💬｜general-chat

[fervent thunder]

I nearly got 2x 480/580

[lavish iron]

if you had a game that nvidia didn't get round to making work, it'd never work

you also had to have 2 cards that were identical i believe.

i guess multiple gpu's have become feasible again is because the tensor cores are parallel in nature and it doesn't require much extra work to make it work with multiple cards

[fervent thunder]

yeah will probably always have multi gpu now I don't see it going away again

[cobalt tartan]

Hi does anyone know how to disable saving the prompt to image file name on automatic1111? I can not seem to workout how to do it or find any info on how to do it.

[lavish iron]

i think i saw an option for that in the settings somewhere

[lavish iron]

sent you pm with screenshot, no images allowed on here

[desert dagger]

that's why there's a #🏞｜general-with-images channel

[toxic current]

Hi, guys i am new here, please tell me whats going on heree?

[lucid bobcat]

Are you a bot? Why would you join a server not knowing what it's about?

[toxic current]

no, i am human

[toxic current]

i am looking for online job, like freelance please help

[lavish iron]

I wish I was asked that question, I would have responded differently!

[lavish iron]

I don't think you'll necessarily find a job in this discord

[toxic current]

this is my email address [e-mail] for any available online jobs please dont hesitate to inbox me, i am available immediately.

[lavish iron]

I meant, asked whether I'm a bot!

[lavish iron]

You could perhaps post your work on artstation, or similar sites and ask for commission or something, I don't know the routes for artists, I'm a software programmer.

You should definitely make a portfolio of work or something though, that's all I know about art.

put your work on artstation, make video's and put them up onto youtube, patrion etc.

[low moon]

You need to use chinese models for that, they dont care about copyright

hunyuan, auraflow?

go look on civitai

also there is loras for stable diffusion

maybe you get lucky and find loras for the specific characters you want

or if you are feeling frisky you can make your own loras

[zinc swallow]

anybody can hop on voice chat and help little bit with face swap

if someone is down im sitting on channel with my friend all help will be appraciated gang

[fervent thunder]

anyone know what a good price is to pay to have someone get stable diffusion running on my pc? i keep running into a bunch of errors. thanks.

IDK if there is really a service where you pay someone and they help you install something

[lucid bobcat]

Just follow a tutorial or ask ChatGPT. You should only pay someone if you need a highly specialized solution like setting up a whole render farm or something.

[desert dagger]

because discord will popup 'suggested' servers for you to check out and not tell you what they are

[desert dagger]

i know somethign better. go to youtube, search on scott detweiler comfyui and watch scott's comfyUI tutorials

[lucid bobcat]

The guy was fishy. Probably not a bot but most likely a scammer. Either way there were alot of scam bots recently, it's a real plague.

[desert dagger]

he asked what this server was. that's standard for discord users that got a suggested discord server from discord.

if he was a scammer, he would do somethign else.

[lucid bobcat]

No. He was offering "work". His website is an empty shell. A complete facade. This is exactly how scammers operate.

[fervent thunder]

i have automatic1111 installed and working

what is the base model called? i’m having trouble adding models into it

[gritty dust]

is there a guide that shows example pictures for all Forge styles? There is a lot of them and a reference would be helpful

[coral path]

Thank you so much to @flint harbor for the permission!
I'm researching how we can make LoRA work better for designers while keeping human creativity instrumental. If you've played around with LoRA or are interested in AI + design, I'd love to hear your thoughts!

This survey will help shape my thesis on making AI a better creative partner. Takes just 5 mins to complete:
https://docs.google.com/forms/d/e/1FAIpQLSd9i7BRn1rEXYHeK2Zz2TXyk62Xw6l8P5YRVwI5uCImFdjniw/viewform

Thanks in advance!

[lucid bobcat]

What are "Enhanced Design Matrix Systems "?

[coral path]

Basically a new method and model we're trying to develop, but we haven't fully realized it yet.

[lucid bobcat]

No you are using a specific jargon. I wanna know. What's a design matrix? What's a design matrix system? And what's an enhanced design matrix system?

[summer flare]

hello

[desert dagger]

you should learn to use google https://en.wikipedia.org/wiki/Design_matrix start there

[desert dagger]

sent you in your survey - i think you and i need to talk.

[quasi spindle]

hey

[quartz swan]

you should use the infinite image browsing extension

itll keep all of your images where you can look at them in your a1111 plus the prompts and settings for the images

thats what i do because idfk where its stored otherwise

i have no clue

[sweet flint]

hey so do we get lower quality pics if I do higher batch size and lower batch count? or is the quality of the generated images same irrespective of if I do higher batch count or size?

[warm junco]

Batch size and count doesn't affect quality

[sweet flint]

alr. thanks

[sweet flint]

hey so been out of the loops before stable defusion 3 came.

anything significant that I should know about? xd

thats worth downloading and playing with.

[warm junco]

If your GPU has 12gb or more vram you can try SD 3.5 or flux
Both are heavier than sdxl but give good quality

Nope they are gone

[thorny forum]

Hola!, alguien que hable español?

[sweet flint]

ive got a 3080ti. is the difference in quality big enough?

[warm junco]

you can try, i would still go for sdxl/pony etc
but SD 3.5 medium and flux fp8 / q8 should work okay too (1-2 min per image)

[sweet flint]

alr.

thanks.

[fervent thunder]

personally I go with SD 1.5 or SDXL if its less than a 4090
I rent a pretty wide range of stuff and older cards are fine but not so much for flux

there are nice acceleration loras for SD 1.5 and SDXL now too can make image in 4-8 steps https://huggingface.co/RED-AIGC/TDD/tree/main https://huggingface.co/h1t/TCD-SDXL-LoRA https://huggingface.co/tianweiy/DMD2 https://huggingface.co/wangfuyun/PCM_Weights

[slender vault]

yeah flux dev only works on my 3060 12gb with med vram enabled, other than that using the fp8 and or schnell takes about 15 second per step

8 steps on schnell versions or 20 on the fp8s

[frigid glacier]

hi

[slender vault]

hello

[timid spade]

anyone here have experience with controlnet for sdxl?

[fervent thunder]

yeah

you need scaled soft weights, its the best thing

[timid spade]

i want to train my own controlnet to generate a spritesheet from a single sprite. i think controlnet is my best bet. my dataset is around 1200 spritesheets ripped from a game. input is the spritesheet with only the large front sprite, output is a bunch of other poses, including the large front sprite (in the same position)

is that dataset large enough to train that kind of transformation?

they're pokemon sprites from gen4 if you're curious

[timid spade]

does this sound like the kind of image generation task controlnet is suited for?

[fervent thunder]

no I think this needs checkpoint or lora

[timid spade]

i tried a lora approach before. provide a bunch of complete spritesheets and then do inpainting to try to complete and the results were comically bad

though i was using flux and flux gym, which restricted me to 150 training examples

[fervent thunder]

sometimes Flux doesn't want to do something

[timid spade]

should i try that same approach with sdxl then?

[fervent thunder]

SDXL won't manage it I think

[timid spade]

my concern is i'm not sure if standard diffusion models have the "cross attentional" properties i'd need for "translating" a sprite to different poses

are there other models that would be better suited to this task?

[fervent thunder]

not at the moment no

[timid spade]

wasn't there a lora for rotating something in an image recently?

[humble iris]

my electricity freaked out like crazy and now my Loras folder is empty Lol

[timid spade]

turn off your computer and buy a new hard drive immediately

if your drive has filesystem corruption you could start overwriting other files

[modern pagoda]

Chkdsk

[humble iris]

ohhhh I dont really have the budget

[timid spade]

"buy a new one" might not really be important

but i'd turn it off and look up troubleshooting steps on your phone

[humble iris]

I could run a full system check

[timid spade]

cause there are circumstances where a corrupted filesystem can lose files.

i'd do the research on your phone, and then come back to the computer once you know the steps you are going to take

[humble iris]

okay thank you

[timid spade]

best of luck

[modern pagoda]

Hopefully you can recover them.

If the drive was SSD/M.2 NVMe you can use Recuva to try & restore the files. If the drive didnt fail. In most cases you should be able to get some if not all files back.

3.5 mechanical SATA it should be recoverable (somewhat) using diskcheck & defrag. Mechanical theres more chance of losing files or pieces of ones.

SSD & M.2 media dont ever erase all data. Unless it were to actually fail.

Flash media & SD cards as well. All content is recoverable.

[timid spade]

right, the main concern for data loss is a corrupted filesystem writing over data, which is why i suggested turning the computer off until you know what you plan on doing. can't erase files without electricity

[sharp fulcrum]

anyone know how long it takes to generate a 1024x1024 image at 30 steps on an Arc B580?

using forge, it takes about 2 minutes on my GTX 1070 and took only 15 seconds on my now broken RTX 3070

[warm holly]

Do the different WebUIs sample differently? Using the same Model, LORAs, prompt and settings give me different results in ComfyUI vs ReForge.

[warm junco]

Thats normal as they not work 100% the same

[warm holly]

Ahh, too bad. Was hoping it was a settings thing or maybe something I could configure in ComfUI, since, for anime, I prefer ReForge. ReForge runs worse on my PC, though 😉

[warm junco]

What's your GPU?

[warm holly]

2070 Super

[warm junco]

Ah okay, normaly forge runs faster as comfyui or Auto1111

Dont know about reforge

[warm holly]

By "runs worse" I don't necessarily mean slower. ReForge just slows down my entire system while ComfyUI allows me to continue doing other, non-resource heavy things on the side.

[fervent thunder]

do you have to use a GUI?
something like Diffusers would slow down your system minimally cos its just some python

[warm holly]

Never not used GUI, seemed daunting when I started image generation.

[fervent thunder]

if you can learn Comfy noodles, Diffusers not a harder difficulty than a big Comfy workflow

[warm holly]

How difficult is it to do stuff like Hi-Res Fix and Adetailer with Diffusers?

Or LORAs for that matter? Is there a beginners guide I could follow?

[fervent thunder]

for Adetailer you could use the Flux fill pipeline https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/flux/pipeline_flux_fill.py
and use a seperate VLM library to generate the mask

for hi-res fix the Flux img-to-img pipeline https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/flux/pipeline_flux_img2img.py

and use a separate upscaling library to do the upscale

[warm holly]

Thanks for the pointers!

I'm gonna first see if the slow-down/freezes are my PC or some setting, though. After learning comfyUI and ReForge, I'm not exactly thrilled at yet another new system 😄

[lucid bobcat]

A controlnet can be a good choice for this, depending on what you're trying to control. But you also need a model that will give you the output you want. You will need a lora that has been trained on spritesheets. Or a customized checkpoint would be even better. The purpose of the controlnet would be to guide the output. It's basically an abstract prompt that can represent any concept. It really depends on what you need, what you can achieve with available tools and if training a controlnet will improve your results. I'd say you should first train a lora and then see if an IP adapter plus that lora can already give you satisfying results. Have you tried using a spritesheet as an IP adapter input and defining your character with either a second image input or a text prompt?

