🤝｜tech-support

[wet flame]

what python version are you using @ornate elk

3.11?

[ornate elk]

3.10.11 64bit

[viral knot]

is python even available in 32 bit these days? lol

[ornate elk]

yea 😮

[viral knot]

damn

[gentle eagle]

when I try to use the text encoder clips on SD3 model it gives out of memory error... anybody have experienced this ?

[ornate elk]

What's your gpu?

[gentle eagle]

T4

15gb

[fossil dew]

welcome back cs n_n lmk when you got the time to help me out ❤️

[ornate elk]

what sd3 model did you tried?

[gentle eagle]

the medium with 4gb one... it's on comfyui btw

it works if I don't use the triple clip thing

if I use --lowvram it proceeds but the images are generated black(with the triple clip that is)

[fossil dew]

heres a random question, just for curiosity's sake

can you download other sampling methods? 🤔

[ornate elk]

Try the 5gb model it has the clips included

[viral knot]

yes like this https://github.com/redhottensors/ComfyUI-ODE

[gentle eagle]

it works alone but I wanted to use the clips separated.. and then it's not gonna have all the clips I think?

it's strange that it's not working

[ornate elk]

which text encoder do you have? the t5- fp8 or the fp16 one?

[wet flame]

i think sd3 is not worth it and not worth the time investment in it as it is now

wait for community release for it from civtai

meanwhile models for xl like the luma is good enough

or fooocus

[gentle eagle]

I tried only the clip_g.safetensors and
clip_l.safetensors ones.. I don't have storage for t5

now I tried a dual instead of a triple with pure clip_g and it proceeded but the generation continue to come out all black

[tame yoke]

what do i do if my dev branch a1111 hangs at 95% in image generationand doesnt let me apply settings as its stuck at ‚loading model weights’ or smth like that

[ornate elk]

i think you need all 3

• sd3 4gb model

[hollow prism]

ive changed no settings its allways worked now suddenly this what is wrong with it? https://gyazo.com/b8ccce929bcda840d66fb28b903a111c

[ornate elk]

what are your txt2img settings?

whats your gpu?

[hollow prism]

3060 ti

https://gyazo.com/0e2b642db797b96ef45b867152e21f6f

also it dosent seem to change even if i change the width and height

nothign has changed besides like a windows update

also i did start using sdxl last week

idk if that is messing it up

[ornate elk]

whats in your webui-user.bat?

[hollow prism]

https://gyazo.com/fad008a0c46c00b49b8040533c299ce8

[ornate elk]

change --medvram to --medvram-sdxl

then set the hires steps to 10

upscale by 2

[hollow prism]

ok i dont usally do upsacles tho

[ornate elk]

okay den disable hires fix

and try a normal gen

[hollow prism]

this is like happening on normal text2img

ok 1 sec

also i never use hires fix to

[ornate elk]

in your screenshot its enabled xD

with wrong settings

so maybe thats your issue

[hollow prism]

oo just having it dropped down turns it on

your --medvram- thing worked

was it allways typed wrong and the program diddent care till now?

thank you for the helps

[ornate elk]

yea because your on an older webui version (in the new version it has checkboxes)

best would be to use or update to the latest auto1111 version if you dont rely on old extensions

[ornate elk]

--medvram is for every model type and --medvram-sdxl is only for sdxl. your gpu (8gb) only needs medvram for sdxl

[lapis oak]

Anyone ever seen an extension called "stable-diffusion-webui-simplefix" ? It appears to be malware and I'm having trouble finding anything about it.

[gentle eagle]

I managed to fit all of them and it worked, thank you!!

[ornate elk]

nice, np 🙂

[half island]

What's the proper syntax to make a venv which i call "comfyui" but creates a venv named folder in comfy's root folder?

[ornate elk]

here is my code for my custom Comfyui Launch bat.
https://github.com/CS1o/Stable-Diffusion-Info/blob/main/Resources/Start-ComfyUI.bat

it creates a venv and installs the requirements in there too. also it installs torch+cu118 (you can change it to your liking)

[hollow prism]

but i still use sd1.5 should i also put normal medvram?

[ornate elk]

nope, --medvram slows 1.5 down, so you wont need that

[hollow prism]

o ok

[half island]

Thanks! Knew there was a step i forgot about. This shit xD

venv\Scripts\activate Just activate that shit, and i'm in it's venv where only shit for it appears :P

[ornate elk]

no problem ^^ yea always some little steps that are missing

[ornate elk]

never seen it, very strange

[half island]

Yep lol. Now to copy pasta that one to all my A.I stuffs so they don't affect eachother :P

[lapis oak]

one result on A1111 issues, and talking to someone else who has been affected by it. Not certain it comes from A1111, might be from another malware he has.

[ornate elk]

i noticed this message yesterday, but dont know if its related:
#💬｜general-chat message

[lapis oak]

yes that one

[ornate elk]

and the github discussion is this:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/15931

we have to keep an eye open

[lapis oak]

yes, there was a discussion on reddit over if using docker would help against malware, this guy has everything in containers, so not quite.

[ornate elk]

ah well, guess thats answers it xD

[mossy oar]

whats the default ENSD?

[ornate elk]

0 (disabled)

[half island]

@ornate elk For a tool that only has "setup.py" and "run_gradio.py", what changes should it have for the run part?

https://github.com/Stability-AI/stable-audio-tools This is what it's for

[ornate elk]

you want that also in a venv?

[ornate elk]

Use this: (it has a venv)
https://github.com/diontimmer/audio-diffusion-gradio?tab=readme-ov-file
its made for stable audio tools

[half island]

I want every A.I tool i have in a venv like this for conveniency, just some i need to alter as they are more just setup.py and gradio run ones :P

[half island]

Ah, so the one i linked to is jsut the barebones, and the one you linked to is the more "here's all you need" fork?

[fair oxide]

loading the clips incorrectly would give you black image

[ornate elk]

the one i linked is more like auto1111, and the one you linked is the source like the first sd code (more cli etc)

[half island]

ah ok, thanks :)

[fervent pier]

@fallen quail I'm new to comfy, so I could have done something wrong, but I tried installing your custom samplers but got an error on MPS during execution: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead. Is this a known issue? I can give you the full traceback if it would be helpful.

[half island]

@ornate elk checked the one you sent, it's almost a year outdated :S

[fallen quail]

Uhhhhh. That's something that we'll have to fix because torchdiffeq prefers to work in double precision floats. Could you file an issue so we don't lose track of it?

[ornate elk]

he made changes last week, but yea the rest is 8 months back :/

but it works

[fervent pier]

For some reason, I'm not seeing the Issues tab on your repo. Just Code and then Pull Requests.

[fallen quail]

Fuck. Uhh, I don't own the repo, I will write it down somewhere

[half island]

Nice. So as i got these, launched it though it's bat file, but it complains it has no models, even if they are in the models folder

[ornate elk]

you need to download the .ckpt version of the model

i tried it too but the safetensor wont worked

[half island]

Ah shit. The insecure one

[ornate elk]

yea xD

[half island]

Though, is the diontimmer's fork as "unsencored" as the one i linked to? Or does sencoredness depend on the model itself? :P

[ornate elk]

it depends on the model

[hexed pine]

Hi. I am using stable diffusion webui with chrome, recetly it suddenly stopped working - its saying that I dont have enough vram. The model has about 10 GB, my vram is 16 GB. I tried reinstalling it, all together with python, torch, git and cuda but its all the same. I cant even load 4 GB version of models, it just says I dont have enough vram. I am honestly lost at what is the problem. Tried reinstalling drivers. Deleting possibly corrupted files, but as I said I even tried fresh automatic install

The error is

AttributeError: 'NoneType' object has no attribute 'lowvram'

[ornate elk]

hey, whats your GPU? whats the model? and whats your webui version

[hexed pine]

its 4080, the model is 10GB version of newest stable diffusion v3 (tried redownloading), webui version is 1.9.4, python 3.10.6, torch 2.1.2+cu121

[half island]

Gotchu. So for this, do i just install cuda toolkit, and rerun? Or as it's in a venv, i need to add cuda toolkit to that?

