world view

[chilly vector]

most people in social media can be divided into two really, those who are and those who aren't dumb.

most people would figure out when something is AI sooner or later

[nova zodiac]

But this is wandering into #🌶｜off-topic territory

[nova zodiac]

Itll get to the point where you need forensic level interrogation to figure it out

[chilly vector]

I don't think so.

But yeah this could be continued on the other channel

[marble steppe]

I keep asking in tech support but there doesn't seem to be much activity there, is there a guide on how to install SD on linux with an AMD card? I'm seriously lost and have never used linux before

[wraith forge]

Hello

[light pagoda]

What happened to the bots?

[floral linden]

Perhaps you can search for 秋叶 aaaki on Bilibili (a Chinese video website), who has issued a document to solve your problem and also provides an integration package。夸克：https://pan.quark.cn/s/19a36cab36ac

百度：https://pan.baidu.com/s/1QDqo2uEoUS_NY1olb4vmVQ?pwd=aaki

sha1校验码：c7c5d497360c7ec3fe9af5ada1624842341d8275 作者：秋葉aaaki https://www.bilibili.com/read/cv26557731/ 出处：bilibili。 This is the download link、

[solemn marsh]

bots still down...

[bleak matrix]

Good morning! How is everyone today?

[woeful oxide]

hey I've questions and need some suggestions. . am i able to design shirt using this. already got the logo and i need to insert my logo in the generated photo . like a human wearing my shirt

[jolly karma]

I need some help.
I'm getting ready to train my first model/loRA, but I'm a little confused.
I'm wanting to generate very specific realistic physical features on random characters. For example say I want "large noses" when tagging my dataset images should I be tagging things that I do not want to see in my desired output, or tagging the things that I would like to see?
Also would a model, or loRa be better for specific physical traits?
I desperately need help. I've spent many hours researching these questions, but I can't seem to find any straightforward answers.
Thanks in advance!

[velvet fossil]

I need some help please

[charred thistle]

does anyone else find mild entertainment adding contradictory prompts and watching them fight it out?

[nova zodiac]

#1047610792226340935

[nova zodiac]

#1047610792226340935

[nova zodiac]

I generally make sure that I caption as much as I can. That said there is a thing called masked training in the OneTrainer software that allows you to mask the input image so it knows what to train specifically on @jolly karma

[nova zodiac]

It is occasionally entertaining - lighting prompts are good for that

[nova zodiac]

Photoshop/Krita still best bet for that stuff atm

[nova zodiac]

Good thanks

How bout u??

[velvet fossil]

I have a problem

"RuntimeError: Not enough memory, use lower resolution (max approx. 896x896). Need: 0.5GB free, Have:0.4GB free"

what do I do?

[warm junco]

Hey what's your GPU? And what's inside your webui-user.bat?

[velvet fossil]

I didn't run that

I just ran webui

GPU 0

NVIDIA GeForce RTX 2070 SUPER

Driver version:    31.0.15.3623
Driver date:    6/8/2023
DirectX version:    12 (FL 12.1)
Physical location:    PCI bus 1, device 0, function 0

Utilization    1%
Dedicated GPU memory    7.7/8.0 GB
Shared GPU memory    1.2/15.9 GB
GPU Memory    9.0/23.9 GB

[nocturne pewter]

Looks like we’re getting an update, looking promising https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.8.0-RC

[tropic frost]

curious about something

when you use a real picture in openpose

will the overall body of the original image affect the end result?

like if you use a large person, whoever you use the pose wtih will get affected and have a similar frame?

[nocturne pewter]

Nope, OpenPose should get the pose only. With other ControlNet models you can copy appearance too.

[naive frost]

i never heard of it but i just started using photostructure as an image gallery server and its just chugging htru my 10tb no problem

supposedly it has facial recognition but i havent seen a signle thing about htat in the UI so.. hopefully it does?

[jolly karma]

@nova zodiac I installed OneTranier and ran it a bit. It's incredible!
Thanks for the suggestion.

However i can't seem to figure out how to even use masked training. Any tips there?

[nova zodiac]

No sorry, Ive not tried it myself - maybe someone in #🔧｜finetune can help?

[latent rampart]

Im using stability.ai API and getting alot of "CONTENT_FILTERED" responses. Even though there is nothing wrong with the image. For example the prompt is "An image of a person standing still, their face a mixture of deep reds and blues, capturing a tumultuous blend of emotions. The person's fists are clenched, and their eyes are wide with a sparkle that suggests tears, lit by the harsh, contrasting light surrounding them." It blurs the result and gives CONTENT_FILTERED

Has anyone experienced this and know of any tips or solutions?

[opal hedge]

Does anyone know if stable cascade is trained on LAION

[nova zodiac]

They havent announced what it was trained on, but if it wasnt a subset of Laion the gasps will be audible

[opal hedge]

I apologize if I'm barking up the wrong tree, but do you know if they typically use the original meta data on LAION images or do they use taggers or some sort of context model to improve image descriptions?

I ask because cascade is a clear improvement over sdxl but the prompt understanding doesn't seem much better.

[lofty shell]

can anyone recommend a model, lora, etc to generate 16x16 pixel art for inventory icons? something very similar to minecraft icons

[nova zodiac]

No idea sorry

[nova zodiac]

Youre best bet is to use a pixel art lora at 512x512 then downscale the result (search civitai for pixel art)

[lofty shell]

what model do I use?

and which lora

theres a lot

[frosty oriole]

SDXL Controlnet vs SD ControlNet, is there a major difference within actual control of the images?

[hasty olive]

Is there a resource to have the text/tags used to generate the image be simplified? I am not good with stable diffusion yet so I don't know how to organize the details

[trail lion]

The prompt can be as simple as you want, "a photo of a dog", the complexity is just trying to nudge the interpretation in a different direction

[hasty olive]

I dont really mean simplified, I mean optimized. can I send you a dm of what I am trying to do?

[trail lion]

No need to dm, just state more plainly the actual question

[hasty olive]

I am trying to make a bat character for a campaign I am working on but no matter what I do to push the image in the direction I want, stable diffusion pushes back and screws up something like not including the wings, making her hair short, ect. this is my prompt
1girl:1.4, bat girl, (pitch black skin color:2), (very thick thighs:1.4, small breasts:2, long legs:1.4), (adult:1.4, mature body:1.5), (large black bat wings on back):1.5, (large pointy elf ears1.5), (long dark gray hair, messy hair, hair covering both eyes:1.5), (light freckles covering whole body), (golden eye color:1.4, dark circles under eyes), wearing a black leotard, black thigh high leggings, full body, simple background, SFW

[hasty olive]

The pictures are close but it isn't what I am looking for. Also when I put something in the negative prompt it seems to specifically add what I don't want

[trail lion]

well, this is more of a generic statement, but the very nature of text to image generation is getting something unique every time. prompt engineering is what you're asking about, but it can only go so far ,sometimes you need more tools in your belt, like controlnet, or using other images to create depth maps, inpainting, outpainting, sketch models, etc

so if you have an image that's most of the way there, you can use that image to maintain the parts you want to keep, using for example a depth map or a canny in controlnet

from there you can change the prompt to add your subtle variations

[trim magnet]

also this prompt sucks,too many words,too many weights on words and a bunch of redundant sentences

[hasty olive]

I have no idea what any of that means😅

[trail lion]

good, then you know what to research next

[hasty olive]

That is what I mean. I have no idea how to simplify it to make it more concise to what I want

[hasty olive]

I will do that next, thank you!