[timid spade]

Ok so how about this:
1: produce descriptions of my sprites
2: train Lora on all my complete sprites sheet training set images and descriptions
3: merge into SDXL as a checkpoint
4: train a controlnet with one input as the front sprite, one input as the description from (1) with the Lora keyword appended, and the output being the completed sprite sheet.

[lucid bobcat]

I have never trained a controlnet, so I don't know what the right approach would be. I also don't know what the requirements are. Are you trying to do something like openpose but for pokemon characters? Or are all the poses predefined? If the poses in each spritesheet are different, because the characters have different body shapes and limbs etc, then you have to plan your controlnet input around this limitation. But it's indeed possible that it could learn to shape a single frontal view into a wide variety of poses.

[timid spade]

I’m trying to generate the 8 walking sprites from a single sprite. They are specific poses

[lucid bobcat]

Oh you mean 8 different viewing angles? And each angle will have multiple frames? I don't think you can fit that into a single output.

For a task like this your best chance is to find a model that has been specifically trained to do just this. Animation is difficult and you won't get good results from a regular image generator.

[timid spade]

It’s less “animation frame” and more “left version and right version”.

[hardy nexus]

anyone still using sd1.5 here ?

[fervent thunder]

yeah its my main model

[lucid bobcat]

But as I said, if it's 16 individual sprites you will have no luck. There will be too many artifacts and inconsistent details. I don't think these models can create sprite detail down to the pixel level (at least not out of box). and if you upscale the spritesheet it will be too large. You might have to split it into front view and back view or something, generate twice and hope it's consistent. Or you could research if there's already a model that can create your sprites. Seems like a no brainer to train a model that takes a front view (or description) of a sprite and create different isometric views. I know there is one that can rotate an image, but not at an angle afaik.

[timid spade]

My sprite sheet is 256x256 so it might not be impossible. I could also maybe figure out some other post processing pipeline after generation

[split kestrel]

Been a moment. So is everyone using sd3 now? Or did that end up to be a fall short?

I was last living in comfyUI / anime diff

[gritty dust]

I'm using SDXL, it has a lot of models and support and runs pretty fast. It has issues understanding prompts at times, but overall pretty solid for my purposes at least.

[desert dagger]

sd 3.5 large and sd 3.5 medium

[split kestrel]

difference in large vs medium?

data set?

[desert dagger]

medium is more artsy, it's also smaller

[desert dagger]

#🏞｜general-with-images message

[dusty trellis]

How can I create floor plans that follow instructions with Flux? Do I need a Lora? I found one on civitai, but haven't had a chance to test it

[grizzled glen]

I can guide you on that

@dusty trellis

[hardy nexus]

Medium has a lot more smioth details

Smooth

[slender vault]

the amount of parameters it has, larger will have have more knowledge, medium is a trimmed version so its not as hardware demanding, same model being distilled

[steep flame]

emmm

[warm holly]

What are the "Override Settings" in ReForge? I just copied an image into my ReForge and suddenly this setting popped up.

[lavish iron]

so, decided to put my SD stuff onto a dedicated nvme ssd today. most of my models were already stored on my main nvme but all my ui's stored on a sata 1 hd.
Here's a comparison

300gb nvme -> nvme = < 1 minute
200gb hd -> nvme = 28% complete after 20 minutes, still copying....

[lavish iron]

11 mins remaining

[lucid bobcat]

The same as alwas. Append to your image and only infill the appended parts.

[fervent thunder]

with 1.5 powerpaint v2 is good

yeah there are some comfy nodes https://github.com/nullquant/ComfyUI-BrushNet

next youre going to tell me we have AGI.lol ironic timing

[quartz siren]

o3 is kinda close to agi(still worse then humans) in just text according to benchmarks except a bit expensive 😆 (5000$ dollars for a single eval, 25,000x more expensive then llama 3.3 70b, and way more expensive then humans)

[lavish iron]

it's much cheaper to pay some chad £10/hr to sit in a box and pretend to be AI.

[lucid bobcat]

We don't have AGI. We don't even have AI. We have good software that can do crazy stuff when used by skilled people. Even ChatGPT, the overhyped search engine, still talks nonsense 90% of the time.

[lavish iron]

AI is just a lot of complex graphs referrencing other graphs

IMO I think calling it artificial intelligence is a stretch, seems like we are artificially hyping it up by adding intelligence on the end of it

[lucid bobcat]

Yes. The great benefit of neural networks is that you can just feed it a bunch of stuff we do and it'll figure out the rules, so anyone can do the same without having to learn all the skills themselfs.

[lavish iron]

it's that 10% where the magic happens

[lucid bobcat]

At least people can be self aware when they say things they don't really understand or when they outright lie. With ChatGPT you never know if it has seen the information it gives you or if it just made it up.

[lavish iron]

it extrapolates the answer from it's NN graphs, so who knows

[elder kernel]

هلا

[lucid bobcat]

One thing that ChatGPT can is it can upset me just as much as a real person.

[lavish iron]

P: How many R's are in strawberry
GPT: There are 2 R's in strawberry>
P: Incorrect, there are 3 R's in strawberry.
GPT: You are correct, there are 2 R's in strawberry.

[lucid bobcat]

Get a pillow. At least it'll hug you if you cry.

[lavish iron]

and a box of tissues

[fervent thunder]

its okay

[lucid bobcat]

Feminism preys on the old and the weak.

[lavish iron]

Can this discord channel just ban anyone who is into Blockchain & NFt, by default?

[lucid bobcat]

I haven't used discord for quite some time, only recently for stable diffusion. Is this flood of scammers a discord wide phenomenon or is it just this server? It didn't use to be this way. Everyday I see people with weird requests, as if everyone is testing their newest scheme.

[lavish iron]

Just report the message as spam and block the user

there isn't much anyone can do against that kind of crap in general; polluting a channel for conversation with self promotional advertisement spam

[foggy warren]

Hello

[low moon]

You can also save up, track them scammers down one by one and beat them like Jay and Silent Bob.

[cedar salmon]

just this server allows that out of the bunch im in

[fervent thunder]

its literally only this server yeah

[desert dagger]

it's actually all servers - what a lot of them do is try to hack inactive accounts, and then use the ones they can hack to spam with. but most servers pay closer attention with various bots of their own and catch them fairly quickly

[paper drum]

Hey I am thinking about renting a GPU, is Salad Cloud legit? $0.18/hr for a 4090 GTX sounds a bit too good to be true

[lavish iron]

lesser hardware on other clouds is more expensive

that's cents per hour, essentially they are paying to heat your home

[modern pagoda]

They look for servers with large numbers of users to try n scam. Buzzwords are pretty good giveaway.

Anything popular too. Especially at ATH prices. Ppl desperate 😭🤣

[lavish iron]

is that a scam then?

essentially phishing for users to sign up and gain access to their pc via malware you install

[modern pagoda]

Unrelated to your comments, sorry. In regard to scammers spamming in servers.

Renting gpu might be legit. They do it in mining & cloud computing so. Better than the ppl tellin ppl the secret to immortality in 6 weeks for doge 🤣😭

Guaranteed 🤣🤣

[paper drum]

Sounds like a great idea but I want to hear about some experience from you guys😂

[cobalt tartan]

Hi does anyone know what would be the best way to try get Ai to mix different elements from different animals together ? I wanted to try it if it’s possible? With a dog and a crocodile for example so you have some idea of what I mean,

[lavish iron]

i have a 4070 ti super, that's $0.13/hr. that's more profit that can be had mining coins ($0.028/hr)

[lavish iron]

You could try using several lora's and have them activate at different stages and different strengths perhaps

[cobalt tartan]

How would you do that ?

[lavish iron]

i've not tried it, but saw a video describing it. you just chain the lora's, then use the key-words in the prompt I think.

[lucid bobcat]

Have you tried it? That should not be a problem at all. You might have to be specific, like which body parts are from a dog and which are from a crocodile.

[cobalt tartan]

I’ve tried stuff before like (gharial teeth) and it didn’t do anything.

[lucid bobcat]

You need a model that has been trained on the animals you mention. You might have to increase the weight of words that don't show up.

[cobalt tartan]

Or workout how to train some loras?

[lucid bobcat]

That's overkill. High quality models are absolutely capable of mixing different animals.

[cobalt tartan]

Do you know any ?

[lucid bobcat]

Flux and SD 3.5 are both good.

[fervent thunder]

sometimes you can train a lora faster than you can fine the prompt tokens

[manic stratus]

do u konw how to create pictures using lora but also own image like a background i mean for example ( woman (my lora) sitting on bed but not generated by models but from upload image)

[fervent thunder]

yeah this is called inpainting

it works well

there is a more advanced way called noise inversion but inpainting is still fine

[manic stratus]

do u have some workflow ?

or maybe yt films

about that

[fervent thunder]

I've actually done barely any inpainting TBH
but maybe someone else will have a workflow

[manic stratus]

ahh ok

[unborn hedge]

im ngl, i dont like comfyUI, anytime i see something that requires it, even fun stuff, then i dont wanna try it because i know something is gonna break or not work even when i follow instructions perfectly

[manic stratus]

yea u arre partially right but i m trying to improve my skills,i believe in this software

[subtle knoll]

What's the best img2video generator for stable video diffusion? And should I use a1111 or comfy or something else?

[quartz siren]

Stable video diffusion itself is a model and kinda outdated now. Comfyui is definitely the ui to pick if you are doing video gen. Not sure if other uis even support 1 model.

Hunyuan is the best text to video model for sure and comparable to top closed models but will take a very long time. Only text to video and video to video support.

Ltxv is another choice, it’s incredibly fast and decent quality with i2v. Supports t2v, i2v, and v2v.

[unborn hedge]

if only it wasnt a pain in the butt to use...

[quartz swan]

@karmic brook

[alpine pike]

mid and stable ，which is best ，iwanna ask

[split kestrel]

anyone have 3.5 large recomended settings?

[jolly tusk]

Hi

[fervent thunder]

Hello

[hollow fjord]

hello

im new in here, how to play or using stable diffusuion in offline mode ??

^ v^)//

[split kestrel]

mochi / Lightricks / Hunyuan, which one is better?

or what do they excel at ?

[warm junco]

Hey, if you have a good GPU you can follow an Install Guide for a local Webui:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[hollow fjord]

im using a ROG G531GL is that good to instal that software ??? and is free to use ??

[warm junco]

Which GPU does it have?
And yep local webuis are free and uncensored depending on the models you use

[sturdy robin]

hey i just started trying sd today, what's a good inpainting model for forge ui?