"warning: User provided device_type of 'cuda', but CUDA is not available. Disabling
warnings.warn("

[hexed pine]

Just tried medium version - 4GB, but with no change

[ornate elk]

SD3 wont work in auto1111 right now

we all wait for an update ^^

[hexed pine]

ohhh

[ornate elk]

it only works in comfyui, stableswarm and sdnext

for now

[hexed pine]

thank you, ill try the other ones then

[ornate elk]

no problem 👍

[ornate elk]

i dont think you need cuda for it

should be installed via torch in the venv

[white tide]

I got around to testing out an LCM model on ComfyUI with Zluda just now

[ornate elk]

ah nice

[white tide]

1.8s for a 10 steps 512x512 image 😄

7.8s for 1024x1024

[ornate elk]

awesome!

[white tide]

Now to look for some LCM workflows I could try out

[white tide]

Generated with LCM 🙂

[half island]

Got cuda working, sorry i didn't edit, forgot to

https://github.com/diontimmer/audio-diffusion-gradio/issues/2#issuecomment-2151127309

[hexed pine]

are you still here? tried stable diffusion webui with other stuff that worked before and I got the same result. I tired AbsoluteReality and EpicRealism_Natural, both about 2GB for my 16GB card and it still says "low vram"

[wet flame]

idk why it keeps giving me this error

[cerulean hound]

how to add --lowvram to ComfyUI on Pinokio?

[wet flame]

and i can't run foocus even with 16 gb vram

idk what happened but it was running all good then everything starts crashing

[cerulean hound]

all i found was this, but cant find it in Pinokio https://github.com/comfyanonymous/ComfyUI/discussions/3158

[wet flame]

what is pinokio?

why you need it?

[cerulean hound]

https://pinokio.computer/

[rancid nymph]

i cant switch checkpoints for some reason

[rich garden]

When I generate an image it gets stuck at 99% but the command prompt window says “total progress 100%”

After waiting a little bit, the image is finally complete but the quality is really bad like 360p for example. It’s pretty much the same quality as if it was still being generated. I can’t really send the picture here cuz I feel it’s a tiny bit on the NSFW side but not extreme

[solemn moon]

i have this too

[wet flame]

OutOfMemoryError: CUDA out of memory. Tried to allocate 1.01 GiB. GPU 0 has a total capacty of 16.00 GiB of which 9.74 GiB is free. Of the allocated memory 4.75 GiB is allocated by PyTorch, and 331.25 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[solemn moon]

check yr drivers are up to date?

[wet flame]

yeah all updated

reinstalled everything today

did not had this issue but it started randomly

and sometimes i get no errors

i should not get this error at all with a 16 gb gpu

[solemn moon]

maybe the full log will help someone help you better

[low basin]

Sorry went AWOL and theres far too many messages to read lolol, did you figure out the issue? I think you said you were using an external ssd/hd to hold the models? If so what is the read and write speed of it? That would 100% cause your astronomically long model load times...

[low basin]

You're the second person in 24 hours to post the exact same issue. The person I just responded to was trying to troubleshoot theirs last night. I have no clue what it is but I feel for you guys

[wet flame]

it was not giving issue for long time

even on my 4 gb laptop with med vram

it started randomly now i can't use AI to make game art assets lol

[low basin]

What GPU? 12GB VRAM I assume?

[wet flame]

back to manually doing things 😦

[low basin]

Dur I cant read 16GB right there. Yeah something isn't right

[wet flame]

on desktop its a 4060 ti

and laptop its 3060

i had no issue for long time on laptop but it just started randomly last week

[low basin]

Is this the first install you've done on the desktop?

[wet flame]

issue started last week bascially

[wet flame]

no i did before too

[low basin]

So it didn't just start after a new install? You just reinstalled to try to fix it?

[wet flame]

no i did a fresh install of everything on desktop

and doing that on laptop too

even reinstalled python and git

[low basin]

So you did the manual install steps as opposed to the quick install? (just ruling out the install method as a cause here)

[wet flame]

but as far i know python or git should not cause the problems

i followed his guide
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides#nvidia-automatic1111-webui-stable-diffusion-webui

although i can do it manually

but takes long time

[low basin]

Which fork did you use (A1111, Fooocus, etc...)

[wet flame]

A111

i can't run focus at all idk why

[low basin]

Hmm, well if you're willing you might try to use this guide (complete quick install) and see if it works. It's what I downloaded a few weeks ago on 4070 and worked just fine. https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs

Again out of curiosity what Nvidia driver and CUDA version do you have? You can quickly see by typing " nvidia-smi " into command prompt

[wet flame]

I think on desktop i got cuda 121

n 555.99 latest drivers

[low basin]

If you do try another install method though I would probably purge any other installs prior just in case it causes issues. You can copy and save model and lora folders before if you have any

[wet flame]

also for some reason it keeps reducing my c drive space

i had like 70 gigs free on my OS c drive

now its down to 60

[low basin]

Depending on what you select when running it may just be from it downloading needed models in the background. I have no clue but yeah it eats storage for breakfast

[wet flame]

darn lol i got OS on a 250 gigs sdd

around that time SDD was super expensive

[low basin]

That's rough... I'm already 600gb down on an ssd I got 2 weeks ago. Long gone are the days when 64gb was all you'd need for a lifetime 😔

[wet flame]

yeah the error is random

also it started after the error resume download depreciated

[hollow prism]

so i updated to the new version of automatic 1111 and i cannot find the directory in the settings where my lora and checkpoints go to. Where is the setting?

[errant tundra]

Hi, I'm hoping I can get some help with ComfyUI, specifically trying to upscale the latent image for a second pass of ksampler. I've watched a few videos but I can't seem to get ANYWHERE close to what is expected.
I'm trying to basically run through my first KSampler at a lower res say 360x640, which does give decent results, then upscale the latent image to a higher resolution to run through another ksampler with less steps since it's a higher resolution in order to add more detail.

My problem is that I do a preview image of each step and the first ksampler comes out fine, I then upscale the latent image then preview that which then gives me a MUCH blurrier version of the first image instead of retaining the same or better quality, then I preview the output of the second ksampler image and I get a COMPLETELY DIFFERENT image of which every single time is completely malformed. I even tried upscaling the actual image of the first ksampler output then encoding it back to latent and running that through the second ksampler but I still get absolutely awful results.

basically I'm trying to do a HiRes Fix but in comfyui

@old sky do you as the developer have any suggestions I could try? Sorry if I am not supposed to ping.

[old sky]

https://comfyanonymous.github.io/ComfyUI_examples/2_pass_txt2img/

[errant tundra]

I always get crap results

like if I do a 15 step first ksampler then on the second sampler I do a 1 step shouldn't I get a pretty similar image to the first? Cause I actually just get a crap load of noisy garbage that looks nothing like the original

[old sky]

do more than 1 step

[eternal jungle]

Hey, I have a question regarding images from social media, are these generated with SD or another image generator? https://www.tiktok.com/@societiesr

[low basin]

Models should be at C:\Users\Name\Desktop\ sd.webui\webui\models\Stable-diffusion and lora's are same path but "lora" as the final path instead of Stable-diffusion Disregard I'm an idiot and that was not at all what you were asking LMFAOOO

[restive sparrow]

Does anyone know if you actually need this CUDA stuff installed to use xformers? "CUDA Runtime"/"CUDA Development"
thank you

I installed them last year and still ahve them

[split comet]

this prompt (tongue out, tongue:2.0) is 2.0 to everything inside the backet right? or no?

[fair oxide]

Feedback: Please make support for Stable Artisan to be universally available in Stable Assistant, there is a Javascript discord error some of us receive when having to log in the current way that prevents us from accessing Stable Artisan on Discord because it continually does a failed login loop. I've tried everything, and now I will have to drop my subscription because I can't access Stable Artisan specifically on Discord.

