💬｜general-chat

[median swift]

Just an upscale at 0,3 denoise, but still...

[fervent thunder]

ah yeah that would do it

fancy slow samplers is another way

[desert dagger]

been there - have kicked off renders that too several days, and that failed right at the end

[marble crown]

Hi guys, to install sdxl on automatic 1111 is there any video link?

[desert dagger]

https://youtu.be/cTNwSAlY8bg?si=CBsJk2MJFSMu1Aeg

[median swift]

Finished and it turned out allright. Pic posted in the images-chat.

[low moon]

Where is the failed diffusions room? 😦

[wet grotto]

T2I adapters (controlnet) are only for SDXL?

it's how is called the controlnets for XL?

[fervent thunder]

mistoline

[wet grotto]

I don't understand it... SDXL Demo? is not only change the checkpoint to xl models?

[wet grotto]

my god it's infinity, new models

[median swift]

Why is every model I've looked through so absolutely worthless on rendering weapons? Axes, swords, pistols, rifles... Just about everything is messed up way beyond recognition. The only Lora I've found on the subject refers to lightsabers, and it works well. But trying to make a hunter with a rifle on his shoulder, or a medieval swordsman... Absolutely impossible.

[desert dagger]

he wanted a tutorial on how to install sdxl on auto1111

[desert dagger]

becaue stable is very weak in that area - that's what loras are for

[median swift]

Do you know of any loras that might do it?

[desert dagger]

nope. i think, quite frankly, you might have to make one

people tend to make overly well endowed anime girls, rather than realistic weapon loras

[fervent thunder]

https://civitai.com/search/models?modelType=LORA&sortBy=models_v9&query=rifle

[median swift]

I suspected as much. Unfortunately I have neither the time nor the knowledge. I can photograph guns (since that is part of my job), but someone would have to build the lora.

But gun and sword pics are plenty on the net.

[desert dagger]

it's not that hard, you could learn it easily. but contact DucHaiten - look for him on civit or ping him on his discord - and maybe he might make a model

[wet grotto]

I didn't know you could install that

[wet grotto]

you can use ip-adapter too

[fervent thunder]

there's around 70 rifle loras

in the link I sent

[wet grotto]

using the image reference to copy the style

[median swift]

Thanks. I got 12 results by that link, all of them hentai based...

[fervent thunder]

you've got adult filters on
and I suspect it is blocking the weapons

there should be exactly 68

[median swift]

Adult for me is allowing weapons and blocking nsfw... :p

But I get it. Will adjust.

[wet grotto]

man, what is diffusers?

I dont understand the theory

[fervent thunder]

diffusers is a library made by Huggingface

[wet grotto]

everything is a diffuser?

[desert dagger]

https://huggingface.co/docs/diffusers/en/index read through this page

[wet grotto]

checkpoints, controlnet, loras

[fervent thunder]

the name is a joke
its a reference to their previous library called Transformers

[wet grotto]

thanks

[still ridge]

Hi all, I am new here.

I have a question. When I rent a GPU that has a AMD CPU, interestingly SD uses mostly CPU (GPU works partially and very little) but if the rented device has an intel CPU then SD uses GPU fully.

I am thinking that there will be something special for AMD cpus. Any idea/thoughts about this issue?

[desert dagger]

you've jumped into the deep end of a swimming pool before learning how to swim. take a deep breath, you're overwhelmed right now.

[wet grotto]

moar moar rss

Ok, thanks

[fervent thunder]

how are you measuring usage

[wet grotto]

later I wil read the article

[still ridge]

I can see these usage on the website's dashboard. And I also uses this command watch -n 1 nvidia-smi

[desert dagger]

i would check your settings, actually - you want the GPU to be the most utilized - its much faster than the CPU - AMD or Intel

[still ridge]

what kind of settings for example? I could not find anything. There was a settings in SD which were CPU, GPU and NV. I tried GPU and then NV but still same..

[desert dagger]

where are you renting from?

[still ridge]

vast.ai

[desert dagger]

why them and not AWS?

[still ridge]

I cant send here a SS. I would like to show what I was seeing

[still ridge]

price?

[desert dagger]

use #🏞｜general-with-images

[plush vale]

gm

[cyan matrix]

Hellooo ✨ I'm using AUTOMATIC111111 Stable Diffusion on github. How can I reduce the video rendering time? Is there a faster version of Stable Diffusion?

[desert dagger]

use a faster gpu

[fervent thunder]

when you say video
what model are you using?

[cyan matrix]

Thank you for your response ✨ What should I use? What do you recommend? Also, when creating a video, the duration of the first video is very long. it decreases even more in the second video. what is the reason for this?

[still ridge]

maybe cache?

[desert dagger]

are you trying to generate on github itself?

[cyan matrix]

I try it on local

[fervent thunder]

could you let us know which model it was

[desert dagger]

oh. well - in that case, if you're generating local, you buy a better machine.

[cyan matrix]

what do you recommend as a machine?

[cyan matrix]

beautiful realistic

[desert dagger]

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

[cyan matrix]

I want to rent a machine. what do you recommend such as NV 4090, A100? and which company is the most reliable that will work smoothly? where can I rent?

[warm junco]

Hey, what's your GPU?

Also do you use xformers for better performance?

[desert dagger]

i'm afraid as far as renting goes, that's probably a question for @warm junco

[lilac flume]

understood, tks

[ancient urchin]

Anyone here knowledgeable with comfy ui

[fervent thunder]

if you need it set up for you then Rundiffusion.com
otherwise sites like Vast ai

the following thing helps:
tick the box "Secure Cloud (Only Trusted Datacenters)"
and in the drop down, select "On-Demand" not "Interruptible"
and make sure Unverified Machines is not ticked
this will get you much more reliable machines

colab, azure, aws are a middle ground between vast ai an rundiffusion

[grave forge]

Hey I'm trying to generate some 3D versions of characters but it keeps giving me 2D how do I get the specifications of what I Want

[desert dagger]

are you trying to generate 3d mesh objects?

[grave forge]

I'm trying to generate 3D character images based on the prompt given but it's only giving me 2D

[desert dagger]

and you used "3d" in your prompt?

try adding "3d shading, 3D character" to the very start of your prompt.

[grave forge]

K

[grave veldt]

who here use amd in stable diffusion?

[desert dagger]

hopefully no one

[grave veldt]

i have amd but its so bad in stable diffusion bro, 2 hours for generate a image

[mortal delta]

have you tried comfyui and directml, that is if your on windows?

[grave veldt]

i used zluda

[mortal delta]

zluda dosent always work for every amd card is only supports certain cards (manly newer ones) ive tried comfy + directml and i found that that worked the best on an amd 580 with 8vram, my amd 580 isint supported by zluda. some modified version of 1111 also work from what ive heard.

[grave veldt]

mine is xr 580 too

you have a tutorial how to installe comfy?

[mortal delta]

well i found that directml worked the best on windows, im not quite sure what works on other platforms but ive heard good things about linux and rocm.

[grave veldt]

mine is windows

[mortal delta]

that you will need to fine on your own, sadley im trash at giving suggestions, but avoid the one click installers and use git clone then you will need to open it with a python command something like python main.py --directml --lowvram.

[grave veldt]

comfy and directml is the same?

[desert dagger]

that's casue your cpu is trying to do the grphics processing instead of letting your GPU do it

[desert dagger]

go to youtube and watch scott detweiler's tutorials

[desert dagger]

no

scott works for stability.AI - watch his comfy tutorials

[grave veldt]

what is your suggestion? comfy or directml

[desert dagger]

comfy is an interface that you run stable diffusion with. directml is https://microsoft.github.io/DirectML/ and has nothing to do with comfy or stable diffusion

[grave veldt]

ohhh

okay

[desert dagger]

and i doubt you need it

[grave veldt]

I didn't find the tutorial to download it

[desert dagger]

you didn't find what tutorial to download what? here https://www.youtube.com/@sedetweiler this is scott's channel. look through his videos till you find his beginner comfyUI setup tutorial

[grave veldt]

comfy bro

[desert dagger]

start with this tutorial https://youtu.be/AbB33AxrcZo?si=qpsTtfYlP0qiguks

[grave veldt]

i just need to download comfy in github, extract, install manage, run cpu and watch this link right?

[desert dagger]

you need to start by watching that tutorial, pausing at every single step, and working along with him. he'll walk you through installing it

make sure you understand what he's doing before you do it, too

[grave veldt]

but he dont teach how install comfy

[desert dagger]

that. tutorial. does. yes

[trail lion]

when walkthrough videos arent enough

[desert dagger]

yeah

[trail lion]

we need holograms

[desert dagger]

normally i would handhold but i'm too tired

[grave veldt]

are you sure? I watched and he just taught me how to configure the options

[desert dagger]

there should be install instructions on the github page, try those

[grave veldt]

ok

the vae of comfy is good or i have to download one

[fervent thunder]

use baked VAE

[grave veldt]

I managed to download comfy, I will generate an image now

[ionic ocean]

so I want to get into ai animated videos. Ive used a bit of contorl net stable diffusion for images in the past. is stable diffusion video the best option? if so what are the other optoins and what are they good at?

[ionic ocean]

speaking of which do people still use automatic 111 webui?

[desert dagger]

yes

[ionic ocean]

like where do i start with learning how the video side of how things work? it seems a little overwhelming / all over the place not sure where i should start

[desert dagger]

you might want to start by watching youtube videos on video production itself. you're not going to learn the basics of how to make good AI videos by just learning how to create them with something like stable diffusion. Video production is an entirely differnet skill set

the AI can act like the video camera, but you have to learn how to be a screen writer, director, etc as well

[plush vale]

gm

[ionic ocean]

Im certainly not a master in that regard but i did study animation, so i have a basic idea of what makes a good story, storyboarding etc. i have a background on art - particularly 3d. i can assure you I'm not some techbro trying to make a quick buck lol

more just trying to understand the technical side

[desert dagger]

you've got a leg up on this then 🙂 most people are rather clueless and just know they want to make videos.