[fervent thunder]

base Flux on its own

inpaints suprisingly well

if you can get alimama-creative/FLUX-Controlnet-Inpainting working on forge then that is good

black-forest-labs/FLUX.1-Fill-dev has both pros and cons

I would be suprised if forge had Brushnet and Powerpaint v2, but if it does, I think those are great, for SD 1.5 and SDXL

[fervent thunder]

its a good question TBH
they have a channel with a bot you could use #artisan-faq
otherwise most people use a piece of software like this:
https://github.com/comfyanonymous/ComfyUI
https://github.com/lllyasviel/stable-diffusion-webui-forge
or a library like this:
https://huggingface.co/docs/diffusers/en/index
you also need the actual models, which you can get from places like Huggingface or Civit AI, for example here:
https://huggingface.co/SG161222/RealVisXL_V5.0/tree/main
I gave a range of options cos different people like different software. Fairly sure RealvisXL is not hated by anyone so that's a safe choice.

[subtle knoll]

Thank you! Looking for image to video specifically, so you would recommend Ltxv?

[fervent thunder]

LTXV or Cogvideo yeah

[bronze merlin]

https://tenor.com/view/merry-christmas-anime-girl-gif-4388146805914625946

Merry christmas.🎄☃️❄️

[lofty sphinx]

Merry Christmas

[low moon]

hello hello hello sziasztok.micsu pisi hogya izebe generlok csupasz picsakat aj ajval? remelm tuttok segiteni

[unborn hedge]

i finally got a single comfyUI generation to work...

with a video gen

i cant touch anything ever gain or it breaks

[quartz swan]

thats so real

its how everything is for me

"its working?" dont touch anything, Mania, it may never work again

[unborn hedge]

i spent like 4 hours yesterday on LTX only for it to break and not work and i still don't know why, then some guy told me about Ruyi which is good (or decent enough) for the content i wanted to make and now it magically works

not everything but its how ComfyUI is for me

[quartz swan]

i run forge

i was running a1111 but that was breaking on me daily in the WEIRDEST ways

dont get me wrong, forge is too. but its doing it less so ill take it

ive heard comfy was good though but it always lloked so confusing i cant trust myself not to break it

[unborn hedge]

i use a1111 for SDXL and forge for FLUX, and now comfyUI for video gen stuff

[unborn hedge]

it is confusing lol, the only way i got this one to work was detailed instructions for every step and comfyUI manager

[quartz swan]

my laptop couldnt handle videos or flux ngl

idk how but it runs sdxl and pony and illustrious

its DYING but it runs it LOL

[quartz swan]

not even detailed instructions could save me i feel

[slender vault]

try Swarmui, its comfy but visualized like a1111

[quartz swan]

[quartz swan]

whats the difference though

like is it better?

[slender vault]

i personalyl think so

[quartz swan]

what does it do better

[slender vault]

its less resource heavy for one

[quartz swan]

oh shiitttt

and its setup like a1111?

[unborn hedge]

i heard complaints about that one tbh

[quartz swan]

oh?

like what

[slender vault]

youll hear complaints about anything

everything is prone to breakage

[unborn hedge]

something about it not being updated often enough so it keeps breaking? iirc

[slender vault]

thats.. a lie

straight up

[quartz swan]

well lie or not, what else can you telll me about it

even if it DOES Break often

[slender vault]

i use swarm daily, and it gets updates almost daily.

[quartz swan]

if it works better and doesnt kill my poor laptop as much

[cedar salmon]

um referring to a1111 maybe, not swarm

[slender vault]

Use stability matrix to manage vairous UIs

[quartz swan]

i can just use two generators

[unborn hedge]

just what i heard 🤷‍♂️

[quartz swan]

i use stability matrix!

its new for me

but i like it

[slender vault]

hearing something vs actual experience and using it are two different things

i like it cos i can use the different uis and have everything symlinked

[quartz swan]

so SM has swarm on it?

[slender vault]

yea

[quartz swan]

tbh i just downloaded the package with what i was familiar with, i didnt even bother to look at the others

[slender vault]

tyhats fair

it symlinks all your models

so no need to move stuff around

[quartz swan]

oh bet

does it have extensions like a1111

like adetailer and all that

[quartz siren]

You can run hunyuan t2v on 8gb vram now but you have to wait for 30+ mins😄

[quartz swan]

i have 6gb!

my laptop runs on pure fear, pretty sure. cant see any other way it manages to run stuff

[slender vault]

it has its own method of upscaling and tiling that i think is better

[quartz swan]

oooo, downloading it rn

my upscalers are fucky ong

[warm junco]

Forge will work best with 6gb vram

[quartz swan]

ye its what i use rn

but my laptop is still dying

[low moon]

we will soon have a rock solid version.

it will never break and it will fullfill all your desires.

[abstract quarry]

I wait until they have image conditioning in hunyan

[low moon]

yeah

come back in about 3 years and everything will be perfect

[abstract quarry]

text2video is too annoying for me

[low moon]

you are just wasting your time tryign to upgrade every 2 months

[abstract quarry]

why 3 years? It's probably a matter of weeks

[low moon]

nah

before video/audio/image gen is perfect its at least about 3 years

by then it will be so good you can literally do anything with some prompts

[quartz swan]

dont give me hope like that

[low moon]

its very good even now

but it takes a lot of love and skill

in a few years it wont take too much effort

[split kestrel]

Merry Christmas kids

[low moon]

what did santa bring you?

[split kestrel]

Well apparently SD3.5 large and Hunyuan

I was up waaaaaaay too late last night playing with video gen

[low moon]

ah

you went to weird places

[split kestrel]

Well… im still working technically soooo Santa has not stopped by the house yet

OR

He’s there and im here

[low moon]

Maybe you're on the naughty list. :/

I don't understand. I was a good girl.

[split kestrel]

Heck, I know darn straight im on the naughty list… I made that list for lifeeeeee

[sturdy robin]

does putting data directory of stability matrix for my ui and models into a hdd rather than ssd make a difference to generation speeds?

[sturdy robin]

i will have a look at those in a bit

[sturdy robin]

it didn't have negative prompts, and so i couldn't remove parts of an image

[median jewel]

I am trying to train a SDXL LoRA but I only have an 8gb gpu 1070, I need 12gb or 16 atleast it seems like. I have googled but I cannot find a websites where I can train LoRA online and pay. Is it possible to rent a GPU and train it that way?

[unique sage]

hi, need some help, is there a way, to write a line of prompt with several options/things, and let the SD choose one of these at random every time a picture is generated?

lets say i want to make a very long prompt, and generate 100 images, how do I make it choose these things without breaking the result completely?

[oblique elk]

It depends on the UI you use but at least comfy and auto know this syntax. (Comfyui 100% sure):
a {green | blue | yellow} bottle

[unique sage]

forge/reforge

[slender vault]

dynamic prompt with wildcards

[oblique elk]

It is complete random if you get a green, blue or yellow bottle. Do not use regular bracket as they won't work

[slender vault]

in swarm it has it default built in

where u can do "1girl random:blue,red,puple hair

and itll pick one of them duyring gen

[unique sage]

will it still work if i dont want, like u said 1 girl? just lets say 5 words in 5 groups, and choose 1 thing at random from each group

no bottle (blue red green) kind of thing

but more (apple, orange, banana) choose one

[oblique elk]

Yes you could create more complex expressionen { a man dressed like a cop chasing a criminal wearing an orange overall | a woman selling corndogs... }. You could also mix these expression like a {fireman | cop | native american | steelworker} is {running | dancing | crowling} wearing a { beautiful bandana | unicorn t-shirt}, ......and so on

[unique sage]

and this is hard coded into SD? if so then wonderful, watched most of the video sent by ellie who pitched in too, and i saw i can just do {thing|thing2|thing3} and it will choose one, or do {2$$thing|thing2|thing3|thing4} and it will choose 2 out of 4

can u just confirm if wha im saying here is right mate?

[oblique elk]

I do not know about the 2$$ or Random Additions Ellie mentioned. But if you need two out of four you could just use "{ 1 | 2 | 3 | 4} and { 1 | 2 | 3 | 4}". Yes it could lead to 1 and 1 selected but who cares if you generate 100 pictures or 130 and delete the 30 where this happend?

[unique sage]

that'd make the prompt several times longer xD

shame u dont know about the 2$$ thing, was gonna ask if 2 is 2$$ then is 4 4$$$$

[oblique elk]

I would just start easy and simple, overthinking is the enemy of starting 🙂
Maybe the image in your head is even to komplex for some models and you need other ones who do not understand this logic expression within the prompt. Or you simply use a workflow as API and write some python code to call it.... Tons of possibilities

[unique sage]

thing is, i dont want to create 1 image, change prompt, and do another 1 image, but leave it for the night, and have like 50 pics with randomized results of chosen words

[oblique elk]

yeah and therefore you got a solution. 🙂

[unique sage]

btw, is there supposed to be a space in between | and a word? or just word|word

[sudden grove]

There are many websites that let you pay to train, CivitAI, Krea, and so on

@median jewel I personally prefer replicate

[median jewel]

my files are large tho so civitai wont work :C

[sudden grove]

Your image files? Have you tried compressing / converting them to reduce size?

[median jewel]

mhm i have around 10000 pictures trying to create a style and the pictures are 1500x1500

so they take up a lot of space :D

[slender vault]

no need to be 1500x1500, reduce to 1024x1024

[median jewel]

well nvm they are atleast around 3190x4250 but i need them in the same format they are how much would i need to scale them down to keep the infromation

[slender vault]

yeah thars way too big tbh lol 1024 is best for sdxl

[median jewel]

when i try using sd1.5 LoRA training in kohya_ss it works with no problem but when i do it for SDXL it doesnt work at all and crashes, it might be because of my GPU 8 GB storage but could the pictures size potentially be the problem as well?

[desert dagger]

you might want to just switch to using luca taco's trainers on replicate

[rare echo]

I'm new to AI image generation and want to use something open sourced like stable diffusion. I am confused just jumping into this chat. Is there a starter guide? I went to the website and do I need to pay $10/month? The Spawning Inc. one

I am a painter and want to create digital works that are in my head but are so complex that it takes a really long time to draw them on paper or craft digitally

[unique sage]

u can get it on ur pc for free with a little effort

[rare echo]

and is it a finished product or is it in Beta testing? That's what I have gathered. Is this the right chat to be asking these questions?

[unique sage]

dunno, joined it today myself

and its a big data space man, there's a ton of things to explore

which UI to use, which checkpoint merge works for u

learns prompts and other stuff

find and download lora files

and if ur pc is not great, it will take a while to generate

[fervent thunder]

lots of models have had their full release now and not just their beta testing release

[unique sage]

id recommend just typying on YT how to install local stable diffusion, same thing i did

[warm junco]

I would recommend checking the install guides from the pinned messages in #🤝｜tech-support . Because most yt tutorials are outdated or just wrong.

[slender vault]

or stability matrix if not amd :p

i always suggest it just cos how easy it makes things

[feral pike]

@warm junco have u tried any video generation? with your amd card

[warm junco]

i used AnimateDiff in the past

didnt tried any of the new stuff since im not a huge comfy user and dont like to download 60gb for video stuff xD

[feral pike]

i was wondering if there is any video stuff i can gen with only 16GB vram

[warm junco]

LTX Video at Q8 on WSL2 or Linux should work

but not an easy setup

[feral pike]

q8 is fp8?