[tame yoke]

is this normal?
i only changed live previews to be jpeg

[ornate elk]

Nope you don't need cuda installed for xformers

[ornate elk]

In Auto1111?
Can you show the full cmd log?

[ornate elk]

Can you show the txt2img settings when you get that error?
Make sure your nvidia driver is updated.
If you changed your gpu in the recent time then you need a complete driver reinstall.

[ornate elk]

Models/stable-diffusion and models/lora

You can't change these paths in settings

But you can define other folders with the launch args:
--ckpt-dir "D:\Path\To\Models"
--lora-dir "E:\Example\Path\Loras"

[ornate elk]

Everything in the ()

[ornate elk]

Can be normal yes

[ornate elk]

Is ist sdxl or 1.5 model?
What's in your webui-user.bat?

[rich garden]

how do i check?

it could also be that i'm using a laptop perhaps cuz it was working again a bit later

[ornate elk]

Check the model file size in models/stable-diffusion
And for the webui-user.bat just right click and edit it

[rich garden]

is this the sdxl model?

idk what tge 1.5 model is called so idk if i have it

[ornate elk]

Yep

1.5 models are smaller than 6gb

[rich garden]

oh i thought u meant there was a file with "1.5" in the name

[ornate elk]

Nope

Only some sdxl models use XL in their name

[rich garden]

what about this?

[ornate elk]

Thats the first 1.5 model made by stability ai

[rich garden]

it seems most of my models are 2-3 GB's

[ornate elk]

These are based on 1.5 then

[rich garden]

ok

i'm still a bit new to this btw 😅

getting there tho

[ornate elk]

No problem at all ^^

[rich garden]

[rich garden]

is this fine?

[ornate elk]

What's your gpu?

[rich garden]

uhh.. does nvidia rtx 3060 laptop help?

i'm not the biggest tech gal xd

like in terms of specs and shit

[ornate elk]

Okay, yea the settings are correct for that gpu

[rich garden]

ok

i do plan to get an actual desktop sometime this year hopefully

[ornate elk]

Thats an good idea 🙂 more power for SD

[rich garden]

yes

cuz i don't think my laptop can use "pony" from what i've tried

[ornate elk]

But for your issue from above, recheck if it happens only with the sdxl model selected or any other model too

A 6gb vram gpu has a hard time with sdxl and pony (which is based on sdxl)

[rich garden]

it happened with a model only 2 gb's

earlier today

which is when i came to the server

[ornate elk]

Ah okay. Then make sure you don't run an other heavy programms in the background, like games or Wallpaper engine.

Whitelist the webui in any browser adblocker

[rich garden]

yeah i don't normally do PC games and i don't have wallpaper engine for now

i have the ad blocker extention for chrome tho

also another question, i tried generating some images earlier today and when i tried some negative prompts like ||bare skin||, ||uncensored breasts|| etc. it didn't actually to what i asked sometimes which means it had the opposite of what i wanted in the pics

shit

[ornate elk]

Also did you used upscaling, hiresfix or any extension while getting that issue?

[rich garden]

i pressed enter on accident

[rich garden]

yeah i tried hiresfix sometimes

[ornate elk]

If you use that make sure to always set hires steps to 10

If not it can error out

[rich garden]

ok

[rich garden]

i fixed the message now

[ornate elk]

It depends on the model. Dont use the base sdxl or 1.5 model as they are not trained for it

But other models do fine

[rich garden]

i used "yesmix"

[ornate elk]

Okay that should do it

[rich garden]

it still had the opposite of what i wanted most of the time tho

[ornate elk]

You can add
Nsfw, nude in the negative tag

Or skimpy

[rich garden]

i didn't even think about "NSFW"

i did add nudity and naked tho

[ornate elk]

You can also add sfw into positive prompt ^^

[rich garden]

true

one weird thing that also happened is when generating images it somehow brought 2 girls into a pic when i believe i only used either 1 character lora or "1girl" so idk where the other one came from but it actually looked sick as in the end. so basically i got lucky but i would've liked for results to appear as often and accurate as the random pic i got

[ornate elk]

That can happen if the resolution is set a lot higher than what to model got trained on.
1.5 models are trained on 512x512
So doing a 1024x1024 can cause duplicates and deformations.
Thats why hires fix is needed for 1.5 to gen like 512x768 and upscale by 2

[rich garden]

oh and that reminds me, i tried using 2 character loras to get them into one image but most of the time it would only add 1 character and not the other. i was using the "BREAK" method thingy and the regional prompter from this guide https://civitai.com/models/339604?modelVersionId=403892

[rich garden]

ah ok cuz i did 1280x720

[ornate elk]

Yea that likely caused that behaviour

[rich garden]

this?

wait lemme double check

ur saying if the dimentions are too high on a 1.5 model then it can cause 2 characters to appear for example?

[rich garden]

cuz i had a bit of a brainfart and thought u were referring to this 😅

[ornate elk]

Yes

[rich garden]

alright thanks

[rich garden]

also do u possibly have a solution for this?

[ornate elk]

For that I would use the tiled diffusion extension which has regional prompt control

[rich garden]

this? https://github.com/shiimizu/ComfyUI-TiledDiffusion

just looking at these sample pictures it seems very complicated

nvm i assume u mean this

[ornate elk]

Yes that one

[rich garden]

how do i use it? 😅

oh wait i missed the bottom lol

i'm still not 100% sure tho

[simple raptor]

guys i need help with a model and a lora don't showing in the UI aftere refresh and restart it. Did i do something wrong??

[ornate elk]

whats the models file size? or whats the model name

[rich garden]

is every region supposed to be 1 character each? cuz i generated an image with 2 regions and both sides had 2 of the same character doing the same thing and it seems the background has been split in half just like the red and yellow boxes

region 1 and region 2 were both 2 different character loras

[simple raptor]

https://civitai.com/models/377060/retro-fantasy-style
and the relative model suggested anything else v4

[steep ivy]

Hi, I am facing an issue when I try to change a model, it returns this error:

attributeerror 'nonetype' object has no attribute 'lowvram'

I saw other people mention that here before but it seemed to me that the solution is GPU specific. My GPU is rtx 2080 super and my args are --listen --medvram --no-half-vae

[ornate elk]

thats a lora not a model

it goes into models/lora folder

[ornate elk]

add --xformers and change --medvram to --medvram-sdxl

then try an other model

[ornate elk]

you need to make a region for the background, and 2 other foreground regions 1 for each character

[rich garden]

ah ok

[simple raptor]

I've put it in the lora folder and the anything else model in the modello folder but, Nothing. I'll send you some screenshot

[steep ivy]

still the same :(

[tame yoke]

is comfyui better then a1111 or and is it also deprecated like forge?

[ornate elk]

whats the model you try?

[solemn moon]

better if you have a hand on nodes

[steep ivy]

None of the models work, it's stuck on sdxl turbo if it matters

[ornate elk]

how much RAM do you have?

[rich garden]

is it also possible to make 2 or more characters interact with each other? like lets say i would want one character to place their hand on the other character's shoulder or something else that's more advanced? if it's possible then i'm not sure what i should type as a prompt and in which box

[ornate elk]

yea its possible by overlapping the zones

[rich garden]

so like this?

[ornate elk]

the background needs to be the full space

but yes

[rich garden]

ah ok one sec

so what do i type in region prompt 1 and 2?

just any example is fine

i would assume in both region prompts i have 1 character lora each at the start?

[ornate elk]

you need 3 regions
1 for background
1 for character 1
1 for character 2

[rich garden]

ah shit yes my bad

[ornate elk]

and in the character prompts just add the loras and describe the char

[rich garden]

describe it as in looks and what it should do?

[ornate elk]

yep

[rich garden]

ok

i should probably mention that sometimes when i describe a character it tends to have some bugs like extra legs for example even tho i've added things like "deformed" "poorly drawn hands" "mutated fingers" etc. in negative prompt

[ornate elk]

yea that can happen sometimes

[rich garden]

is there a method to be more likely to avoid it?