[trail lion]

I have had images, for example that were like 90% of what I wanted, but I wanted the arm in a different place. These things are possible, but not just with prompting

[smoky rain]

I haven't played with the models/tooling for few months. Where are we in terms of reusing specific elements (key character, object) in different scenes?

[nova zodiac]

Much better place - ipadapter and instantID for sdxl are really good

[smoky rain]

instantid has a premade runpod dockerfile, nice

[frigid sinew]

who can help me with whisper?

I am trying to get timestamps

segment

[nocturne pewter]

I'm trying to get a differnet head using maskin in inpainting. I get ok results but most of the time the new head and face are slightly off being too big or too small. Which ControlNet model I could use to help with that? I want different head and face but in the way it sits well and looks natural.

So far Canny has given me the best results. With Canny, the problem is it's making the face look too similar to each other.

[trail lion]

Depth is probably better to maintain the position and just change the look

[nocturne pewter]

Thanks, will test that next

which processor would do faces best, Midas?

[nova zodiac]

Midas or zoe

[nocturne pewter]

Midas seems to do pretty good job, will test Zoe too next

[lime crest]

Hello, what are the best arguments for a 6GB VRAM ? I use --medvram --no-half-vae --no-half

[warm junco]

What's your GPU?

[lime crest]

Nvidia GTX 1660 SUPER

[warm junco]

Then you need:
--xformers --medvram --no-half

Xformers will give you a performance boost

[lime crest]

But I heard it decreases quality

[warm junco]

Nope

It doesnt

[lime crest]

It is worth waiting a bit longer if the quality is higher

[warm junco]

It doesn't decrease quality

Its not worth to not use it

[lime crest]

Thank you, I will try

[warm junco]

Np 🙂

[steel garden]

hey, why does my code not work? How can i load my checkpoint (https://civitai.com/models/271592/big-head-3dxl?modelVersionId=306137)
Code:

from diffusers import StableDiffusionPipeline
import torch

def generate_images(prompt, num_images=5, local_model_dir="/home/yago/bigheadsdxl/bigHead3DXL_v10.safetensors", output_dir="./generated_images"):
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    torch.cuda.set_device(device)

    try:
        pipe = StableDiffusionPipeline.from_pretrained(local_model_dir, local_files_only=True)
        pipe = pipe.to(device)
    except Exception as e:
        print(f"Failed to load model from local path: {e}")
        return

    for i in range(num_images):
        generated_image = pipe(prompt, guidance_scale=7.5)["sample"][0]
        image_path = f"{output_dir}/image_{i+1}.png"
        generated_image.save(image_path)
        print(f"Image {i+1} saved to {image_path}")

if __name__ == "__main__":
    generate_images("a sweet cat")```


Error:

Failed to load model from local path: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/home/yago/bigheadsdxl/bigHead3DXL_v10.safetensors'. Use repo_type argument if needed.

[edgy oyster]

I haven't been here for a while. Is it no longer possible to use stable diffusion via the channels in this server?

[nocturne pewter]

ah crap, I hate hands and arms. Been trying to get a decend looking hand for an hour or more now, still getting all sorts of weird stuff. ControlNet tested etc.

[fervent thunder]

Is dream studio down? I cant every time i try to create images i keep getting something went wrong on our end please try again later

[tropic frost]

got a question. Do you think using koikatsu pictures in image2image will get you more accurate designs of the characters in it? (under the condition that you use a koikatsu image and the lora of the same character you want)

[marble steppe]

other than Arch Manjaro, what's the best linux distro for running SD with an RX6800?

manjaro is frustrating me that I can't use pip to install stuff

[warm junco]

yes that could help, but mostly a lora is enough

[tropic frost]

im thinking of consistency issues

like if you have a character with a more detailed design

[serene depot]

hey guys, if i have an image of a shirt, and i want to create a similar art style shirt just with a different material, what kind of prompt i should use?

or settings that im missing

[trail lion]

you might be able to do that with IP adapter, but essentially you are telling it to use an image as the basis, so it's a controlnet thing vs a txt2img thing

try loading it into img2img, and into controlnet with ip adapter, and use a prompt like "make it cotton" or "make it silk". I've never tried it, so not sure if it'll work

start with a lower denoise and increase if nothing happens

[fervent thunder]

hey guys

why can I not use finetuning?

[trail lion]

bad luck?

[fervent thunder]

the bottom row of options are all greyed out

[trail lion]

in kohya_ss?

[fervent thunder]

no in https://platform.stability.ai/sandbox

[trail lion]

oh, not familiar with that

there's a tech-support channel though, fwiw

[fervent thunder]

where?

nvm thanks

[serene depot]

whats ip adapter?

[trail lion]

https://huggingface.co/h94/IP-Adapter/tree/main

there are plenty of youtube tutorials and such

[serene depot]

yeah im looking right now to even how to install it 😛

[serene depot]

any chance you have few mins to guide me trough it on disc in the voice channel here? ill sharescreen

[trail lion]

do you have controlnet installed? or know what that is?

[serene depot]

nope, i have stable diffusion and able to generate some images with prompts atm

[trail lion]

ok, so then I would recommend watching some basic videos on what controlnet does, working without it, kind of like going into battle without a gun frankly

[serene depot]

alright, will do

[abstract ibex]

where do i start generating images ?

[trail lion]

depends, you can install something local on your computer, or use an online service

[abstract ibex]

how do i start generation here

[trail lion]

there are bot channels

I dont use them, but check them out, you'll see others generating stuff

[abstract ibex]

actually i am new here so can you let me know where specificslly

[trail lion]

hold on, I have all those channels hidden, let me undo that

[thorn willow]

bots are down since a month I think

They should be back soon but idk when

[abstract ibex]

what should i use till then ]\

[hoary oasis]

Hello to everyone!! 🙂
I need to be useful to my wife. She is an aesthetic doctor and needs an image (before and after) for a gummy smile treatment. You know. A smile with a very thin upper lip and when smiling, the gum and frenulum are prominently visible. This is a very simple and beautiful medical treatment.

I'm trying to make the first image, the before. I'm using "Fooocus", because I saw that it is an AI specialized in faces.
It should be noted that I don't have much of an idea and I have managed to run it because there is a .bat on github. Otherwise I wouldn't know how to do it.