[desert dagger]

you might look into animate_diff for stable

you might talk to @warm junco

[lusty relic]

It is my or SD3 models are crap?

[loud solar]

Basic models aren't finetuned, yet ...

[vocal leaf]

Yeah cuz of the licensing shit

im just waiting until I can run face detailer and basic animations

so far im getting errors

so pytorch still aint supporting the AMD cards :-:

but everything else works perfectly on ZLUDA

just that some nodes dont work

[median swift]

In the wake of Suno, I'm hearing about some people using a music version of Stable Diffusion.
Anyone know a local platform music AI?

[bleak matrix]

Good morning, everyone! How are we today?

[rough apex]

hello guys im new to this stable diffusion thing, can i ask questions here?

[sudden ruin]

Please use #1092446741984444416

[mystic siren]

Does anyone using automatic1111 know If with the new update the speed generation is as fast as forge???

[warm junco]

Reactor node works in comfy but requires a manual fix

[low moon]

Bring back the failed diffusions channel. (and cascade and sd15) XD

[desert dagger]

it's you

[worldly sand]

Hi guys, where will I be able to make a logo? Sorry new to stable diffusion

[rose cargo]

Not sure if this is the place to ask, but is anyone else having issues with inpainting in the newest Forge version? Everything comes out as a blurry blob

[desert dagger]

when was the newest forge version released?

[rose cargo]

Just a few hours ago I believe, I ran the update file today and it did some things

[desert dagger]

might send the devs some screen shots and a bug report

[rose cargo]

How does one do that?

[desert dagger]

where did you download the update from?

[rose cargo]

When I downloaded Forge it came with an update file that you double click and it automatically updates

[desert dagger]

where did you download forge from?

[rose cargo]

From Github I believe

[desert dagger]

go back to that page, click on issues (at the top above the code) and open a new issue

[rose cargo]

Alright thanks

[rain pecan]

Hey!!, I read that SD3 is unbanned on Civit and now is also compatible with A1111 web ui. I can't seem to select SD3 as a checkpoint in web ui?

[versed saffron]

hello. how do i make stable diffusion make art of a person in an oil painting style without drastically changing their facial features? I use control net canny but it just doesnt give the leevel of accuracy i need

i have a photograph that i want to convert into an oil painting style

[bold haven]

hi everyone

[pale latch]

It's tricky. I put the sd3 checkpoint into the models folder and the three tenc models into the clip folder. Which is weird since t5 isn't clip but i digress. I couldn't get it to use t5 anyways since it wants to hold the thing in memory the whole diffusion so it breaches my 16gb

There's a setting in the stable diffusion category to use t5 or not

[rain pecan]

I did not have a Clip folder under Models? Is that normal

[pale latch]

I'm not sure. My models folder is so old and legacy. I might've created mine at some point. It's the folder that a1 will look in for the tenc weights

[rain pecan]

Oh, even I have decided to make a folder and put the checkpoint under that folder

[pale latch]

It should download the clip models if you don't have them already

The checkpoint goes in models/stable diffusion

https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/16030

[rain pecan]

This is what I was referring to

Yes, I can see the checkpoint in web UI, I just can't seem to load it.

[desert dagger]

got enough vram?

[bold haven]

guys if i use 6 go gpu how slow will the image generation be compred to 8go or 12 go

[desert dagger]

check back to see if it's finished in a week or so

[bold haven]

what's the min perfs to get results in few seconds

[desert dagger]

that totally depends on how many calculations you're having it do. how many nodes you have loaded, etc

[bold haven]

imagine i want to do image expansion.
how many seconds it would take on average with 6go, 8go, 12go .Etc

or generative fill

i'm going to buy a new pc this week so i'm trying to find the best perf that fit my budget

1000 euro

[neon jungle]

Hello ,

so i'm an architectural designer , i need someone to give me one to one coaching on the whole SD thing , i've been searching around and trying to figure things out , i just feel lost .. who is intrested in coaching me and answering my questions , ofcourse i'll pay for the service 😄

[desert dagger]

you're not likely to get anything that will render quickly for that. you want an Intel CPU, and an nvidia gpu. and a min of 8gig of vram, but 16 is better, and 32 is even better

[bold haven]

what bout ram does it affect the generation a lot ?

[bold haven]

i found a pc for around 1000
gpu 3080 8go vram. how fast an image expansion will be on average

[desert dagger]

there is no way to tell that. will you have a short prompt with one ksampler, and almost nothing for the machine to do? or will you have a long, convoluted prompt, a lot of negative terms in the negative prompt, ksamplers chained together, other complex nodes involved?

there is no such thing as 'average'

[bold haven]

i found what i was looking for https://www.youtube.com/watch?v=4YLHGEVeVpk

[pale latch]

Memory size allows you to process more parameters and pixels at a time. Things can go extraordinarily slow if you don't have enough memory. If you're working with sdxl resolutions, you want more than 8gb to fit it all. But how much the card can manage a time isn't what defines the speed of the process. A 4080 is faster than a 3090, but it has less memory.

I might be wrong about that last point, but the idea is the same

AMD cards have a lot of vram but are quite slow

[desert dagger]

and they like to take over the graphics processing

[keen kraken]

hi im trying to work with a style lora, but it doesnt seem like the lora is always activating

[pale latch]

Look at the prompts the author uses in examples and try to copy those.

Speaking of that I need to remake lora example images with metadata

[keen kraken]

i noticed that he uses source_anime in all his examples, could that be the thing'

let me try that, remaking based off his prompts

[trail lion]

You just reminded me of the "got milk" campaign. We need T-shirts that read "got vram?"

[bold haven]

what do you think of this pc https://www.cdiscount.com/informatique/ordinateurs-pc-portables/pc-portable-msi-creator-m16-a11uc-850fr-16-qh/f-1070992-9s7158242850.html#mpos=0|mp

[desert dagger]

where's the nvidia GPU?

[bold haven]

PC Portable
MSI Creator M16 A11UC-850FR
16" QHD+
i7-11800H
RAM 16Go
Stockage 512Go SSD -
RTX 3050 - Windows 10 - AZERTY

[desert dagger]

yeah - no. it doesn't have nvidia AND it's windows 10

[bold haven]

it says rtx 3050

i thought that means it's nvidi

[keen kraken]

are you looking to buy a laptop with a dGPU

[desert dagger]

get something like this https://www.newegg.com/abs-aqa14700kf4060ti16g-stratos-aqua/p/N82E16883360436 unless you absolutely must be portable

[bold haven]

i'm looking for a pc that allow me to run all the automatic 1111 models

[keen kraken]

so i tried remaking with the lora data from the makers example and its similar but the lora style didnt activate

[bold haven]

i'm planning to integrate them in websites i will code

[keen kraken]

dont go for anything lower that the 60s

[pale latch]

i think its a notebook and not a pc

[keen kraken]

im able to run on a 2060 but my gpu is better that a 3050

[desert dagger]

automatic1111 is just an interface - you're looking for a machine that can run stable diffusion

[pale latch]

check the console to see if there are errors about the lora

[bold haven]

it's a pc it says it in the page

[pale latch]

it's a companion device

[desert dagger]

it's a portable pc - i.e. laptop - and it's windows 10

[keen kraken]

can i post the console info, dont know what im looking at

[bold haven]

yeah i know i mean running all the popular model except the llms

[pale latch]

i was asked for my opinion on it. it's not a pc. that's my opinion

[desert dagger]

it's also a page that's not in English - they mean the same thing you do

[pale latch]

if all you've ever known is mobile components, you'll be fine with it. just don't taste the true nectar cause it'll ruin it forever

[desert dagger]

if you actually want to do something serious, not just play with stable as a hobby, you need a desktop and a decent sized monitor

[bold haven]

i meant if it would easily run the popular models

[desert dagger]

and a couple terrabytes of storage

[keen kraken]

it could be a possible error with the actual file when i downloaded it, because with all my tests ive had it work really badly once

[pale latch]

i think it's a good companion device for managing models while not at home, but i'd sooner run a server on a more capable device and remote into it

[keen kraken]

your gonna need alot of storage if you want alot of models

[desert dagger]

and you're going to want to store them on an external drive. portables like that are NOT designed for serious graphics work

[bold haven]

i can just delete the models i don't find intresting. i want to test many of them till i find the best workflows to generate thhe content i need

[desert dagger]

jsut generate with them on civit then

[pale latch]

if the file was corrupt it likely wouldn't load into memory and the console would show that the lora didn't work. it wouldn't load and then cause a bad generation. though once loaded, the ui code can corrupt memory and you'll have to restart the UI. happens to me a lot when swapping loras on different models

[bold haven]

online they're token based

you get out of tokens super fast

[keen kraken]

yeah but i only ever used 4 or 5 models when i generate, based on the type of things, like one model better for humans and ones better for realistic

[keen kraken]

ill try restarting my ui

[pale latch]

i realize that it is technically by many definitions a personal computer , and that it runs windows. I just find notebooks to be sub par and it's a market wher manufactuerers will cut important corners and do shit like shared electric lanes

[desert dagger]

you asked our opinions. our opinions are you'll be waisting your money if you buy that

[pale latch]

in my opinion, a notebook isn't a pc. not in today's market.

[keen kraken]

on another note, what are good negative prompts, i feel mine are bad

[bold haven]

ah. why people buy that type of pc then what are they using them for

training models ?

[pale latch]

not just a browser refresh i should note. console reset. i use ctrl+c to end the process usually. that's an oldie terminal trick. write that one down.

[bold haven]

i copy mine from seaart ai

[desert dagger]

they usualy buy those to 1. play games on and 2. do schoolwork or job work on. not graphics, and not serious graphics

[pale latch]

lots of reasons why people buy a notebook. a lot of it comes down to marketing

marketing lies

[keen kraken]

i just close the webui when my generation tasks become to much of a lovecraftian blob

[desert dagger]

you are not training a model on that, or anything close to that

you're not even training a lora on that most likely

[pale latch]

"closing the webui" is an ambiguous thing to say. it's a terminal program that is running a local web server. you then access that local web ui through your browser.

are you closing the terminal, or the web browser?

