💬｜general-chat

[floral umbra]

Took me a while to memorize what's needed where just for basic image gens :P

Wait, there's malware in jsons now?

[inner grove]

What would be the difference between Stable and Comfy practically, you still have to connect these nodes, don't you ?

[floral umbra]

And if that's the case, just make your own json. Type in whatever node people use, and study the wiring between them, and save it.

[floral umbra]

The big difference as far as i can tell between them, is that auto redoes the entire process for each gen, while comfy only redoes the steps that has a change in variables.

Plus auto is broken if you model hop. Leaves broken gens. While forge, although apparently well outdated fixes that.

I only use auto/forge for controlnet poses, as the poser in comfy is highly annoyingly limited

[honest mica]

No, you don't. Swarms User interface is not node based. You can change the flow in the backend, but you don't need to.

[inner grove]

Do you mean Openpose or ControlNet in general ? You are scaring me, I use ControlNet Depth Anything in 90% of my projects, having nodes for Depth Anything v2 (no preprocessor yet available for Forge/A1111) is one of my main motivations to do the Comfy jump

[fervent thunder]

omg why doooes fooooocus dooownload juggggernaut

[trim magnet]

ikr,should download pony instead 🐴

[fervent thunder]

haha!

fo real yo!

I have never used Pony... I mean, what is this Pony you speak of?

[trim magnet]

small horse

[inner grove]

Forge did that too a few months ago, and it was an outdated version, just put any model in the models folder and no other model will be downloaded I'd say

[fervent thunder]

I never tried forge much, or swarm, I kinda like simple stuff like comfy, 111111111, and foooocus looks nice but I am almost at 2 minutes for changing presets now

some of us have to be casual observers in this new AI World that is taking over

the new AI world order

[inner grove]

Simple stuff like Comfy, wow, that's a bald affirmation 🥳

[fervent thunder]

People too often get effective confused with complex in my experience.

[floral umbra]

Openpose.

For instance, i can't drag and select the whole stick figure in comfy, but i can with auto's

[inner grove]

Ah yeah, I don't use OpenPose either, doesn't work well with SDXL

[floral umbra]

I got it to work with SDXL, you just need SDXL trained models. Doesn't work well with animations iirc though

[inner grove]

Depth and the new xinsir Canny work better with SDXL than the XL-openpose models imo, but perhaps my settings aren't right

[dull quail]

?

where everyone go

hello ?

[storm shard]

I get the impression they're slowly killing this discord

[vast ingot]

the channel deletions are giving that impression to me too, not sure why. I don't think it's intentional, I know most of those that mod or admin around here and I wouldn't think it was that, but I'm failing to get the good logic behind what's happening

[storm shard]

It looks to me they want us to forget the other models and just focus on SD3. I sure hope I'm wrong, because that would mean a somber future for SDXL too.

[vast ingot]

if it is was the case, they would not use the "lack of participation in those channels" as reason for removal. #✨｜sdxl is still quite alive and has not been discussed at all yet. The winner's galery removal does not make any sense to me though, even without events currently. And even without activity in there, it would be the prime candidate for archiving, especially now that the archives have been purged and are empty.
I'm surprised it has not been announced at all either

[loud solar]

The last one closes the door...

[vast ingot]

well, to be fair, archive deletion wasn't announced either before we asked

[storm shard]

tbf it feels like stability has lost interest in its community @vast ingot

[vast ingot]

I do want to stay positive. I've given too much energy to this discord and its community to believe that.
But I do feel like that when I let myself go on the negative slope

[potent merlin]

Hello @hidden dagger @swift dagger @gritty scarab @viscid burrow and other Team Developers, Nice to meet you.
Can you help me create a storyboard for a movie?
I am currently using OpenAI's DALL-E model to generate them, but they are completely inconsistent.
Is it possible to generate consistent storyboards programmatically using Stable Diffusion API?

[loud solar]

What is the Bot doing here? https://discord.com/channels/1002292111942635562/1004159122335354970

[honest mica]

Thinking

[loud solar]

🤣

[vast ingot]

it's not a Stability/SD bot

it's just an app someone seem to have called

(new user that joined today)

[loud solar]

Strange they can do that 🙂

[shell tendon]

that's definitely the impression i get too

[vast ingot]

it's something new on discord that got added not long ago. if the admin don't remove that right, then people can just do that, yeah. Seem like something that didn't get moderated yet

[fervent thunder]

oh goodness

[shell tendon]

i'm all on board with trying to fix SD3 and i haven't given up yet, but without any major progress with training 2B, that's a dead end

it may still be doable yet, but right now, no one has been able to get particularly encouraging results

myself included

[vast ingot]

did they change anything on the license yet ? because SD3 feels great and I've defended it since it came out, but that point is hard to defend for now
because a ban on civitAI will be hard to circumvent. I'm not finetuning until this gets sorted at least

[shell tendon]

which is why it's kinda crazy - why not diversify a bit? if SAI wants something new that is working right now while we figure out training for 2B (assuming that does actually work out), that something new is cascade

[sudden ruin]

I reported it, maybe it will be fixed

[shell tendon]

especilally now that i've implemented the exact noise type that SAI used to train stage B of cascade... the results are really really good now

[loud solar]

It didn't harm yet. I was just wondering ... 🙂

[shell tendon]

not that i'm aware of... i'd imagine that's bogged down in lawyer hell

[fervent thunder]

#🌶｜off-topic

[vocal leaf]

WAIT WHAT you use SD?

ur from PCMR

[floral umbra]

I use pretty much all of the fun A.I's out there, yes

And yep, i is from pcmr (i also made this emoji )

[fervent thunder]

thats quackers!

[vocal leaf]

give me more ones to use

im getting bored with comfyui

(most because of having an AMD card that cant run specific nodes)

ehm FaceDetailer

[potent merlin]

is there any member of Team in this channel?

[forest trout]

The stability.ai developers here generally don't troubleshoot individual users questions. I would try posting your questions in #🤝｜tech-support

[floral umbra]

Sadly, the most recent fun i had was with a model that iirc can't run on non cuda, which is a image to 3d model gen

[dawn mulch]

Anyone have prompts that can generate good city images realistic, 3d, or artistic with the view similar to those in older resident evil games.

Iv tried using the default sdxl model but it has really plain, often brownish buildings and images.

[bleak matrix]

Good afternoon, everyone! How are you all doing?

[brave vigil]

Question about lora training, resolution, and downscaling

Firstly - if I'm training a lora for an SD1.5 model, would training at higher resolution with 1024x1024 images matter?

Or does the model's base 512x512 render the higher resolution of training images moot?

[pale latch]

you'll get there and it'll be a timeshare pitch

[nova lily]

whats the deal with those "source_whatever" prompts. where did they come from? do they actually work

[vocal leaf]

Im not well versed yet on AI image gen, but I think its because the nodes I got or the nodes we got were made by various people at various skill levels. which that alone reveals the amount of compatibility issues with various systems 😦

[loud solar]

There was a ComfyUI Summit where they talked about standards ...

[pale latch]

its a genius idea for quality sorting that a model refiner had but it just made it so every prompt for a ponyxl model looks identical

[floral umbra]

For regular image gens, on amd, you can just use torch-directML. That's what i did on steam deck

[pulsar venture]

Hi, i'm need an help, can you link me a tutorial for learn to animated with SDXL and automatic1111?

[low moon]

Video is just no there.

Come back in 2 years.

[graceful bramble]

Is there somewhere or someone I can ask some really basic and or stupid questions about getting started with stable diffusion? I have used other website based ai's but doing everything myself is quite different.

[desert dagger]

sure. just poste the questions in here and a lot of people can provide answers

[graceful bramble]

Perfect thanks!

So first of all. I'm starting with pony. What is the purpose of the word "BREAK" in the prompt?

[desert dagger]

can i see the prompt?

[graceful bramble]

score9, score_8_up, score_7_up, BREAK zPDXL and then the rest of the prompt

[desert dagger]

https://www.reddit.com/r/StableDiffusion/comments/13a9avh/quick_question_does_putting_break_in_a_prompt/ this might answer that question.

[graceful bramble]

Interesting.

[desert dagger]

personally, i don't use prompts long enough to need to worry about whether i've hit the token limit or not

[graceful bramble]

I've also read that pony is based on the danbooru tag system or something like that? So when I use those tags things go fine but anything outside of those it starts to struggle. How can I use concepts that aren't in those tags?

For example. Narrow hips. I can't for the life of me find a way to do that.

[desert dagger]

don't use pony. Use a different base model. or perhaps use lora's or checkpoint models created from the pony base model

[desert dagger]

that sort of goes against the pony model's core idea...

[graceful bramble]

Hmm ok i just chose pony because there are so many loras for it for different characters.

[desert dagger]

@frail sonnet might know more about how to use it right

[graceful bramble]

I'll have to look for some other ones then.

[desert dagger]

most of the models and lora's on CivitAI created for SDXL do a very nice job

[graceful bramble]

Ok what about image sharpness cleanness focus whatever the right word for that is. I always end up with things just a little fuzzy and the eyes are a mess.

[desert dagger]

there are a lot of settings that can cause that if not set right. how are you running stable diffusion at home?

[graceful bramble]

Automatic 1111

[shell tendon]

We should probably archive/delete the research channel... Only one post in there in the last three weeks

[frail sonnet]

Perhaps try thin hips. Though if you are using a particular lora, that could mess with things. Also, you can play with the Score_9 up etc. Lots of people leave out score 6 on down.

Also, there is a learning curve with Pony, since it prompts differently than SD 1.5. To make life easier with Pony, you can always start with Autism mix, or DucHaiten's pony models, WAY easier!

[graceful bramble]

I was actually using the autism mix. And i did try thin hips, also slim slender narrow small skinny and petite.

[frail sonnet]

Fortunately those Pony loras also work with Autism mix etc.

[graceful bramble]

Yup! I quite like the mcfg too I think just gotta learn how to make it do what I want.

[frail sonnet]

Are you comfortable sharing your entire prompt?

[graceful bramble]

I don't have a set one yet I've been messing around alot. I'll make on up though.

score9, score_8_up, score_7_up, rating_safe, BREAK zPDXLxxx, 1girl, full body, narrow hips, detailed in focus face, realistic eyes, toned, isolated on solid white background, bright colors, high contrast, dark background, vivid lighting, lora:intricate_details:1

[frail sonnet]

full body is probably getting in your way

[graceful bramble]

Is that not how I should avoid a close up shot?

[frail sonnet]

wearing leather boots

bare feet

high heels

or perhaps entire boday, I"m trying some, it's just slow 😄

I forgot to put rating safe, but you get the idea:
"photo of an anime girl, thin, slender, tall, wearing a corset, tall leather boots".... I can't add image, just imagine her lolol

[desert dagger]

i'll prompt for something on the head and something on the feet. the AI has to draw the entire person then. so like: long hair, red slippers,

[graceful bramble]

Yeah and I can get the whole photo but the hips are still pretty big most of the time. They are average now and then.