The fact is that I can't get it to display the image correctly. No matter how detailed the prompt is regarding the topic of gums, it only makes normal smiles. Normal lips + pretty teeth. And it is precisely the opposite what I am looking for. A variant of that. But I don't know how to do it anymore. Can anyone help or guide me?

infinite thanks in advance community!

I used this: https://github.com/lllyasviel/Fooocus?tab=readme-ov-file

veeeeeeeery grateful in advanceeee

[abstract ibex]

how do i run or use this

[hoary oasis]

I used this prompt: "
A close-up view of a woman in her 40s, capturing her wide smile with her thin upper lip, clearly exposing the gum tissue and the frenulum that connects the lip to the gum. The texture of the gum tissue is pronounced due to her wide smile and the tension of her upper lip, which draws the viewer's attention directly to the space between her lip and teeth. "

What i have said wrong????

[trail lion]

you're right, I wasnt aware because I never used it, but the post about it is here #1047610792226340935 message

[serene depot]

got the control net installed and running, which ip adapter model should I get?

[trail lion]

so according to that link, you'll have to install SD or find another way to generate images with an online service

[thorn willow]

There are a few options, first one is to download stable diffusion but you need a powerfull computer. Else you can use midjourney, craiyon, runwayml, canva, ...

Adobe firefly is pretty good too

[trail lion]

sorry, that link I posted above was only for 1.5. Are you using sdxl? if so go here lllyasviel/sd_control_collection

on hugginface

just grab the ip-adapter_xl.pth, and here https://huggingface.co/h94/IP-Adapter/tree/main/sdxl_models

[serene depot]

And I put it under the same folder models?

I have AAA1 so I installed the control net trough the extension tab

[trail lion]

stable-diffusion-webui/extensions/sd-webui-controlnet/models that's on my pc, but I'm linux, same concept though

[serene depot]

Thanks a lot buddy, I'll try it!

[trail lion]

good luck

[serene depot]

Looks like I'll need it 🙂

[latent scarab]

Hey all

[echo kite]

any good models that capture the feel of early AI image generation? (dall-e mini etc)

[fervent thunder]

are you also interested in changing shirt to underwear too?

[radiant frigate]

Can anyone give me the link through which I can generate images with sdxl in huggingface?

[radiant frigate]

and also runway ml

[fervent thunder]

guys for deepfakes, what are the state of the art methods?

[radiant frigate]

(I don’t want to run them locally)

[serene depot]

Not really no, I'm developing a game and I bought some art asset like a shirt, pants shoes.. I need more assets for different class and level

[spring marsh]

off topic, but i generated an image on a core 2 duo with 2gb ram in only 22 minutes

this is actually crazy tbh

[upper rose]

off topic, but does anybody know how to faceswap in a video?

[tepid yarrow]

What would be more interesting: buying credits on dreamstudio or becoming a membership on the stabiliy.ia website to create realistic photographic images?

[nova zodiac]

Think of membership as acquiring a licence to use it (and thats free if not selling anything); credits on dreamstudio will allow you to generate

[nova zodiac]

Impressive patience!!

[spring marsh]

and pain

[nova zodiac]

Search the huggingface spaces for sdxl

[nova zodiac]

Runway ml has to be done on the runway website

[nova zodiac]

Just find the earlier model (ie stable diffusion 1.4)

[spring herald]

Let's make a RIP on Dream's generations, presenting our favorite best works in a video slide to remember. And let's share in this chat. This will of course take time... but looking at the work that came out in the dream, I can’t compare it to anything else. Whatever engine, etc. standing in this chat it was impossible to repeat! or let’s share the works that impressed you and became irreplaceable forever... I’m completely inspired and I’m so glad that I was part of this project!

👍🏻

I just now saw that I can’t attach an image or video, maybe I should create a separate chat?

[torn coyote]

Hi! Has anyone here experimented with cascade? I'm just getting into it, and I'm trying to figure out if there's a specific prompt format that works best.

[spring herald]

Dream の世代に敬意を表し、思い出に残るビデオ スライドでお気に入りの最高の作品を紹介しましょう。そしてこのチャットで共有しましょう。もちろん時間はかかりますが…夢に出てきた作品を見ると、他のものと比べることはできません。エンジンなど何でも構いません。このチャットに立って、繰り返すことは不可能でした。または、あなたが感動した、かけがえのない作品を永遠に共有しましょう... 私は完全にインスピレーションを受けており、このプロジェクトに参加できて本当にうれしいです！

画像やビデオを添付できないことがわかりました。別のチャットを作成したほうがよいでしょうか?

[woven owl]

Someone got an Idea of how to implemented real fashion to an ai picture like and sweater from Nike. Is there an add on to automatic1111?

[clear oyster]

are the stable diffusion video requirements same as normal sd?

and is there a demo

[tepid yarrow]

Thanks!

[nova zodiac]

general with images

[nova zodiac]

there are krita and photoshop plugins for both a1111 and krita

[nova zodiac]

nope, video requires way more vram - I'm pretty sure theres a space on huggingface that has the demo

[nova zodiac]

cascade and sdxl share clip models, so full sentences in the prompts should go better

[clear oyster]

yeah thats what I thoujght...

[serene depot]

Can you elaborate about the photoshop plug in? I have aaa1

[nova zodiac]

https://www.youtube.com/watch?v=Y3KJli8ohKI

[serene depot]

Thanks!!

[stark hornet]

can i use comfyui with the new stableCascade yet ?

it always tells me missing modules, but i cant seem to find them

[nova zodiac]

yeah, it is built into comfy now, will need to update the whole install

[stark hornet]

just a git pull ?

hmm, i didnt even install with git, but the portable version

how do i best update it ?

[nova zodiac]

find new version of portable installer I think

[stark hornet]

hmm, i got a thousand addons and stuff installed, you gotta start again from scratch, yes ?

i think i just lack the new nodes for cascade, any idea where to get those ?

[woven owl]

Thanks a lot! Do you have some links for more information?

[nova zodiac]

https://civitai.com/articles/4008/is-krita-the-better-tool-for-you

[nova zodiac]

Nope, should be able to install over the top I would have thought...

[stark hornet]

can you give me a link of what to get exactly by chance?

[nova zodiac]

https://www.reddit.com/r/StableDiffusion/comments/1aswzcd/native_stable_cascade_support_for_comfyui_just/

[stark hornet]

thx

[stark hornet]

where do i get that clip_g_sdxl.fp16.safetensors from and where do i have to put it to?