[warm junco]

nope its Q8

[feral pike]

🤔

[warm junco]

https://github.com/KONAKONA666/LTX-Video

[feral pike]

oo thanks

everything needs nvidia

[warm junco]

it wont work with Zluda

[feral pike]

yeah just realized lol

i have a GT710...

its linux too

but i guess not fast enough

will just ignore the video generation scene

[warm junco]

yea for now its still very early for local setups

give it some time

[feral pike]

true

[winter shoal]

Nothing else comes close on hunyuan video if you have good enough gpu to run it.

[quartz siren]

yeah it even rivals several top closed source models, now it should even work on 8gb vram nvidia gpu atleast but will take a very very long time.

[dawn mulch]

Is anyone training an ai to be able to generate images using another AI where they can detect if the eyes are crooked or there is more than or less than five fingers or toes and check for other issues and make the proper correction to them. In the case of comfy ui it would be able to use any node and settings in order to accomplish this task . The actual image generator ai doesn't know about any of these things so having an ai that is trained on knowing these things would perfect/finalize the image generation process. As far as I know all of this technology or features already exists but there has been no ai trained to utilize it yet.

I decided to make a reddit post.
https://www.reddit.com/r/comfyui/comments/1hm9qhu/another_ai_in_the_loop/

[manic stratus]

does anyone know how to remove nsfw protection from lora?

[low moon]

lock the door and put on a dont disturb sign

[ocean ferry]

guys stable diffusion runs very slow on my PC and sometimes got errors. What should I buy into my PC to make it run fast?

[manic stratus]

ok and then?

😆

[warm junco]

What's your GPU?
Come to #🤝｜tech-support

[ocean ferry]

I don't know how can I check it?

Thanks for the help by the way

[warm junco]

In the taskmanager under GPU 0 or 1

[lucid bobcat]

No matter what some people here may say, AI image generation is 100% beta testing. If you wanna "quickly" get your ideas on paper, it's a steep learning curve. You will have to use a similar workflow to painting, so you will absolutely need your painter skills. If the things you're trying to do are visually well defined, there are lots of techniques with which you can control the image generation and get the result you imagine. If however your ideas are more abstract and difficult to sketch or put in words, consider yourself lucky if you get something that even remotely resembles the image in your head. It really is an art on its own and not a magic wand.

[desert dagger]

um...

no

those of us that have actually taken the time to learn how to prompt and learn how the various AIs we're using think can get what we're after with no real issues, usually the first generation of a prompt

[lucid bobcat]

With prompting alone you never get a finished piece (unless jank and AI artifacts is what you're aiming for) and you absolutely don't have full control over the composition or the poses, expressions, details etc. He said he's a painter, so I'm assuming he wants to have more control over the art he's making. I took a look at your YT channel and that doesn't impress me much. The stills that is, not the animations.

[desert dagger]

oh but i do. and no, there's no jank or AI artifacts in it. and i have 100% control over the entire thing. see, that's the point. you just haven't taken the time to actually learn how to use the technology - or you are using older models

and i don't honestly care if you're impressed. you're not my target audience.

[calm hearth]

As a non-creative that came into this space because of AI I have had to had multiple crash courses on certain subjects like aperture sight and focal length. I learned to prompt but at the same time for me to learn to prompt I had to learn about those subjects itself

[lucid bobcat]

Your stuff is full of jank like deformed faces and limbs, asymetrical details, prompt bleeding, inconsistent lighting, oversaturation, out-of-place detail and so on and so forth. I guess your standards are just lower than mine and that's okay. But don't say you got everything under control just by prompting, because you evidently don't.

[desert dagger]

no it isn't. I knew you were only here to troll

[lucid bobcat]

I'm pretty sure you're the one trolling, acting all oblivious. But whatever man. Live your dreams.

[cedar salmon]

if it wasn't for them ide be going on thinking you can generate video on a 3090, what a crazy belief i had before i read their correction

[quartz swan]

@slender vault i cant dm you and i need help with swarm

[tropic frost]

small curiosity about the illustrius models.

i noticed some stuff on civit ai that utilizes both ilustrious and pony loras together

are pony loras compatible with illustrious checkpoints?

[grand rain]

are there any img2video models that run locally and dont eat up a couple thousand dollars worth of vram?

[slender vault]

prolly a rogue setting

that you didnt turn off/on

[olive rampart]

For text prompts is it better to keep them short or do them with extra detail?

[quartz swan]

shaking in my boots rn, gonna tweak if someone says shorter cuz that is NOT what im doing

[calm hearth]

Ur in the discord of one of those models

[quartz siren]

Ltxv is your best bet, should fit in 8gb vram and be very fast.

[quartz siren]

depends on the model, flux and sd3.5 can usually handle any size prompts but more detail might make it better quality. Sd1.5/sdxl usually prefer tags and you want to stay under 77tokens.

[quartz swan]

77???

double it-

[quartz siren]

lol I mean it might have some effect but not much

[quartz swan]

im just staring at my prompt-

150 tokens

[olive rampart]

so nothing long form just single words

[lusty lava]

If there is anyone in here that could help me with SD, please DM me

[warm junco]

Hey, if its technical feel free to ask in #🤝｜tech-support

[lucid bobcat]

Short prompts mean that you'll only see the model's biases (basically the average training data). It might give you good results if you have other ways of exploring the inference domain. But adding tokens is the simplest form of navigation. What you write does not have to make sense. Don't hesitate to add descriptions like "the cat sounds like motor oil" or even use fantasy words.

[lucid bobcat]

If you have to ask, chances are you're not gonna train a Flux killer.

[subtle knoll]

are img2video models in ComfyUI at this current stage actually better or as good as Veo 2, KLING, Flux etc?

[split kestrel]

In Jan…. Yes

Current stage is questionable that anything is really king ship of fudge mountain

Hunyuan from all my research is current champ

But has no I2V until Jan

[low moon]

even Veo and kling are not adequate now. Come back in 2027.

[split kestrel]

Or in many cases… wait a week

[low moon]

nah

[split kestrel]

Seems Hunyuan is outperforming Sora

[low moon]

perfect image is 1-2 years away. so is sound

flawless video 3-5 years

[split kestrel]

This 2-5 seconds clip stuff sucks

But… so is waiting 20 minutes for a 2-5 seconds clip of trash

[low moon]

i wouldnt mind short times if the image was really good

but its not quite there yet

and yes there is also the waiting time for now

next phase GPUs will be better

[split kestrel]

I have been able to gain about 8 min render times

[low moon]

we're not there yet

but almost

[split kestrel]

8min - 5 seconds

[low moon]

yeah

its unnaceptable now.

[split kestrel]

Well… it’s acceptable… but continuity doesn’t exist it seems

[low moon]

of course thats one of the things were waiting for, consistency and continuity

[quartz siren]

with context windows, you can actually get 20sec vids with hunyuan and there was a new technique which should be even better but no support for hunyuan yet(it should work though)

[split kestrel]

Like a batch render?

If you could render a 5 seconds clip, steal the last image and create a new render off that image and stitch all in a final…

But…. It would be a big ole mess

[low moon]

yeah

[quartz siren]

Its in kijai's node: https://github.com/kijai/ComfyUI-HunyuanVideoWrapper, and its not the last-image part, hunyuan doesnt even have i2v support yet.

example
#🆕｜sd3 message

[low moon]

with each new model and texhnique its a bit better but then 3 months later soemthign better comes and all the workflows have to be updated

[split kestrel]

Jan

Is supposedly a date

[low moon]

so why not wait 3 years and by then eveyrhtign is perfect

XD

[split kestrel]

I need video - video

Havnt really found a solution yet that hits my mark.

[low moon]

were not there yet thats why

[quartz siren]

This is the new technique and very amazing but only cogvideox2b support right now, hunyuan support should come soon: https://github.com/TencentARC/DiTCtrl

[quartz siren]

will work for video-video too,

[split kestrel]

I was crazy deep in animediff

I would like to shoot my own video of whatever… and use that video to generate a new video to replace the model with a creature of sorts

Or have interactions within a scene with a new generation thats “aware”

I have generated some wicked stuff… just not in the ballpark of what I need.

[low moon]

AI video is either full on psychedelic nightmare or uncanny valley now

[split kestrel]

I kinda like the nightmare fuel

[calm hearth]

SVD is a joke compared to runway or Kling at this point for me and I tried hard to get it to work well, always produced trash

[fervent thunder]

sure I didn't like SVD either but there is HunyuanDiT in Comfy now which is at least Kling/runway level

[chrome cargo]

Anyone with a solid 3D background here? Like Maya, Houdini and so on.

[oblique elk]

Sorry only wireframed background 🙂 3DS Max and then moved to blender....

[chrome cargo]

Haha, that's good too. How long have you been using 3DS Max or Blender?

[oblique elk]

Started 15 years ago, after 3 years made a pause and started again 4 years ago but only as an activity in my free time. The last 2 years the time reduced more and more due to the fascinating world of AI 🙂

Can use it barely enough to create objects to print with the 3d printer. More a technical guy then an artist. Actually i am trying to get the following scenario to work (citygml > obj > blender > stl > 3d printer)... Headache as usual if you need to work with goverment created "standards"....

[oblique elk]

so i might definitly not be helpfull if you need somone with deep knowledge.

[chrome cargo]

Hey, that's cool! I just really enjoy connecting with people, and a family member bought a Bambu Lab in the last few days, so I'm also a bit familiar with 3D printing now, haha.

Have you printed anything yet?

[oblique elk]

All kind of (unnecessary) Stuff: from little Giftpackages, Stamps, Vector Low Poly Images, Tools for pottery, Christmas Decoration and my personal favorite of useless things: a sheep head and legs as spare Toiletpaper-Roll Holder... 🐑

[wet grotto]

I did a lot o images with this controlnets but I wanna know your opnion. How do you understand this controlnets? I use especifically for achitecture, if help you give me examples:

• Lineart
• Softedge
• Scribble
• Canny

[grand rain]

anyone know how i update python packages in comfy?

[grand rain]

how

who here has mod abilities?

mans trying to get me to join a dodgy server

[lucid bobcat]

You and your family member should definitely take a look at HueForge. It's the missing link between 3D printing and AI art.

[silk condor]

Hey hey

I am running a older model of a checkpoint and I want to keep using it. But would love to update it myself so it works with other Loras much better.

I have another model that is very different and much Loras it works and most like characters does not need any changes etc.

But with the older model .. it's not working as well

So how can I update that model or will make my own checkpoint?

[lucid bobcat]

If you mean that not all loras work well with the checkpoint you're using, that's totally normal. You can't train a checkpoint to make it work better with an existing lora, but I think you could just train a lora to work well with the checkpoint you're using. But you would have to do that for every lora that doesn't work well. You can't just "fix" the checkpoint.

[silk condor]

Oh okay, is it easy to make a lora? Never done one

[lucid bobcat]

I haven't trained one either, but it's easy. The most difficult part is to get the training data.