[ornate elk]

more negative tags, other resolutions, using sdxl or pony models

[rich garden]

ok

[ornate elk]

and when using hires fix lowering the denois can help too

[rich garden]

denoising strength?

i generated an image with 3 regions but both the characters look like twins when it's 2 different loras

[steep ivy]

16, I managed to fix the issue by deleting the venv folder. But now it gives me 1.5s/it :(

[ornate elk]

make sure you always launch the webui-user.bat and not the webui.bat

and if your using sdxl with an 8gb gpu, thats pretty normal speed

[steep ivy]

nope, 1.5 dreamshaper

[ornate elk]

ohhh

can you show which settings you used?

[steep ivy]

args or screenshot of the webui?

[ornate elk]

both

[rich garden]

what would i have to type in a region prompt to get one character interacting with the other?

[steep ivy]

With that I now get ~2it/s is that normal?

[ornate elk]

you have batch size on 4

thats slows down because it generates the images at the same time and using more vram

[fierce hollow]

Images began to be generated for a long time. What could this be related to? 3070ti + 32 GB ram

[steep ivy]

with 1 I get ~4.8it/s. Would you consider that normal for 8GB VRAM RTX 2080 super, 16GB ddr4 RAM?

[ornate elk]

make sure you close every other programm like games or wallpaper engine etc

is your webui updated ?

check the version at the bottom

[rich garden]

@ornate elk

[ornate elk]

any interaction, like touching, kissing, hugging etc

[rich garden]

oh i meant like only one character interacting. for example lets say one of them is pregnant and the other character rubs or hugs the belly etc.

[ornate elk]

then only prompt the interaction for one character, like hugs.. or waves etc

mostly the interaction will come automaticly when the regions overlapping

[rich garden]

so what could i type for something like this? (this is what i meant by more advanced btw since it's only one character interacting)

[fierce hollow]

help

[ornate elk]

more information needed

long time? how long?,
what model is used?
what settings are úsed?

[fierce hollow]

An 832x1216 image with Hires has been generated for 5 minutes already. fix that is, the final image is 1248x1824
Model - realvisxlV40_v40LightningBakedvae
The settings are as follows: Sampling steps - 5
Sampling method DPM ++SDE
CFG - 1

It took me about 7 minutes to generate one image

[rich garden]

i'm only getting 1 character in the image now. here are my prompts and settings to give u an idea

[ornate elk]

SDE samplers can take double the time then non SDE samplers, so try DPM++ 2M or Euler a.
Also when using Hires fix, set the hires steps to 10

[rich garden]

should i maybe not have posted this cuz it's too lewd?

forgot about this btw

[ornate elk]

nah its fine

[rich garden]

ok

what should i type instead?

or do

[ornate elk]

checkout this guide:
https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/wiki/Regional-Prompt-Control

but i cant help much with difficult positions

its mostly more luck

you can also try/use the regional prompter extension, but i find it a bit difficult to use

[fierce hollow]

Milking this model requires exactly such settings to create a realistic image. It used to take less time

[thorny dragon]

hello there, i started using stable diffusion about a month ago with automatic1111 but for some reason suddenly it was giving me "cuda out of memory" error. Even after fixing that error, a1111 was still running really slow, like 30 mins per image. Then i installed forge UI which is crazy fast but i am facing troubles in img2img. Even at low denoising strength it is distorting the body and instead of upscaling the image loses detail and quality. I am using rtx 2060 6gb vram graphics card. Can anyone help me fix this?

[ornate elk]

hey, i can help you fix auto1111

[ornate elk]

you can screenshot them, maybe there is some wrong setting there

or try a restart

[thorny dragon]

thank you!!!...plz tell me what screenshots i need to provide...

[fierce hollow]

https://imgur.com/a/hHn8neB

[ornate elk]

your content of the webui-user.bat
and your txt2img settings when getting a slow image.

[ornate elk]

looks normal, whats your gpu again?

[wet flame]

@ornate elk yes its a new install of gpu drivers and sd A111

where can i find the full logs of errors?

[ornate elk]

in the cmd

copy the whole text

[wet flame]

ok

[rich garden]

how do u change to dark theme? i've always wondered that

[fierce hollow]

3070ti

[fierce hollow]

Oh... somewhere in the settings, I'm sorry, I don't remember anymore, I switched over a long time ago...

[ornate elk]

add --theme dark
to the webui-user.bat

[rich garden]

thx

[ornate elk]

and you have --xformers --medvram-sdxl --no-half-vae in your webui-user.bat?

[wet flame]

my errors started appearing after this error btw
J:\AI Apps\SDA111\stable-diffusion-webui\venv\lib\site-packages\huggingface_hub\file_download.py:1132: FutureWarning: resume_download is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use force_download=True.
warnings.warn(

@ornate elk

OutOfMemoryError: CUDA out of memory. Tried to allocate 1.01 GiB. GPU 0 has a total capacty of 16.00 GiB of which 9.23 GiB is free. Of the allocated memory 5.50 GiB is allocated by PyTorch, and 81.35 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Time taken: 1 min. 10.6 sec.

A: 6.81 GB, R: 7.09 GB, Sys: 8.3/15.9961 GB (51.7%

[fierce hollow]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--autolaunch --xformers --medvram-sdxl --no-half-vae

call webui.bat

[ornate elk]

what are you trying to do?
it seems like your using a tiling extension in img2img or seomthing

[ornate elk]

okay looks good

[wet flame]

yes i use that always it did not give errors before

creating seamless tiles from my images

[ornate elk]

can you show the whole img2img settings then?

[fierce hollow]

I started having these problems after installing ComfyUI (not immediately, after some time ...)

[ornate elk]

you can try to delete your venv folder and relaunch the webui-user.bat
that can remove unwantedd changes

[ornate elk]

you need to remove --always-batch-cond uncond and instead add --xformers --no-half-vae
then relaunch the webui-user.bat

[fierce hollow]

https://imgur.com/a/Pr5ZwwO
?

[ornate elk]

delete the whole venv folder and relaunch the webui-user.bat

[rich garden]

how come some of the words in the prompts are in brackets?

[fierce hollow]

Sorry, but what is this folder responsible for?

[wet flame]

nothing fancy just making some textures from images

[ornate elk]

thats for getting more attention to the word
can be used like (((word))) or (word:1.3)

[rich garden]

ooh ok

maybe that'll help

[ornate elk]

it stores librarys and stuff from extensions. also the whole torch and xformers stuff.
when launching the webui-user.bat it will get recreated automaticly

[thorny dragon]

okay, i will try that....but can u help me with forge tho? i would really like to use forge but the img2img is generating weird bodies even in low denoising strength..any fix for that?

[fierce hollow]

I got it, I'll try it now, thanks

[ornate elk]

Forge is deprecated for normal users. with the next update it will break. so i dont recommend using it

but you can use img2img in auto1111

no clue what could be wrong here :/

[wet flame]

i got no clue either

[fierce hollow]

By the way, a question about SD3, has the automatic111 update been released that supports these models?

[wet flame]

and i feel bad for being a computer techniciain and i can't fix it sigh

[ornate elk]

its in development but not stable release yet

[fierce hollow]

Is there an approximate time when to expect? At least approximate dates

[ornate elk]

nope, but my guess is a week

[fierce hollow]

Understood, thanks. Will this be notified in #📣｜announcements and #1010934719455707218 ?)

[ornate elk]

nope

[rich garden]

i tried ((hand on shoulder)), just as an example and it didn't work. here's my regions

i read through that link u sent btw

so far it seems to be the interaction that doesn't wanna work

[thorny dragon]

@ornate elk i was getting cuda out of memory errors beacuse of cuda fallback system on new nvidia drivers which enables the gpu to use my pc's ram as extra vram...according to some reddit post this causes massive memory leak and was the reason of slow genaration..they recomended turning it off...but when i turn it off it gets even worse, u got any opinions on this?

[ornate elk]

Dont turn that off!