[desert dagger]

or: a simple country girl. She has shoulder length blonde hair, bare feet, and is holding flowers

[frail sonnet]

#🏞｜general-with-images message

or if you really want thin hips, you can prompt a guy, then add "pretty female face, covered breasts" to the prompt. Guys have thinner hips than girls. Sounds counter intutitive, but it works.

[graceful bramble]

Ok i think i figured it out after seeing yours

Some of it was the model. Autism does better than mcfg. And i was running the xx pdxl version instead of the regular which really likes wide hips for some reason?

[frail sonnet]

Try Duc Haiten's Ponys, his models tend towards thin hips

[graceful bramble]

I can do that.

Now how do I get it to produce an image that doesn't look like this? #🏞｜general-with-images message

[eager shard]

i think ur suppose to use the safe tag at the end atleast if ur using pony

[graceful bramble]

safe?

[eager shard]

"score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, just describe what you want, tag1, tag2" from the civitai page

[eager shard]

rating_safe

[graceful bramble]

Ah. Ok. thanks!

[frail sonnet]

Specify the type of clothing/style, as well as the background. The background also influences the main character.

[graceful bramble]

I mean the focus. How do I get it not to be all blurry.

Especially the eyes.

[eager shard]

"Other special data selection tags include, 'source_pony', 'source_furry', 'source_cartoon' and 'source_anime' and ratings of 'rating_safe', 'rating_questionable' and 'rating_explicit'." these are the tags it says on the page, no clue if theres more

[frail sonnet]

Try different config numbers. Are you running it on your own computer, or?

[graceful bramble]

Yes

I've messed around with the steps and the hires fix some. but mostly I've used the recommended settings from pony

[desert dagger]

DuchHaiten has a discord if you're interested

[graceful bramble]

I'm looking at the models now but I might try something other than pony.

[eager shard]

remove the detailed in focus face

replaced it with detailed face and it stopped generating cursed images for me atleast

[graceful bramble]

That helps a little thanks!

I guess I'll figure out more as I go. For the moment though loras.

[eager shard]

its still generating monsters ( for some reason ) but now they're feminine monsters lol

before and after in #🏞｜general-with-images

[graceful bramble]

You have to include the blabla.lora:1 and use the trigger words? Does the order matter?

[eager shard]

order and trigger words matter yes

i only use loras when the model isnt doing what i want, usually just have lowra sometimes

[graceful bramble]

I saw one article say the <lora> part doesn't matter but when I was messing around it did change things. Also they said it doesn't need a comma after it because the whole term is removed instantly as it activates?

[eager shard]

lora:LowRA:1 score_9, score_8_up, score_7_up, score_6_up,

should be something like that

u can just do lora:LowRA:1lora:LowRA:1 and itll be fine

[graceful bramble]

Always at the beginning?

[eager shard]

no clue about the lora but the score 100% should be at the beginning

[graceful bramble]

Yeah I mean the lora

[eager shard]

idk! the trigger words would make sense because the sooner u type something it priortizes it more unless u do stuff like () (())

so you usually type in the most important things of ur prompt first

i do () when it isnt working for whatever reason

[desert dagger]

i just use the keyword, if the lora has one. I use comfy and if i'm using a lora, i'ts connected into my workflow

[graceful bramble]

Yeah I understand prompt weighting and priorities but Loras are new to me.

So you don't type in the <lora> part? and I'm sorry but I don't know what a workflow is.

[desert dagger]

a lora is like an errata sheet. it's a small handful of weights that are revised. so the model it's for is critical. when generation kicks off, the prompt is run with the base model first. and then if you call the lora, the data is revised based on what the lora knows. if you try to use a lora for a model it wasn't created for, you'll get garbage

[frail sonnet]

Loras can be quite awesome, for specific characters or themes.

How many steps are you using btw?
Also, are you using Euler? While DPM is amazing for 1.5, Euler on average works better with Pony

[graceful bramble]

Eula a 25 steps

I understand the basic idea of a lora. But how exactly to use it is a little more difficult.

I've looked at other peoples prompts and everyone seems to use them different ways.

[frail sonnet]

You don't need them necessarily, but they can be handy. Are you using Comfy, or?

[graceful bramble]

Automatic 1111

[frail sonnet]

Btw, I had to many windows/projects going at once, that image wasn't actually pony, but this one is, exact same prompt, and I challenged myself by using the straight up ponly model: #🏞｜general-with-images message

[frail sonnet]

there's an add lora button somewhere?

[graceful bramble]

Yeah and it adds the lora:1 part but some peoples prompts don't thats why i'm confused.

[frail sonnet]

So the catch with Pony is that you sort of need pony specific loras only or the image sucks

[graceful bramble]

and then embeds are just the file name right?

[frail sonnet]

Some good Pony Loras are Vixen's set (from memory, I hope that's the name lol)

[eager shard]

if ur looking on civitai i think thats just a bug, sometimes i see loras in the loras used area but when i look at the text it says completely different loras

[graceful bramble]

Ok well that clears that up. It was really giving me a headache.

[eager shard]

was confused what you were talking about for a bit lol

[graceful bramble]

So if I was going to try something outside of pony what is a good base model? Whats the difference between sd and xl, turbo and hyper and whatnot?

[graceful bramble]

Sorry I'm sure I wasn't explaining it the best. And I have done some basic ai stuff before but it was all on a subscription website that handles all this so this is all Way outside my comfort zone. Heck just getting automatic 1111 downloaded and dealing with some of the errors and junk was pretty much space magic as far as I am concerned.

[eager shard]

yeah automatic1111 was my first too

[graceful bramble]

What do you use now?

[desert dagger]

SD = stable diffusion - every version. you modify the SD with the version so people know which one you're talking about. i.e. SD 1.4, SD 1.5, SD XL, SD 3

[eager shard]

forge because automatic1111 still isnt too good for high resolution sdxl, ill be switching to stable swarm in a couple weeks though because theyre making forge more experimential/ probably more bugs

[graceful bramble]

Whats the benefit of stable swarm?

[eager shard]

no clue! just discovered about it yesterday

its apparently comfyui with bearable ui from what ive read

[desert dagger]

it's an interface that allows you to do a lot of different things as well as running the image generator

[graceful bramble]

So which version should I be looking at 1.5, 2, 3, hyper or turbo. I assume lightning is faster less quality?

[desert dagger]

lightning is optimized to be FAST

[graceful bramble]

Yeah I don't want that. I want higher quality.

[eager shard]

just stick with sdxl then

[desert dagger]

1.5 has a lot of concept issues - but with the right prompts, and they tend to be fairly long and convoluted, you get nice results. SD XL is the next level up, and it runs with a refiner to create better images.

[graceful bramble]

You were right about duchaiten's it has much smaller hips. Just got it downloaded.

[desert dagger]

you want the link to his discord?

[graceful bramble]

No I have it.

[desert dagger]

okay. that's the place to ask questions about his models

[graceful bramble]

So xl then do I want xl hyper, turbo or lcm?

[desert dagger]

lcm isn't a model. it's - something that tells the AI how to do the math

[desert dagger]

you're just getting started. you want SDXL and a couple of models for it perhaps. just to play with and get a feel for how thigns work

[graceful bramble]

Sdxl 1.0?

[frail sonnet]

Also, or which checkpoints, it depends a lot on what your theme is. They all have specific themes.

[desert dagger]

to start with, sure.

you need to just play and explore. see how things work, try different models and get a feel for this

[graceful bramble]

I used to use Albedo alot.

[frail sonnet]

What's that one? I haven't heard of it

[eager shard]

i keep trying to make ur prompt to make humans but for some reason it just keeps making female monsters 😭

[graceful bramble]

Ha don't worry about it. I appreciate the persitence though!

[frail sonnet]

sounds good to me, which model and lora? 😄 lol

what was the prompt again? Going to try some horror ladies!

[eager shard]

the prompt he gave was ( going to scroll up for a bit )

score9, score_8_up, score_7_up, rating_safe, BREAK zPDXLxxx, 1girl, full body, narrow hips, detailed in focus face, realistic eyes, toned, isolated on solid white background, bright colors, high contrast, dark background, vivid lighting, lora:intricate_details:1

[graceful bramble]

https://civitai.com/models/140737/albedobase-xl

[eager shard]

and i changed it to this, which seems to make it work but 90% of the time instead of a human female its a monster female

score_9, score_8_up, score_7_up, 1girl, solo, dynamic pose, hipbone, detailed face, realistic eyes, toned, solid white background, high contrast, dark background, torso, legs, rating_safe, source_anime

[graceful bramble]

This is what I used when I was with a subscription website. They had their own proprietary models too but I liked this one the best.

[eager shard]

pony btw

[frail sonnet]

I remember that one now, it never worked well for me for whatever reason

[graceful bramble]

It took a while to get used to.

[eager shard]

does pony not automatically work w albedo?

[frail sonnet]

If you mean Mage, that was one of my least fave models on there lololol

[eager shard]

oh i thought u meant albedo from overlord lmao

[graceful bramble]

No I used leonardo.ai

[frail sonnet]

This prompt gave me a nude muscular girl, who looked sort of like those SD3 lady on grass warped images lolol

[graceful bramble]

Well thank you all so much for the help! I will try some things out and look at some different models. You have saved me a headache and a half.

[eager shard]

made it do albedo and i think it works WAY better if u specify a character

#🏞｜general-with-images

[graceful bramble]

Thank you!

[desert dagger]

let us know how it goes

[graceful bramble]

I'm sure I'll have some more stupid questions before too long.

[frail sonnet]

There are no questions, trying 10,000 generations can't answer! 😄

[desert dagger]

there are no stupid questions - unless you do know the answer and are trolling

[eager shard]

its ok i only know how to use two extensions lmao

[graceful bramble]

Hey thats two more than me!

[eager shard]

i learned them because i wanted to try out making videos, controlnet & animatediff

[graceful bramble]

I will eventually. But just a clean basic image of various different characters comes first.

[eager shard]

also ur prompt works INSANELY better if u specify a character

[graceful bramble]

Good to know.

[frail sonnet]

This prompt worked fine for me, but I'm not sure if Pony is capeable of doing sfw enough to post on Discord

[eager shard]

i just put nsfw in negative and it works fine for me

or you can just always do otoko no ko

[frail sonnet]

I use SD3 for my SFW 😄

[eager shard]

i havent since i read it got banned for something

[frail sonnet]

loras and checkpoints on civitae specifically, temporarily

They are still working on a legal agreement non-layers can read 😄

[eager shard]

is there any model good for 3D? i was making wholesome max caulfield and chloe price wallpapers and i notice pony messes up alot for 3D

[frail sonnet]

From what I read of it, if you make money from the art you generate, you pay them $20 per month. If you don't make any money, you don't pay them

No idea on the 3D thing

[low moon]

They were stoned and drunk when writing that license.

[obtuse goblet]

Do you guys host/run your own SD instances or are you using some service?