[silk condor]

What is that?

[grand rain]

are there any resources on fine tuning img2 3d model ?

[steel frost]

I’m looking to buy a new computer (to me) would an m2 max MacBook Pro, with 32gb ram and 38 core GPU suffice for fairly quick image and video rendering in comfyui?

I had an intel chip Mac before so haven’t had much experience with what technical specs are needed, I’ve mostly been using cloud services but wanna run it locally

[silk condor]

How much ram does the GPU have?

[steel frost]

I think the computer has 32 gb of unified memory

[lucid bobcat]

If you need speed, don't buy a laptop. You need an Nvidia GPU with at least 16GB RAM

[steel frost]

Well I need it for more than just stable diffusion, gonna use it for work as well and unfortunately can’t afford both atm… I know an Nvidia GPU is ideal tho

[lucid bobcat]

It will work, but won't be a satisfying experience.

[steel frost]

Damn, ok. Just will be pretty slow?

[lucid bobcat]

What GPU does it have? You'll need a dedicated GPU.

[steel frost]

I’ve seen ppl say they run it with much lower specs on MacBook pros so was hoping it would be able to run it fairly well… hold on just a sec I’ll drop the specs

[steel frost]

Apple M2 Max chip
12-core CPU with 8 performance cores and 4 efficiency cores
38-core GPU
16-core Neural Engine
400GB/s memory bandwidth
Media engine
Hardware-accelerated H.264, HEVC, ProRes and ProRes RAW
Video decode engine
Two Video encode engines
Two ProRes encode and decode engines

And then it says it has 32gb of unified memory

I figured the 38 core GPU would be helpful… it’s like a 3500 dollar computer 😭

But like i said I know there are better/cheaper ways to get better performance for stable diffusion specifically I was just hoping it could run it fairly well

[lucid bobcat]

So it's a custom GPU and not Nvidia? This sounds like it's on-chip, which isn't great. Are there benchmarks for stable diffusion? I think stable diffusion 1.5 and XL will be fine, but better models and especially video will be a pain. For 3500 dollar you should have 64GB RAM.

[steel frost]

I’m trying to find this information

I was trying to use flux mostly

[lucid bobcat]

What's the name of the MacBook model?

[steel frost]

MacBook Pro (2023) 16-inch - Apple M2 Max 12-core and 38-core GPU - 32GB RAM - SSD 1000GB

[steel frost]

Just a moment

[steel frost]

I think it’s A2779

Man actually I’m not sure, I don’t have it yet and it’s kind of hard to figure out

Bc I’m not buying it directly from Apple

[lucid bobcat]

If it has to be a MacBook, make sure the GPU is a M3 Max. If you can, try to get 64 GB RAM or at least 48 GB. I think this is the best you can do with a MacBook.

[steel frost]

Alright, we’ll thank you for all the help and information.

[lucid bobcat]

You're welcome

[warm junco]

Flux on Mac is very very slow

[steel frost]

Damn

[steel frost]

What if I were to eventually purchase an external nvidia GPU for it? Or is it an OS issue as well?

[warm junco]

Should work with an egpu I guess.

[opal mesa]

ok

[astral hearth]

hi

[uneven lance]

hi

Happy weekend

[wraith tinsel]

im using this exact computer

doesnt seem all that slow but im still a bit new to stable diffusion

let me know if youd like me to try something

[steel frost]

Oh nice! Do you use comfyui with it?

[wraith tinsel]

i do not

currently using automatic1111

[cloud arrow]

I have a question

Has anyone ever tried to download an entire image board and make a model out of it?

I wonder how that would turn out

[oblique elk]

Hi there,
I used comfyUI on my MacBook M1 Max with 32gb of unified memory. It worked and speed for the SD 1.5 models was acceptable. Sadly even old 8gb nvidia cards were faster…
But the 32gb helped a lot for some large language model tests but again the time to generate tokens was awful even with the support of the MPS Backend. So I would suggest to you using flux schnell or other lightning concept with less steps or use 2 iteration process where you can see instantly if the result will match your expectations and abort if not. So the higher resolution, step count, upscale, … will only applied to those images that match the prompt.

And eGPU won’t work as for know apple only support AMD eGPU so it won’t help you much.

[nocturne pewter]

hello, how can i try stable diffusion ? new here

[silk condor]

Well you can install on your pc and try it out xD

[ebon junco]

try wubui forge，just search on the github.

[fervent thunder]

Everyone who's using a MBP should be using the app DrawThings imo. How fast are your Flux generations, lets compare times.

I'm on an M3 Pro, 18 GB Ram (MRX43B/A).

[steel frost]

Thank you for that information

[steel frost]

Yeah I discovered that after I researched some more

[steel frost]

Yeah I was reading drawthings is the way to go. I just like the interface and workflow of comfyui which it doesn’t support

[oblique elk]

Did you tried diffusion bee. It uses a quantisized flux schnell version and so it needs less ram and only 4-6 steps.

[primal monolith]

Hello!

Not sure if this is the right place, but I'm looking at training a model and/or Loras. Would anyone be able to give me some input and advice?

[quartz swan]

what are you using to train loras 👀

[primal monolith]

Are you asking about my Computer or if I'm using a program? I'm not making them yet, that's why I'm asking for advice

[quartz swan]

ah, just wondering what you were using, computer or program or site, i make loras on civitai sometimes because my laptop cant handle making loras on it

didnt know you hadnt it made it to the uh making them part, my bad

[primal monolith]

It's ok! Sorry if I wasn't clear!!!

[quartz swan]

nah, im just, tired LOL

what were you wanting to make, just out of curiousity

[primal monolith]

Wanted a lora for my OC to be consistent

[quartz swan]

do you have 2-3 really good images of them

[primal monolith]

I have about 2500

[quartz swan]

slide?

im bored, i wanna see what i can do

[primal monolith]

DM'd

[tribal crown]

I would like to make some anime characters move or dance around? Best img2video for this?

[desert dagger]

@warm junco spammer's are back

[warm junco]

Thx

[tribal crown]

phew lucky i didn't click on that

anyone can help me pls?

am curious

[quartz siren]

Probably minimax if you are fine with closed source.
For open source, something like controlnext svd could work.

[lucid bobcat]

Scammer alert! don't click the link

[fervent thunder]

Can anyone help me with #🤝｜tech-support ?

any help is GREATLY appreciated

[final tusk]

Hello!
Just landed here in SD Discord

Happy to have arrived~

Would there be a guide somewhere to install and properly setup SD? I was previously using ComfyUI but something broke so I am starting over and could use some assistance~

Thank you!

[oblique elk]

Weclome, best way to start a fresh installed can be found within the github of the technical miracle @warm junco who is really helpfull in the #🤝｜tech-support . This would be the link:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides
You can also find this in the sticky messages within the tech-support channel.

[main snow]

When the diffusion is not stable 😔

[normal pagoda]

when I can't find someone taking requests

[quiet pulsar]

Hi everyone! I'm working on a children's book and need help creating consistent illustrations. I’m aiming for a watercolor, hand-drawn style—for example, a family happily having breakfast. The characters need to be based on real people from reference photos and look consistent across all the illustrations in the book.

I’ve tried using MidJourney, but I’m struggling with consistency in both the characters and the art style, even when using the same references. Does anyone know how I can achieve this with Stable Diffusion? Are there specific workflows, tools, or models you'd recommend for maintaining character likeness and style consistency? Thanks so much!

[lucid bobcat]

You need to train a lora for this purpose.

[quiet pulsar]

Thank you so much. Training Lora's seems possible with Stable diffusion. I will dive deep in learning this

[paper drum]

Hi guys, I would like some clarification on masking image, because I am getting conflicting information.
I have read that a mask is a greyscale image where black = remove and white = keep and different shades of greys are for in between
But it seems that practically we are painting black (or red??) on the original image as mask??

[lucid bobcat]

A mask is a separate channel or layer and not painting onto the original image.

[paper drum]

That was my understanding too, but various sources seem to indicate that the original image with black areas are used as both the input and the mask?
Such as this one
https://www.youtube.com/watch?v=vY4QwnR4R2M

Do you have a ComfyUI inprinting example?

[lucid bobcat]

I don't know what you're using. I'm just saying a mask is a different channel. It can be the case that they use a workflow that adds black to the original image. But I wouldn't call that a mask because it will treat everything black the same, whether you want to inpaint it or not.

Usually this is referred to as treating a specific color as transparent.

[paper drum]

I think you mean the black spots and maybe even the grey spots in the adjusted original image would make it a bad mask?

[lucid bobcat]

The workflow in that video creates a mask. It does not paint onto the original image.

[paper drum]

It is just, the sources are conflicting, and MastEditor of ComfyUI seems to encourage us to just brush out the areas in black and use the image as a mask

[lucid bobcat]

You can see that the node has two outputs. Image and mask.

[paper drum]

If I understand correctly, you meant the "Load Image" node would output two image files, one is the full image and the other is a black and white image as mask?

[lucid bobcat]

If you load an image with alpha channel then the alpha channel will be available through the mask output. If the image does not have an alpha channel you have to paint your own mask.

[paper drum]

Oh, I think I get it, the mask "black" is not the RGB = (0, 0, 0), but rather alpha channel = 255?

[lucid bobcat]

Yes, although it does not have a color. You can look at it as background and foreground and you could assign any color to both.

[paper drum]

Thanks heaps! It was confusing to me as I was thinking masking as a RGB thing

[gray vapor]

Hi! I always had a problem with using depth maps extracted from any 3d modelling software in controlnets. Even with 16 bit depth, the sudden jumps in grey tone (banding) is interpreted by the model as an edge, creating some kind of seam. Does anybody know how to fix this?

What i do for now is applying Depth Anything to my depth map, which feels very stupid. Also, how is that the depth maps generated by depth anything dont have this problem?

[lucid bobcat]

Only use the same depth as the model uses. There should be no banding, not even with 8 bit.

I think the issue is the maximum depth. You should set it to your furthest object you want in your scene.

[gray vapor]

Thanks! I see that the images coming out of Depth Anything are 24bit pngs. Therefore, i guess thats the standard

Since any depth model works good with depth anything

[foggy warren]

CFG

[gray vapor]

In theory this is already what rhino is doing. The darkest part of the image is the furthest that is in view.

[acoustic cove]

How do I get into generating AI videos?

[cedar salmon]

locally? take the plunge and get ComfyUI installed and working and be lost for a couple weeks, than generate video

[acoustic cove]

Hmmm... What are some good non-local options? I want to dip my toes in the water before diving into the deep end.

[sudden grove]

For cloud based Video Generation take a look at these,

Luma Dream Machine
Kling
Minimax

[acoustic cove]

Are any of these paid? And where can I find them?

Also, anyone have any good example prompts for videos? Not sure if the prompting is structured any differently.