[keen kraken]

ok lets ask fully, what exactly will you be doing on that laptop, generating models, remote server, training, playing games. what exactly @bold haven

[keen kraken]

so with my webui the terminal and the webpage are always open, i close both

if i open another with the prior console open my history becomes a second instance

[desert dagger]

1. close everything 2. do what he told you to do

[bold haven]

i will create a better UI to create content faster. the faster you do the easier you trend on social media

the ui will combine many models to create a specefic result i want

[pale latch]

if all you've ever known is notebooks (using them on lap gives you nut cancer, jk but kinda not really) , you wont care about the boosts you can get on a desktop thats built to purpose. same way a person driving a civic doesn't ever think they'll want a truck. i'm not sure, is a desktop pc a truck or a lambo? maybe a production drafting cad machine is a truck, and a gaming pc is a lambo. so an ai pc is a cybertruck?

[keen kraken]

but thats just any UI

[pale latch]

my analogy fails sorry

[bold haven]

nope i will code mine as it fit my needs

[pale latch]

closing both works. the actual program is the terminal. that's what uses all the gpu resources

[keen kraken]

i understood it dont worry

[bold haven]

for exemple if you want to change the outfit of person and you need to draw a mask one the part you want to hve the best result you can't do it perfectly on automtic 1111

if i code that functionnality allowing a better segmentation of the region you want to modify it will be an improvement

[pale latch]

i felt bad about the cybertruck example since the cybertruck isn't a midpoint between a lambo and a truck, but only has the worst of both and does nothing that either can do

[keen kraken]

honestly you dont need a top of the line thing, my laptop it getting on in years in terms of pc spec

[bold haven]

imagine you want to combine a model generating a 3D model and live portrait + animation models

[keen kraken]

and i can generate decently

[bold haven]

what's the spec of your pc

[keen kraken]

ive never actual generated any ai vids

[pale latch]

can do

[bold haven]

i'm a content creator nd that shit is trending hard is posted at the right time

[pale latch]

if you've got a 3d model though, it'd be easier to map the facial expression data that liveportrait or live emoji would use, and just map that directly onto the 3d rigging

[keen kraken]

Ryzen 7 4800H
24gb RAM
256gb NVME (main drive) (OS)
2Tb NVME(second drive)
RTX 2060 6GB

[pale latch]

the expression transfer stuff is so legit

[bold haven]

that was just one exemple. code + ai models = insane combination

you can do so many things

[pale latch]

comfyui lets you create node graphs that are essentially code, and you don't need to learn code

[keen kraken]

if you are planning to generate 3d stuff which can take longer, maybe the 70s models would be better

[bold haven]

in the video i posted above it showed that the 3080 is almost twice as fast as 2060 when generating details

[keen kraken]

especially if you make your own proper 3d stuff in blender and all that

[pale latch]

you can do an animate diff on someone's portrait, make their face move a little around, something easy that animate diff can do well, then use live portrait to sync their face to a new performance. ugh its so crazy. i really hope amateur production takes off and we have a new all your base period

[keen kraken]

yeah but i can still generate SDXL stuff in below 2 mins time with High res fix

[bold haven]

that was just an exemple
one of the main thing i want to do is animate manga panels as soon as they come out
or have workflow making it easy to color them

[bold haven]

how much that kind of pc cost

i'm wondering if your 24 go ram played a big part in it

[keen kraken]

at the time i bought it during silicon shortages and all that so it was around 1200

[keen kraken]

that 24 is just normal ram, you need to look at Vram

[bold haven]

i know it's normal ram that's why i was wondering if it played a role in the performance

you only got 6 go vram and i saw lot of videos on youtube saying that's just the minimum and it will be slow

maybe their definition of slow is different

[keen kraken]

yeah but most of the rtx 50s model all sit at 4gb

maybe some of the newers are higher

and ai generating is all in the Vram, not really the normal ram

thing of it like games, High Vram will make games faster

as a bad analogy

[keen kraken]

@pale latch it seems it was just the webui being difficult and it just needed a restart

[desert dagger]

okay. sounds like a good projct. the UI isn't going to do much, though - it's whether the machine can run the program that matters

[pale latch]

i get that all the time with lora's on newer versions of a1. my guess is when loading them in and out of memory, something goes haywire.

[keen kraken]

i guess after constant generating it needs a reset before you do something new

[pale latch]

something is getting unsynced and going wrong. i blame pony models. it always seems to be when i'm testing loras on various pony realism models. the realism models take sdxl loras and controlnets a lot better than base pony does, since they're just merges with existing realism models in most cases.

[desert dagger]

system cache not clearing...

[plush vale]

yo i'm looking for a mod for my project paying $600 a month

[pale latch]

yeah something like that. trainwreck ensues. maybe i'll try the reforge project

[pale latch]

anyone thinking this might be easy money should take note, common scam. they pay you to do work, but the money never actually comes through. or they pay you too much and say some has to come back so you send some back, but the original payment was spoofed and never existed. $600 sounds a bit raph. a little too raph.

[woeful quest]

How many gb does the whole thing take

And does it increase hh if I generate images??

[plush vale]

You're wrong

[desert dagger]

got taken by those guys and don't want to admit it, did you?

[woeful quest]

Please

[plush vale]

What are you talking about?

[keen kraken]

how would you react when a new guy randomly joins a discussion offering money for a small job\

[desert dagger]

one that's a fairly well known scam

[slim jewel]

Then how do you actually get people?

I'm similarly new to the server and looking for someone to help build a stable diffusion pipeline for me, the awkward reality is both sides have to risk something

In my case if I get someone with a portfolio, I can pay partially up front as a good will gesture, but not many people have those

[keen kraken]

can i ask what a pipeline is? not to sure on all the technical terms

[hasty hornet]

kinda funny how every scammer now does web3, crypto, nft on their bio these days

yet never seen one explaining properly what his job is on public, it's always general smart words meaning nothing in particular

[desert dagger]

https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/overview read through this page

[desert dagger]

first of all, most mods are users already, and are invited to moderate because they've shown themselves to be passionate about the community. no one just goes out and says "hey, i'll pay you to moderate" - also, he said he wanted a "mod for his project" - really? that's a very odd thing to say. not a dev, not a customer support agent, nothing like that, a 'mod' - a moderator - for what? nothing he wanted to explain. and he wants to pay - 600 bucks a month? mods are almost 100% volunteer. even for huge companies like adobe - it's volunteer

those are huge red flags for those of us that have been doing this a while

[desert dagger]

yes, but see what you said? you are looking for someone to help develop, you went into detail about what you need - someone that can help you build a specific thing. And you're up front about the fact that you'd like to pay but might not be able to pay much. that's a whole lot different than what he said.

[hasty hornet]

I think either way there are better places to hire \ fidn a job, random messages on discord rarely ends well

I'd straight up ban this

[slim jewel]

pipeline is just a sequence of tasks to complete a goal

In my case I need to generate arbitrary characters in similar poses

[pale latch]

pipeline to deez balls

[desert dagger]

that's what controlnet openpose is for

[low moon]

No but seriously where is the failed diffusions channel?

😢

Was eveyrone posting their SD3 failures there and emad pulled te plug? 😄

[desert dagger]

thre's a facebook group for that, if you want it

[high flint]

So, I've been away for a bit, (used 1.5) and now I see SDXL and Pony? as an option. I'm still set up on Automatic1111 but I'm not sure if that's the smartest choice to go as far as generating images goes.

[low moon]

Come to Noodletown.

[high flint]

Which UI should I go for overall?

[low moon]

Don;t make my mistake...

I onyl regret resisting the noodles for a few months

[high flint]

noodletown?

[low moon]

Comfyui.

[high flint]

ah

What advantages does it have over automatic?

[low moon]

Yes now Pony and SDXL are the in thing. And Kolors and Cacade/hunyuan/pixart sigma if youre more into it deep

SD3 is ok - ish for backgrounds but nothign else for now

It has many advantages

it runs faster and its more optimized

and you cna combine things almost endlessly

[high flint]

well, just the fact it runs more optimally is a big one off the bat

[low moon]

yep

[high flint]

and by combine things endlessly, what do you mean by that?

[low moon]

Well you can start a generation with pony, refine it with sdxl, hook up 5 controlnets, refine the result with sd15.

its really infinitely more versatile than A111/forge and foocus

[high flint]

How's the UI in terms of organization?

[low moon]

you get nodes which you hook up with "wires". kinda of like an old Moog synth

[high flint]

I should also be able to just copy and paste my already downloaded files over to comfyui as well right?

[low moon]

You mean models?

[high flint]

yeah, and lora's

[low moon]

yes you can use them in comfy as is

to use a lora and checkpoint you just need a very basic workflow

the default one will handle checkpoints

[high flint]

Yeah... I kind of like to mess a ton with different art style and building loras haha

[low moon]

In fact you can use any model in there already, while with A1111 and others it takes a lot longer for new models to become available

[high flint]

I'll give it a shot and see how it turns out

[low moon]

😉

[high flint]

I honestly love to use AI for sci fi images

if you keep at it, you can end up with some pretty cool results

[low moon]

yep

[desert dagger]

welcome back.

[high flint]

which should I go for, pony or sdxl?

[wet grotto]

pony is to good as sdxl?

[low moon]

if you want overall better quality than SD15 then use SDXL

if you want anatomical correctness then Pony

[wet grotto]

and about kolors?

[high flint]

I see, so if you have characters or persons, Pony is better. SDXL is better for landscapes.

[desert dagger]

what about?

[desert dagger]

actually - yes/no/maybe - you should really just try all the models and see which you like best

[high flint]

oh, does ComfyUI have remote hosting?