[frail sonnet]

All of the above

[eager shard]

yeah cause they look super off #🏞｜general-with-images

[graceful bramble]

Ok I thought of one last question. Is there any way to save or see from a previous image what the prompt/settings are so I could reproduce something I made a few hours or days ago?

[eager shard]

if you saved it yes, open the image with notepad

it'll take awhile to load but ur prompts will be there

[graceful bramble]

Oh! There it is. Awesome thank you!

[eager shard]

#🏞｜general-with-images itll look like this

[frail sonnet]

Get comfy. Yuo can load yours, or others images, and the entire workflow is embedded right in there.

Well OK or use some exif into thingee online

https://pics.io/photo-metadata-viewer

[eager shard]

does anyone know what steps restart sampler is good at? was looking for the highest quality one (found) but now im looking for something that pumps out decent images fast

[frail sonnet]

I run a randomizer and go to sleep, or out to do errands 😄

[cinder cedar]

Hi everyone, i'm new here

[orchid kiln]

Hi

[neon hawk]

stable diffusion vs decryption ai, who wins?

why are so many bad imagines left in the training datasets?

obfuscation doesnt remove anything

what happens when the model gets decrypted again?

[vivid shale]

anyone here with knowledge of stable diffusion openvino?

i have a couple questions but i think i can get not much help even thru tech support, dude there says they dont know much about openvino or running sd on cpu

do i need to install optimum intel via pip first before i can take advantage of acceleration like https://github.com/openvinotoolkit/stable-diffusion-webui ???

[fervent thunder]

would recommend getting an Nvidia GPU and using CUDA

[vivid shale]

send me money?

[frail sonnet]

try glif or huggingface, there's several SD3 ones 🙂

[fervent thunder]

the problem with using non-CUDA is
even if you get the base image generation working
eventually you will likely want to add lots of other models on top in the workflow
and everything gotta be manually troubleshooted

[icy sonnet]

Is there any comprehensive guide on how to build and run stable diffusion XL with docker? Not the web ui, just docker

[vivid shale]

how can i find no one to help me even in tech support with simple questions about how to properly set up stable diffusion with openvino acceleration? i only want to know if openvino is something i have to install before i can benefit from these builds promising boost

[trim magnet]

because almost no one uses openvino cpu only mode because the speed sucks

most people here use their own gpu or online services like glif / seaart / shakker,etc..

[vivid shale]

is there way i can openvino and directml together i wonder

[trim magnet]

try to ask in #🤝｜tech-support i saw some ppl there talking about openvino

[vivid shale]

cuz i have shit computer (was ok 7 years ago but) but i have intel onboard cpu/gpu and amd radeon graphics, only in total, my ram is 16gigs and my vram says 4 gigs for amd (tho i swear i have more vram somewhere)

that person talking openvino in tech support was probably me, lol, i am dumb and desperate, lol

[trim magnet]

yea old onboard intel gpus suck and on cpu u probably gonna get like 1 image per 30 mins with cpu only mode whether you use directml or openvino

[dawn mulch]

I need help with upscaling and adding details. For some reason all i am getting is upscale with little to no detail added. Its as blurry as it was with the smaller image. I posted the images in the image room.

I am using comfyui and Ultimate SD Upscale and ControlNet Tile

[vivid shale]

i get 20 steps done in 20 minutes but i think i am capable of much better, only idk if im getting acceleration via openvino cuz idk if i even using it (despite build says to use, idk if its installed on my pc and idk how to install it so can be utilized) also, i know there other things like models specifically tailored for use with openvino, etc, etc, only idk where to find all this stuff and download/install, etc, im dumb as fuck when comes to anything but download/install exe files (videogames) idk even how to use python much cept following simplest step by step pip install shit

[vivid shale]

ima retard at this stuff, i need simple explations how to do things

[dawn mulch]

@vivid shaleIts going to take a good amount of time and effort to become familiar with all this. But I dont know about openvino myself so i cant help. But you can check if you have it installed using 'pip list' in cmd, but only if you have pip and python installed. Im no expert either but thats how id check for openvino

[tight mason]

https://youtu.be/Azj9Kkpif0M jailbreak gemma 2

[median shore]

WebUI was banned because too many people used it in colab, now you can still use Stable Diffusion https://colab.research.google.com/github/R3gm/SD_diffusers_interactive/blob/main/Stable_diffusion_interactive_notebook.ipynb

[tawny halo]

hey

guys created a new ai book dm me if interested

[sage sky]

anyone know how to put stable diffusion on my taskbar???????

[hasty hornet]

I don't think that's allowed too, that's still user interface.
Probably just not enough people do that yet for google to bother and how do you even detect and ban something that is another question..

But...I don't think google banned anyone for using webui's , just warnings so far I think

[full lark]

Nice!

[median shore]

that could explain why focus and Invoke-Ai still work in free colab but these are better known so it's contradictory

[full lark]

Someone got hacked lmao

Wait its just a bot, mods ban this fucker

[trim magnet]

its free money sunday 💸 🎉 🥳

[loud solar]

;-D

[mortal sphinx]

Is there a way I can make Stable Diffusion or anyone know a tool that can clean up and make hair HD?

[loud solar]

Washing?

[desert dagger]

you could use blender

[fervent thunder]

lora + ip adapter style transfer + canny control net + depth control net + face swap

this combo works for 99.99% of difficult images

[loud solar]

One point I just can't get. Why is SAI unable to release a model that is at least acceptable? How long have they been working on them? The models come out and nobody wants to use them. Than it takes the community a few weeks? months? to make them usable?

[fervent thunder]

that's actually not a bad thing

what is happening is that

in order for the model to be able to be finetuned for many different uses

the weights have to be "loose"

in other words "undertrained"

otherwise it will be like the Playground 2.5 model

it looks better out of the box but you can't fine tune it

[loud solar]

I'm getting the idea ... but dismorphed bodies and hands ...

[fervent thunder]

do you remember when Juggernaut massively improved the hands
by adding 10k+ hand images to the fine tune?

[loud solar]

And I think SAI has at lest 10 times the computer power to work on a model ... compared to a guy/girl from the community ...

[loud solar]

To many news every day ... I don't remember, soory!

So you want to say if it's overtrained people can't create Aliens or Simpsons (with only 4 fingers) any longer?

[fervent thunder]

if a model gets too overtrained
then you can't easily shift the output away from the inputs that it was overtrained on

but when we fine tune, shifting the output away from the original input is what we are trying to do

in machine learning its very different to train a model that is intended to be fine tuned, compared to a model that is not compared to be fine tuned

if you are training the final model end to end, and you will not touch the weights again
then you make the weights very tight
and go as close to overtraining as possible

whereas if you want the model to work well with fine tuning

you pull back a bit from the limit

so you leave some headroom for the fine tuning

[loud solar]

Thank you for the information! But how is MJ doing that? For sure I think the pictures are a bit boring ... but it's the other extreme to me.

[fervent thunder]

midjourney don't allow fine tunes
so they can push the model as far as they want

[loud solar]

Well they started with kinda new unfinished models sometimes, too. Thanks for the information. I don't know you but they sound ... right for me now.

Sounds like Germany missing a goal here 😄

[hardy sedge]

What is the fastest stable diffusion gui other than comfyui?

[fervent thunder]

not actually sure

I feel like comfy advantages are too big

[loud solar]

I would think Stable Forge ... cause the latest Update for A1111 is pretty old, too ... but I can't say that's 100% correct.

And there's also Stable Swarm ... I just couldn't become a friend of ...

[hardy sedge]

It's a pity that comfyui is not available with a different GUI Like a1111

[loud solar]

Swarm is kinda trial to do that ....

But not working for me ... sorry to the brave programmer!

[fervent thunder]

yeah this is what Swarm is all about

to give a nice front end to comfy

I use Huggingface Transformers a lot too

LOL I mean Diffusers

[loud solar]

Swarm don't really have the TXT2IMG, IMG2IMG, Extras ... you have to configure it by yourself ...

Sounds like Germany made a goal 🙂

[hardy sedge]

What is the fastest video maker? Faster than SVD?

[loud solar]

I have no idea about how much ttheir servers have to work at the moment. But Luma is pretty cool

Local I only know some morphing workflows ...

[fervent thunder]

Luma is amazing

there is opensora

but have to rent A100 every time

so its like 1.5 dollar per hour

[loud solar]

I'd pay Luma to make some of my old Videos now ... just to safe time ...

[echo sentinel]

3 am and I'm generating

[fervent thunder]

its the potential to fine tune opensora
that is exciting

[loud solar]

Addicted you are my friend 😄

[echo sentinel]

Yes, it’s boring, I still can’t sleep

[loud solar]

Enjoy it! You can be creative!

[echo sentinel]

I'm thinking about the prompt, I'm out of ideas

[fervent thunder]

I don't actually write prompts lol
I just give image to vision model

[loud solar]

If you are out of ideas ... do something complete different .... and they will fly to you ....

[fervent thunder]

also just ask claude for random prompts

for sci fi, it makes better ideas than I do

[loud solar]

Hmmm... I think a picture needs a real story or at least idea behind

There's a difference between MJ random crab and real ideas .....

Can someone delete RonalSilve?

[echo sentinel]

It seems to me, or has everyone forgotten about Kandinsky?

[loud solar]

Same in General with Images, please

[loud solar]

Not really working with it .... and didn't heared any news ...

[echo sentinel]

That’s what I’m saying, no one says anything about him, that’s how the new model came out, everyone was talking, and still nothing was heard about him

[loud solar]

I'm willing to try if I get Information ...

[echo sentinel]

Well damn, he’ll still be worse than stable

[abstract linden]

Hello!

I have a question or two.
I wanna no more about A1111 and if it has any connection to SSD and SD in general.

[desert dagger]

it's an interface to run Stable Diffusion with

[abstract linden]

I see, is it common among those who use SD?

[desert dagger]

Lots of people use it. Lots of people use something else, too. what sort of machine do you have?

[abstract linden]

Well, I was thinking of using both my desktop and my phone

[desert dagger]

youre not running stable diffusion on your phone. what sort of desktop do you have?

[abstract linden]

Must require a lot of recourses, I have a 8 core processor with 16 gb of ram and Windows 10 installed.

What info are you looking for precisly?

[desert dagger]

this page https://archvizartist.com/article/how-to-install-stable-diffusion-on-windows-automatic1111/ has the steps to install automatic1111 and the min hardware requirements.

your system sounds like it should handle it fine

[abstract linden]

Excellent,a nd thank you!

Do you use A1111?

[desert dagger]

i use ComfyUI because i prefer that interface. I believe @warm junco uses auto1111 though

[abstract linden]

Hmm. Okay, broader question, it appears from a quick Google search that A1111 and Comfy do seem to be the two most common

Is comfy free?

[desert dagger]

yes. and there are very good tutorials on youtube for installing and using it. just search youtube for how to install comfyui

[warm junco]

hey yea, if your new to SD i would recommend starting with Auto1111

[abstract linden]

Alright then, I think I will start out with Auto1111. Seems like a good way to become more familiar with SD in general.