[nova zodiac]

huggingface (or it should download automatically?) you might be best asking in that reddit thread sorry

[stark hornet]

hmm, it works with model.safetensors instead of the clip_g_sdxl ... but doesnt seem to make too good results. but i cant find the other one anywhere

[frigid sinew]

Hello all. Basically in open ai whisper I am trying to change this python script instead of console.logging I would like it to save the segmented results into a .srt file. But currently when I console log it its just in pure text, non segmented format

import whisper

# Specify the full path of your audio file using raw string notation

audio_file_path = r"C:\Users\USER\Desktop\MikeQuinnPostVoiceRemoval.wav"

# Load the model

model = whisper.load_model("base")

# Transcribe the audio

result = model.transcribe(audio_file_path)

# Print the transcribed text

print(result["text"])

Which I call using this python interpreter

& C:/Users/USER/anaconda3/envs/whisper/python.exe c:/Users/USER/Programming/AI/Speech-to-Text/Transcription/transcribe.py

[slate dust]

This is Stable Diffusion discord

[frigid sinew]

ask me if i care? @slate dust

why dont u mind ur own business

with ur shitty ai dinosaur pfp xD

lol ill stop

[slate dust]

Ok you rude pos 🤷🏻‍♂️

[frigid sinew]

u started it

[nova zodiac]

@frigid sinew - whisper is not a stability AI tool, you might luck and and find someone in here that is able to assist (possibly in #🎵｜stable-audio ) but there are probably better discords/subreddits to be asking that question in, or possibly even asking chatgpt by putting in the code, the result of the print statement and then asking it to edit the code to get it into a format you want

[frigid sinew]

well i figured it out

i just had to read a blog post

[nova zodiac]

nice 🙂

[frigid sinew]

yup

do u know how to use sd-xl

[nova zodiac]

I have used sd-xl

what would you like to know?

[frigid sinew]

how do I add it to my stable-diffusion-webui

I dont even know what version I have 1.5? 2.2?

when I look in ./models it is just controlnet and other stufdf

[nova zodiac]

so you have stable diffusion webui (automatic 1111)

ok cool

[frigid sinew]

yeah

[nova zodiac]

if you open it up, you should see a version number at the bottom

very bottom of the screen.. should be a 1.x number, likely to be 1.6 or 1.7

[frigid sinew]

How do I start it again?

which batch file

[nova zodiac]

webui-user.bat

[frigid sinew]

1.7

[nova zodiac]

if its above 1.6 then to run sd-xl all you need to do is download a sd-xl model and put it in the ./models/StableDiffusion folder

[frigid sinew]

I thought there are more better versions though

like 2.5

[nova zodiac]

ignore the web interface for now, you've got a system that will work

[frigid sinew]

ok. current my models are

v1-5-pruned-emaonly

realisticVisionV60B1_v51VAE

what models do u recommend?

[nova zodiac]

you can download a model from civitai that is using the SD-XL base (I recommend https://civitai.com/models/229002) shove that in the models/StableDiffusion folder and away you go

(assuming you have 8+gb of vram)

[frigid sinew]

yep I have 12gb

stable diffusion xl is just inpainting?

do u have any tutorials that would resemble my interface

when it finishes downloading

[nova zodiac]

stable diffusion is the whole method of making art from text prompts

[frigid sinew]

they said sdxl combines inpainting and image generation

[nova zodiac]

sd-xl is a full model, so can do text2image, image2image and inpainting

[frigid sinew]

whats so special about it

[nova zodiac]

it's trained on a much higher base resolution than 1.5 (1024x1024 rather than the old 512x512)

also behaves much nicer on full sentences in the prompt rather than the word salads that 1.5 prefers

[frigid sinew]

oh cool

I thought it also combined image2image and inpainting into a single step

rather than you having to do it in two

or am I thinking about a different webUI

[nova zodiac]

different..

[frigid sinew]

do u know what i am refering to?

I am interested in doing that

[nova zodiac]

inpainting is just image2image for a small part of the image, as opposed to doing the whole thing

[frigid sinew]

how do u do that?

inpainting in your ai art workflow

[nova zodiac]

you make the image, then send it to the inpainting tab using the button below the image preview, and then you draw a mask over the area you wish to inpaint, choose settings (selecting "only masked") and then hit generate

[frigid sinew]

thanks

[nova zodiac]

make sure you update the prompt describing what you specifically want in the inpainted area, otherwise it will try to redraw a whole image in the area you just wanted a new hand

[ivory kindle]

What are some good tips for getting good realistic images? My faces aren't turning out very good... I'm using a realistic model, but still kinda blurry and fuzzy with a deformed face. Thanks so much

[nova zodiac]

which model, what resolution and sampler, what user interface?

[ivory kindle]

Model : epicrealism_naturalSinRC1VAE.safetensors (something I found on civitai) sampler :DPM++ SDE Karras UI: SD1.5

im using 30 steps and 512 x 512

[nova zodiac]

epic realism is a good start.. try using DPM 2M Karras at 768x768 - are you using comfyUI, automatic1111 or an online generator?

[ivory kindle]

Okay cool thank you, automatic 1111

[nova zodiac]

ok.. in that case, your best bet is to do 768x768 and then use hires fix with upscaler real-esrgan 4x+ upscale by 1.5, 10 steps at 0.4 denoise as well

also reduce the cfg

[ivory kindle]

okay cool cool I was wondering how to upscale

whats the cfg

[nova zodiac]

cfg scale, below height and width

[ivory kindle]

oh okay thank you. What does cfg do?

[nova zodiac]

tells the ai how creative it can be with 1 being just make me whatever and higher numbers being more true to the prompt (but at the risk of the image going weird)

Most models will be happy in the 6-8 range, but the realism ones can go from 2-6 with 5 being a fairly good middle ground

#🏞｜general-with-images message for what you should be able to get

[weary summit]

Guys where is the bots did they removed? İ wasnt online much

[nova zodiac]

#1047610792226340935

[spark tinsel]

using the a1111 webui, how do I create a matrix where the same lora is used at different weights, without the lora at a given weight being inserted into the prompt next to itself at another weight? I don't want "<(MY PROMPT)>, |<LORA_A:0.25>, |<LORA_A:0.6>" to happen;

worst case scenario, I have to make a matrix for each and every weighting I'm looking to use, which would be 0.0, 0.25, 0.6, and 0.9; with five different loras

[nova zodiac]

use the x/y/z plot script at the bottom, and use Prompt S/R (which is search and replace) - format is the text you want first, then a comma, then the text you want to replace it with - you can keep adding commas and more replacements

[spark tinsel]

Thank you! Can you give me a brief example prompt with explanation of how to modify it?

hm, I see the settings that option reveals. huh.

will look for a tutorial

think I know how it works, or at least how to test it!

only supports 3 loras at a time, huh.

[nova zodiac]

you can do the same lora many times though

[pale latch]

no it supports 3 axis. i'd do prompt s/r on lora names on one axis, and strengths on the 2nd axis. like the X s/r field is <lora:AA,<lora:BB, <lora:CC and then on y axis do s/r like :1>,:0.8>,:0.7>,:0.6>

then you get a grid of loras and their strenghts nicely correlated

you can even do 1 grid for each lora. like x s/r is the strength, y set it to cfg, and z s/r the lora name

[spark tinsel]

what I figured, thank you!

[mortal delta]

So the only way to get consistent characters is thru loras? or are there better methods, also i use comfy-ui because its what my hardware will run.