[fervent thunder]

kolors actually has no commercial license at all LOL

its enterprise only

[high flint]

because I may run that on my other server that's running a lightweight text only os of debian

to maximize the performance

[wet grotto]

it's more used to what?

[low moon]

Yeah Kolors is very nice, in many ways almost as good as SD3 could be, but only for personal use really.

[desert dagger]

it's just one more new ai image generator. it's nice, it's not the best, i'ts not the worst.

[high flint]

Oh yeah, I can work with this in the comfyui passthrough

super helpful readme file

shoutout to the dev for making a good readme file

I've seen bad readme's so often that I have to appreciate the good ones when I encounter them

going to be fun to set this up on my servers that are connected to my Arista switches

[fervent thunder]

I love kolors style so much

but its only one style

[hearty grail]

is anyone having issues opening forge today? 😦

[low moon]

I haven't used forge in a while but I heard it's not being developed anymore.

[high flint]

Does Comfy UI use the system file paths?

[low moon]

It's complicated.

I recommend installing the Manager first.

[hearty grail]

I've used it after that with no issues, but today doenst wants to open anymore :/

[low moon]

that will help you install things.

[high flint]

I meant such as uh,

[low moon]

Lemme try

[high flint]

Oh, cant attach images here.

posted in the other chat

[rose cargo]

Can running out of storage space bug out SD generations?

[desert dagger]

yes

[rose cargo]

I'll try cleaning up and deleting a bunch of stuff then, hopefully that fixes it

[low moon]

I can run Forge still. Version 0 1 7

[high flint]

nevermind, last command didnt run. lmao

oh well, seems to be installing properly now.

Also, does/should python be updated as well?

my current ver is 22.2.1 vs the 24.1.2 I could update to

[desert dagger]

you want python to be higher than that. you might want to as in #🤝｜tech-support for recommended levels

[abstract linden]

What's the general consensus on workflows? Are they super important/Valuable or are they just kinda nice to haves?

[fervent thunder]

its more important to know how to build workflows than to have super complex ones

because think how the workflow itself is the abstract brush that paints the entire potential output space

even if you don't know how to make super big and crazy automations, you can still navigate and build things that interest yourself

[high flint]

Which python file should I run to run comfyui?

I'm not sure if I'm doing this right but I'm trying to execute the "main" file in the comfyUI

to launch the webgui

[desert dagger]

you run the .bat file

[abstract linden]

Interesting. Currently (to my knowledge) I don't have any workflows. I am unfamiliar with the concept, so I was wondering if it's something that would create a dramatic difference, but if it's more advanced, then for now I will keep fine tuning my skills.

[high flint]

Funny, I dont see said bat file

I'm guessing the install went borked

[desert dagger]

load this playlist https://www.youtube.com/watch?v=AbB33AxrcZo&list=PLIF38owJLhR1EGDY4kOnsEnMyolZgza1x

and watch his tutorials

[high flint]

I'll gladly watch those, however those seem to not have the install process

Yeah, the install from terminal missed the batch files

grabbed the download version of that and just moved the gpu batch file over and ran it and it works like a charm

[rose cargo]

Even after freeing up a lot of space on my disk, inpainting still only gives blurry results 😔 Did a new update change the way inpainting works or something?

[low moon]

From what I know how much free space you have should not affect your inpainting results...

[desert dagger]

can you post your settings?

[rose cargo]

Inpaint tab settings?

[desert dagger]

at least that

[high flint]

going to take a bit of time to get used to comfy, but already I can see huge upsides

[low moon]

[high flint]

what's the new name for nested nodes?

Group nodes... of course.

[high flint]

I'm just going through the videos trying to understand exactly how and why he's using each node here, it's kind of wild

[desert dagger]

Not really

[high flint]

Yeah, found out they are not very useful anyway

too big and limits what you can do

[nova lily]

So when you leave hires steps at 0 it does wateer your base step count is IN HIRES?

Is leaving 0 the best? I heard people if they do 30 steops, they put the base step at 20 and hires steops to 10 to total 30. and stuff like that

[rain pecan]

Can someone help me understand what is the difference between A1111 and Forge Web Ui ?

[neon jungle]

Hello ,
so i'm an architectural designer , i need someone to give me one to one coaching on the whole SD thing , i've been searching around and trying to figure things out , i just feel lost .. who is intrested in coaching me and answering my questions , ofcourse i'll pay for the service 😄

[wet grotto]

hey I'm architect too, I can help you

for free

what is your questions?

I have a channel on yt about sd and achitecture too, it's in portuguese but you can see the workflow and use subtliteds

https://www.youtube.com/channel/UCissXBD6LDXmf2WNxE_F15A

[wet grotto]

forge is more optimized for low vram

[neon jungle]

Hey dear @wet grotto , amazing , can i DM you

[wet grotto]

sure

[cobalt harbor]

hi everyone

could someone help me fix the blurryness in tech support ill be eternally grateful

[rain pecan]

I would suggest go on YouTube there are many mentors who can guide you for free one you know what you are doing look into patreon artists from your niche and pay. Most of the stuff is free and surely you’ll find a way among them no need to get scammed

[wet grotto]

@neon jungle I think he is right. It's better we won't make the call now,I have fear too, try start search videos in yt, looks for and search is part of learning

You can see my channel too is only about sd and architecture is what you ask for

Later you can dm me if have some questions

Try experiment first, is in this way that you will learn

[fervent thunder]

maybe a year ago you could get away without workflows
now you kinda need some combination of FreeU/PerpNeg/PAG/SAG/CADS/thresholding, CFG scaling/scheduling, IP adapters. control nets, multiple refiner passes, inpainting or automated detailing, noise injections and a single or multi-stage upscale

I made that list quite carefully to only include things that definitely work

those are the things that could be considered for any image

and then experimental stuff goes on top of that

[pale latch]

you can still get away with basic workflows or the default swarmui interface or the a1111 interface or foooocus.

Tools are as useful as their operator. Use what you're experienced and comfortable with. all those extras, not needed. Useful when applied properly, but not necessary.

likely won't be a general consensus on how to diffuse

[wet grotto]

I don't recognize almost nothing there rss

[fervent thunder]

Swarm is probably the ideal system for most people
because you could load a big workflow but then just never look at it
and do your generations in the A111-style GUI

its probably the closest thing to a best of both worlds

[wet grotto]

man I underestimate swarm, maybe I should give it a chance and try it out

[fervent thunder]

swarm has comfy inside it as a tab

but you don't have to look at the comfy tab

you can just use the regular generation tab

[wet grotto]

I will install and experiment

you show interesting points

[keen hornet]

Jesus Christ die for our sins on a cross, then three days later rose from the dead. If you believe in your heart and confess with your mouth that Jesus Christ is the Son of God. You receive the free gift of eternal life.

[fervent thunder]

otherwise try comfy itself but just start really slow

it has a default workflow you don't have to add anything at first

you only have to add a handful of nodes for Comfy to overtake other interfaces LOL

[wet grotto]

I started yestarday use comfy, I came from a11

comfy is really powerfull

[fervent thunder]

ye its really fun

[pale latch]

not really. large workflows with custom nodes, the don't integrate into the swarm ui so well. you can't use loras with them or a number of other options. often changes like that have to be added into the workflow and then imported back to the tab. while it is useful, i wouldn't call it "just like a1111"

. with sd3 workflows, the ui doesn't understand the triple text prompt node, so isntead of using swarm's prompt field you have to scroll the left hand menu, find the drop box for the node that takes 3 prompts, and prompt from there. often i find that depending on the workflow, nodes will be in different positions in the menu too. small annoyances. it's a false promise to tell people they'll never have to leave the swarmui tab to figure out workflows.

[fervent thunder]

I disagree I think there are a lot of decent workflows that will just run without leaving the swarm tab

you could setup the majority of this list and then just leave it at default settings:

cos a lot of this stuff doesn't really need input once the workflow has been saved into .JSON

[west pulsar]

Depict a gritty and poignant scene in black and white of the black african slavers celebrating their freedom

[fervent thunder]

#artisan-faq
you can generate here but you need to go to this channel

[cedar salmon]

hum auraflow 0.2

[plush vale]

yo i'm looking for a mod for my project paying $600 a month

[pale latch]

does it still do the safety images from the other model?

[cedar salmon]

no idea will poke around a little later

[fervent thunder]

they should leave it in as a joke

easter egg

[pale latch]

except change the sign text to "safety? what for? did you fart?"

[high flint]

Which Controlnet Preprocessor do you recommend? (I'm assuming this is a custom node?) I dont see that block at all yet I have the apply controlnet and the specific stuff like, Midas depth map preprocessesor provider (SEGS) and it's cousins

[raw bough]

is there free access to SB3

[pale latch]

smash bros 3? like smash bros brawl? sometimes if you go to a dentist waiting room they might have a wii u you can play on for free. might have sb3

[fervent thunder]

there is some new fancy union control net

I haven't learnt how it works yet

[high flint]

which models do you all like best for controlnet?

[fervent thunder]

mistoline for SDXL is good

but you should definitely look into the new union one

its meant to be the next generation of control net

[high flint]

well, I will once I get the hang of how controlnet works in comfyui

trying to learn all the steps bit by bit instead of just leapfrogging all the way

[fervent thunder]

yeah I haven't tried it yet either

[rose cargo]

Is there a way to revert all A1111/Forge settings to default?

[low moon]

Turbodong is the Spielberg of AI movies.

[high flint]

ok, I think I finally understand the proper way to use upscaling in this comfyui

I really like the way you can apply a lot of control on this

and automate a large sequence of things

[fervent thunder]

I don't follow movies

which movie is this?

[low moon]

Search for Turbodong on youtube...

[fervent thunder]

LOL I found it

[abstract linden]

@pale latch@fervent thunder Thank you both for the great information!

[fervent thunder]

no problem

[low moon]

People really hate noodles huh? XD

[abstract linden]

May I DM with a couple of workflow related questions?