[warm junco]

Here you find all the install guides for different webuis for nvidia and amd:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides

[abstract linden]

Oooh thank you much!

[warm junco]

for any install question feel free to ask in #🤝｜tech-support

[abstract linden]

Thank you! I will move my discussion to there.

I'll try to install it now and see what happens. Rn I just got the Teaser sub for SD to see if it's the way I want to take my future AI art endeavors

[warm junco]

ahh alright
With the local SD installation you can do much more than with any cloud based solution

but you need a good gpu ^^

[abstract linden]

My computer SHOULD be up to snuff, it's got a lot of power to it. Though if it isn't would be a good excuse to upgrape 🤔

[desert dagger]

also, once you start using SD on your machine, there are a lot of channels to post in and share your images 🙂

[opal hedge]

SAI is basically dead now, right?

[abstract linden]

Id be delighted to!!

It's installing now so we'll see what happens soon enough! 😁

Actually, I do have a question that shouldn't be too much on the technical side. Someone mentioend that SD is heavily reliant on the community styles and Loras and so on, with A1111 will it be explicitly obvious how or where I can get those community driven things or will I have to follow a different tutorial?

[warm junco]

ah yea, you get the models and loras from Civitai.com its the biggest community model database

then you download the stuff there and put it into the right folders in auto1111 and they are ready to use

[abstract linden]

Very intersting, I've visited Civitai a little bit, I didn't know it was the one-stop-shop for models. Great info! Thank you!

[desert dagger]

i would advise you not to worry about using the fine tuned models and loras to start with. first you need to find out what the base model does. then you can look for fine tunes that will adjust it where it doesn't work the way you want

[fervent thunder]

idk
the SDXL base model is pretty bad

[desert dagger]

you still want to learn what the base model can do first. foundations.

[fervent thunder]

actually yeah that does make sense

[abstract linden]

Is there a channel for prompt engineering? I'd like to become more familiar with SD syntaxes

Oh nvm I see it I could be dumb

I guess I could also just look up documentation 😅

[brave matrix]

Hi all!

I am looking for people to test my new AI LLM powered livestream on twitch where you can talk to the characters in realtime

If anyone is curious to help test

https://www.twitch.tv/timelesstakes

[fervent thunder]

I think I saw that stream on reddit before

[nova lily]

Why did they remove the restore face ticker from the A1111 menu? I suppose its still working though

[pale latch]

its in settings now. inconvenient change

[low moon]

very inconvenient. If its not broken it needs tweaking.

ship it, the peasants will love it

[thorny hemlock]

why does it seem like every revision of A1111 gets worse? Features removed, menus changed (not for the better), tiling taken away or hidden someplace I can't find it, the "add random artist" button removed..

[low moon]

Yes. Tis is why I used Forge but Forge is abandoned I hear.

We all must go to Noodletown. This is the future. this is the way.

[iron imp]

Just use SwarmUI? Sleek UI with a backend of Comfy so you can noodle it up if you want.

[low moon]

I just embraced the noodle... I don't need swarm. I am the noodle.

Took me 4 months tho. So I do not hold your apprehention against you. Noodletown is intimidating. It is a very wretched hive of scum and villainy. But the best freighter pilots are to be found there.

Forgeland, Autotown and Fooocusville are a joke.

The pilots there are amateurs.

[iron imp]

Sometimes I just want to have a simple UI where I can clicky clicky and art comes out.

[low moon]

Yeah I did that too

But then I compared what comes out there compared to Comfy and I was shocked. =0

Esp. Fooocus. I mean it's cool for a quick generation but the quality really suffers.

[iron imp]

Swarm literally uses Comfy its just a front end it cannot gen images on its own.

there is even a tab for comfy UI right there on the page so you can always switch to it.

I always sound like I am promoting Swarm lol Im just a fan of the UI.

[low moon]

I know.

But I am noodle addict.

[iron imp]

noodle it up!

[low moon]

Once you go noodle you never go back. I promise you.

[iron imp]

just dont try to eat them in your gens

[fervent thunder]

fooocus is also noodle

[proud burrow]

Good evening, just getting into Stable Diffusion here, this is above my paygrade but we're taking a crack at it, looking for some guidance on where to start. This computer is setup with 5 GPU cards and an 8 GB RAM card. My wife and I like to watch animes and want to make pictures of characters in this style, we are curious to see more characters closer to our age!

Do you have a suggestion on a suitable starting program for this situation?

[fervent thunder]

could you name the card please?

or at least name the best card

the most important thing is to have an Nvidia card and not an AMD card

[prisma junco]

上方

[proud burrow]

Hello, thanks for the response! I'm not sure the cards seem to be Nvidia. They are Intel Cores. In the Windows About they are "Processor Intel(R) Core(TM) i5-4670K CPU @ 3.40GHz 3.40 GHz"
Looks like they have 4670 byte VRAM. I understand this may limit our options of which program is ideal

[fervent thunder]

Intel i5 4670k is your CPU
for Stable Diffusion you need something called a graphics card, also known as a GPU

do you have a thing that looks like this:
https://www.trustedreviews.com/wp-content/uploads/sites/54/2020/12/RTX-3060-Ti-fans-e1606826648576-920x614.jpg

[proud burrow]

That is a clever way to describe it! It has a very similar device. I imagine 3 fans versus 2 is still a graphics card?

[desert dagger]

no. if you had a GPU it would be listed as a GPU

[proud burrow]

I have searched the computer and it has AMD software. It looks like it has an AMD.

[desert dagger]

it sounds like you ahve an intel card with the graphics on both the CPU

[fervent thunder]

can you upload a photograph of the inside of your PC

[proud burrow]

"GPU AMD Radeon R9 200 Series Primary/Discrete"

[fervent thunder]

ah

[desert dagger]

You need something like this https://www.newegg.com/abs-aqa14700kf4060ti16g-stratos-aqua/p/N82E16883360436

[proud burrow]

it has this in the AMD software.

[fervent thunder]

you do have a GPU

[proud burrow]

"VRAM 4096 GDDR5 1250"
Looks like 4096 byte VRAM. Is there program we should look into?

[fervent thunder]

well I have good news you can definitely run stable diffusion

[proud burrow]

Thank you for your help. This has been an exciting look into a new world. We will have to read the materials on how this works, then.

[fervent thunder]

the issue is
you have an AMD graphics card
and most people use an Nvidia graphics card for this

it will make it more difficult to setup

I don't know how to do it with AMD but maybe someone else will

[proud burrow]

I see. From looking at the programs we see on Reddit they have mentioned DirectML being a popular aspect for AMD. A few others that were experiments. We'll start by reading about Stable Diffusion with DirectML.

[fervent thunder]

hopefully it will be okay
doing AI on AMD can be tricky
especially for new stuff

[proud burrow]

Well, this is just a venture for fun and exploring a new time. All is well if we're unable to operate it. Thank you both very much for your help. You are very pleasant and informative.

[desert dagger]

the problem is that if you want SD to run well, you need to run CUDA - and it won't run on amd - you can try and see if you can get AMD to run python

[nocturne venture]

sofa Greenery and flowers Modern Chinese an expansive view of zen a room with wooden furniture and a wooden table, in the style of neo-geo minimalism,subtle color variations, minimalist figures, textural detail, light beige and light amber, dada-inspired constructions Modern Chinese Traditional elements Modern design concepts Simplicity Elegance Serenity Natural materials Minimalist lines Functional design Tranquility Comfort Blend with nature .Cinematic shot FHD 18K high detail --ar 4:5

[desert dagger]

you can't generate in this channel

[proud burrow]

We had begun to gather this from looking around prior but this is a much cleaner way to state it. Thank you, we will read on this as well.

[graceful bramble]

So i have a random stupid question. Are embeddings always pt files or can they be safetensors too and in reverse can loras be pt's or only safetensors?

[warm junco]

Hey, checkout my SD install guides for AMD.
They are linked in the pinned messages of #🤝｜tech-support
There is also Auto1111 with Directml.

[warm junco]

Embeddings can be both.
Loras can not be .pt

[graceful bramble]

Got it thank you!

[tight comet]

hello

[keen granite]

Hey I have little experience with stable diffusion and want to create fantasy images. On civit I saw a lot of people using pony v6 XL for fantasy images is this the one I should use?

[plucky linden]

anyone making that wouldn't be selling courses on ebay

[river sparrow]

If you want highly sexual fantasy images, yes

[atomic mortar]

Hmm you can use pony for good SFW work too. Pony prompting requires some more work then normal though

[river sparrow]

yeah, that is true. But if you only want sfw, there might be better models

[atomic mortar]

Yeah pure sfw theres probably some better ones but pony has a advantage of knowing certain characters and concepts really well as its advantage

[lilac forge]

Hi everyone,

I've recently started using Stable Diffusion XL models after having experience with older models, LORAs, and checkpoints. I've noticed several differences and have some questions, especially regarding some issues I'm facing with corrupted images.

The older models and LORAs seemed more flexible, especially for adjusting proportions. For example, it seemed easier to change the proportions of a subject model with older models. In XL models, the subject often stays closer to its default look, creating some friction. What other significant differences are there between XL and non-XL models? Also, is "Pony" a subset of XL, or are they entirely different? I've been using a lot of "Pony" models as they seem to be super popular on Civitai.

Dynamic Prompts, XL Models, and Corrupted Images (Most Important):
With older non-XL models, I could successfully use dynamic prompts with a list of LORA activation words and supporting words in a wildcard format. This worked well with Stable Diffusion Automatic 1111 UI, which is what I'm using. However, with XL models, I often get corrupted or glitchy images. I was gonna include an example, but I can't post images here. Suffice it to say, they are very random in texture and color, but this one looks like green and black clouds mixed together. This can happen even with a single LORA and without the wildcard system. For example, changing the weight of an adjective from (adjective:1.2) to (adjective:1.4) or (adjective:1.0) can cause this issue. It just feels like the XL models are super finicky and fragile. While testing out many of them, I found that when they work, they are worth it, but sometimes I'd get a corrupted image for no apparent reason. Why is this happening, and how can I avoid these corrupted outputs?

Usage on Civitai:
When browsing models on Civitai, example images sometimes don't list the LORAs in the prompt using the < > symbols. For example, this prompt uses them:
score_9, score_8_up, score_7_up, rating_safe, 1girl, Sheik (Ocarina of Time), serious, lora:SheikSDXL:1
But this does't
score_9, Score_8_up, Score_7_up, Score_6_up, Source_game, Source_furry, Rating_safe, 1boy, JBowser
Has there been a change in how LORAs are integrated or displayed in prompts?

Score and Underline Notations:
Many example images include notations like score_9, score_8_up, score_7_up, etc.. Does this mean the AI selects images rated 9/10, 8/10, etc.? What does the "up" signify? Also, what does rating_safe mean? Does source_game and source_furry mean the AI should only draw inspiration from non-humanoid game characters?