[spark tinsel]

only way to get consistent characters is to use the same seed every time

consistency is, well, you'll get a feel for it; you can use very specific prompts get the same pose, expression, angle, background, shading, and so on

[wet roost]

After no crypts Collab got banned are there any free alternatives?

I'm gonna buy a pro membership but I run a side hussle using stable diffusion and I really need some alternatives

[mortal delta]

so seeds are a better/easier method to get similar looking chacrters? i was wondering about that.

[spark tinsel]

once you get something you like, keep that seed, adjust lora and prompt weights to change the 'size', 'figure', and 'style' of it

afaik

[wet roost]

Guys?

[nova zodiac]

instantID, ipadapter, roop, reactor - all tools in comfy that can do faceswap for consistency

[pale latch]

loras are pretty effective. well trained checkpoints can be too. specialized to the characters you want

[mortal delta]

do they do outfits aswell or only faces?

[nova zodiac]

both

[mortal delta]

i just dont know if i have the hardware to create a lora and what if i want two different characters in 1 scene? or something.

[pale latch]

face swaps are great for swapping faces. ipadapters can often do style transfer, but wont be consistent with clothing. sometimes pants somets shorts

[nova zodiac]

if you are running a side hustle, probably should pony up some $$ 😛 cheap (and some with free tier slow gpus) gpu available at kaggle, paperspace, runpod, vast.ai, aws sagemaker

[nova zodiac]

regional prompting is what you need for that

[pale latch]

loras are pretty lightweight to manage and create. i believe in you!! you can train 2 characters to one lora if you build the data set right. or you can train seperate loras and do regional prompting

[mortal delta]

thank you guys/all for the help and beliveing me.

[pale latch]

we're all some random guy

[mortal delta]

i aksed this questions because ive been interested in creative media such as games, and comics and what not.

[pale latch]

its a fun new techh

in 2 years, it wont look anything like it does today

[mortal delta]

Probably so.

[nova zodiac]

it'll be similar, but the quality will be mindblowing in 2 years time

[pale latch]

i'm trying ot figure out what do to for loras on cascade. the popular projects have half built branches. there are lora scripts in the cascade repo. i just.. i'm not that good

[mortal delta]

may i ask what do you all use ai such a stable diffusion for anyways?

[wet roost]

I got the cash I just don't have a credit card or PayPal to pay for a service

My sister's getting her card made but until then

I need to use something

[pale latch]

i float around different uis. right now been using forge. got it on the 1.8rc branch now

[nova zodiac]

sd-webui-forge

[mortal delta]

intesrting ive tried many ui's and i find comfy works the best on my hardware.

[nova zodiac]

civitai has a generator thats totally free, but can only generate in limited resolutions and no upscaling/inpainting

tinybots.net/artbot is free as well, but as you would be an anonymous user it'll be slow (unless you join the horde at aihorde.net)

happyaccidents.ai has a free tier

[pale latch]

forge uses a few memory patches from comfy code and now all the major uis are using the same pytorch library. really depends on your hardware and how it gets set up too. i have a 4080 so i'm sailing

[nova zodiac]

comfy and forge are the fastest ones by far

[mortal delta]

i wish i had ea 4080

[pale latch]

stable swarmui puts a great interface on top of comfyui node graph. it's nice.

[mortal delta]

i bet so but for me and my hardware comfy works out.

[mortal delta]

i should look into that then.

[pale latch]

https://github.com/Stability-AI/StableSwarmUI i mean to play with it more often. its very customizable like comfy

[mortal delta]

cool and you said this is all comfy based...

[pale latch]

by default. can use others as a back end

[mortal delta]

that is interesting

also sense we were talking about ai sidehustles, i wish i could start one but idk where to start or even begin i manly just use ai for fun and random stuff/testing.

[nova zodiac]

You can make ai art for stock images, do commissions, use it to make dropshipped items; or train AI models for others

[mortal delta]

ooh stock images and commsions seem like a good start. are there any good videos/resources on how to start with them.

[nova zodiac]

search youtube - every man and his dog has a video on how to make an AI influencer! but there will be stuff there for stock images as well.

[mortal delta]

cool thanks guys

i just want to do something different from other people, you know?

[supple oxide]

hey guys, hope you all are doing good, I am new to stable diffusion, although my purpose super-clear in regards to using text to image generation in life and i.e. UI elements, have installed and tried the vectorstudio extension of automatic111 for logo, illustrations, art etc
I will break down my requirements in bullet points here. So I am looking for any suggestions related to

1. better arrangement, if any, for the kind of outputs I am looking for
2. Is there a place to look for finetuned models and trying, comparing their outcomes? I mean like a community where people post their work, I know looking for similar tags in hugging face search results could be a way
3. Any documented guideline for positive/negative prompts for best results

Note: I am looking to use TensorRT versions of any apt model, I am assuming that any model I pick will be a derivative of stable diffusion XL. You can comment on this choice too

[nova zodiac]

tensorRT is best self compiled and doesn't give massive speed boosts from what i've seen, and for everything else, look on Civitai.com

[supple oxide]

yeah joined their server too

But I think if one makes sure what model one wants to use, it's still better to have boosted generation

[nova zodiac]

it is boosted to a fixed resolution though innit

[supple oxide]

I think there's a way to set dynamic resolution while generating the tensorRT version

[nova zodiac]

right..

but yeah.. civitai is absolutely the site for looking through models, they have a generator for trying most of the models (still in progress/setup mode), and most of the images there have prompts on them so you can see what works for each model and what doesn't

and you can filter the model by type so you only get sdxl models

eh.. haters gonna hate...

there is some validity to the "artists weren't asked before their data was included in LAION" argument, but the way AI works is no different to you looking at all of the images by the masters and then figuring out how to recreate it yourself; AI is just really good at that

power usage on GPU, yeah thats a thing, but it's tiny compared to a lot of things, and as for crypto, proof of work coins are really bad, but proof of stake coins are fine for the environment

haters gonna hate...

[supple oxide]

will look through

[pale latch]

there's a few times things are over fit in sd 1.5 base model. and then other times where people were purposely training models on specific artists to piss them off (see a civitai contest)

[nova zodiac]

probably Stable Cascade

[forest trout]

Depends on what you want. There's general models available but in reality, every model has a default style that sort of can be seen throughout the model.

[void shoal]

What the best site to host Stable Diffusion / pinokio.

not only in terms of price but also bc of use as i dont feel like / cant go through setting up EC2 again

[summer minnow]

I'm sorry but why is anything 5 a safetensor?

Are Lora files also safetensors? I don't get it

[nova zodiac]

Safetensors just refers to the fact it is a set of static numbers that cant be executed

[spark tinsel]

find myself hammering the interrupt button and it does nothing lately.

I don't want to close the app, so just wait for a1111 to finish processing bad images.