[desert dagger]

hint: if you want to see what's in a model's data training set, then kick off a generation of at least 4 images, and do not give it a prompt or negative prompt. just leave them either blank, or use just a . as the prompt

[abstract linden]

Do people typically ever run with ComfyUI And say AUto1111 at the same time?

[fervent thunder]

that's kinda what swarm is

[fresh kernel]

Any good tutorials on how to start with Stable Diffusion? I've heard it's a very powerful tool, but I'm running into the newbie problem of bad anatomy on characters and etc.

[fervent thunder]

if its SD3 then we can't help

we get anatomy problems too

if its an older model then we can help

[desert dagger]

sure - tutorials all over youtube.

[fresh kernel]

gotcha, ill download sd2 or whatever older version you guys use

[fervent thunder]

SDXL

would be a decent choice

[low moon]

Pony is good at anatomy, and stuff.

[fervent thunder]

I don't like the internet drama surrounding pony and I don't do NSFW but it is indeed the best model for anatomy and probably will be for a long time

[fresh kernel]

Just trying to DnD settings really.

[fervent thunder]

the nature of how it was made makes it better for that

[fervent thunder]

you can run Juggernaut XL at 26 steps CFG 4 with DPM 2M++ sampler and that will do fine

[fresh kernel]

would 3gb of vram be an issue for that or could i just run it slower?

[fervent thunder]

oh this will be a huge issue

its possible to run stable diffusion with 3GB but it is specialist and I don't know how to do it

its definitely possible

if I remember rightly even 2GB is possible

but you have to use very specific methods

[fresh kernel]

Hmmm.

[desert dagger]

it's an issue

[fresh kernel]

I'll try to exhaust my options with the cloud-based stuff then if I have the time look into running SD low performance.

[desert dagger]

unless you don't mind sitting there for a week waiting for an image to generate

[desert dagger]

try mage.space

[fervent thunder]

Cloud has insanely good prices sometimes
like a server with ten A40 48GB GPUs for $2.303/hr

I actually have a gaming PC I built but I use cloud anyway for that reason

to take advantage of high VRAM you have to run a very high batch size

and high tile sizes for things that tile (mostly upscalers)

otherwise VRAM can kinda go to waste

[nova lily]

Hm how can I put this... in A1111 you know how you can save entire prompt presets.Is there some sort of feature or extension that lets you store banks of words and just click add them?

[fervent thunder]

in comfy yeah

you can do this

[rose cargo]

Is really no one having any issues with the new Forge inpainting? It's driving me insane

No matter which sampling method I use, which checkpoint I use, which resolution I use, which CFG scale I use, how many sampling steps I use, what mask mode or masked content mode I use, the seed I use, the output for inpainting is always the same same blurry blob

[desert dagger]

not sure anyone but you is trying to do that in forge

[rose cargo]

Why would no one else be using inpainting? Isn't inpainting used pretty much all the time?

[desert dagger]

didn't you say the 'new' forge inpainting? didn't they just update that?

[rose cargo]

Yeah, it's been like this for a couple days, has no one else updated it or something?

[desert dagger]

have only seen 3 people ask about forge since the update, and all were having issues. ...

but none mentioned inpainting specifically

[rose cargo]

Yeah, it looks like the new update screwed many things up ☠️ Now the extra stuff like SVD and Z123 are nowhere to be found, and batch upscaling images no longer works
Is there any way to revert to an older version?

[desert dagger]

as a suggestion, how about don't use forge? use comfy

[rose cargo]

I tried Comfy but didn't like it, honestly

[desert dagger]

all i know is that people have been having problems since the update, and maybe auto1111 or comfy is a better choice.

[rose cargo]

I saw that there was a new update to A1111 like today or something, should I give that one a try?

[rose cargo]

A1111 doesn't work with my PC specs, trying to generate anything with an XL model gets stuck at 0% forever

Ok it worked but it took almost 12 minutes to generate 1 image lol when in Forge it takes around 35 secs on average

[pale latch]

a1 has a few things that aren't enabled by default. like sdp attention or xformers. probably it.

[rose cargo]

I use --xformers --no-half --medvram --opt-split-attention --upcast-sampling

[royal field]

Hello, is there anyway to interact with sd programmatically? I have a python script that generates prompts for me, I want it to pass that into sd and generate that way. I only use automatic so not really sure. Thanks

[nova lily]

Can you put Lora's the main adetailer prompt?
im guessing it just follows it in the main overal prompt
I forget that its a second pass

[fervent thunder]

you can in comfy but not in A1111

[warm junco]

It was announced that the new forge update breaks everything for normal users.
Better switch to Auto1111 like the forge dev suggested or use the Fork of forge or the old working commit

[warm junco]

What's your GPU?

Remove --opt-split-attention if your using --xformers

If you have only 6gb vram you can enable fp8 mode for sdxl models and then you can use them easily

[fervent thunder]

I generated an image of someone with a very specific facial structure and would like to replace it

Even on the same seed changing anything about the promt ruins it

Is there any way to "save" this so I can put it into different images

[warm junco]

You can try using the Reactor extension.
Its for faceswap

[ionic ocean]

is comfy ui better? all the really high quality vids seem to use it

[sudden ruin]

You have a lot more control and extensions, better is subjective tho. Its a lot more complex for sure

[ionic ocean]

thats a yes for me lol

I studied 3d animation so im used to using complex software

[hardy sedge]

Forge needs twice as much vram as comfyui or fooocus. It's also much slower. Would like to use forge but it has too many disadvantages.

[sudden ruin]

Same as me then lol

[topaz cipher]

Hello I am looking at getting a license but not sure if I need community or enterprise license. I am a business running less than $1million a year

[fleet condor]

Is it possible to implement video object recognition functionality similar to track-anything (link: https://github.com/gaomingqi/Track-Anything) in comfyui or web ui?

[astral sigil]

Hello, guys

Any of you guys has experience with removing things from image using inpaint?

I'm having a lot of trouble

It can be various things making my result bad

Prompt, inpaint configs, denoising strenght, but I have no experience with this, so, yeah, anything I change is a shot in the dark

[sudden ruin]

What are you using to inpaint?

[hardy sedge]

Try Foocus Best in inpainting. Easiest way.

[topaz cipher]

been trying to get a hold of someone to get an enterprise license is there anyone in here I can talk to?

[storm shard]

It's in the https://discord.com/channels/1002292111942635562/1002292398703001601 https://stability.ai/news/license-update

[topaz cipher]

This doesn’t actually give me the license I need the actual license

[storm shard]

I'm sorry but what is unclear?
"Our new Community License is now free for research, non-commercial, and commercial use. You only need a paid Enterprise license if your yearly revenues exceed USD$1M and you use Stability AI models in commercial products or services. "

You mentioned you are running a business less than

[topaz cipher]

I filled out the information for community license but they never sent an email back.

[storm shard]

You don't need one sigh

[topaz cipher]

sorry looking into this I actually need enterprise as it is for commercial products along with needing to develop our own API

[storm shard]

You mentioned you're not exceeding the $1 limit, so it's free. RTFM.

[topaz cipher]

do you know if you can develop your own API in the community license?

[storm shard]

Only commercial users need to self-report** If you integrate Stability AI’s models or derived products into your own commercial products or services, and your annual revenues are below USD $1M, you don’t need to pay anything to Stability AI. ** We do ask that you fill out this brief form, and indicate what models are you most interested in. Once your annual revenues exceed USD $1M (or local currency equivalent), you'll need to contact Stability for a separate Enterprise license.

[topaz cipher]

understood thanks!

[storm shard]

It's all on that License update link I sent you, you could have read it before asking. Nm. I'll shut up now.

[fervent thunder]

Hey guys

[delicate cedar]

Hi

[fervent thunder]

I'm having problems with the installation can someone help me?

[rose cargo]

What's the fork of forge?
I saw that the dev posted an older version of Forge so I'll try that and see if it works. Did the dev really suggest switching to A1111? Does he not plan on fixing/updating forge anymore?

[rose cargo]

1660 Super

[warm junco]

Come to #🤝｜tech-support

[warm junco]

He will update forge but it will go into a different more experimental direction. Thats why he said to switch. You can read that here:
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/801

There should be a "Download previous version" button at the repo. To get the latest working version. But better use Auto1111 if your GPU has more than 6gb vram.

[sudden garnet]

a panda hugging a dolphin

[rose cargo]

I'll try the latest working version, since A1111 is significantly slower on my end no matter what I do

[warm junco]

when trying auto1111 your card needs --xformers --medvram--no-half
and then you should also enable fp8 mode for sdxl in settings

[unborn raptor]

where do i go to buy a subscription? I used this in the past, like 2 years agho and it was free

[rose cargo]

Would it be faster than Forge? I'm using the older version now, which seems to work without any issues

[warm junco]

if you run it localy its still free. if you want to subscribe you can do it here:
https://stability.ai/stable-artisan

[warm junco]

for sdxl models maybe

[sudden ruin]

What the heck

[fervent thunder]

discord doesn't automatically collapse long replies?

even chatGPT has that feature

[sudden ruin]

Usually messages are character limited I think

idk how they bypassed the automod

[fervent thunder]

maybe some hidden character trick

like some characters from Chinese keyboard

[tawny halo]

guys can smbdy tell me how to do advertisement and how to get peoples attention ?

[tawny halo]

i am trying to gather a programming team to create products and be able to join the market if u are interested dm me cuz i want to create a well structured team that could perform great outcomes !!

[narrow kernel]

this isn't the place to do it, take it to a platform or subreddit that allows it

[abstract linden]

@warm junco Question if I may, I was looking on Civitai and found a checkpoint I had a couple of questions on. If I wanted to post the link to the check point and ask those quetions, would here be a good place to do it? Big reason I am asking is that it's a Pony check point and I didn't think I should post about it in the Sd3 or Sdxl tab.

[warm junco]

If its not nsfw its okay to post and ask for example in #📝｜prompting-help

[abstract linden]

Perfect, thank you.