Any insights or advice on these points would be greatly appreciated. Thanks in advance for your help!

[fervent thunder]

blinks

I am going back to sleep

[trim magnet]

the score_# tags are only used by pony models(a finetune of XL),some images dont list loras because pony models dont need a lora to make some characters like Jbowser you only need to type its name and a basic description of its attire

[river sparrow]

that might be unrelated to the models, but linked to difference between a1111 and comfy. Using loras in comfy does not required the lora:1.0 part in the prompt since the lora weight can be controlled in a node outside the prompt.

[lilac forge]

lol sorry for the wall of text. But you made me laugh

[lilac forge]

Okay, that makes sense. It makes it kinda hard sometimes, though, because when things were always within a <lora> format, I could tell that they were models, not part of the description... something that I might want to look up and download if I wanted to accuratly replicate their image. Now, with no such markers, all I can do is see if something looks odd (in the Bowser Jr case, it's JBowser... that's easy, but what if it was just, like... jacket ?"

[river sparrow]

why would you want to exactly replicate already existing images anyway?

[lilac forge]

I want to start from where they left off to have a solid base. Like if I love an image of a person in a yellow shirt but want it blue, I'd want the same data to keep the look, pose, texture, and background the same. Everything that I like that I want to keep aside from changing the shirt from yellow to blue.

[novel trail]

Does somebody now how I can replace an animated avatar on a green font by an image of a person ?

[velvet patrol]

Hello guys

[desert dagger]

hi

[velvet patrol]

Nice to meet u bro... I just started learning ai art.. is there a manual book/refference for prompting tecnique?

[desert dagger]

not really. and everyone has their own way of doing so. The best thing you can do is start with one word prompts and just see what the AI does when it sees that one word. The add a couple of words. For example, just use the word apple. generate that a few times. Then use the word dog. generate that a few times. look at what you get. then try both words in the prompt: apple, dog <-- geneate that a few times then maybe try: a dog eating an apple. doing this will let you learn how the AI you are using thinks, and how to communicate with it so it understands you. And all the AI's are different

[craggy void]

hello m new here. can someone please tell me about architectural visualization with SD and m already using it but need help to improve

[vast ingot]

Some positive feedback too, since I did the negative feedback the last few times I came around : nice to see the #🏆｜winner-gallery back in the archived communities. Thanks admins (Fruit I believe, thanks)

I still hope we can see some of the previously archived channel back in those archives though.... but I do seem to be the only one so I get it

[forest trout]

I want the old channels back too.

[mortal delta]

is there a place on this server where i can talk about llms? or nah...

i know there use to be a channel for stable diffusions llm but i believe its not there anymore.

[hybrid edge]

Hello

[fervent thunder]

hey guys is there any neg prompt list?

[loud solar]

As less as possible ...

[proud burrow]

Hello, thank you for these excellent guides. This would have been an ideal starting point, they are very concise.
Seeing as our graphics card is only capable of DirectML rather than ZLUDA, will Stable Diffusion technology models give us a better experience than Stable Diffusion XL models?

[warm junco]

No problem, and yea please try to use 1.5 based models these are usually 2-4gb in size.
2gb models should work good for your gpu

[cerulean knoll]

hey guys, just wanna ask if I can use High res fix with SDXL models??

[warm junco]

Yes, depends on your GPU

[cerulean knoll]

pls tell me if anyone know, would be a great help

[cerulean knoll]

3080??

or I need the 40 series?

[warm junco]

Yep should work. But you need to have
--xformers --medvram-sdxl in your webui-user.bat
And maybe the tiled vae extension installed

[cerulean knoll]

already using those arguements

[cerulean knoll]

pls tell the settings , I mean the steps and res?? 2X 0r 0?

[warm junco]

Make sure to use 10 hires steps

[cerulean knoll]

https://civitai.com/models/261336/animapencil-xl Was trying to use this model

its really great for anime artworks

[warm junco]

Yea thats nice

Try with 2x on a 768x1024 resolution

10 hires steps, 30 normal

[cerulean knoll]

any tips for anime arts, I know I'm going overboard

[cerulean knoll]

okay, thanks a lot

again

[warm junco]

Using adetailer extension can improve the face and the eyes

[cerulean knoll]

I create scenery stuff

[warm junco]

Using Boorutag autocomplete extension can help for anime prompting

[cerulean knoll]

backgrounds mainly

[cerulean knoll]

gonna try

[cerulean knoll]

is there any new scheduler? Align your steps??

[junior thistle]

how to change a1111

model, lora file paths?

is it even possible

[loud solar]

in the start batch file ...

[mortal garden]

I love you.

[ornate flame]

Please move this discussion to the confessions channel

[junior thistle]

Damn

I'll have to check it when I'm on later

[warm junco]

By adding
--ckpt-dir "D:\Path\toModels"or --lora-dir "C:\Path\toLoras"
To the webui-user.bat

[hollow salmon]

hey newbie here, can anyone tell whats up with sd 3, can i use it?

[river sparrow]

yes you can use it

but at the moment you cannot merge it or make lora, or fine tune the model

civitai published the model, then took it down cuz of license doubts

because it seems all derived works, such as fine tuned models, loras, etc, are under control

[delicate oxide]

[last oyster]

why does mahe not generate unique faces? every roll of the model generates identical faces?

[river sparrow]

did you prompt for different faces?

[desert dagger]

try using names in your prompt

[nova lily]

why are inpainting companion models seem less prominent? not worth making? main model does good enough job?

[loud solar]

I can just talk from my point of view ... learning prompting 2 years now ... and so I have the stupid idea of creating something new that way ... and not to change something that still exists ... it's just my point of view and I know might not be the cleverest one ... 🙂

[forest trout]

Hoy.

[nova lily]

Im still dont get the wat adetailer goal is. I feel like its for automaticaly fixing broken parts instead of faving to inpaint is that correct.

[fervent thunder]

IDK I use comfy

in comfy you can use an object detection model to mask an object automatically
and then inpaint automatically

as far as I know adetailer is similar

but this sort of thing needs a lot of tweaking, I wouldn't want a premade tool doing it

[woven canopy]

Hello is this a crypto server?

[desert dagger]

nope

[floral umbra]

How many images do you guys reckon would be needed for a decent lora/lycoris? As there's lack in tribal lora's for a project i have, so decided i wanna train my own

[forest trout]

I've seen people do Loras with a single image and get decent flexible Loras with pretty good results. It really is subjective though.

[desert dagger]

people suggest 30. i'm working on a dataset that'll have 50

[fervent thunder]

one image is ok

[uncut night]

Anybody here have any experience with creating 3d meshes from stable diffusion images locally?
I've gotten pretty good results from instantmesh, but I can't run that locally right now (it errors out saying that it tried to allocate 15 gb of vram when I only have 11)

[floral umbra]

Gotchu, i ended up getting a few 100

[desert dagger]

🙂 i may actually wind up doing that too, the way i'm going

[vivid shale]

anyone here familiar with openvino? specifically how to install/run openvino in preperation for https://github.com/openvinotoolkit/stable-diffusion-webui

[shut owl]

So guys when are we getting stable diffusion 4? Asking for a friend

[fervent thunder]

well with their new investment

maybe soon

[desert dagger]

they're not finished with 3.

[thorn wigeon]

__

[harsh linden]

Hello everyone

I am very happy to communicate with you.

Now I am going to fine tune the stable diffusion 3

how to prepare the dataset and how much amount is needed?

please help me. I am looking forward to getting guidance from you.

[hallow nacelle]

Is there a way to find out what checkpoints are used to merge in a checkpoint file?

[desert dagger]

SD 3 medium is an unfinished model that you're going to have a lot of problems making fine tunes for

[harsh linden]

As I know, SD 3 was released in the last year

[pearl ocean]

I hear SDXL is better than SD3, that true?

[loud solar]

SDXL has community models ...

[next zealot]

chat thats wild

[junior urchin]

SD3 has a handful of community models too

That they're bad, is something else

Pretty sure Base SDXL was better than the current SD3-Medium base

Or as comfy would call it, SD3-Failure Base, as it wasn't meant to be released

[cyan nest]

hey ppls 😄 Can someone send me over the link to github or similar to look at the different install options?

[warm junco]

Its in the pinned messages of #🤝｜tech-support

[fervent thunder]

The fact that people think they can come into any channel and spam anything blows my mind.

[wheat mulch]

Hey! I'm new here, and don't worry i wont spam or anything

[fervent thunder]

Entitled people in general blow my mind still. And I am getting old. People are actually raised by their parents to believe they have a certain right to things; without earning that right.

[frail sonnet]

I'm debating about some additional internet services. How much data do you folks use per moth on average? I mean those of you who also DL far too many checkpoints and loras. Also there's the backup of thousands of images, etc. etc.

I currently have 200gb per month high speed. This apparently isn't nearly enough.

[fervent thunder]

oh goodness, thats a good question

But I play games sometimes

[frail sonnet]

I travel a lot, so I've just bene using my phone data as a hotspot, it's so fast! But only 200gb 😦

[fervent thunder]

I did that when I traveled too, but I did not have that much data. I had to upload photos to a ftp site for work. I have no idea how much data tho. That's a good question.

[frail sonnet]

Everytime I hit an airport, I get out my list of all the checkpoints I want, and start the downloads 😄

My phone data is 1gb speed, which I think has made me spoiled

[fervent thunder]

That's tough because I imagine those checkpoints eat your data up.

I am not sure with big daddy AT&T if I can even check my data usage. I am not sure I want to see it 🤣

[harsh linden]

Sorry to interrupt your conversation

And don't worry, I'll never spam

The only thing that I want to is just learing and exchainging experience.

nothing else

So please help me, what is the latest version of stable diffusing and how to train or fine-tune it?

There are two ways:

• Using API
• training model by myself

right?

[fervent thunder]

I just use it locally, but I am a casual user. I see others that use api's, etc.

[sudden ruin]

Limited bandwidth? Thats a thing?

[frail sonnet]

They all claim unlimited, but if you want the high speed version, then there are limits eventually.

[sudden ruin]

Where are you from?
Cause over here thats only a thing for mobile phones

[frail sonnet]

"then reduces to slower speeds after 500gb" for eg.

[cedar salmon]

dont download every checkpoint 🙂

[frail sonnet]

I use my phone as a wifi hotspot

[sudden ruin]

Oh that makes sense then, but why

[frail sonnet]

You have such restraint! 😄 lol
I only download the good ones, but that somehow ends up being a LOT lol

[hasty hornet]

Some good models aren't popular models, can't tell if it's good until you try it.
One day I downloaded some weird model with barely any info and examples, which turned out to be the model I used the most in the end...or one of the most.

[frail sonnet]

Less expensive, and I travel a lot. But the cable internet caps at 500gb high speed.

[frail sonnet]

At least I can try them out first on that iste everyone hates!