[tender cove]

Prefix with eicar string and it'll be executable

Also, will get deleted by antivirus

[vapid tangle]

how do i get a person to look down with their eyes

i cant get it to work

[nova zodiac]

Inpainting and a whole lot of luck

[rich kestrel]

inpaint sketch works amazing for that

but eye positions in general are hard for AI. A lora also helps

[dim thorn]

hi

[shell tendon]

Think about what it costs in terms of energy to grow and transport and refrigerate the food required to keep you fueled up while tediously grinding enough drawing every last detail by hand

Bet it's a lot more than a couple minutes of juice to power a GPU even a big one

Yeah

Ppl make some desperate arguments

Oh yeah ik 🙂

I was backing ya up

[civic jolt]

which could be the best LR scheduler for a face LORA?

[rich kestrel]

just use prodigy

[civic jolt]

and the unet and text encoder wich values u recommend?

[dusty onyx]

hello everyone

[crimson acorn]

Where do I report a user in this server? Like do I DM a community mod or what?

[karmic brook]

#1010934719455707218

[crimson acorn]

I can't even attach an image when I used that? Should I just link to a catbox?

[karmic brook]

You can once you send the initial msg inside

[crimson acorn]

Ah okay thanks

Btw Gigabot seems broken, after ticket closed, it gave a transcript, but it's empty. Not that I personally care, but figured I'd let you know incase you don't already know

[obsidian current]

I can't draw this instructions now

[manic narwhal]

Just saw this https://www.reddit.com/r/StableDiffusion/comments/1attell/cascadesdxl/

I don't use Comfy / Automatic only diffusers, but my guess is that the latent output from StableCascadePriorPipeline is fed to the SDXL refiner?

[nova zodiac]

not quite - it is sent to a second model in the same way the refiner is, but its based on the wuerstchen architecture

[nocturne pewter]

Trying to oupaint a room, what would be the best method to it? I have base image of 768x1024 which is then expanded to 1700x1024 in GIMP, basically just added empty canvas to both sides. In the base image there's a person standing and I want to fill the rest of the room. I've earlier just masked the empty areas, tweaked denoising and basically hoped for the best. This process is really slow as it's mostly purely luck to get something that would fit the base image.

Masked content fill / orginal, Inpaint area: whole pic / only masked, different denoise values. If denoise is high enough to get more of the room generated, it usually does not fit the base image at all. Or if the denoise is low enough, I get a door or wall blocking the view, or just some color or white or what ever. at around 0.71 seems to be the sweet spot, still far from good haha.

Also poorman's outpaint and outpaintmk2 give bad results.

Could I get some help with ControlNet?

[visual sparrow]

Hello everybody!

Please, I need someone to help me install Stable Diffusion on my PC. I've been trying for weeks and for some reason, I always end up getting an error. I have an AMD processor and graphics card (I think this is the problem). I've watched tutorials and tried numerous possible solutions, but I can't find the solution. I'm sure it's something simple that I'm overlooking.

If someone could help me privately, I would be very grateful.

[nova zodiac]

Installing on windows? which amd graphics card?

if ubuntu linux - scroll up to about 10 hours ago, I went through it with someone this afternoon

[visual sparrow]

Windows 11. 6700XT

[nova zodiac]

try #🤝｜tech-support message

[sterile torrent]

i have a question to ppl who have 6-8gb cram, how works you guys SD forge or normal SD from a1111 u can generate sd 1.5 and SDXL image bigger then 1024x1024?

[nova zodiac]

tiled VAE - automatically done in forge, or needs the multidiffusion extension in a1111 - splits the numbers up as they go into the last part of processing to not swallow up all the vram at once

[sterile torrent]

ok but its possible? and how log time take generate one image 1024x1024 with 6-8gb vram

[nova zodiac]

Depends on card and model - a 1060 6gb will generate much slower than a 3070 8gb

Definitely possible though

[sterile torrent]

ok thanks

[chilly tangle]

https://github.com/kijai/ComfyUI-Diffusers-X-Adapter

Anyone seen this?

[untold flax]

Hey all, does the speed and amount of compute power determine what kinda image will be spit out by stable diffusion ie. 1070 vs 4070

you could use the same prompt but it will give totally different output?

[spark garnet]

fix bot please sir

[warm junco]

No, you can generate the same image on a 1070 and a 4070

[still glacier]

It depends of the settings used. There are some extensions and settings (mostly --xformers) that makes SD undeterministic and/or change the result depending of hardware. But no the quality of the output does not depend of the hardware used.

[untold flax]

thank you

[bleak matrix]

Good morning, everyone! How are we all today?

[gusty oriole]

I am fine on this President's Day

[quaint echo]

Hello everyone!

Im quite new to generative AIs. I was wondering if the base model, like v1.5 is necessary to have in your model folder when you already have other models trained on this version, like dreamshaper or epicGasmPhoto or so? Or for example, to use JuggernautXL, do i have to first download SDXL also and have it in the models folder? Or these base models are mainly relevant for people to build on these models, and afterwards they are independent

I currently have e.g. SD 1.5, epicrealism, dreamshaper in my models folder and in the UI epicrealism is selected as checkpoint since I play with it. Do I have to keep SD 1.5 itself or can those checkpoints can work without it?

[still glacier]

no, each community trained model is independant and can work by itself.

[bleak matrix]

Good to hear!

[woven lagoon]

guys im getting this error

RuntimeError: Could not allocate tensor with 2047360052 bytes. There is not enough GPU video memory available!

I am on a 6700xt with 12 gigabytes of vram

on directML windows

using no-half

and low-vram

it generates the whole image fine but when it gets to upscaling

it insta-crashes

[young ivy]

damn at this point, being able to get a 3090 24gb or something like 4070 ti 16gb is the only thing that's gonna save my interest in doing this. Maybe evn more vram ;/ Hopefully AMDs (or guys from AMD) making ZLUDA will eventually result in that being able to work for communicating quickly between cuda and non cuda

AMD are just worth so much more at this point, esp if they release another batch soon

[woven lagoon]

so, got any solutions?

[young ivy]

nay

[still glacier]

you can already use zluda with SD

[young ivy]

I've been gone for too long to remember how to fix things

[still glacier]

cf pinned comment in #🤝｜tech-support

[young ivy]

or how toget SD to work with an AMD

and how toget an AMD

currently have a 3070 ti

sitting at 8gb vram, which is a bit of a bummer

[woven lagoon]

does it work well?

[still glacier]

Looks like it, there is some problem with some extensions and other stuff of course. I don t know much about it as I m an nvidia user.

[woven lagoon]

i was shocked to see the whole thing written in rust

[still glacier]

Careful.... #✍🏼｜rules-and-tos message rule 4
(missclick)

@indigo wasp

[indigo wasp]

sorry 😦

can i do that or is the bot limited

[warm junco]

It works very well for 1.5 models, loras but sdxl and some extensions are bugged

[potent snow]

Image d’un quartier animé de Paris le soir

[still glacier]

Dont try to do that kind of stuff with the bot in here (whenever and if it comes back online), keep it to local instances. Otherwise you ll get bonked

[still glacier]

#1047610792226340935

[indigo wasp]

i do it in installed sd

but it just doesnt work

theres still the bra

[still glacier]

It depends on your model, what you prompt for, etc. By default there s no censorship on local instances.