[rain crest]

How do I keep character consistency with strong checkpoints that controlnet has no effect on? I want to use IP adapter with this Pony-based checkpoint, but it's just useless.

[fervent thunder]

have you tried the mad scientist one?

its stronger

[fervent thunder]

its an ip adapter node

that goes block by block

[rain crest]

I should've said I was looking for solutions on A1111, but I might try it if I don't find anything else👍

[pale latch]

reforge has more samplers and schedulers now. wtf i suppsoed to do with all these

time to do xyz grids again i guess

[low moon]

How's Monday treating ya'll?

[desert dagger]

you're supposed to make balls with all of them

[sleek scarab]

Looking for fastest free AI tool to remove 500pic's background simultaneously without decreasing resolution like Online ones , uses pc power , or paid software can work offline . Any ideas?

[warm junco]

Hey, try this tool:
https://fyrean.itch.io/bgbye-background-remover

But for local offline I would use Automatic1111 with the RemBG extension installed.

Then you can use img2img batch input

[sleek scarab]

how can you work with this 500pics simultaneously ?

[warm junco]

I don't think there is any tool that can do it simultaneously

Its always one by one

[sleek scarab]

remove.bg does but it paid and not offline :/

there is must be a way , thanks for helping tho

[warm junco]

Automatic1111 rembg extension in img2img batch input is also one by one but it takes seconds so it could be fast enough depending on your GPU

[wide acorn]

hey guys so i have tydiffusion inside 3ds max now, is there a resource to explain all the special characters and symbols and what they do for sd. Like [ ] --1, things like that not needing a list of how to prompt but how to usde the special charachters and symbols to add more dept to your image, hope this makes sense?

[desert dagger]

who's the author of tydiffusion?

[wide acorn]

tyflow

[desert dagger]

you should probably ask them

[wide acorn]

simulation sytuff, liquids, fires etc

[desert dagger]

you should still ask the author how to use those symbols and what they're for

[wide acorn]

well im after a more general list not for 3ds max

[desert dagger]

okay, in general, all of that stuff isn't stble diffusion, it's midjourney

[wide acorn]

oh

so weights and what not arent used in sd?

[desert dagger]

you need additional code running for stable to make those mean anything to it other than just random dice rolls for data

[wide acorn]

gotcha

[desert dagger]

not like that. any site that does use that has additional code to deal with it. stable, natively, assigns weigth to tokens as they show up in the prompt - the closer to the start, the more weight, and repeated tokens add additional weight to the token

(but not a lot)

[wide acorn]

answer i needed thanks so much crystalW

[desert dagger]

welcome

[wide acorn]

yeah id check out tydiffusion its pretty epic but only for max

[drowsy mountain]

hi guys, how i can automate this ?

[cedar salmon]

its uh comfyui implemented in 3DMaxx, ide just look for any stable diffusion prompting guide, better to be specific to what model you choose like 1.5 or SDXL

[wide acorn]

yes its a great one click install then pick what models u want

[thin abyss]

May I ask what happen to the image generation? Came back after a while away and the channels are gone.

Nm found it and it's not free anymore sigh

[trail lion]

Chatting? It's easy, there are all kinds of bots out there

Although that does beg the question of why

[silk condor]

Hey, is it any difference if I would go for a RTX 3060 with 12 GB or an 40xx Serie with 12 GB?

[fervent thunder]

looks like AMD made the best app for stable diffusion but its censored lmao

[high flint]

censorship is really great! /s

[pale latch]

the bulk of the internet and mass media have been censored in some form or another and we've made it this far due to it still. it'll be fine

[fervent thunder]

what app is this?

[high flint]

Can ComfyUI output images to another program on a local LAN Programmatically?

[fervent thunder]

amuse 2.0

its too censored u cant even render a woman curves

its based on onnx models which is not good news for me cuz i use safetensores with comfyui

[fervent thunder]

https://youtu.be/1FXrk9Xcx2g

Intel has them ai app too, its based on saftensore

[high flint]

For ComfyUI, is there a way to view a preview image of how the checkpoint or lora image could do?

Because it is kind of nice to be able to have a quick visual refrence when looking through all of them

[fervent thunder]

yeah see civitai for that

[high flint]

in other words, not possible within the program? that's a shame.

[fervent thunder]

ithink there examples in comfyui github page

https://openart.ai/workflows/all

this too has examples

[high flint]

I mean, there's a big difference in looking it up vs having a handy refrence when scrolling through a lot of checkpoints/loras

[cedar salmon]

there is a browser addon but you have to add all the info yourself

[fervent thunder]

just go to civitai search for loras and u will see whats all about

[high flint]

I dont mind doing that, what's the name of that addon?

[cedar salmon]

https://github.com/hayden-fr/ComfyUI-Model-Manager

[high flint]

cheers

[fervent thunder]

https://openart.ai/workflows/home
have alot of examples just download .json and drag it to comfyui web

[fervent thunder]

u mean for civitai ?

[high flint]

I think you misunderstood sand land kid....

[cedar salmon]

wasnt specifically but that can link to civit also

[high flint]

https://github.com/pythongosssss/ComfyUI-Custom-Scripts this also looks promising

ok, the one from Pythongossss is supre helpful. it's a unique checkpoint loader but w/e it's really helpful in terms of finding the lora/checkpoint you want as well as getting the details

[dawn mulch]

I am fine with either. There wouldn't be much of a speed difference between the 2. Maybe a second or 2 for the 4000

[verbal osprey]

anyone have any updates on the next SD3 model release?

[tropic frost]

say, do you guys keep all your loras together or sort them out in different folders depending on what they are, you know characters, styles, etc. ?

[warm junco]

Amuse is not censored.
It depends on the models you use

They even state that at the installation that its not censored

You can use any model but it needs to be converted to onnx

But Zluda is still better in terms of usage and compatibility

[hardy sedge]

What is the best method to make a face more detailed when it is further away?

[rain crest]

Anyone know how to have SD generating in the background while playing games without it being choppy?

[solemn ermine]

hello，friend

[warm junco]

Its not possible because both uses the same GPU. And you can't limit the vram SD takes

[hardy sedge]

Is forge with the new Update faster and Not so vram hungry Like comfyui?

[rain crest]

24GB isn't enough to run both?

[warm junco]

The Adetailer extension

[warm junco]

Depends what you generate and what you play

Without upscaling it could work fine

[rain crest]

I play Overwatch and I only upscale to 1.5x. The game stays at 600fps but it's still choppy and stutters..

[narrow kernel]

It is possible: get a second GPU. As long as the game doesn't demand a ton of CPU, it shouldn't lead to too much hitching. Oh and as long as the diffusion going on in the background is all staying loaded on the other GPU and isn't model swapping a bunch(ties up the CPU and sysram)

Obviously not the most elegant solution, but I know plenty of people that do it

[frigid tendon]

For anyone on AMD, they've just released this nice software based on stable diffusion. https://www.amuse-ai.com/

[fervent thunder]

its intersting one yet its censored and uses less famous model type

[frigid tendon]

Not censored, runs locally and you can install any models, loras etc you want as normal.

[warm junco]

Its okay but zluda is still better in terms of performance/usage

[stark tusk]

Stretching Each Dollar: Diffusion Training from Scratch on a Micro-Budget

using only 37M publicly available real and synthetic images, we train a 1.16 billion parameter sparse transformer with only $1,890 economical cost and achieve a 12.7 FID in zero-shot generation on the COCO dataset. Notably, our model achieves competitive FID and high-quality generations while incurring 118× lower cost than stable diffusion models and 14× lower cost than the current state-of-the-art approach that costs $28,400.

How is Stability.AI going to prevent people from seeing women laying in the grass now?

[frigid tendon]

I've found it to be really quick, quicker than zluda

[warm junco]

Yea but less functions, no extension support, every models needs to be converted.
It has the same issue as the directml fork of auto1111 sadly

[fervent thunder]

what is zluda ?

i have rtx 2060 and amuse works like a charm

[warm junco]

Zluda emulates cuda stuff to rocm so you can use Auto1111, Comfyui, forge etc with the normal installation and torch+cuda with AMD cards on Windows.
Its the best between usage and performance

[warm junco]

With your card I would use Auto1111 instead of Amuse

[frigid tendon]

There is an advanced mode which opens it up, only played with it briefly this morning so need to check out more of it's features. Seems nice so far though.

[warm junco]

Yea I tested it too yesterday. It has some cool functions but its no alternative for zluda rn for me

Mostly because of the missing extension support

[fervent thunder]

are u comparing ai sd apps with an acceleration model for amd gpuz ? i dont get it

[warm junco]

No sry, Amuse uses Onnx. Auto1111 can use Zluda.
So I compare both

Auto1111 can use onnx too

But the convenience is not great

[fervent thunder]

yeah not sure why amd used onnx for amuse ?

or maybe safetensore will not work ?

[warm junco]

Onnx can't use .safetensor files
They have to be converted first

[fervent thunder]

can i convert them myself ? or it require tools and coding ?

if u search civitai u only find one post about onnx XD

[quasi dust]

Where should I go ask if I need help with how to use Stability APIs?

[fervent thunder]

Amuse is insanely optimized, taking 3gb ram while rendering a video

[warm junco]

Amuse should convert them. But its not needed for nvidia

[fervent thunder]

i notice that amuse uses alot of ssd usage during render lol

the ssd is maxed out am wondering if other apps have the same usage but i am sure they store the model in vram

now that i have 6gb of vram am wondering if i can use sd xl in amuse since it dont store to vram

i will have to try it

[warm junco]

Auto1111 and comfyui using RAM to store the models.
And with 6gb you can use sdxl in both

[bleak matrix]

Good morning, everyone!

How are we all today?