Like you said it makes it worse

[thorny dragon]

btw, this kinda helped to reduce the genaration speed a little..tnx...

[ornate elk]

You need the tiled diffusion extension if your going for higher resolutions

It has an Option called Tiled VAE

Thats what you need for higher resolutions or upscaling

And its normal that it takes longer for your gpu if your using sdxl

For sdxl models you also need to increase your windows pagefile if you only have 16gb RAM

[wet flame]

i get the cuda out of memory with 16 gb gpu

@ornate elk
OutOfMemoryError: CUDA out of memory. Tried to allocate 1024.00 MiB. GPU 0 has a total capacty of 16.00 GiB of which 8.21 GiB is free. Of the allocated memory 5.03 GiB is allocated by PyTorch, and 1.56 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Time taken: 2 min. 28.0 sec.

A: 6.79 GB, R: 7.12 GB, Sys: 8.3/15.9961 GB (52.0%)

[thorny dragon]

ohhh! i see....i will install that extension then..tnx for all these valueable tips man!!!!

[thorny dragon]

how do i do that? the windows pagefile thing....

[fierce hollow]

It didn't help

[fierce hollow]

95% processing takes a long time (1 image) Without increasing the quality. The standard 832.

I'm sorry, it was the first generation, I tried the second one - everything is OK, now I will try to zoom in

[ornate elk]

No problem.
Its under extensions tab, available, load from, then search for Tiled.
Then only enabled tiled vae and leave its settings on default.

@thorny dragon for pagefile look here:
#🤝｜tech-support message

Make sure its enabled for C drive and disabled for any other drive.
Then set the min to 16000
And Max to 24000

And make sure you have 15gb free space on C

[simple raptor]

Here where i put the models, is it ok??

As you can see it doesn't show them

[ornate elk]

your webui version is to old

and remove this file from the models/stable-diffusion folder

[simple raptor]

ahahah damn, can i update the UI from the UI itself?

[ornate elk]

not with the ui. you have to do some more steps.
first in the webui click on the extensions tabs and click "check for updates" then let it load.
after that.
close the whole webui and cmd.
right click to edit the webui-user.bat
in an emtpy row you have to add Git pull
then at the line Commandline_ARGs= you add --xformers --no-half-vae
if it isnt there already.
Then save the file.

After that open up a cmd and type
Pip cache purge
Then delete the venv folder.

And finally launch the webui-user.bat

Then it should download the updated files

[simple raptor]

ok i've done, i'm not sure of the next steps, i don't understand well

[ornate elk]

looks good, dont forget to save

[ornate elk]

whats your gpu btw?

if it has 6gb vram or less you need to add --medvram

if it has 8-11gb vram you ned to add --medvram-sdxl

[ornate elk]

how much dedicated vram? scroll down a bit to check

[ornate elk]

okay 6gb

then add --medvram to the webui-user.bat

like you did with xformers

[simple raptor]

just next to xformers?

[ornate elk]

yep, like
--xformers --medvram --no-half-vae

[simple raptor]

perfect

[ornate elk]

Then the next step is to Open up a cmd.
click on the windows symbol and type cmd
then launch the cmd.
in the cmd type
pip cache purge and then press enter.
that will remove old temp files from the first webui install

[simple raptor]

ok

[simple raptor]

done

[ornate elk]

and then delete the venv folder

that folder will be recreated at the next launch with the new files

[simple raptor]

this one?

[ornate elk]

yep

[simple raptor]

ok, then i restart?

[ornate elk]

yes, now launch the webui-user.bat

[simple raptor]

still the same issue, it didn't show the lora and the model

[ornate elk]

make sure the lora is only in models/lora

and not in model/stable-diffusion

and then selecte and 1.5 based model

and delete the 1.4 model from your models/stable-diffsuion folder (its really old and bad)

[simple raptor]

here the models, i need to delete sd-v1-4?

[ornate elk]

yep

[simple raptor]

ok, than as you can see should be all ok, right?

[ornate elk]

okay and what model is selectet at the top left?

in txt2img

[simple raptor]

still mantein the old setting

not showing the one i need and showing the deleted one

[ornate elk]

the lora should be on the lora tab

also click the blue refresh button

then the 1.4 should be gone

[simple raptor]

i've clicked multiple times but nothing happens and the lora tab is empty

[ornate elk]

and extensions installed? that could mess with loras?

can you try download an other lora to test if the lora is the problem?

[simple raptor]

yeah, now i try

[ornate elk]

for example this one:
https://civitai.com/models/58390/detail-tweaker-lora-lora

[chrome mica]

It's been like this for a while now, it doesn't give me any errors but it doesn't do anything. Does anyone have an idea, or do they just need more time?

[ornate elk]

what even is this? 😄
what tool are you trying to run

[chrome mica]

Srry XD Im trying SD with Zluda in windows with AMD, text to image

[ornate elk]

ohh, so your trying SD Next i asume

whats your GPU?

[chrome mica]

Yes

[chrome mica]

RX6600XT

[ornate elk]

then i would recommend not trying it with SD Next and instead following my Automatic1111 AMD Zluda Guide.
im on AMD too and ive made many guides.
you can find them all here and feel free to ask if any step isnt clear.
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides

[ornate elk]

now reload the lora folder

click on the reload icon on the right

[shy charm]

ok so i'm moving back to a1111 since forge is apparently going to shit the bed
it told me to update xformers, so i did, and now it's saying this shit

[ornate elk]

hey whats your GPU?

[chrome mica]

I can try, which one do you recommend?

[shy charm]

RTX 3070, i was previously using a 1070 though and moved to forge when i got the 3070, so i never used the 3070 with a1111

[ornate elk]

the Automatic1111 with ZLUDA guide

it covers everything you need. i helped like 3-4 guys with a 6600xt getting it to run ^^

so it should work for you

[ornate elk]

okay, whats in your webui-user.bat?

[shy charm]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --reinstall-xformers --xformers

call webui.bat

it said xformers was out of date so i put that arg in there

my A1111 is also up to date on the stable release, i did a git pull before anything else

[ornate elk]

okay add Git Pull into the second line thats empty
then remove --reinstall-xformers
then add --medvram-sdxl --no-half-vae to the --xformers
then save.
Now install Python 3.10.11 (because 3.10.7 has some bugs)
https://www.python.org/downloads/release/python-31011/
after that delete the venv folder
then relaunch the webuio-user.bat

[chrome mica]

I tried it but from what I see is that you have a few more steps added than another one that tried to install. I have tried several times to install it, is there anything you would recommend I clean?

[shy charm]

man why did the forge dev have to ruin it, this is so complicated

[ornate elk]

you can remove the sdnext (automatic) folder and in a cmd run
pip cache purge
then your ready

[chrome mica]

I'll do that, thank you very much for the help!!

[ornate elk]

and make sure your AMD driver is updated

and for any question feel free to ask

[shy charm]

i'm just going to keep using the current version of forge until there is a genuine reason to move on

[ornate elk]

why not fixing your auto1111 install?
most stuff from forge will get into the next auto1111 releases, thats why he forge gets back to experimental state, to improve auto1111

for example SD3 will not be usable with Forge

[shy charm]

someone said the dev branch of a1111 has the same performance improvements but the comments on the forge announcement say otherwise

[ornate elk]

you have a 3070, thats work great with auto1111 preformance wise

until last year i used a GTX1080 and had no issues even with SDXL

[simple raptor]

nothing my man

[half island]

Isn't torchmcubes gotten from pip? As i'm missing that module for a node

[shy charm]

i feel like maybe i should just completely re-install a1111 because this install is from like 2022 and i've only kept it up to date using git pull

[simple raptor]

would you raccomand a fresh install?

[ornate elk]

yep, that could solve it.
because your install was really old

follow the guide from the pinned messages

[simple raptor]

so, there's something i can keep?