[cedar salmon]

ya i like trying the more unique ones

[hasty hornet]

oh I didn't see we can try models on civi now

[desert dagger]

Depends on what you are trying to accomplish. Personally, whil I love SD3 and SDXL, I still find 1.5 better for some things I want to create

[frail sonnet]

1.5 is my fave still! Though the crispness of those SD3 images is pretty wonderful as well

[frail sonnet]

Well you can create images using models people have uploaded. THere are also example images, but as someone pointed out to me, those are cherry picked and sometimes edited

and my fave model hasn't been updated in a couple of years 😦 I may have to take matters into my own hands soon LOL

[desert dagger]

Which one is that?

[frail sonnet]

I asked Gemini for advice on this and it asked me some uncomfortable questions ROFL "Given your usage patterns, an additional 500GB of data might be sufficient, but it's difficult to say for sure without knowing more details:

AI image creation: How many images do you create per month, and what is their average size?"

[fervent thunder]

oh goodness

[frail sonnet]

lololol

perhaps local creation mostly, and backing up to a physical hard drive instead of google, is in order for me!

[fervent thunder]

I think google photos hates me sometimes

[river sparrow]

I got those "warning, your google account is on 80% of it's storage limit" ... and Im not yet at 80% of my life (I hope)

but thats only some photos, emails and audio. No AI stuff on my google drive

In may I generated 10427 png files, total size 10284 MB

wait, I may have deleted some images, those are not counted

June was about 8959 png files, total size 9551 MB

[dull quail]

hi guys

[river sparrow]

hi

[desert dagger]

Make a private discord, upload your images ther rather than taking up you drive space

[river sparrow]

no, i rather keep most local thank you

oh wait, you are talking about the photos and stuff, not the ai images

[supple plank]

which channel can I ask question about how to achieve a ceratain pose?

nvm I guess #🏞｜general-with-images lol

[next hedge]

Hello guys, I’m looking to buy a PC whose sole purpose will be to run SD. 

I want to be able to use any SD version and model in a reasonable amount of time. I also want to be able to work with videos so probably it should be able to run ComfyUI as well.

After discussing with some of you yesterday (since I am not a PC guy at all) I came up with this setup:

• NVIDIA GeForce RTX 4060 EAGLE OC 8GB GDDR6 (329€)
• AMD Ryzen 5 - 4600G / 32GB RAM / 1To SSD M.2 NVMe (429€)
• beQuiet PURE POWER 11 700W (129€)

I would appreciate hearing you guys opinions about it
(also given the current speed of how AI progresses, I do not want to need to upgrade everything in like 6 months lol)

[frail sonnet]

I fell for the double my storage offer at that point lol

[frail sonnet]

Are they 100 res and no limits?!
Well I guess I might have 10k mj ones on one if my discords still!

[hollow cloud]

Hello, I am having problems installing web u automatic 1111

[desert dagger]

when you upload an image to discord, it stores the full sized image. it displays a low res image but you click on the image, then use the 'open in browser' link and open the full sized image, which is in the format you uploaded, in another tab. and if it happened to be a comfy created image, it also has your entire workflow in it. just download it, then drop it into comfy to open the workflow. discord provides unlimited storage (at the moment) and doesn't charge for it

[cedar salmon]

ide aim for a 16GB video card minimum, IMHO. and i think people may have to jump through additional hoops with AMD for some things, im very unclear on this as i dont have one

[next hedge]

Mmmh interesting about AMD, let's see if someone who have an AMD processor can confirm this

[next hedge]

Just to be sure to understand, the 16GB RAM on the graphic card will mainly make things go faster?

[desert dagger]

Gen 3 is now open to everyone on Runway

[cedar salmon]

TBH maybe i was remembering seeing things with AMD video cards, ide disregard me lol but yes on the video card memory

[next hedge]

haha alright, I see the GeForce RTX 4060 Ti Advanced Edition got 16GB RAM

[frail sonnet]

That is seriously amazing!

[next hedge]

is there a difference between the brands? Inno3D ASUS MSI Gigabyte they all have their version of the "GeForce RTX 4060 Ti 16GB"

[desert dagger]

yeah. Asus tends to be better at creating computers for graphics work

[next hedge]

alright! and I guess they are all equally compatible with AMD Ryzen 5 CPUs?

[frail sonnet]

I bought a brand new computer with only 8gb gpu, it's kinda obsolete already 😦
It runs SD 1.5 and SDXL just fine, but SD3 is a challenge, especially with complex workflows. 16gb NVidia gpu if you can.

[atomic mortar]

im on a AMD ryzen 9 5900x and no issues with SD

[next hedge]

This is exactly what I wanted to know and the reason I posted here before buying anything, I'll definitely go for 16GB! Thanks for the info

[frail sonnet]

I definitely should have posted here before I bought mine! Fortunately it runs great for most stuff, and there are many online options now for the other things!

[trim magnet]

yes but a 4070 12gb would be faster because the 4060 ti only has a 128 bit bus

if you really want 16gb with no handicaps get the 4080

[desert dagger]

you want more than 16gig if you can afford it... what's coming over the next year is going to require more, all the way up to 32gig

[next hedge]

I think I can push to go to the 4070 with 12gb but 16gb is out of budget I'm afraid :/

[frail sonnet]

why is Runpod 30% more expensive than its competitors, (same specs), is it better, or?

[trim magnet]

whats your max budget for a gpu

[frail sonnet]

Perhaps make sure it's upgradable gpu wise?

[next hedge]

well I see the RTX 4070 12GB is around 600 eur and I could afford this but I would not want to go much higher than that

[trim magnet]

well if i had 600$ for a gpu id just buy an used 3090

[next hedge]

mmmh interesting, i didn't think about going for used stuff

[next hedge]

3090 is older but more powerful right? I see 24GB ram, so not yet 32 but better than 12 haha

[cedar salmon]

has more cuda cores than some 40 series cards also

alot more

[trim magnet]

yea only the 4080 is slighty faster but 4080 has 8gb less vram so only card who can beat it is the 4090

[frail sonnet]

I'll take one of these for noe ROFL https://www.dell.com/en-ca/shop/cty/pdp/spd/alienware-aurora-r16-desktop/caneahctor16i29?tfcid=54598298&gacd=9683519-3041-5761040-266312346-0&dgc=ST&SA360CID=71700000110509372&gad_source=1&gclid=CjwKCAjwp4m0BhBAEiwAsdc4aFaxXPaGOY7RUGtcehOf23hiGnkrJs8-of5SZSQG6xHwO4Yq5XFQEBoC1sIQAvD_BwE&gclsrc=aw.ds

[next hedge]

mmmh i see, and longevity wise is it gonna keep up vs the like of 4070 or 4080?

[trim magnet]

yes plus if you have money u can buy a second or third one since its last rtx gpu to support nvlink

[frail sonnet]

2nd or 3rd 4090..... cries

[cedar salmon]

whats the 5090 gonna be like $4995 lol

[frail sonnet]

That's what my AT cost me in 1986 (obo)

[next hedge]

very cool! and nothing to worry about buying this kind of hardware used?

[trim magnet]

i only buy stuff on ebay with sellers that have a lot of feedback,offer free returns,tested the card and put lots of pics/videos of the product

[next hedge]

I just found this one for 550eur https://rog.asus.com/graphics-cards/graphics-cards/rog-strix/rog-strix-rtx3090-o24g-gaming-model/

[trim magnet]

that one looks great

[desert dagger]

you do have to keep in mind that a used card MIGHT have been 'abused' - allowed to run too hot or in other ways damaged - and the damge might not be something the seller knows about

[keen dirge]

Hi Stable Diffusers,

Do you happen to know a platform (or a source) where I could find the latest finetunes available (mostly looking for SD3 rn)
Or where I could look for them when and if they come out?

[next hedge]

good point, maybe I should ask the seller what he used it for and for how long and depending of his usage maybe I can infer if this is safe to buy it or not?

[desert dagger]

i would ask him about its history at least

the question i always have is "if it's working, why are you selling it" ?

[river sparrow]

same for me, when can we train sd3 loras?

[frail sonnet]

SD3 isn't fully complete yet AND, they haven't released the necessary info for training. Soooo you can make SD3 checkpoints and loras, supposedly, a few ways, but they are going to suck compared to if you just wait a few weeks or whatever. (excuse the bluntness lol)

[next hedge]

yep, I already have another seller (in case this first one is not convincing) who sell it because of upgrade (its a bit higher in price though but still ok)

[keen dirge]

alright!

[frail sonnet]

Gemini advanced is skeptical as well, I asked it 😉

I've seen and tried a few models so far, but they are very subtle differences, nothingn like the differences SDXL and 1.5 models and lora make on their base mosel

There's some perturbation code laying about to use on SD3, but it is so very suble

Or just do 10k prompts and you'll know it well enough you can make nearly anything on SD3 (my method lol)

[desert dagger]

people really don't want to make loras and stuff for SD 3 medium. Everyone is jumping the gun. people need to wait till 8b is out and make loras and stuff for it IF it even needs those - and it might not

[frail sonnet]

So very true, 8b is amazing!!! That and it really does know a LOT

could compete with MJ on its own a lot <ducks>

[river sparrow]

But, can I run it local with 12 gb vram ??

[frail sonnet]

Lora or checkpoint? Lora maybe? Checkpoint I doubt it, everydream2 trainer wants 16gb vram. The others are probably similar. Though you can always try, I made an SDXL one on my 8gb system, it was the tiniest checkpoint ever produced, took 4 hours 😄

[river sparrow]

im using stableDiffusion3SD3_sd3MediumInclClips and I think it beats sdxl on many way, but not all!

I just wish it had an easy way to train lora, cuz it def misses something

It really needs my lora

3d fractals as in the style of 3d renders by mandelbulber2

at the moment it can only do the popular semi 3d fractal stuff

it just needs a bit of depth on areas of specialization

[carmine edge]

setup the chat

anyone wants to test?

[full lark]

Me, i have sd3 opened

[desert dagger]

SD 3 medium is missing a LOT of stuff. it's unfinished and a lot of the fine tuning was skipped in order to rush it out to release. i wouldn't waste my time on fine tuning stuff for it

[carmine edge]

its a app

gate-app.com

login and text something

[full lark]

Aaaa

[river sparrow]

they seem to put some effort in hiding nipples, to give one example

oh no, I think I wrote a wrong word ...

[full lark]

Banned in 3..2..1... 😁

[river sparrow]

was about to say that SD3 does not have any issues with the area between a women's face and top - even though it prevents unclothed female chest

[carmine edge]

what do you think i should do

[river sparrow]

I personally think you should stop spamming on this channel here

[full lark]

For faces and simple body photos it works fine, for a little more complex like lying, sitting or things like that it breaks

[carmine edge]

wdym

[full lark]

For architecture it works way better than the other sd AIs in my opinion, also the lighting can be so realistic

[river sparrow]

it also works fine with line art, cartoon style, a bit of anime, horror, portraits, nature shots, landscapes, etc

probably much, much more

[full lark]

Too bad idk how to train a lora on it, I would love to train it using screencaps from some games like dark souls, classic doom or others

[carmine edge]

im just looking for non mushbrain people

where do you think i can find them

[river sparrow]

Well sorry cuz I happen to love the biological world of fungi, as some have the capability to expand the brain and make you see that all is one, a pretty heavy but ultimately a nice thing. I have absolutely no idea where you would be able to find any non-mushbrain people, as all or not most are in some ways mushy. Sorry.