[fervent thunder]

for me all good , busy using sd ai all night lol

[river sparrow]

I once had a party so good, my friends thought I was dead for days. Later I heard about a woman who got pregnant, but swore the lord did. I just cant remember, but the kid is called billie. BJ

[fervent thunder]

oh for once i thought this is the offtopic channel

[unkempt depot]

Hi, I tried to generate a larger video with AnimateDiff and Controlnet and got an error: In Stable Diffusion it says: "OutOfMemoryError: CUDA out of memory. Tried to allocate 1.10 GiB (GPU 0; 8.00 GiB total capacity; 12.37 GiB already allocated; 0 bytes free; 14.65 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF" Does anyone know what I could do? The error message suggests I should "set max_split_size_mb", but what even is max_split_size_mb and how do I set it? Any help would be very appreciated 🙂

[fervent thunder]

@unkempt depotuse cuda fall back pilicy to ram in nvidia control panel

[river sparrow]

Buy bigger card or download vram 😛

[fervent thunder]

what is ur gpu ?

[unkempt depot]

I have a nvidea gtx 1080

[unkempt depot]

How exactly do I do that? 😅

[fervent thunder]

https://youtu.be/XWuHPMvO-ps

see this

[unkempt depot]

Ok, ill check that out thanks 👍

[minor inlet]

Hello, what is the best model for generate roleplay character avatar please? I have tested dreamshaper mais same or close prompt generate everytime the same character

[unkempt depot]

I now tried running the thing again with CUDA set to "prefer no system fallback" and basically the exact same thing happened as before. After around 50min I got this error: "OutOfMemoryError: CUDA out of memory. Tried to allocate 564.00 MiB (GPU 0; 8.00 GiB total capacity; 6.52 GiB already allocated; 0 bytes free; 6.96 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF"

[fervent thunder]

why u set it to no fall back lol ?

set it to system fall back so that when u run out of vram ti goes to ram.

[unkempt depot]

Cause the guy in the youtube vid you sent me put it to no fallback, didn't read that you said I should use cuda fall back lmao

My bad

Welp, let's try again

[fervent thunder]

if it dont work at all n u need to use sd1.5 based models and not sdxl ones. also try not to use too many loras and control net.

https://youtu.be/zzA1iUgtiEs

in this video there goes your issue, hopfully its fixed

[narrow kernel]

Yeah it means you've got some kind of memory leak or some kind of history cache not clearing

[brittle rock]

Hello, is anyone try to install stable diffusion on local on a windows and meet problem with pytorch ?

[fervent thunder]

would recommend dual booting with linux

[frank shadow]

Hello, there is someone who using SD Forge and making videos in SVD tab ?
Well i don't know what happened, or why but when i updated the SD forge the svd, z123, training tab is dissapered.

How can be solve to reapper again, they are still in extensions tab with a checkbox.

[unkempt depot]

I already use a sd1.5 based model and only use an openpose controlnet and no loras. Though despite now setting CUDA to "prefer system fallback", I still got an error. In A1111 it says: "OutOfMemoryError: CUDA out of memory. Tried to allocate 564.00 MiB (GPU 0; 8.00 GiB total capacity; 6.58 GiB already allocated; 0 bytes free; 7.35 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF"

Maybe my graphics card is just too bad or maybe my disk space doesn't have enough memory

[unkempt depot]

I guess I'll try downloading tensorRT and try it once again

[foggy yacht]

can i train lora with selfies only?

[fervent thunder]

you can get train lora with just one image

[nimble ridge]

Hi 🥂, we've been granted $2500 OpenAI credits with access to the API and GPT-4o and fine tuning, currently we've no valuable use case for them, so I can give it to the Stable Diffusion community as I belive it'll be way more useful. I offer a 70% discount, you can DM to sort this.

[fervent thunder]

What world did i land on?

[foggy yacht]

yeh, but i want some decent lora i have 30-40 images

[unkempt depot]

Btw I put the entire error log in #🤝｜tech-support , incase it could help anyone (The error is very very long 😅)

[foggy yacht]

but selfies work? someone on a reddit thread said you have to invert the image

[fervent thunder]

firstly the contents of the image don't matter
so you don't need to ask if selfies work
the AI doesn't know what images are, it will work with everything

and secondly, yeah inverting the image is one method for getting more training data

[foggy yacht]

aight thats what i was wondering ty for help. so its like just flip them horizontally?

[fervent thunder]

TBH I am not a huge fan of that method

[foggy yacht]

ohh what would you suggest then?

[fervent thunder]

what I would do is make a lora with just one image
and then make 1,000 generations with it
and then cherry pick the best ones out of that 1,000
and then use those to train a better lora

[foggy yacht]

ohh, that's really cleaver! thanks I'll go with this method

[fervent thunder]

you can use control net and IP adapter to help, if you know what those are
but they are not neccesary

[foggy yacht]

like to stage the poses?

[warm bane]

Do u guys know of any good AI animation website that just does minimalistic animations, like very slight dynamic movement and keeps the image like 99% the same?
I've been using Runway and its okay but looking for something even better ofc 😄
I just want very small dynamic movement animations for images I have already, fantasy stuff

[fervent thunder]

yeah control net for pose and IP adapter for style

which can include colours

[woven rivet]

https://liveportrait.org/ ``https://github.com/KwaiVGI/LivePortrait

[warm bane]

Is this good for fantasy images tho? I'm not looking to animate a face

just very slight movements of the image i already have

[woven rivet]

Ohhhhhh. Sorry. That only works with faces my bad

[warm bane]

yeah no problem

i can see it needs a reference 😄

[woven rivet]

does exactly what you ask for besides that haha

[warm bane]

I mean runway is pretty good but lowers the quality a bit and I haven't gotten to try the new model cuz u can't even test it without paying, idk why

[unkempt depot]

Does anyone now what version of Nvidea TensorRt I should download for StableDiffusion / A1111 / AnimateDiff / ControlNet?

[woven rivet]

Check out fotor. It was decent before I got into local stuff. They do it but ive not used it.

[foggy yacht]

btw with one picture how many repeats would you suggest?

[fervent thunder]

I don't wanna give advice on training parameters or scripts cos I haven't kept up with modern methods

there will be others that can give advice for that though

[foggy yacht]

Ohh ok, thanks !

[fervent thunder]

oh i forgot to mention that u need more ram if u have 16gb like me increas virtual memory to 30+gb

[unkempt depot]

Yeah I have 16gb too, but I have never heard the word virtual memory before lol

I'll look a a couple of youtube vids and try that out though 👍

[unkempt depot]

Btw would you recommend downloading a specific version of Nvidea TensorRt for StableDiffusion / A1111 / AnimateDiff / ControlNet or should I just download the newest version?

[low moon]

Eventually it'll cost even less hehe, personal weird models for everyone. XD

the hodge podge of privately cobbled together models floating around everywhere will be a sight to behold

If you think the stuff on civitai is weird thats just the tip of the iceberg of what lurks in the human mind

[fervent thunder]

tensorrt is pain in the S.

[low moon]

it gives 30% performance boosts according to mythology.

[fervent thunder]

its twice as fast not 30%

[warm junco]

TensorRT dont work for GTX cards.
Also your card needs --xformers --medvram-sdxl --no-half-vae in the webui-user.bat for the best performance

[steep pebble]

anyone doing architectural rendering enhancements with artificial intelligence?

[abstract linden]

Kinda out of curiosity more than looking for improvement, how do people usually launch Auto1111? The video I watched said I should click my webui-user.bat file. And that works for me, but I was wondering if that's the norm. I tried launch.py out of curiosity and am waiting on that currently.

Kinda a silly question, but I am just curious

[loud solar]

webui-user.bat is the normal way ... you might need some changes ... and for AMD graphic cards more workarounds ...

[abstract linden]

Gotcha! So I am not weird! Lol thanks haha.
And was that second part for me?

[loud solar]

I don't know your graphic card ... if it's an AMD it's for you 🙂

[loud solar]

Oh.... and I was thinking of a Windows PC, cause I run Windows ...

[abstract linden]

Lol I run windows as well

[loud solar]

So looks like you are doing fine

[desert dagger]

@bleak matrix you've got spammers on the discord

[rain crest]

Hey, is there a IPAdapter face preprocessor specifically for SDXL?

[iron shadow]

hello everyone how are u doing, so i have been trying to get consistant character on a1111 i used ip adapter nothing used reference controlenet nothing tried to mix them nothing as planned. i used roop too did not do the job. i wonder if u know anything about this topic and how to achieve it. please tag me if u reply thank u

[rain crest]

Show me your controlnet settings on ipadapter

[wicked sonnet]

What do you do at night when your sweating cold sweats

[iron shadow]

PREPROCESSOR ipadapter face id plus ..... model face id plus v2 ...... starting control step 0.45 .... end control step 1 ...... control weight 1...... pixel perfect enabled

[iron shadow]

going to sleep is the best u can do

[frank fable]

Guys, could someone help me on #🤝｜tech-support ?

[pale latch]

sam 2 looks cool and neat

https://ai.meta.com/SAM2/

gonna help out animation remixes a lot

pretty huge give away for facebook. no reason tehy couldn't release with commercial restrictions, but they're building a vision and this is part of it. yay free tools

[nova lily]

Most check points in 2024 still need face resto, but I never see any mention in description and face resto is hidden in A1111. 🤷‍♀️New people must be baffled by bad results.

[abstract linden]

What's face resto?

[low moon]

i need that thing in comfy

[pale latch]

codeformer is a cheap hack and it's very codeformer ish. you can pin the setting to the top if you want. i just use it through adetailer if i think it'll help

also is in extras tab i think

[low moon]

is it accessible in comfy?

[abstract linden]

Question for you since you use comfy, I've been thinking of grabbing it and giving it a try, but rn I use Auto1111. Do you use both?

[low moon]

i use comfy and forge mostly but i have auto and foocus too

u should try comfy.... its more versatile

[crisp delta]

Hello, I’m new here, I don’t really understand. Can someone tell me how can I make a realistic dragon? 😔

[high flint]

I've made the switch to ComfyUI myself just a few days ago, it's a lot more resource efficient, it also really gives you a lot more control over what the AI is doing to the image as well

[abstract linden]

Wow, so you use a variety! People do say I should give comfy a try. I guess I'm wondering like, if I try comfy, does one typically just go using comfy exclusively? Or lile, if they pick their favorite, do they stick with one or is there a reason to keep both?