[indigo wasp]

which model should i use

[quaint echo]

thanks 🙏 then I can simply delete SD 1.5 then

[warm junco]

nope, currently the bots are under maintenance. #1047610792226340935

like i said the bots in the bot channels where you can generate images with, are offline right now

[stuck gazelle]

Is there a tutorial to generate images with 1 object/body part always in the same spot/location?

[frigid sinew]

@tidal adder hello sir

[tidal adder]

hello?

[frigid sinew]

@tidal adder can u help me out

[still glacier]

probably not as there are many ways to do so.

• using some mask
• using controlnet
• partitioning the output into multile zone with stuff such as "Regional Prompt Control"
• using a reference img and altering it with img2img

@stuck gazelle

[frigid sinew]

does this have a local api? https://github.com/Mangio621/Mangio-RVC-Fork

I didnt know you made that thats kinda crazy

yup

agreed

[stuck gazelle]

Thank you, which method would you deem best? I tried mask but it's inconsistent unless I missed something

[frigid sinew]

who can help me out what is the front end that is used for SD

If it is just a front end that runs on a browser and it doesnt have a api. can I just reverse engineer the requests

[still glacier]

it depends of what exactly you re trying to do, try asking and be more specific in #📝｜prompting-help (got to run for now)

[stuck gazelle]

Will do, appreciate it!

[frigid sinew]

@tidal adder I guess u dont care :/

[tidal adder]

i don't do stable diffusion stuff

i specialize in language models

oh wait

you had two questions

[tidal adder]

no

it's gradioware unfortunately

[frigid sinew]

@tidal adder Is there any way to programmatically interact with it?

[pale latch]

anyone able to use this for controlnet tile on sdxl yet? https://github.com/showlab/X-Adapter?tab=readme-ov-file#thank-kijai-for-creating-a-cumfyui-warpper-node-here

[old hatch]

why does the ai not make a ai image by my reference image?

[humble trench]

Hi when are bots coming back? Bot 9 Is my best friend :/

[trail lion]

last I saw it was still no eta

[burnt jasper]

hi there, do all checkpoints use SD1.5 etc as a base when creating their own?
or does anyone build their own checkpoints from scratch (w/o the base)?
i want to study more how the whole process works overall

[trail lion]

almost nobody is creating checkpoints from scratch, with consumer hardware...what most of us do is finetune a base checkpoint using dreambooth or similar

you would need a farm of computers, and months, and $$$$

[burnt jasper]

oh? so the base models are the most advanced i guess?

hasn't any one made alternatives to the base model for specialised purposes?

[trail lion]

sure, you can finetune them, merge them together, etc

add new concepts using new images, all that is possible

[burnt jasper]

the purpose would be to analyze each layer of the process

like can't you make your own base using the same checkpoint creation tools

[mild grove]

What is the process for training a model in Comfy? Suppose I already have the base model, how do I reward it if it's right, and correct it if it's wrong?

[trail lion]

nothing is stopping you from creating your own base checkpoint, but reaserch it for a while , and you'll realize the massive amout of data you need to feed it, it's not the process that's hard necessarily, but the data and resources needed

consider there are big companies doing this now, stability, google, meta, with teams of people and billions of dollars at their disposal

[wicked gust]

Hi — we are looking for AI Engineers entry level that can help us finetune Stable Diffusion model. If you are interested please DM me with any related work you have done professionally or personally if oyu have no work experience.

[burnt jasper]

i don't need to make a replacement per se--
i was wondering though if it could just substitute it with a small model to see what i can output w. my own training
is it similar to in kohya where it takes a ton of images and captions?

[trail lion]

start researching LAION

[dry tendon]

Hi guys, anybody know one Upscaler model with have similar results to https://magnific.ai/ ?

[trail lion]

if you want those results, use that

can you upscale with other methods, yup, lots of ways to upscale, and SD is really good at it actually

SD is a little qwerky though in some ways, it has resolutions in prefers, there are limitations based on your resources, etc

[burnt jasper]

what tools do people typically use for creating their checkpoints--and how do they differ from lora's
is that what the finetuning tab in kohya is for?

[nova zodiac]

lora edits small parts of the model, finetuning edits the whole model; similar results - what tools: kohya or onetrainer

[stuck gazelle]

Just updated to Stable Diffusion Forge. Is this the current best option? What do ya'll use?

[nova zodiac]

It's a good option, best is personal preference but generally forge/a1111/comfyUI are "best"; I'm running forge currently

[burnt jasper]

so are you saying the lora's outputted have a small part of the base model injected into them relevant to your training data?

so it's like a patch for a portion of the base i guess?

[nova zodiac]

nope.. the output is a set of numbers to swap into parts of a base model to make the model generate in a way that is relevant to the training data

patch is a good word for it yeah

[burnt jasper]

i see

so all checkpoints are usually just finetuning the base right?

do you know what tool typically artists used for 1.5 checkpoints like on civit

is kohya pretty useful for that or mostly for lora's

[wintry stirrup]

anyone else having issues with dreamstudio?

[nova zodiac]

a checkpoint is a full model

[nova zodiac]

kohya_ss (bmaltais GUI) or OneTrainer are common tools, or if making loras you can do that on the Civitai website as well

[burnt jasper]

yea but most of them use the base model still

[nova zodiac]

kohya is for any training

[nova zodiac]

it is always best to train off the base model, and will have the architecture of the base model, it just changes the numbers as it does the maths to make the new model resemble the training data

[lucid bane]

is there a real performance difference between storing the webui and content on an HDD instead of an SSD?

[burnt jasper]

i don't think so--you are not outputting very large files
i think that your real gains come from ram, gpu and vram

one other thing, anyone know about the checkpoint merger in A1111
is weighted sum option still useful if using a tertiary model?
the example i've seen suggests that a 3rd model is mostly for "add difference"

[nova zodiac]

only time you will notice a difference is when swapping models

[nova zodiac]

yep - weighted sum not useful for 3rd model in checkpoint merger - there is some use in supermerger extension - I would suggest grabbing that extension and reading up on the settings available in that one as that'll get you 99% of the way there

[burnt jasper]

okay--i am guessing without the extension the 3rd model isn't used right?

also one other thing, how do i know what config option to select
what does the config affect

[nova zodiac]

that's to do with yaml files, if merging different types (i.e. inpainting models), I've never touched it

[burnt jasper]

do you by any chance know what the json files outputted in kohya are for?
should they be added to the webui with my loras?