[ornate elk]

you can keep your models and loras

and outputs

[shy charm]

i don't remember if i installed python using bucket or scoop or whatever it's called

[ornate elk]

yea then start fresh ^^ youll find an easy guide for that in the pinned messages of this channel (it will take like 10 mins max)

[simple raptor]

ok, i change folder for those and then i delete the webUi1111 folder and start from 0?

[ornate elk]

yep

[simple raptor]

can you link me a new guide please?

[ornate elk]

Sure here:
#🤝｜tech-support message
or here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides#nvidia-automatic1111-webui-stable-diffusion-webui

[shy charm]

i don't remember where i installed python 3.10.7, can't find it in the start menu, and it isn't in the default install location for 3.10.11

[obsidian gull]

guys any idea why my images are taking 5 minutes to generate? i have a rtx 4060

[ornate elk]

then uninstall via system settings /apps

or open up a cmd and type
where python to get its path

[ornate elk]

hey, whats in your webui-user.bat?

[shy charm]

ah, that's where it is

[obsidian gull]

im sorry but how can i find whats in there?

[ornate elk]

right click and edit it

[shy charm]

i don't remember why i ever installed python using scoop

[ornate elk]

yea never saw anyone here using scoop xD

most people use python or conda

[obsidian gull]

this one?

[ornate elk]

the rxt4060 has 8gb vram right?

[obsidian gull]

yep

[shy charm]

i'm just going to delete the whole scoop folder, surely nothing bad will happen
if i did install it for a reason, it's probably irrelevant now since i can't remember it

[wet flame]

put med vram

[ornate elk]

at the line commandline_args=
you need to add:
--xformers --medvram-sdxl --no-half-vae

then save and relaunch the webui-user.bat

[wet flame]

you need that sdxl command to run sdxl models?

[ornate elk]

no but for 8gb vram cards --medvram-sdxl is better than --medvram

because --medvram slows down 1.5 models

[ornate elk]

yea probably ^^

[shy charm]

is the venv folder on my a1111 install supposed to be 10 GB

[ornate elk]

yes

[obsidian gull]

hey i got a new problem now

😔

photos are turning out like this

[shy charm]

what's your cfg and steps

[ornate elk]

whats the model your trying?

[obsidian gull]

its a model i got from civitai

[ornate elk]

yea check if its made for sdxl or 1.5

[shy charm]

one time i accidentally cranked the cfg to 30 and didn't notice and i was like "wtf why is everything turning out like crap"

[ornate elk]

and yea make sure cfg is set to 7

[obsidian gull]

it says 1.5

[ornate elk]

okay can you show your txt2img settings, like steps, resolution etc?

[shy charm]

it honestly really would have helped to have the image with the metadata

[obsidian gull]

like when it is generating it is fine, looks normal, but when its at the end with the final result it fucks it up

[ornate elk]

make sure you also use an 1.5 vae

or when its sdxl based you need to use an sdxl vae

[obsidian gull]

is this a bad one?

[ornate elk]

nope thats the 1.5 one

how big is the model file?

[obsidian gull]

whats bis?

[ornate elk]

big

[obsidian gull]

oh it was around 6gb

[ornate elk]

Then its mostly an sdxl model

That needs an sdxl vae

[obsidian gull]

can you send me a link for one?

[ornate elk]

Like this VAE here:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

[obsidian gull]

thaks

thanks'

[ornate elk]

The second from the bottom is the right one

Np

[shy charm]

i don't think i've personally encountered a checkpoint that didn't have a baked vae
i use mostly pony derivatives though

[obsidian gull]

thanks man it worked

[simple raptor]

is it normal? or do i need to write something special in the command line?

[shy charm]

what file contains the settings again? i want to move them from my old install to my new one

[ornate elk]

You need to edit the webui-user.bat like stated in my guide

[ornate elk]

I wouldn't do that. But its ui-config.json and config.json

[wet flame]

someone help me even after having 16 gb vram it says out of memory

[shy charm]

do you think i can copy and paste my extensions at least?

[ornate elk]

You can but if they are that old I would recommend to just install them fresh

They dont contain much data anyway

[obsidian gull]

hey @ornate elk how do i generate more than one photo?

do i change batch count ?

[shy charm]

i launched it once and it worked, closed it, launched it again, and now it's doing this

[obsidian gull]

or batch size?

[ornate elk]

batch count

batch size will slow down the process

[ornate elk]

did you moved an extension over? or the configs?

[shy charm]

i did move the extensions, but i tried deleting them and it's still doing this

[ornate elk]

you then need to delete the venv folder

because the extensions can break the venv

most likely what happened here

[shy charm]

is there like a specific part of the venv i can delete to fix it? i have really slow internet

[ornate elk]

dont worry, the files got cached at the first download so it wont download big stuff again

there isnt a specific part, as the extensions could mess up any folder in there

[shy charm]

i think it most likely messed up because i copied my extensions from forge and forgot to exclude forge couple

[ornate elk]

ah well xD

[shy charm]

i'll try copying them over again without it and if it breaks again i'll just reinstall them

[wet flame]

hey i am unable to use sdxl and also it is giving me cuda out of memory can you help 😅

but i got latest driver and everything for the 16 gb 4060 ti

using A111

[ornate elk]

whats in your webui-user.bat?

[wet flame]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --medvram-sdxl --no-half-vae

call webui.bat

i did the one without the medram

same problem

[ornate elk]

forge has specific modified extensions that wont work in auto1111 because they were made for auto1111 and modified for forge xD
example is tiled diffusion and controlnet (please dont copy them)

[ornate elk]

okay looks good, what yre your settings

[wet flame]

same as before

settings in settings tab?

[ornate elk]

no txt2img settings, like resolution steps etc

[wet flame]

I am doing image to image

[ornate elk]

the full settings here, any script selectet?

[wet flame]

none

[ornate elk]

okay then please try txt2img first

[wet flame]

gave me connection timed out and press any button to continue

was there something happened with some update

everything was working just fine before

[ornate elk]

Nope, are you on the latest Version ?

[wet flame]

yes

A111 webui

[ornate elk]

Make sure any other programms are closed, like games or Wallpaper engine.

[wet flame]

nothing open besides a folder

out of memory again with a 16 gb gpu this is crazy

[ornate elk]

Yea something isnt right

We had some people with the same issue two days ago. They fixed it with a complete nvidia driver reinstall

You can also try to delete the venv folder

[wet flame]

ok

i did a fresh install yesterday including drivers

sometimes it works but most of the time it goes cuda out of memory

[shy charm]

what is the sechedule type dropdown for?

[wet flame]

just did that still cuda out of memory

[shy charm]

this is comical
it took 6 minutes to spit out the image

i am using a RTX 3070 with these commandline args
--xformers --no-half-vae --medvram-sdxl

what was it doing?

it took like 2 minutes just to start the hires steps

[young bridge]

hey guys

can anyone help me install comfyui pls

[shy charm]

forge does the same thing in way less than a minute

[restive sparrow]

ah thank you.

[subtle current]

forge might work better for your configuration then. use that.

I tried offering help but you argued. now i'm just annoyed

[shy charm]

i'm just confused and a little frustrated because i've been told to move back to A1111 and i just spent the last hour doing that

[wet flame]

why can't i use sdxl on my automatic a111?

it seems i can make 512x512 images fine but not 1024

[half island]

Hoi, any idea why comfy manager just stopped working after i updated it?

[shy charm]

how do i switch to the dev branch of A1111, and can i switch back to the normal one?

[karmic crown]

git switch dev
git switch master

[round glacier]

hayo, bois!

I have some serius trouble using A1111 AnimateDiff.

Whenever i try to use it, i'm welcomed by this error

and i seem to be unable to fix it.

I tried on my setup, as well on a new fresh install.

no diffrence.

Now, clearly, i'm doing something wrong. but i do not know what

[ornate elk]

Hey, what's in Your webui-user.bat?

And what's your gpu?