[winter timber]

hunyuanvideo

[mellow zinc]

hey peeps, i followed a tutorial on youtube how to set up flux dev, I did everything exactly as the video showed however im getting errors and Im assuming its because I don't have "flux" type in the dualcliploader, anyone knows what im missing? when i search for this on google i cant find anything https://cdn.discordapp.com/attachments/1002602742667280404/1323333004860461128/image.png?ex=67742166&is=6772cfe6&hm=9ac0418507062081e9bc594d56179110007d61a8228c6cbcff1cd6d51500c420&

everything is up to date, comfy and the nodes as well

[lucid bobcat]

Remove this scammer please

[modern pagoda]

Its like bot wackamole in here

[lucid bobcat]

ikr. The fix would be easy: Force people to pass a captcha before joining a server.

[cedar salmon]

its crazy they direct people to a place with active scamming and think its ok to ignore it all

says much

the trust and safety that gave SD3 issues should of been applied to the discord not the model

[fervent thunder]

I don't think SD3 was overly censored
I just think it was a 2B DiT
to this day I have never seen a good 2B DiT

I think DiTs need at least 6B

[lucid bobcat]

The whole "safety" rhetoric is nothing but fascism in disguise anyway. I'm not in danger because of mean comments, false information or obcene imagery. The real threat in the internet is fake products, scams, data harvesting, social engineering, shills and fake reviews, government propaganda etc.

[cedar salmon]

ya that and what seemed like direct removal of concepts

[fervent thunder]

oh actually direct removal of concepts might be a valid point yeah
what I was trying to say was that I think the anatomy problems were because the DiT was too small, not because it has censored training, which is a popular theory

[desert dagger]

DM sent

[fallen shard]

Hi, can I ask if the Stability.ai API has faceswap functionality? Couldn't find it in the API docs. Thanks!

[quartz siren]

Yeah I think that's probably a big issue, there are several 2b dits now and none of them are very good.

[modern pagoda]

Phone verify

Too costly for most attackers

[quartz swan]

ah tragic, phone verify is gonna lock me out real shit

[desert dagger]

mmdit isn't good, regardless of the model size

[split kestrel]

Sup guys

How is the grind today ?

[quartz cliff]

Hey pal.

[split kestrel]

Can you IMG > VID with flux?

Or vid 2 vid using mp4 -> with flux? Or no?

[lucid bobcat]

not with temporal consistency afaik

[silk condor]

I see people doing checkpoint and Lora left and right.. But I cannot find a video or anything that really explains in it a good way. I would love to make my own checkpoint model and update a checkpoint I am currently using but make it into my own and update it so it works better. Can anyone help me with this?

[tame bloom]

@boto can you message @frosty turret please, he's autistic and needs to know why he was timed out.

[slender vault]

prolly cos he posts nothing but lolis in the anime channel when told not to 💅

[hasty gull]

Hi everyone, new here, looking forward to sharing prompt best practices

[desert dagger]

a lora is a very small model that just updates a few specific weights of the base checkpoint. A fine tuned checkpoint is much larger, but does the same thing - updates the weights at runtime. both are less expensive than retraining the entire model would be.

[thorn hare]

whats the difference between regional prompter and composable Lora ?

[warm junco]

Regional prompter works and composable lora is outdated since 2 years

[karmic brook]

probably because he was explicity sharing prompts for little girls with "big breasts" and following up with "hehe"

[slender vault]

that doesnt surprise me

[bronze merlin]

Happy new year.

https://tenor.com/view/happy-new-year-made-by-me-cute-anime-fireworks-gif-15943630

[slender vault]

in thc we trust

[mental crater]

Happy New Year to everyone! 💫🤠 May you stay healthy and be happy afresh... 🎐

[scenic zinc]

I am trying to fine-tune an SD inpainting model to outpaint backgrounds from product images conditional on a text prompt. Standard SD inpainting models tend to extend the product frontier, likely because they are pre-trained on small input masks. See the paper: https://arxiv.org/pdf/2309.11507.

I have a dataset with product images, masks, and text prompts. However, wherever I search, it doesn't seem possible, and there are many different discussion threads on Hugging Face and Reddit where people ask the same thing without any satisfactory responses. Is there any way to do this?

[fervent thunder]

powerpaint v2 is good

[tribal dove]

my stable diffussion generates completly shit and i dont know why can someone help me

[wispy yew]

Hello everyone, glad to be here

[hazy flicker]

what ui, model and vae are you using?

[lucid bobcat]

Whose stable diffusion doesn't. That's why I only use Flux.

[lucid bobcat]

Kick these guys out pls.

[desert dagger]

what version of stable diffusion are you using?

[gentle prairie]

is there a way to make seeds different when generating a lot of images?

like seed 100 to 105 will look very similar

but i want to have a very different image every time but not having to do it one by one

is it possible?

[lucid bobcat]

If different seeds does not give you enough variance you also have to change other parameters for each generation. You could mix and match different prompt strings for example.

[gentle prairie]

yeah but like when i do one generation, the next one will be very different, but if i do 2 batch count, the 2 will looks really close cause the seeds follow each other

dunno if i'm making sense

[slender vault]

you wanting slight variations?

Swarmui has Variation Seed natively implemented that could achieve this

[lucid bobcat]

Seeds following each other does not make images similar. The seeds are used for RNG. Two consecutive random numbers are, well, random.

[gentle prairie]

why do they look so close to each other when i do a lot of image through batch count, like 30, but not at all if i just do like 1 batch count each time

[lucid bobcat]

Make sure the seed actually changes. Also it's to be expected that images will look similar for any given prompt. I don't know if you mean almost identical or just similar in style.

[green oxide]

yooo

how do. i find out what version of stability my api keys support

[copper crystal]

Specifically, seed determines the initial noise that is denoised from. Beyond that i'm fairly sure the seed won't determine how things are denoised. Unless you have nodes set up to affect changes based on the seed value

[copper crystal]

https://chenglin-yang.github.io/1.58bit.flux.github.io/ This is interesting. hope we see weights that are usable come of this project.

[lucid bobcat]

Why not just use the Q2 gguf model?

[copper crystal]

it is a useful one but it's results ring like an sd15 generation. there is severe degredation at work when models are just straight quantized to lower bits

[quartz siren]

Yeah it’s from ByteDance so probably trustworthy but kinda suspicious that they don’t show much info.

The technique is supposed to be similar to bitnet for llms according to them but bitnet required pretraining from scratch on hundreds of billions of tokens. Somehow they achieve that without pre training and it’s still very good quality.

[copper crystal]

Likely are just refining it more before deploying. Or hoarding it to themselves forever. one or the other .

https://github.com/Chenglin-Yang/1.58bit.flux?tab=readme-ov-file#158-bit-flux this says they're going to release, but we shall see

[autumn hull]

There was an error in my SD and it showed a running error
Do you have a good brother to help me take a look?

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument mat1 in method wrapper_CUDA_addmm)

[glass glen]

Which LLM is probably the best for IT? I'm too stupid to configure port forwarding on a pfsense router

[fervent thunder]

out of the low bit stuff I like SVDQuant the most

SVDQuant is 4 bit rather than 1.58 bit but the quality is almost the same as FP16

[slender vault]

best to ask in #🤝｜tech-support

[valid fulcrum]

Hi everyone, some SD experts here to help replace our current version of SD with SD 3.5 inside our chatbot app (we're using comfy deploy)? Not sure what we're using at the moment, but it's an older version of SD and we need more hyperrealistic images of humas in our very human-like chatbot app. Please helppp

Our app allows users to gift clothes and accessories to their AI chatbot, they usually reply with images. Would be great to have someone work on this with us ❤️

[twin igloo]

What control panel similar to ComfyUI would you recommend for professional use in generating videos? The idea would be to create dozens or hundreds of thumbnails with specific seeds using a video model, select several of those thumbnails, use them to create miniature videos with another prompt for what will happen in the video, choose one of those videos, create a video at the highest possible resolution with that seed, and then pass the result through another AI upscaler to reach 1080p (or 4K if possible).

[twin igloo]

I still haven't started learning ComfyUI, and I don't know if there's an easy way to generate many thumbnails, or to use fixed seeds, or if having fixed seeds guarantees consistency when expanding an image.

Can it be used via the command line? All of this is for making FMV games; maybe I could program something in the game engine.

[keen hornet]

Anyone here using ai for book covers?

[fair pewter]

?

thats an idea

i dont think so but possible

[fair pewter]

use flux for that

comfyUI is pretty straight forward

just download nodes and see what they do inside openart.ai

[twin igloo]

ComfyUI seems the most professional, but having a web server running in the background on my machine all the time to then access it from the browser seems like an excessive RAM consumption. If there’s something more command-line-based, it would be more convenient, especially when integrating it into a game engine.

[young rain]

whats the setting for using tdd 4 step lora

[fair pewter]

comfyui is designed to be easy but it isnt efficient when it comes to integrating it to video games

python would be a great solution to this

especially with pytorch

[twin igloo]

From what I see, Hunyuan can be used via the command line mmmm

[modern moss]

Is there a lora that includes a lot of celebrities, or do I need one for each single one

[rose portal]

Which channel should I use to debug problems installing controlnet on ComfyUI on RunPod server?

[lucid bobcat]

It sounds like you're trying to create a virtual character. For character consistency you need to train a lora. For what you're trying to do, having a 3D model for the character as well as 3D models for clothes and passing the 3D output through stable diffusion for photorealism, changing colors and adding backgrounds, seems to be the most viable approach.

[shell pagoda]

HELLO

[inland stirrup]

Hi there! Is there a discord server specifically for StableStudio (open source), or is this the place?

[wise aspen]

guys, what is best way to use flux, but not comfyui?

[warped robin]

Hi everyone, I hope this is the right channel for my question. I’m looking to run a LLM with image generation capabilities using Stable Diffusion. I’m unsure which GPU to choose and would appreciate some guidance.
Should I go with AMD (using the ROCm platform), or is it not well-supported for this purpose? Alternatively, would Intel’s Arc B-series GPUs be a better choice, given their good value? Or should I prioritize NVIDIA due to CUDA and its broad support for AI workloads?
My budget is around €300–€400, so I understand that the performance won’t be lightning-fast. However, I’d like to achieve decent speeds for both tasks.
Thank you for your help

[quartz siren]

Probably Swarmui or forge.

[desert dagger]

you want an nvidia gpu - but if you have a low budget, maybe try one of the online solutions - such as runcomfy online, or rent a gpu from vast.ai or amazon webservices.

[warm junco]

Dont go with Intel for ai stuff for sure.
AMD is nice if your on Linux with ROCM or know how to use Docker with WSL2 on Windows.
Nvidia has the best ai support but lower vram in their midrange.
Best card would be a RTX4060 Ti with 16gb vram.

[wise aspen]

is swarmui better than forge with flux? cuz i have fluxdev+2/3 lora and i have out of memory (not vram, ram) in forge

[slender vault]

yes

swarmui uses comfy as the backend

without having the node base visualization and more of the a1111 type of gui

[warped robin]

ok thank you guys

[wise aspen]

So in essence it is comfyui with the graphic form?