[trail lion]

you can save information about the loras, prompt shortcuts, image previews, etc

[burnt jasper]

is that just for kohya i guess?

well doesn't seem to be needed for A1111 and i wouldn't want to distribute with lora because it has my system file paths etc

[nova zodiac]

not needed, can be removed unless you want to figure out what you did last time to recreate a training

[burnt jasper]

i figured it out it's no different than the ones manually saved

[neat current]

I'm new to stable diffusion, and I got SDXL turbo working for txt2img and img2img. I was wondering how it's possible that the same set of weights can work for two different tasks, cause the if the model works for img2img, then it should be trained on text and image conditioning. Then shouldn't it break down/not function properly in the txt2img setting, since the image condition is no longer being provided?

Is the image condition being set to some "null value" (like in classifier free guidance) when I use the model for txt2img?

[lost quarry]

where doo i generate images with tabl diffution?

stable*

[still hare]

If you use Google Colab paid you get 100 compute units for $10. No where can I find how this actually relates to something real. If I train a lora (koyha_ss) with 15 minutes and it takes 20 minutes, how many "compute units" is that?

[fervent thunder]

is there already some prompts sample

to have direct prompt stock

?

[still hare]

If you are using a model you downloaded from civitai, then on the page you downloaded if from, if you scroll down will be lots of example images with the positive and negative verbiage.

Even if you aren't using civitai, might be a good place to get some starting verbiage.

[fervent thunder]

Ok thx bro

You're the best

Have a good night

[analog juniper]

Drawing a traditional Chinese painting against the backdrop of Lingang New City in Pudong New Area, Shanghai. The work can showcase the following aspects of content: Showcasing the governance and protection of the ecological environment of rivers, lakes, and seas; Showcasing a waterfront open space with clear water, green banks, and pleasant ecology, an ecologically clean small watershed, a waterfront landscape at the entrance of a home, and a beautiful river and lake landscape; Display scenes of river leaders at all levels, volunteer teams, and civilian river leaders patrolling and protecting the river; Display themes related to water knowledge, water conservation, water conservation, and hydrophilicity

[lucid badge]

Hi guys, I'm new in stable diffusion. I'm thinking about maybe using the stable diffusion model for a thesis with imaging applications. Do you think it is possible to use it, I would like to ask you what recommendations you have? That is, when a research is proposed, improvements are usually suggested in some aspect of the model and comparisons are made. In this case, what could it be with stable diffusion?

[grave crown]

cat

[nova zodiac]

#1047610792226340935

[nova zodiac]

What would your thesis be? What hypothesis are you testing?

[manic latch]

does anyone know how to setup a good face swap? willling to pay $

[copper spoke]

can i ( Hires. fix ) in img2img ?

[fervent condor]

has stable cascade been implemented in any front end like comfy ui yet ?

[mild gulch]

Bots STILL not fixed yet?

[nova zodiac]

they aren't broken, just disabled

[nova zodiac]

no.. hires fix is essentially img2img under the hood

[nova zodiac]

look up instantID or ipadapter

[manic latch]

does anyone know how to setup a good face swap? willling to pay $

[nova zodiac]

there are proper comfy nodes for it, and a temporary a1111 extension

[opal hedge]

Hello, does anyone know a good resource for comfyui that checks for doubled or duplicated pieces of a prompt?

For example, it might check if I have 'blurry' written twice

[nova zodiac]

Not that Im aware of but it sounds like a good idea

[acoustic solstice]

hi all, is there a quick intro guide to getting stable cascacde running locally on my GPU? i saw the github but didn't really see a step by step guide for someone new to this. i am familiar with python though.

[rich kestrel]

what makes stable cascade different than 1.5/xl. excuse my laziness but I really dont want to waste time watching videos on something that can be summarized in one sentence

[hallow sable]

what ive seen is that cascade does a much better job with words

[nova zodiac]

Under the hood it is doing maths on blocks of 24x24 pixels instead of 64x64

[mystic imp]

Yes. Since the weekend I alway get the error „something went wrong on our end“ I already opened a ticket but so far no reply

[cosmic cape]

Hi there,
Can somebody what' that model used in AI to make a face psoe at different angle in real time?

[nova zodiac]

That was the DragGAN paper if I remember rightly - hunt that down and you may have your answer

[cosmic cape]

yesh... really really htnaks

*thanks

[wintry stirrup]

Same here. So frustrating, especially with no communication around what’s going on. I have a good amount of credits in there.

[nova zodiac]

@karmic brook @finite cloak - you know anything about the dreamstudio issues/who to chase up in case you dont??

[fervent thunder]

Dont use google colab, they are openly against uncensored models. Id use vast ai or runpod

Google also intentionally has sabotaged SD before, and i couldnt get the webui running due to resource limitations even after upgrade

[nova zodiac]

This - and youll get a better gpu on runpod/vast.ai

[nova zodiac]

They do remove accounts circumventing the ban on a1111 and comfy

[urban elm]

How do/are you making money with AI art?

[nova zodiac]

There are generally 3 groups - those running a service (like mage.space owners), those doing commissions (either model training or making art), or those making products (dropshipping style putting their ai art on things)

Is there $$ to be made? Maybe but I wouldnt quit your day job to persue it given barriers to entry are low and its not a new space anymore

Two others I hadnt thought of were Youtuber (eg Matt Wolfe, Sebastian Kampf, Olivio Sarikas), and software dev/in house expert in the corporate space

Hope that helps @urban elm

[placid void]

Hi guys, im not that much of an ai user, I use it from time to time. Been creating a startup recently, logged in to generate some logos for my startup but that maintenance on bots ruined my plan. And it seems like it wont be available for a while too.

Can you guys suggest me where to create a decent logo for my start up ?

[fervent condor]

Is there official Stable Cascade workflow ?

[nova zodiac]

I think there was one floating around on reddit linking to one

[nova zodiac]

Try the create button on Civitai.com

[radiant tusk]

My images generating in what used to be 10 seconds is now taking like 3 minutes to generate, restarted everything made no changes to models or settings

How do i fix this

[radiant tusk]

Okay just narrowed the problem down

for some reason automatic1111 is only using 30% of GPU where before it used to be like 90%+

Cuda usage is also 30%

[nova zodiac]

what resolution and what graphics card?

I'm assuming nvidia, and a reasonably large resolution for the card

[radiant tusk]

Earlier this morning things generated 5-10x faster and I would lag just watching youtube videos while it was generating

[radiant tusk]

512x512 rtx 4090

[nova zodiac]

I mean youtube at high resolution does suck some gfx capacity, are you using batch size or batch count ?

[radiant tusk]

Batch count

[nova zodiac]

you got a controlnet stuck? have you done the cudnn file fix?

[radiant tusk]

I don't have control net installed

[nova zodiac]

have you turned it off and on again?

[radiant tusk]

I'll try this

[radiant tusk]

Generation speed without loras are now a lot higher but with loras it goes from 20it/s down to 1.5it/s on the same settings

the moment I remove the lora goes 10x

same speed with 1 lora or 10

VRAM usage is 5.8gb without any loras although there are random spikes to 16gb

[nova zodiac]

that'll be while the vae is loading.. but shouldn't be that high

weird with the loras though

[radiant tusk]

I think that issue started today

The instant I even use a single lora goes down to 1it/s