[desert dagger]

well - i wouldn't suggest you look for them on reddit

[frail sonnet]

Espcially horror! 😄

[carmine edge]

oh so youre a psychedelics guy

[river sparrow]

yeah lol, since SD3 I have been adding "body horror" to the NEG prompt

yes sir!

well, not in the last .. hm, idk maybe 12 years

[carmine edge]

damn i hate old ass

[river sparrow]

wait! it's loading . . . see https://discord.com/channels/1002292111942635562/1004159122335354970

[carmine edge]

youre like 40 playing with ai stuff right

[full lark]

And you´re 80 trying to bait us loging in your page using our google account so you get all our info 👌

Stop the ragebait and insults

[river sparrow]

yes

So, PTB Godgather, who are you, where did you come from and why are you here?

[carmine edge]

why tf would i need your info bruh

[loud solar]

https://tenor.com/view/michael-jackson-eating-popcorn-enjoy-i-like-nom-nom-gif-11040065238845078056

[carmine edge]

who tf even goes around sharing stuff to get info for no reason

[river sparrow]

me

now you

[carmine edge]

im a bad guy

[river sparrow]

love u2

[carmine edge]

scary monster

[river sparrow]

hell no, I ment the band. what did you think?

[carmine edge]

?

[river sparrow]

and this is how we communicate

past eachother

will be generating and posting some images now

[carmine edge]

lmao

[frail sonnet]

Anyone ever used Shakkar to create loras etc.? Not so sure I trust it, but I'm also wondering how it can create checkpoints for free??

[river sparrow]

no idea, since ive only used comfyui for checkpoint merging, and civitai lora training

[bleak matrix]

Good afternoon, everyone! How are we all doing?

[river sparrow]

just generating some more ai images

oh, are you still here?

[fervent thunder]

Hi everyone 🙂

I have a silly question, and please let me know if I'm asking in the wrong place or direct me elsewhere if possible 🙂

I'm on my way towards creating a "character" that is entirely virtual. I've been working on refining my prompts, using different tools (Fooocus, comfy workflows, different models from civit.ai, etc.) to get the character built. My vision is to somehow have a specific character defined - for example: Timmy, age 62, 5'10", light skin, slightly overweight, scar on his face and be able to call that specific character into a scene. I feel like this should be doable, but I'm green and unsure which avenue to take in order to do this. So my questions: Is this currently possible? Will I need to rely on image-to-image every time in order to recall a specific character?

thanks to whoever reads! 🙂

[broken smelt]

I think creating a lora of that character is what you want

[brazen vale]

Hello everyone, I'm new here. Can someone guide what is the command to use in rooms to animate images?

/ After this what to write?

[desert dagger]

read the information here #🗣｜artisan-support-feedback

[brazen vale]

Thank you. I need to create video using image i.e. image to video

[desert dagger]

okay, with any discord server, as long as you are in a channel that has been configured to allow something like an image generation bot - on here that's the artisan channels - just type the / and then look through the commands list to find the commands you need to use. they'll all be listed under the names of the bots that can run them. once you find the command you need, click on it and then fill in any boxes that display. if there are options that don't fit on the line or that aren't required, you'll see something like +3 more at the end of the line. click it to get the list of options and click the ones you want to use. hit enter when you have everything filled in and selected

if you've used a command a few times, it'll show up at the top of the list when you type /

[brazen vale]

I found dream command but it only generates images l guess. I need to animate images

[desert dagger]

you generate an image first. then after it's generated, you choose to edit it, and then you can animate it

[brazen vale]

Awesome, will give it a try dear friend

Thank you 😊

[desert dagger]

welcome 🙂

[vestal ridge]

Hello there!
I am looking to train the stable diffusion model to generative images with the art style I have in mind
I looked up a video and it told me to give the sample images generic names like ghconf0
But I want to somehow tell the AI what is in each sample image to hopefully get a better outcome
Is this possible and is it even a good idea? thanks for the help 🙏

Using this repo btw:
https://github.com/TheLastBen/fast-stable-diffusion

[fervent thunder]

Thank you! I will look further into this.

[desert dagger]

you use lables if you want to do that.

[vestal ridge]

Thanks for the help 🙏

[mossy panther]

Yo

[lyric nova]

Nice to meet you,
I'm currently using stable diffusion forge, is there an extension that separates Lora's description area?
I've used forge-couple, but it doesn't work well for very complex things. If there are other extensions, I would like to know about them.
Sorry if this is the wrong place to ask this question.

[eager shard]

what the hell, for some reason i get higher it/s using 8 batch size instead of 1

ok nvm slightly lower

nvm its the same

???

is lower or higher it/s good

0.92s/it vs 2.88s/it

[pale latch]

it/s is iterations per second. s/it is seconds per iteration

[eager shard]

how do i exchange the numbers

[pale latch]

it happens automatically. if you have more than one iteration a second it'll display it/s. at a low enough threshold it switches to seconds per iteration.

1 iteration is 1 step

[eager shard]

ahhh ok

tyty

was always wondering what any of that meant

[eager shard]

using swarmui if i change from 2 to 10 batch size its the same generation speed per image but theyre all faster than 1 batch size lol

[violet bobcat]

Hello guys

[eager shard]

just got to 3 iterations per second woooo ( switched from forge to swarmui )

nooooooo swarm ui is bad like automatic1111 for high resolution:(((

sigh i guess ill reinstall forge

nvm it only lagged at the end thats not too bad

[shrewd frigate]

niace

hello everbody

[simple marsh]

Hi all
New here

Need help in deciding if I should purchase a gtx 1660 super for stable diffusion? I currently have a 4gb quadro k2200. It's very slow but does give a decent output.

[golden arrow]

I hope moderators will allow this once-off promo, even though (strictly speaking) it is a wall of text. I don't like clickbait, that's the reason I summarised it. Apologies for any inconvenience. 🙏

[junior urchin]

#1092446741984444416

[woeful basin]

Looking for AI/ML engineer based in Pakistan.
With high salary and for long term project.

[vital gazelle]

Iam looking for a high salary

[bleak matrix]

🌞 Good morning, everyone! How are we all doing this fine summer morning?

[golden arrow]

Bookmarked! Thank you 🙂

[golden arrow]

Not too bad, thank you. This winter evening is pleasant indeed. 🙂

[bleak matrix]

Good to hear! Up to anything interesting?

[desert dagger]

walls of text look like scams. not the reaction you want

[fervent thunder]

yeah it looked exactly like a scam
a very short comment would have worked better

[native heart]

guys anyone know how i can use or find the model PonyXL but i want to use it with FP16

[river sparrow]

what do you mean? Pony is already fp16?

search for pony on civitai

[simple heron]

I use Stability Matrix and I have Forge and Comfy installed. I want to add Fooocus but there is 3 different versions. Advice appreciated.

[floral owl]

regular one works just fine

[pale latch]

dont' actually do this . you'll be inundated with grooming material

[river sparrow]

ok, ill give the direct link in that case, posted by user PurpleSmartAI : https://civitai.com/models/257749/pony-diffusion-v6-xl

[pale latch]

yeah that wouldn't help. Most of the kiddy porn posted to civit happens in that model's example gallery

[hardy nexus]

16ch open source vae dropped : https://huggingface.co/AuraDiffusion/16ch-vae

[pale latch]

i don't understand why all diffusers models have identical non descriptive filenames. diffusion_pytorch_model.safetensors says nothing about the actual file. We get it, you use diffusers. OOoo.

It's a really bad standard and it's like they do things differently and incompatible with everyone else's implementation purposely.

what are the actual benefits to diffusers other than they're the first to rush code support out usually?

[tropic frost]

out of curiosity

is there anti AI art software?

had this crazy idea

basically, have a software that real artists can use

when they make their art, they put it in the software and it become encoded with anti AI program

[foggy halo]

Hi i needed some help...i have an image of a cartoon rapper nugget, and i want it in a different pose, such as a side profile view..is there any ai that can help me with this?

[hardy nexus]

i don't think it could work, there will always be workarounds to your solution. Again, if it can be done.

glaze for example does not work with all architectures

[tropic frost]

i just think of the artists who art is always used to create new loras,
would be funny if they could encode their art with an anti AI coding that would cause the lora making ot fail,

[trim magnet]

a lora takes like a couple of hours to make,the program that was supposed to poison artist imgs is a failure because it can be defeated pretty easily with img2img at low denose

[hardy nexus]

they can't, there are workarounds for the time being.

[hardy nexus]

that too, but it does not work for all architechtures. They have been designed primarily for stable diffusion

[tropic frost]

just saying, would be a good business if there was a software like that, pretty sure a lot of artists would pay for it.

[trim magnet]

impossible to make unless you can invent an img format where pixels move around

[hardy nexus]

there is already nightshade and glaze to do that, but as i said there is workarounds. The artists think they are safe because they exist, but they are not

[trim magnet]

and even then can be defeated by taking a screenshot of your img

it would be like trying to prevent ppl from stealing your NFT

[hardy nexus]

exactly. Artists should make art for art sake. An ai image generator, if it's a good one, does not replicate the images it has seen, because then it will be overtrained and not usable

[desert dagger]

they've been trying to develop something that would do this sort of thing. so far, nothing's worked

[pale latch]

that's what nightshade attempted to do. Any possible version is easily defeated though. You could train an ai to recognize and remove the markers from images that break lora training, but before that there are simpler solutions too. Resizing it. Adding a slight blur. Taking a photo of it and then training the photo. Training data is all just pixels and any information that would defeat a training algorithm would have to be specially constructed and specifically formatted over that arrangement of pixels. So to defeat it you'd just shift the pixels a bit. Easy.

[fervent thunder]

We had watermarks at one time that were supposed to protect you from theft

as a former dmca agent for many years I can tell you that shit dont work

[frail spire]

Ok

[desert dagger]

plus using nightshade on the model actually made a lot of the images it geneated, better, in the tests

(and for any of those to work, you have to get access to the model someone's using and infect it...)

[velvet patrol]

Hey guys... How you prompt a text correctly?

[abstract linden]

Prompting does seem a little tricky. I'm learning too. I think SD uses alot more one word type tokens and short references like "apple, red, long stem, juicy".

There is more documentation, but that might be a good place to start.