[round glacier]

3070ti

[ornate elk]

Thats strange

[round glacier]

That's my my fresh one

[wet flame]

also i can't load any sdxl models

[round glacier]

and that's my standard one

[wet flame]

before in my laptop with 4 gb ram everything ran smooth even sdxl

[ornate elk]

Ah okay. At the line commandline_args=
You need to add
--xformers --medvram-sdxl --no-half-vae

Then at AnimateDiff set the batch size to 8 or 4

[round glacier]

Lemme see, one sec

[ornate elk]

How much RAM do you have?

[wet flame]

vram or ram?

vram 16 gb on desktop and 24 gb ram ddr4

[ornate elk]

RAM, sdxl needs a lot to load

[shy charm]

it sucks that forge is going to die because it would instantly solve this issue methinks

do you know if forge has any equivalent to --share=True?
i'm going outside of my home network for a bit but i still want to be able to generate on my PC from my laptop

[ornate elk]

And the models are on HDD or SSD?

[wet flame]

its on hdd on desktop

sdd are kind of expensive

[shy charm]

SATA SSDs aren't that expensive these days and they are still way faster than HDDs

[round glacier]

No change. the settings.

[ornate elk]

Okay then check your Windows pagefile.
It needs to be enabled only for C and not for any other drive.
You can increase the pagefile by setting it to 16000min and 24000 max

[shy charm]

i haven't been able to find if forge has any way to create a public link

normally on my home network i just RDP from my laptop to my desktop but i can't do that from outside the network

[ornate elk]

For real? Can you recheck?
5 to 1 minute is huge

I never saw that decrease in forge

[round glacier]

any ideas CS1?

[shy charm]

forge takes like 45 seconds for me to generate a 1024x1024 image at 30 steps with 15 hires steps

i feel like something is probably just wrong with my A1111 install

even though it's literally a brand new install

sure enough, only 48 seconds

[ornate elk]

Yea thats the strange thing. But we had that yesterday too with someone that upgraded his gpu.
And his auto1111 thought its still using the old GPU with limited vram until he reinstalled the nvidia drivers

[ornate elk]

Is your webui updated?
Is your extension updated?

[round glacier]

we should not use the gaming ones, right?

[ornate elk]

The gaming ones are fine

And I would recommend using them

[round glacier]

kay, let me update

couse druing my mesxsing tryigng to get the Animate working

[shy charm]

i think my old install from 2022 that i had kept updated with git pull actually performed better

[round glacier]

i get only garbage outpout out of all 111's

xD

[hexed pine]

Sure thing. When launching I at the end get this error

At should I try to generate, it lists error with "low vram"

[wet flame]

btw for upscaling i suggest using a tool called upscayl

instead of wasting time using A111

its free and its local install

[ornate elk]

its a nice tool but it cant use latent upscale

so upscaling via hires fix is still better in auto1111

[round glacier]

^

Hires fix, sometimes, make it the image too... clean.

in my experience.

[ornate elk]

your trying to use SD3 in auto1111. Its not supported yet

[ornate elk]

true, but lowering the denois can help a bit

[round glacier]

yes.

still, when you try to make a image that is gritty, as in

like, damaged helmet or dirty something

it does tend to clean it out a little bit.

okay, newest drivers installe.d

let's try again

[hexed pine]

strange. I actually was using stable diffusion v3 for like 2 days. But randomly at one generating it stopped working with this error. So now nothing with SD3 can be used until its updated. Does it update itself or do I check the website from time to time? Thanks!

[ornate elk]

if you have git pull in your webui-user.bat it will auto update

if not, then you have to manually run it

[hexed pine]

nice, thanks

[ornate elk]

you can also add --update-check to the launch args to get notified (this wont update, just notify for an available update)

[round glacier]

Okay. Got a clean install, new drivers, and installed the animateddiff and control network. Should be good to go, with the command liknes you recommended, right?

[ornate elk]

yep

[round glacier]

the model for animateddiff (mm_sd15_v3) goes into i\extensions\sd-webui-animatediff\model

right?

stderr: ERROR: Could not install packages due to an OSError: [WinError 5] Access is denied: 'C:\new sdg\stable-diffusion-webui\venv\Lib\site-packages\cv2\cv2.pyd'
Check the permissions.

...and it crashed when i try to make an animation again.

[ornate elk]

for the sd3 model you also need to use its lora

try the sd2 model

[round glacier]

roger.

[ornate elk]

it should work easier

[round glacier]

downloading.

[steady thistle]

So I'm looking into a desktop in the $1300 range without peripherials, and I want it to run AI stuff smoothly. Do you have any build suggestions? Didn't know if this was discussed elsewhere also.

[round glacier]

no change.

all that console said.

C:\actions-runner_work\pytorch\pytorch\builder\windows\pytorch\aten\src\ATen\native\cuda\IndexKernel.cu:92: block: [89,0,0], thread: [66,0,0] Assertion -sizes[i] <= index && index < sizes[i] && "index out of bounds" failed.
Shows up quite alot.

[ornate elk]

what resolution did you tried?

and which model

you need an 1.5 based model (2gb) and the resolution should be 512x768 max

[round glacier]

512x512, mm_sd15_v2 o arUniverse_V8,

wait, ti's doing maths now

like, mega slow, but it's not crashing

[ornate elk]

ahh good

what was the batch size now?

[ornate elk]

what was your gpu again now?

[round glacier]

Contex Match size?

1

[ornate elk]

batch size

[round glacier]

i assume you mean this?

hmm... should it be taking multiple minutes per frame? Alot more then a normal image, of that size?

i guess, makes sense.

[ornate elk]

yes its very slow and takes very very long

[round glacier]

the output... is worrying, as well

yeah, those outputs feel like a stroke vision

[ornate elk]

you need to enable a setting in auto1111 to make animatediff work

[round glacier]

of course, i'm a dum dum

...how

[ornate elk]

and better try with 24 frames to test

[round glacier]

good test values?

[ornate elk]

in the settings go to Optmizations and there enable this:

[round glacier]

didn

didn't help

fuck, what the fuck is going on here xD

[ornate elk]

disable FILM

and check webm

then try an other model

best working for me was Photon

sampler euler a or dpm++2m

[round glacier]

it more or less works, now. Thanks, CS1.

you the man

[ornate elk]

no problem 🙂

[steady thistle]

Radeon and AMD and it's a laptop

[ornate elk]

ah you had the integrated gpu

[steady thistle]

AMD Ryzen 9 5980HX with Radeon Graphics, 3301 Mhz, 8 Core(s), 16 Logical Processor(s), Radeon RX 6800M GPU with 12GB GDDR6

[ornate elk]

okay, are you located in US or EU ? price related question xD

[steady thistle]

US! N.W. Arkansas to be more precise.

[ornate elk]

okay

[young bridge]

hello

I am getting this error

ImportError: numpy.core.multiarray failed to import

[steady thistle]

https://www.amazon.com/Skytech-Blaze4-Gaming-4-6GHz-NVIDIA/dp/B0D1XM7C9G?th=1 Thoughts?

[ornate elk]

Nope

Dont buy anything that has a GPU with only 8GB Vram

You want 12GB or more

The more the better for Ai stuff

[steady thistle]

Ah right, gotcha.

[ornate elk]

As processor go for AMD

[steady thistle]

Cool, is there an Amazon seller you like, or should I consider Best Buy?

[errant tundra]

is there a way to make comfyui fallback to shared gpu memory if dedicated gpu memory fills up? working with 6GB dedicated but 16gb shared

[ornate elk]

Have you enabled the fallback in the nvidia control panel?

[ornate elk]

Idk sry, im in EU so its very different
But I wouldn't buy complete pc builds on Amazon

[errant tundra]

cuda fallback yes

yes I have

[ornate elk]

Are you trying to use sdxl or sd3 ?

[errant tundra]

idk I'm using a different model. it's an NSFW model. Pretty sure it's SDXL and not sd3

since not using any sd3 specific latent images or things like that

[steady thistle]

Oh cool, danke sehr. 🙂

[karmic crown]

i guess you can just replace that with RTX 4070 TI SUPER 16GB, everything else looks fine to me

[fierce hollow]

Good day to all, please tell me the analog of the model realvisxlV40 for creating REALISTIC images of people.