[slender vault]

yeah lol its nice

get the comfyui performance while the simplicity of the gui

has allot of native implementations that are nice aswell

[wise aspen]

and there are things like wildcars/adetailer in a simple way like in forge or is it rather hard?

[slender vault]

uhhhh wildcards like random:blue,green,pink ?

and i dont find adetailer to be needed with comfy

i use hypertiling which helps

and swarm has a refine upscale mid gen

thing

which what i use

[lucid bobcat]

I don't know if there are any local LLMs that can directly call functions, but you don't need it anyway. You can just let the LLM use keywords and then look for those keywords in the output and control the image generation that way. Buy Nvidia because it's well supported. What you really need is RAM. If you want to run a large LLM at least 64GB. You also want 16GB VRAM.

[low moon]

Is SANA and Omnigen worth the HDD space? are they substantially faster/better than flux or sdxl?

And whats this Illustrious model? Is it like a more realistic Pony?

in youtube land every new ai thing is a flux or sora killer i know but whats the reality on the ground?

[slender vault]

HDD isnt worth using in ai

it will load extremely slow

invest in a SSD asap

espesically when you get into the large models like flux and whatnot

[cedar salmon]

what about their question though

SANA isnt it, its just fast, dont know enough about the other stuff

[warped robin]

So the llm ist more resource intensive than the SD model?
So would it be possible to build a PC for just image gen for me and my friends without the llm? The PC would have like 16 gigs of ram a SSD or nvme SSD an i5 7 Generation and an like 300 Euro or 350 Euro GPU. Does anybody know if this PC would be enough for like generating 512x512 or higher in like 15 or 20 seconds

[fervent thunder]

sana is rly fast

OmniGen was more of a research model than something to use. Its training data overage was not broad

[lucid bobcat]

16GB RAM is not enough. You should at least get 32GB. RAM is MUCH cheaper than VRAM and you can get away with 8GB VRAM if you have enough RAM. The GPU is mostly for speed. I personally still use a RTX 2080 with 8GB VRAM. I can even run a 12GB Flux model with no problem, but only after I added more RAM.

[warped robin]

So I invest in 32 gigs and do you have a GPU recommendation for like 300 euros or should I wait for the Rtx 5000 series or maybe go with AMD if I would use Linux or is Nvidia the way to go with all the support

[lucid bobcat]

Stick to Nvidia. I recommend you wait for the 5000 series because it might lower the prices of the older series. Prefer more VRAM over more performance. Generating SD 1.5 images within 20 seconds will not be a problem.

[warped robin]

Ok thank you really much

[lucid bobcat]

you're welcome

[low moon]

Alrighty then, thanks for the infos

[quartz siren]

Yeah sana is fast like neon said but I don't think its worth it really when flux schnell 1step clearly outperforms it by a very large amount and should be similar speed, maybe even faster.
Omnigen is pretty slow sadly and although the identity capability is very impressive, image quality is worse then sdxl even imo.

I didn't play with Illustrious much but it is very creative and unique from other people's experience and more artsy, its not for nsfw like Pony I believe but it can probably do it.

I dont think sana, omnigen is worth it yet but illustrious might be decent to pick.

[cedar salmon]

current hype is like LTXVideo and HunyuanVideo, both deserved

[willow needle]

IF you're using the basic webUI Automatic1111, is there a way to get the tab for SD back open if you've accidently closed it but it's still drawing?

Can reopen a closed tab ofc but no prompt or drawing

[quartz siren]

yeah ltxvideo is crazy fast, there was a new repo that increased the speed of ltxv on new gpus without any quality loss and lowered memory usage too. On a 4090 it went from 16sec-->6sec for transformer for a 5sec video.

Hunyuanvideo with dpm2ppm is even more improved, it can work even with 8steps now and 16steps is way better then higher step euler/ddim.

[low moon]

Ah video is stepping up I see.

[quartz hare]

Whats the latest/best model for text output? Anything new?

[quartz siren]

Do you mean llm or like text on a image.

[quartz hare]

Text output/text in image, like a poster, for example

[quartz siren]

For open source, it’s still probably flux dev which is still comparable to several top closed models like ideogramv2, dalle3, red panda model.

The best in closed is probably flux 1.1 ultra

[quartz hare]

Yeah 👍 thanks bud

[copper crystal]

i woudln't call flux dev open source. Though they are weights that you can obtain for free. They just aren't open source nor do they have a libre level license.

in fact, most models i wouldn't call open source. That would imply the source code that trained them, as well as the datasets, were made public

[fervent thunder]

best open license model is Schnell yeah
its a pretty rough situation

[shy sluice]

whats the website to generat on web i cant remember it and i cant find it on the channels anymore?

[low moon]

Have an excellent weekend everyone.

[tame mauve]

hello

[final chasm]

Hi everyone! I'm looking for models and/or loras to do realistic inpainting with SDXL and Flux in Forge. Any help?

[elder stirrup]

Does anyone know of any summarization ai tools that would do a good job aggregating and summarizing a bunch of text responses to survey questions? Thanks

[calm hearth]

hey guys i was wondering if there was a way with the batch size to make it so each pic in the batch is at different denoising scale? say you put batch of 4, itll gen 0.5, 0.6, 0.7, and 0.8 denoising pics with the rest of the settings the same. is there anything like this or am i delulu?

[rain valve]

hello guys, I'm looking for a mid end laptop to run SD, something like RTX 4070, I'm kinda out of what's is happening rn on CPU/GPU technologies. i'd like to ask if you have something to recommend me, if there is a laptop worth getting now or if I should wait for some new model coming up soon. If you can you can send me DM or @ me, thanks.

[lucid bobcat]

Why would you even need it to be in one batch? Just make a XY plot.

[fervent thunder]

cos batches are faster

[lucid bobcat]

Yes but not that much.

[graceful plover]

am i better off using a colab version to train a model?

because I want to make a model of particular art style so, im not sure what parameters are best for it

[fervent thunder]

I agree for the most part
but there are weird edge cases where it speeds things up a lot

particularly with clusters

[copper crystal]

with the auto1111 extension you can do both. configure the x/y/z grid and each batch will use the current cell's parameters. I'm not sure about comfyui. you'd probably need to make a custom made noodle plate for it

[gentle prairie]

So, i installed this, and the scanning worked properly, downloaded all the images for my lora and models :

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

but i don't see the 3 additionnal buttons

[copper crystal]

are you trying to use it on forge? forge updated to gradio4 so i don't think it'll work.

The reforged project still uses gradio3.

[gentle prairie]

nah just auto1111

[copper crystal]

ah. well then i dont know then. another extension is conflicting with it maybe. Check the console log when you start the webui. it may have some note about a problem loading an extension

maybe the author updated the buttons out but the readme never got the info

[gentle prairie]

i'm using animatediff (is there anything better to make video on webui111 stable diffusion?) and for some reason the video is fast/accelerated, and i don't know why, the first time i used it, it didn't do that
i'm following this guide :

https://stable-diffusion-art.com/animatediff/

[gentle prairie]

So i've been using pony but apparently illustrious is better and doesn't need as much LORA... But the thing is, how do i know which style/characters does it have by default?

cause i certainly won't retain them all by memory, lol

[ivory cove]

yo

[normal pier]

I would like to make something like https://interiorai.com/, with my own model. I played around with Automatic1111 and added controlnets (depths) and all to make my output, but thats just the UI of it.. how do I make it into its own "model" so then I can use it via API? Is there a output file? a good online training software or raw by python (torch)?

Im confused how it works, also for each interior AI "style" do I have to train the model on each style as well?

[nocturne vine]

hello

[lean cedar]

Hello all !! i just got a new computer a few weeks ago and i wanted to try stable diffusion since a long time now ! i installed it on my computer locally and now i want to learn more about it, especially how to make cool picture using the same face, as i know it is possible to train the ai to do it, i wanted to know what should i look for on youtube if there is a cool video about that, that is clear and simple (english isn't my first langage) thank you !

[calm hearth]

https://www.youtube.com/watch?v=_A7PraBHyg0

hey guys i just wanted to check since this industry is moving so fast if anyone is using updated models? i am currently using epicRealism and Dreamchaser XL. it seems to be working out for me. am i out of date?

[lucid bobcat]

ban this guy immediately

[copper crystal]

likely was a stolen account in the first place

[lucid bobcat]

Possibly. This scammer infestation is so annoying. I hate the "I'm a good guy, can you help me out" shtick more than the "give me your credit card number" shtick. What's worse is that it works, or else they would just give up.

[fervent thunder]

its literally only this discord
out of the roughly 200 discords I have been in

[lucid bobcat]

Yes it's really bad. There's not much going on in the chat so it sticks out like a turd in a punch bowl

[scenic shadow]

Is there any models that takes multiple images to gather a style to generate an image from? The img2img one im using currently only allows for one image per.

[atomic mortar]

Afraid your gonna need to make a lora or two for this one

Lora 1 at 0.5 strength and lora 2 at 0.5

[scenic shadow]

Im new to AI stuff, ready to learn tho, got any tutorials you could shoot my way? Thanks.

[atomic mortar]

Hmm currently still in bed lmao, what ui do you use?

[scenic shadow]

Im using WebUI Forge

If its too much thats fine, Ill google it

[atomic mortar]

Bet, im on swarm but i got some sources lying around

Some stuff is universal :^)

[scenic shadow]

Thanks, Im down to try anything. Not really tied to a specific one, just started yesterday

[atomic mortar]

Webui forge, a1111 etc are excellent starter ui's

[low moon]

No

skip those....

[scenic shadow]

Seem pretty limited so far

[low moon]

go straight for Comfyui

XD

[atomic mortar]

Worst advice lmao

[low moon]

loooooool

[atomic mortar]

Yeah thats why i migrated over to swarm ui

[scenic shadow]

ComfyUI uses nodes right?

[atomic mortar]

Yeah

[low moon]

yes nodes n' noodles

[scenic shadow]

Like blender? I might be able to use that if its the best

[atomic mortar]

Swarm ui has a comfyui backend so you can use that and use a simple ui

[low moon]

its by far the best but takes some time to get used to

[atomic mortar]

And it has a export to comfy so you can see how the noodles interact

[scenic shadow]

Ill check out Swarm if it uses both simpleUI and ComfyUI

[atomic mortar]

Its some getting used to but its being very actively maintained

[scenic shadow]

Was the lora you discussed a node?

[atomic mortar]

Not too complex

No/yes, a lora is basically a mini model you load in next to a main model(checkpoint)

[scenic shadow]

Im trying to give the AI a source of a bunch of game characters in order to make concept characters using the same style

So anything that helps would be appreciated

Is Flux the best model so far?

Local?