[tawny halo]

yo

[desert dagger]

think like the computer. remember, it's only going to understand the terms that were attached to images when it trained. start simple - one word prompts. generate a few times and see what it thinks about by default. then a 2 or 3 word phrase. generate a few times. etc

[river sparrow]

for you own ease of use, a somewhat fixed prompt shape can help. For example: "subject, actions, scene, style, extra stuff"

[full lark]

In my opinion its probably a good idea to get an used 3060 12gb, with 6gb there´s some things you can´t run or it will run slow if you offload it to the ram, but if that´s the gpu you can get its fine, it would be faster than the quadro and you can probably train sdxl loras with it

[fervent thunder]

Is it possible to take a current image I have an use the software to edit it? I’d like to take a picture of a truck and use the AI to put a logo on it

[fervent thunder]

It would be easier with photo editing software

You were excited when you created your nickname weren't you

hahah i just like exclamation

more fun

try using a free software like photopea

ok thank u! : )

also it it just me or does literally anything with humans break the rules and it doesn’t let me generate it

[fervent thunder]

It takes a bit of practice

ugh

[vestal ridge]

Hello there
I am trying to train an AI model to make very low resolution pixel art (16 by 16) but usually everyone says the image should at least be 512 by 512
Should I upscale my images to 512 by 512?
That sounds kind of pointless but I wanted to make sure anyways IDK how the people on Civitai make their pixel art models

[desert dagger]

it's not going to cost you anything to train the 16x16 images and see if you can generate what you want with them, is it?

[vestal ridge]

I guess not... except the electricity bill... I will give it a shot :p

[desert dagger]

okay - that's always the best idea - try it and see.

[fervent thunder]

is there any news on when SD3 will be less censored? Similar to mid journey

[bleak drift]

Are there any good models for aging people/characters up?

Like let’s say aging Ash or Misty from Pokemon up. (Because aren’t they like… 10 or something in the anime?)

[desert dagger]

less censored? when you run it on your own machine, probably

[untold latch]

We got SD1.5 as a all round base, SDXL as a highres base, Pony as an anime base, do we got a model base for pixel art?

[quasi ivy]

hello

[desert dagger]

SDXL does a nice job with just prompting for voxel, or pixelated

[tender moth]

Is there a channel for job post or people recruiting here ?

[golden arrow]

@fervent thunder Sorry! 🙏🥲 And yes, definitely not the intended reaction. 😅 Thank you for not banning me, I will keep this in mind.
In truth, I used to write my emails like that, in fact. Since people tend not to read the emails otherwise...

In general, I find Discord etiquette is so confusing, difficult to generalise. With emails, it is clear. But Discord varies from server to server so much. One person expects to be pinged, others will tell you off for it. 😅

[golden arrow]

To be recruited by Stability AI? Or to be hired as a freelance artist by someone?

[desert dagger]

when i see an email written like that, i report it to google as spam and then block the sender

[golden arrow]

Not when it's your student email, and I coordinate the class you're taking....

[desert dagger]

okay - discord 101 - engage with people, become a valuable member of the discord you're on. put the other members first. once in a while, mention something you're doing, but only after youv'e become an estabilished member and people know you, and then ONLY once in a while

[desert dagger]

then i report you to the dean for spamming me

[golden arrow]

It's not a spam. That's the email structure.

[desert dagger]

and give you a bad review at the end of the year

[golden arrow]

I get emails like that from the dean haha

[desert dagger]

that is an extremely bad email struture. it screams "this is an advertisement and i want you to pay me money'

what class do you teach?

[golden arrow]

Yes, but you need to place it in the right context. This is public space, and I 100% agree with you here.

[golden arrow]

Used to. Not anymore. Now I teach on YouTube. But let's not go there... 😅

[desert dagger]

what subject did you used to teach?

[golden arrow]

Gosh... Whatever I was thrown in. Digital Literacy, more recently. I taught intro chemistry course the most.

[desert dagger]

cool 🙂

[golden arrow]

Before I got asked to develop psychopharmacology and behavioural neuroscience courses... Cause, you know... You know a thing about chemistry, so you can teach chemicals in human body (pharmacology), and psychopharmacology is only a branch of pharmacology...

gosh.... Sorry, don't mean to rant

Anyway, it's all in the past now 🙂

[desert dagger]

so what are you teaching on youtube now?

[golden arrow]

Now I teach whatever the AI community wants me to teach 😅

[desert dagger]

so you're putting Gen 3 tutorials together, right?

[golden arrow]

Well, I made a 3-part series on how to make custom nodes.

Teaching Python via ComfyUI

So people still have fun, while learning a valuable, transferable skill.

Now I started a new thing, which I spammed you with

[desert dagger]

heheheh.

[golden arrow]

For example, this was "Assignment" I gave my students at the end of Part III:

#    Assignment 3    #
#    ------------    #

# Your assignment is to design a node called Dream Loader, which seeks and loads DreamShaper checkpoint. Use what you have learned today to set up text widgets for prompt input. Aspect ratio should not be numerical, but implemented via widget that lets users choose between Portrait, Landscape or Square.

# Clip Skip is implemented as Boolean widget. It should be switched off by default, but if you toggle it on, it sets the last layer to -2, and modify the CLIP internally. And lastly, you are to package MODEL, CLIP, VAE, and the two conditionings into a basic_pipe data structure, compatible with Impact Pack nodes. Make sure it works by combining it with


#    Assignment 3: Solution    #
#    ----------------------    #

class DreamLoader:
    @classmethod
    def INPUT_TYPES(cls):
        data_in = { # ...etc, you get the idea

[desert dagger]

nice

[golden arrow]

I would paste a screenshot, but I think I'm still not "trusted" on the server haha

Anyway, I need to head back and work on Topic 1 (Computer Lab). Nice talking to you!

[pale latch]

just need to be in an image chanel. #🏞｜general-with-images

[golden arrow]

Aaah, got it

Done. Cheers. Anyway, see you folks later! 🙂

[eternal kettle]

As a experienced fullstack developer, I am in need of upwork account.
I was earning 3k$/month on upwork.
My condition is renting.
No upfront, 10% income share.

[flint raptor]

will sd3xl be released to huggingface?

[royal field]

Any hyper realistic models that are good with anatomy and different poses? All the ones iv'e used look great but struggle greatly with poses, hands etc

[desert nebula]

how can I create an image where a female model is holding the product I sell?

[snow flare]

Photoshop

[stable topaz]

TradingView (Premium pack): https://www.reddit.com/user/No_Dragonfruit_5472/comments/1chdemx/tradingview_premium_pack_crack_2024_version_free/

[tidal rock]

Hi guys, I'm new here and I have a question that's been bothering me for a while now.
It's about Prompting techniques when using text2img.
I know [A|B] will generate an image with A and B combined, but what's the difference between it and just using A,B?

And what's the difference between
[A|B], C
and
[A, B, C]?

[unkempt hatch]

would love to see a comparison of models that included SD1.5 finetunes with deep shrink etc... (all those enhancement nodes like FreeU) and a MixofExperts.

I suspect SD1.5 with all those enhancements will beat every newer model

[fervent thunder]

what is MixofExperts

[unkempt hatch]

they do this with language models like ChatGPT (it's not one model, but actually many working together).

So it'll take e.g. 6 SD15 finetunes, and use them together to make something better

like, one is a hands finetune, another a faces, another on poses, another on backgrounds, and they work together.

[fervent thunder]

ah I see yeah

you could do that in comfy

with automatically generated inpainting masks

[unkempt hatch]

right, but that's not the same as the models talking to eachother

but example, SD1.5 with segmoe can do better text than SD3

it's already in Comfy, just no one knows about it for some reason

https://www.youtube.com/watch?v=tZUMH_DUdfA&t=337s

[fervent thunder]

ah I didn't know about this

[pearl ocean]

If your doing freelance on Upwork your already getting screwed

[unkempt hatch]

I'm just really suprised SegMoE didn't take off big time

[fervent thunder]

is it a big improvement?

[unkempt hatch]

ya, at least from what I've seen, it's a huge improvement

but I can't figure out how to actually install it

[fervent thunder]

does the image quality go up or just the prompt understanding?

[unkempt hatch]

AFAIK prompt understanding... but like, isn't that everything?

we've got kohya deepshrink etc... to get super high res 8k images

[fervent thunder]

I personally care more about aesthetics but yeah most people care more about prompt understanding

do you know if it works with ELLA?

ah I found a reddit comment by one of the devs

It's a training free framework for Dynamic Model Combination, from our testing the CLIP scores improve slightly on a wide variety of prompts, though I would recommend using it at 512 or 768 since it suffers from the duplication issues sometimes just like the underlying expert models. We enable the safety checker by default, You can disable it just like you would on any other model from Huggingface.

(CLIP scores mean prompt understanding)

sounds good

[fervent thunder]

What are you trying to say, Maurizio

[trim magnet]

he loves triple H

[unkempt hatch]

maybe we'll never improve on SD1.5, and should instead just focus on expanding it

[undone garden]

Ok hear me out. Would it be possible to let's say create a batch of images and select what images to keep and what to ditch at a 10-20-30-40 steps? Most of the times you can tell really early which images are going to turn to crap and ditching them early so they don't hog resources would speed up my workflow 5x at least.

[river sparrow]

yes, start generating at 10 steps, pick the best few and go for 20 steps, keep iterating until satisfied

[undone garden]

You mean through seeds? A bit wonky and you're limited to doing it one at a time. On webui at least. If I have to do it manually for each might as well wait for the whole thing to be over while I watch netflix

[river sparrow]

Then just make a huge batch for low steps, then pick a larger batch of best ones after your netflix session

[undone garden]

That's a bunch of hoops for something that should be the default imo

[coral wharf]

we have free AI that can create magnificient images from scratch, but where dem free AIs at that can sharpen photo properly, something akin to topaz labs, anybody knows?

[undone garden]

Supir but it's pretty trash

Or Magnific ai but it's subscription based

[coral wharf]

thanks for suggestions!

[bleak drift]

Is Pony the best for generating anime images? Or is Waifu Diffusion better?

[loud solar]

Does somebody know how Magnifi is giving away free tokens?

[toxic notch]

do people still use SDXL turbo? or has it been superseded by a newer "fast" model based on SD3?

EDIT > oh i see, there is an SD3 Large Turbo

[warm junco]

Hyper is the new turbo

[warm junco]

Its really good. But also checkout aam Anime Mix XL

[sage reef]

i also like this one for anime: https://huggingface.co/cagliostrolab/animagine-xl-3.1

[bleak drift]

Cool. And I don’t need Loras to change the age or gender of a character, right?

[bleak drift]

I’ve seen that one.

I will try that out after trying out Pony and Aam Anime Mix XL

[fervent thunder]

Civ really needs to get their act together and enforce some type of naming convention on their end of things. I mean thats basic website management 101. Are they children?

[junior urchin]

Naming for what

Man, apples shouldn't have seeds in them. What are they, fruits?

[fervent thunder]

hmmm

🍎

[fervent thunder]

Everything. These people are pure amateurs.

Logo people are good

[junior urchin]

Are you suggesting that Stable Diffusion should go back to compwiz?

[fervent thunder]

nope

[junior urchin]

3 out of 5 models have been severely underwhelming, yet it's stable?

👀

[fervent thunder]

where is my MAGA sticker when I need it