[low moon]

i started with A1111 then trued forge then foocus and now comfy

[abstract linden]

Ohhhh resource efficiency... that alone is a tempting detail...

[abstract linden]

I can give a try to help!

[low moon]

u can also set everything up to use one folder so u dotn waste space

so i have all my models in A1111 and everythign else uses them

[crisp delta]

Thanks!

[high flint]

Yeah, I know of someone who can run SDXL models on comfyUI when they would just get OOM errors on Automatic,

[abstract linden]

Very interesting. Does it rely on shortcuts?

[low moon]

u have to edit a file with the folder paths

[abstract linden]

Actually, rn we're having a discussion you may find as a great starting point.

[high flint]

It grabs everything from Automatic's file path you specify

[abstract linden]

Hmm.

[abstract linden]

OH!

[high flint]

So, you can literally keep automatic's install and install everything for that, and the just boot up ComfyUI and have it use everything from Automatic without having to duplicate anything. More or less becomes an extra tool you can use for a very minimal harddrive space

[abstract linden]

Oh okay, I think I understand. Took me a minute

[low moon]

this feature alone saved me like 300 gb pf space

XD

[abstract linden]

I guess my real question is like, if I make the switch to comfy, is there any reason to keep Auto1111?

[high flint]

Well, Automatic is a bit more user friendly, ComfyUI takes a bit of learning to use effectively

[crisp delta]

I see, I try to keep up with it!

[high flint]

As for dragons, Sopita,

I'd look for a Lora of dragons styles you like

(preferably in SDXL)

[abstract linden]

I do like a challenge.

I suppose I really should just bite the bullet and give Comfy a try 🤔

[abstract linden]

What's your background? Are you starting from square one or are you familiar with AI at all?

[high flint]

https://civitai.com/models/153332?modelVersionId=171702

this may be what you're after perhaps

[crisp delta]

No, I nearly understand what you guys are talking about. 😂

[crisp delta]

Thank you Rai! I was looking for more realistic dragons but those are really good 👀

[high flint]

there's a lot more that look really nice on there, just take a look

[abstract linden]

Alright then, so right now we're talking about two different UIs. (User Interface). Which is how you interact with the software.

It sounds like in short, Auto1111 (what I currently use) is more user friendly and probably easier to use.

Comfy UI is more difficult but gives more freedom too.

[abstract linden]

Civitai is absolutely filled with resources. But you may not want to worry about them quite yet until you gave your UI set up. There are plenty of YouTube tutorials, which I can send if you'd like. (For Auto1111 at least. I shan't recommend one for Comfy until I've gone through the process me-self.)

[low moon]

you can do it!

[high flint]

also better / faster overall as well. Very good if you are worried about resource limitations

[abstract linden]

Hell yes! I can! And I will!!! 🤩

[high flint]

also, when learning the UI, dont fall into the trap of downloading workflows, try to make your own and understand how they work instead. At least until you get a handle on the basics

[low moon]

yeah

grabbing workflows can be super intimidating and confusing for newbies

I spent hours cleanign up nerdy rodents stuff lol

[high flint]

Not to mention they will all have tons of additional stuff added to the manager you may not want.

[low moon]

red node overload

[crisp delta]

Thanks for the help, I will do my research. 😁

[high flint]

I've just started looking at other workflows to see what others did, and honestly, they look cool but also at the same time, not very workable since I dont fully know everything it needs yet

Also, mostly trying to integrate multiple loras into images in comfyUI without having said image become a total monstrosity at this point.

[abstract linden]

Very welcome! Good luck!!

[abstract linden]

Thank yuou!!!

[abstract linden]

Thank you!!

Honestly that's a SUPER huge tip! It's definatly made to seem like workflows are as important as like. Loras or embeddings, so knowing NOT to do that is VERY good to know!

[high flint]

Instead, I'd follow a few tutorials

of where they explain the program

better off knowing of WHY and HOW things are done vs "use x to get y, dont worry about the other stuff"

[abstract linden]

Hmm okay makes sense. I think I saw one that looks pretty credible.

Let me grab real quick

https://www.youtube.com/watch?v=AbB33AxrcZo

Have you by chance seen this guy?

[high flint]

oh, also, use reroute notes

those things are great for keeping your workflow a lot more tidy

[abstract linden]

Re-route notes. I'll remember that

[high flint]

also yes

[abstract linden]

Glad to hear. I'll watch him then

[high flint]

Hrm, when do you want to use upscale latent or upscale image over the other?

Hrm, that's odd, I'm getting blurry as heck photos at 1024x1024 in SDXL on this workflow

but when I scale it up by 2x they get a lot better.

(still not great but, still worse than what I was getting without loras)

... maybe I'm using an incompatible lora?

[unkempt depot]

Oh ok, thanks for that info, btw I already have all those parameters, maybe it's just sadly impossible

I guess I'll look into the max_split_size_mb option and see if anything else is possible

Btw this error 100% has nothing to do with the fact that my local drive has very little space, correct? (This might be a dumb question; I really don't know a lot about memory and ram usage lol)

[high flint]

Here's a question for all of you in comfyUI, What have you had the most success with in terms of Upscaling or "High res fix", doing one larger upscale latent or doing multiple upscale latents in smaller chunks?
and which method for said upscaling?
ultimate SD upscale is mainly for the last step before making it be the final image size? or is that what I should be doing?

[mossy arch]

HI I am trying to use open pose on a existing image to alter the characters position. However, it is only giving me blank results. I was wondering if anyone knows the solution to using open pose on existing images. Thank you

[desert dagger]

yeah, cause it's a lot more userfriendly

[desert dagger]

i don't use either. i upscale with something other than the AI i generate with - usualy Topaz or Magnific, sometimes caput's image upscaler

[high flint]

Oh? I'll have to take a look into those,

are those seperate programs or?

[unkempt depot]

Update: I saw multiple comments on the internet saying I should add "set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:number" to my webui-user.bat file, however some people said one should use set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:512 and some suggested using set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:128. I have a nvidea gtx 1080, does anyone have a recommendation, which of the two options I should use?

[warm junco]

That can be an indicator because the Pagefile is also stored on C normaly. So you should have 10-15gb free space there

But come to #🤝｜tech-support and show your settings when you get the error

[warm junco]

None of them. I have a GTX1080 in my old PC too and can do everything

[unkempt depot]

Do you mean my webui-user.bat settings?

[unkempt depot]

Oh ok, can you also generate big AnimateDiff videos by any chance?

[warm junco]

Txt2img settings

[warm junco]

Thats a rough task because it needs so much resources.
Maybe if you set the batch size of AnimateDiff to 2 or 4

[unkempt depot]

I can do that

[unkempt depot]

Do you think PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:number could help for animateDiff?

[warm junco]

Nope

Try batch size 2 and dont use sdxl motion module

And I mean batch size of AnimateDiff and not the normal batch size

And for resolution use 512x512 or 512x768

With no upscaling

[unkempt depot]

Ok, I'll try that!

Btw I posted the settings in #🤝｜tech-support

[swift sluice]

火影忍者 漩涡鸣人背着他的孩子，牵着雏田的手，站在斜坡上，远远地静静看着远处燃烧的木叶村

[old lichen]

https://civitai.com/models/107842/aniverse

Guys, can you suggest a similar model/Lora to "pony" or a good merge with the style of Aniverse?

[velvet wing]

阿怎麼沒圖片

[sudden ruin]

Please keep it english!

[rain crest]

Anyone know how to use IPadapter for 2 different faces in 1 image?

[fallen hollow]

Hi to community ..

[low moon]

Hi sir.

[velvet gulch]

Hi there, I have a question about which GPU i should buy for better performance, can I ask it here or should I ask it on tech support channel?

[warm junco]

An nvidia GPU with 12gb or more vram is good for Ai stuff.
More vram is better

[velvet gulch]

I know it but im in between RTX 4070S and RTX 4060 ti 16GB, RTX 4070S is better in every way but 4060ti's 16gb of VRAM is kinda confusing me

i will upgrade my RTX 3060 12gb to one of those but idk which one is generally better, at least for AI stuff

[wet grotto]

controllora is a controlnet with lora inside?

[warm junco]

In gaming the 4070S is better than a 3090 xD

But yea 12gb vram is sad

The 4060ti has 16 but a reduced bandwidth

[velvet gulch]

well, if 4060ti is not significantly better than im gonna stuck with 4070s

[low moon]

Aight aight, hows Wednesday treating everyone?

Friday is approaching.

I has 12gb of vram and it is neat. I started with 4gb - it was doable but painful. I assume 24 gb would be even better.

Share your vram GB experiences with everyone on this webzone.

[inland drum]

Hi all

I recently got myself a 4090, any tips on how to use its full potential will be welcomed

[low moon]

stop flexing!

lol jk

use it as any other card, ull just have faster generations

[warm junco]

Hey, if you use Auto1111 you need to add --xformers to the webui-user.bat

[golden coral]

Hey, I want to perform image reconstruction. I am trying to condition a diffusion model using the latent embeddings of the image I want to reconstruct, sourced from a pretrained encoder . Does anyone know how I should approach this task?

[atomic gull]

Hey everyone. I'm pretty new to Stable Diffusion and I honestly don't have enough time to research and learn the proper techniques to bring out proper images, NSFW or SFW. Is anyone here available to teach me in one session about the advanced techniques? I'm willing to pay via PayPal. I already know the basic concepts of using Automatic 1111's UI and I can generate images myself, not just the output that I am expecting. Anyone willing to teach me in one session, please do leave me a message. Thanks!

[crystal bone]

Is there a pretrained controlnet (with canny support) for stable video diffusion.

[wide acorn]

anyone using tydiffusion in 3ds max?

[wide acorn]

if so would u dm me i just need someone to talk the possibilities with tydiffusion