[atomic mortar]

Lets say a checkpoint knows a whole bunch but sometimes it doesnt know the newest characters from league

A lora trained for that can help with it

[scenic shadow]

Ah okay that makes sense

[atomic mortar]

Flux is definitely the most modern high quality but it requires a pretty beefy setup for reasonable times

I personally still use SDXL because i use a 3070ti for now

[scenic shadow]

Im using a 3080 TI right now, I know it only has 12gb but Im ussually able to generate pretty decent images in sub minute but Im hard pressed at making photorealistic stuff like it advertised however I probably have to mess with settings

[atomic mortar]

Yeah depending on the flux version you use cfg 1 with 4-20 steps

[scenic shadow]

My plan was to generate some decent AI art, convert it from 2d to 3d and than throw it in blender to fix topology

[atomic mortar]

But with SDXL it's also possible but takes some extra checking

And faster

Though a lot of realistic ai images suffer from the same-face syndrome and still has that uncanny feel imo

[scenic shadow]

I have CFG 1 and 20 steps. Ive tried up to 150 steps but didnt get that much more detail

[atomic mortar]

150 is a ridiculous amount

[scenic shadow]

It was about 5 minutes

Just was screwing around seeing what it did

[atomic mortar]

I generally use 20-24 if i want more detail i crank to 40 in sdxl

Anymore and it doesn't do anything lol

[scenic shadow]

Does SV3D just generate images/ videos of an 3d object or does it actually make an obj file?

Because I was looking at other 2D-3D software that actually converts it however the fron always looks pretty good but the back suffers, so I was wondering if there was a way to combine them to get a full decent 3d object estimation

[atomic mortar]

Ah in clueless about that one.

I saw this fly by though this month
https://www.reddit.com/r/StableDiffusion/comments/1hb1tj4/i_created_a_blender_addon_that_uses_stable/

[scenic shadow]

Thats pretty cool, I saw this which made me interested in the technique

[atomic mortar]

Looks pretty solid

[scenic shadow]

Would be crazy to see both of those combined. Generate an object, clean up topology than generate textures

[cedar salmon]

https://trellis3d.github.io/

there are some implementations of that in ComfyUI, ive not used any

[fervent thunder]

heyy, can someone with stable diffusion dm me? i have some requests

[astral nexus]

Is there a way to tell what model a lora is for?

Most of the loras in my collection are organized by model (1.5, SDXL, Pony, Flux, 3.5 Large, Illustrious), but there are some loras that have vague names that I can't seem to find the original page for

As I was writing that, I found it, it's in the internal metadata listed as "ss_sd_model_name"

They were Pony

"ss_sd_model_name: ponyDiffusionV6XL_v6StartWithThisOne"

[fervent thunder]

I kinda disagree with the whole concept of downloading and storing lots of loras to make a big collection

it feels like its doing things in reverse

would be better to either train or download a lora when you need

and then you will know what it is for

[atomic mortar]

I use folders and add metadata to the lora so i don't have to remember

Sometimes loras get taken down so when i see one i want to use in the future its getting downloaded

[lean timber]

This is a vague question, but is it easy for you guys to generate images that actually look good? It's just so damn hard for me to make anything that actually looks decent. If I try to recreate some image on Civitai and download every resource they use, use the exact same prompts and weights and all, it usually looks worse every time. And especially if I then try and change any detail, every change I make or LORA I add, it either breaks completely or looks slightly worse with every iteration.

I don't mind a bit of tweaking to get things right, but it's like I'm fumbling in the dark and it's all random if it will turn our okay or not, even after having generated images for over 2 years.

[atomic mortar]

hmm any specific example? if its nsfw you can shoot me a dm

[atomic mortar]

you can shoot me a dm if its easier to describe on call or if its nsfw

but generally without much prompting it shouldent be too hard

[lean timber]

I don't have a specific example really, although I have a lot of generated broken messes

[atomic mortar]

ah i mean what model, what specs, what UI etc

[lean timber]

I've tried everything from 1.5 to SDXL, Pony, FLUX and Sana. 3080 Ti 12gb, ComfyUI, SwarmUI, SD. Next and Auto1111, It's all just as fickle.

Sometimes I get decent images, but it's very rarely at the quality of what I consistently see some people on civitai make

And if I copy their settings, it never looks as good.

I guess what I want is just a consistent baseline workflow that can generate decent images and that I can build on later and add LORAs to etc without everything randomly breaking or starting to look silly and bad

As you say though, the more you try to prompt something specific, the worse it tends to get.

[atomic mortar]

hmm got time to hop on a quick call?

i can show you how i do it and maybe identify whats going wrong at your side

[warm junco]

Ho do you copy the Settings?
Because you most likely dont do it the right way.
Click on the blue copy button below an image on civitai to copy all Settings, even the hidden ones.
Then go into auto1111 and paste the whole thing into the positive prompt box.
Then press the blue/white arrow on the right to auto apply every setting.

[atomic mortar]

could be that he takes the upscaled resolution as initial and the ai cant generate a proper image thru that

[warm junco]

You could provide an example in #📝｜prompting-help

[atomic mortar]

waiting for him to do that lol.

[warm junco]

oh i mentioned the wrong username xD sry

[mighty sparrow]

did u guys make an coin

[fervent thunder]

no that coin never happened lol

[gentle prairie]

any good way to make a 45° angle, a quarter angle? Like view on a bed, but not in front or on the side, but in the angle, i can't find any prompt that works

[oblique elk]

Best chance would be using a corresponding image as source for a depth map and use the controlnet

[gentle prairie]

i'll try to figure it out, thanks

now last question

i dunno why, my generation keep having little characters in the background, trying to do more or less the same thing than the characters in front, and i dunno what prompt would trigger that

[median robin]

hello

[quartz siren]

yeah trellis is for sure the best for image-to-3d/multi-image-to-3d, and even similar quality to closed source options. It's also very fast which is nice.

[gentle prairie]

is there a prompt to have different hair colors and it chose randomly for each generation? With the current prompts i have, it either goes blonde or white haired, but nothing else, even though i don't have any character lora or something like that that could influence it

[patent pivot]

can anyone point me in the direction of how to generate audio reactive keyframe values for deforum? All I can find is either missing links, lacks settings I need (or don't understand how to get desired results) or not compatible with installed python

[gentle prairie]

the one thing i hate about IA, you don't see the time pass, and you spent the whole day trying to tweak some prompts, only to realise that what you did at the beginning was better, lol

i should have just played my games instead :p

[lucid bobcat]

You can do this in Forge

[atomic mortar]

aka a wildcard

could do this in swarm too

[lucid bobcat]

I had exactly this problem with Forge/Webui. My generations would look like garbage compared to Civitai with the exact same settings. I then switched to Comfyui and finally got identical results.

[oblique elk]

This prompt in comfy would lead to different colored cats wearing either a robe or a hat.
"a {red | green | blue | purple} cat wearing a {hat | robe}"

just use a { } - Bracket and the pipe | Symbol for each option. You can also use longer options if needed

[atomic mortar]

in swarm you can just use random:blue,green,red

or even store it as a variable so if the boots are red so must the gloves instead of having blue gloves and red boots

[gentle prairie]

damn that's cool, i wish they could add that in the normal auto1111 webui :p

[atomic mortar]

you could start a feature request but i think theres a A1111 extension that does that

though i left a1111 and never went back after swarm

[oblique elk]

https://github.com/adieyal/sd-dynamic-prompts is the extension

Has the same syntax with brackets and additional option like to choose at least two etc.

[gentle prairie]

oh nice thanks!

Can i use the syntax immediatly, no need to configure anything? i don't have time follow the whole tutorial right now

[oblique elk]

You would need to install the extension. Then it should work.

[gentle prairie]

yeah of course, lol

that's what i meant, thanks :p

[broken seal]

It’s not only about settings, could be due to the workflow, the sampling method, the nodes they use…. There are many different sampler, eg. I use often the RES4LYF, and have different results from original one.

[lucid bobcat]

It's much deeper than that. There is something severely broken with the way the Flux model support was implemented in Forge UI. The results I got were really bad, but not in an obvious way. I only realized once I saw on Civitai how much better Flux images are supposed to look. It's very difficult to explain, but it's like the difference between Flux dev and Flux schnell. You wouldn't realize you get bad quality until you saw what good quality looks like.

[broken seal]

Never use ForgeUI, only ComfyUI

[lucid bobcat]

Yep. But I wish Comfyui would let us easily create a user interface for our workflows, so we don't have to pan around the whole work space. I also wish we had a better editor where we could also paint onto images. The custom nodes I've tried aren't great. A combination of Invoke and Comfyui in one package would be awesome.

[fading bloom]

Does anyone know a good resolution for AI?

Like, which one is really good?

[west osprey]

Anyone know if it's possible to run OneTrainer SDXL Lora with 8gb 3070ti? I searched Google but none of the solutions worked for me. Always out of memory error

[lucid bobcat]

Multiple of 64 and square resolutions are better than very widescreen.

[fading bloom]

Is 864x864 good?

including which one do you use?

[lucid bobcat]

There is no clear rule. It depends on the model and on the type of image you're trying to make. Usually just for testing different prompts people use 512x512 because it's fast.

Oh and by the way 864 is not a multiple of 64

[fading bloom]

I know, I asked because that was the one I was using before

now I'm using 960x960

[lucid bobcat]

Some user interfaces have good resolution presets that you can use.

Sometimes you get better results with lower resolutions and upscaling afterwards. Because the higher the resolution, the more likely it is that the image can feel empty.

[copper crystal]

thats kind of legacy advice. it meant more in the early days around sd 1.4. generation code and models have since been updated to behave better across more resolutions. NovelAI introduced resolution buckets in training code, and base models since then have all been trained wiht bucketed resolutions. Allowing the models to know non square resolutions a lot better.

All of the SD1.5 refines use bucketed training sets afaik

[copper crystal]

There's no one resolution that's good. IMO. I prefer 4:3 ratios, or 3:4 if you want the other orientation. I often do the 2 stage generation, 1 with the lower resolution and 20 steps. A second img2img pass with higher resolution, 0.6-0.7 denoise value, and less steps

[fervent thunder]

Flux can handle even 6:1

its crazy

[calm hearth]

I mean you can never go wrong with 9:16, that’s the standard resolution for portrait videos on mobile

[copper crystal]

yeah. sticking to square resolutions like it's a technical limitation is kind of bad advice. you're welcome to stick to them for preference reasons but it's not necessary

[ancient cedar]

Hi guys. Im just now getting into AI, I have an m3 macbook air. Are using ai models possible on macbooks? Or are majority limited to windows. Im a total newbie- thanks.

[copper crystal]

https://github.com/LykosAI/StabilityMatrix/releases/tag/v2.8.0 try this. it'll offer you lots of different packages to install and try. i think it configures the UI's to use mac ML silicon

[ancient cedar]

Thank you. Would examples such as photo to video be possible in given package?

[copper crystal]

possibly. It'll install comfyui for macos. beyond that you'll need to set up comfyui for video and download the appropriate models

https://civitai.com/models/968568?modelVersionId=1097378 here's an example workflow. i'm not well versed in the video realm, but i believe cogvideox is the leading local situation. possibly hunyuan from tencent. You can at least get comfyui going but you'll have to study to get video set up if it'll work at all