[karmic crown]

it should not be drastically more expensive than NVIDIA RTX 4060 Ti 8GB

[steady thistle]

I was disappointed with the power supply, but it's such a small part. I should rip stuff out of my old desktop.

[ornate elk]

How much normal RAM do you have?

[errant tundra]

32GB

[ornate elk]

A 4070ti is great but pricey.
Also go for 32GB system ram. You'll need it

[errant tundra]

I'm working off my poor mans 1660 super XD

[karmic crown]

https://arstechnica.com/gadgets/2024/01/review-geforce-rtx-4070-super-makes-a-good-gpu-even-better-still-costs-600/ <- check out this, you have comparation between similar cards here

[steady thistle]

I'm being lazy and don't want to build it myself. Been there, done that.

[ornate elk]

Have you tried Auto1111? Does it behave the same there with sdxl

[karmic crown]

you do not need to do that, just find the card that is right for you, than buy something with that card

[errant tundra]

I immediatly moved away from A1111 for comfyUI, I only had the allocation error happen once I was just asking to see if anyone had any info

but would like some help in dms since I am working with an NSFW model and don't know if that's allowed here to post images

specifically with HiRes Fixing

getting crap results even tho last night I did get some pretty good results but now it's just trash

also comfyUI just seems overall MUCH better. on A1111 my entire pc would come to a near standstill and sometimes even lock up entirely. I haven't had any issues like that on comfyUI

[karmic crown]

you can try to generate SFW image by adding "nsfw, topless, nude, naked...." in the negative and something like "girl in white dress" in prompt. that works with some nsfw models, but not all

[errant tundra]

I can probably just switch off this model too but. idk it's whatever. I'm just confused how I was getting pretty decent results HiRes Fixing last night but now it's trash

since this is so garbled I think this is safe enough to post as an example:

[steady thistle]

Finding the right Gb size in the right processor takes some hunting

[errant tundra]

this is my second pass ksampler

comfy said to put start at step to 5 last night which gave me the decent results but then today it was just trash. tried at 7 but same thing

[karmic crown]

for PC - you want as much VRAM as possible and decent amount of RAM, and Nvidia works better with AI than Radeon. for Mac its much easier - just get one with most unified RAM you can afford 🙂

[karmic crown]

cfg is too high

[errant tundra]

but that is so low

[steady thistle]

I have become well acquainted with Radeon issues, I'm sorry to say. This laptop is heading to be the new work laptop.

[errant tundra]

running now with cfg at 1

[steady thistle]

If I'm using a turbo, 2.5 cfg is as high as I dare.

[errant tundra]

wdym by a turbo?

[steady thistle]

SDXL turbo model that allows few steps and gives good results

I'm using dreamshaperXL_turboDpmppSDE.safetensors right now

[errant tundra]

mine doesn't have turbo in the name

[karmic crown]

i like lightning more than both turbo and lcm

[steady thistle]

I stopped DLing models a couple months ago. I'll have to try lightning. What model do you recommend?

[karmic crown]

any with 8 steps

you can try https://civitai.com/models/202108/bluepencil-xl-lcm-lightning?modelVersionId=369896

its a merge of lcm and lightning

[steady thistle]

Nice, but I only use realism

[karmic crown]

https://civitai.com/models/331967/hephaistosnextgenlightning

this one is experimental

lcm version is stable and you can use slightly higher cfg than with "standard" lcm (up to 5)
https://civitai.com/models/200061?modelVersionId=256258

[steady thistle]

Nice

[karmic crown]

i haven't use a1111 for months, didn't have time and was busy at work

[errant tundra]

I just LOVE the modularity of ComfyUI and the fact it doesn't lag my computer at all is a MASSIVE W

[karmic crown]

on my mac a1111 works faster than both forge and comfy (with default settings) 🙊

but i managed to make forge to work slightly faster

[errant tundra]

comfy runs pretty fast for me but I think it's mostly model dependant.

[karmic crown]

with same model comfy should work faster on most computers

[steady thistle]

I haven't even considered forge. I just do a1111, tried comfy but didn't really go deep dive at all.

[errant tundra]

it's awesome honestly. much prefer it to A1111

other than hires fix being super finicky

[steady thistle]

Should I get the 13gb version or is the 6gb version okay?

[karmic crown]

for which model exactly?

[steady thistle]

https://civitai.com/models/331967?modelVersionId=371923

[errant tundra]

usually I tend to go for the larger versions so the model has more data to work with

[steady thistle]

FB32 vs. the BF16

I'm not sure what the actual difference is, and bandwidth issues. 😛

[errant tundra]

FP32 says it's slightly better

[karmic crown]

in general, smaller models need less VRAM, so always download smaller first
if it works, great
if not, get bigger one

[steady thistle]

6.62 gb it is

[karmic crown]

i will download that one when i finally install a1111 properly

i always have like 10+ experimental installations
since i need to test my install script and some issues

[errant tundra]

anyone willing to hop into a VC with me to help me get my HiRes Fix working properly again? NSFW Warning

[shy charm]

did adetailer just get an update recently? i swear it had a checkbox to activate it that was separate from the dropdown toggle

[karmic crown]

i see it on your screenshot

[mortal ember]

Does anyone have a recommendation for where I can learn the basics of installing and learning to use stable diffusions weights? I'm finding it difficult to find a simple step-by-step tutorial and am finding it difficult to wrap my head around it. cheers.

[steady thistle]

How about the 4070 rather than the 4070 Super? 12gb. It's difficult finding the PC I want for the right price

[shy charm]

i said separate from the dropdown toggle, it used to have one inside the menu itself like this
and i remember seeing it like that just yesterday

idk why they changed it

[karmic crown]

4070 super has more cuda cores than 4070, but if that is make price too high for you, go for 4070

[gilded canyon]

I'm facing issues again with a1111, 12gb of vram but I can't do a single generation on SDXL without running out of memory

[gilded canyon]

I'm looking at task manager and when I generate something without a Lora, it works, but when I use a Lora, it shoots up to 12GB of VRAM utilization

Any ideas on how to fix this?

fatcat pls save me

i see you typing

[low basin]

That's normal as it holds the lora in VRAM alongside the model. I don't think there's a way around that other than using Forge or Fooocus

[gilded canyon]

I am using Forge, it was working perfectly up until a few days ago

now i cant generate using LoRA without my a1111 crashing

due to no memory

[low basin]

Did you reenable sysmem fallback policy in NCC after you nuked drivers?

[gilded canyon]

no ill do that rn oops

[low basin]

The 4070 to 4070 Super is the most cost effective upgrade/choice in the 40 series. 10% price increase for nearly 18% more CUDA/Tensor cores. Regardless, you'll be incredibly happy either way and both have 12GB VRAM so it's just generation speed that is affected (although the 4070 Super can pull some serious O/C...)

[gilded canyon]

enabled it and lora is still mkaking it crash

Moving model(s) has taken 0.64 seconds
To load target model SDXL
Begin to load 1 model
Reuse 1 loaded models
[Memory Management] Current Free GPU Memory (MB) = 4211.08544921875
[Memory Management] Model Memory (MB) = 0.0
[Memory Management] Minimal Inference Memory (MB) = 1024.0
[Memory Management] Estimated Remaining GPU Memory (MB) = 3187.08544921875

then its gone

[low basin]

What is the file size of the lora it's trying to load

[gilded canyon]

65,400 KB

is the size of the .safetensors file

im about to just switch to comfyui atp

[low basin]

Are you missing a 0 in there by chance?

[low basin]

Huh that's a tiny lora

The other day I think you mentioned something about using an external drive or similar? I wonder if that may have something to do with it

[gilded canyon]

yup i have literally no clue why stuff is breaking

yea so

my C drive is where my python is installed alongside other windows stuff and pip

and my D drive is my main drive

where I have the stable diffusion folders

will that cause an issue?

[wet flame]

i really dislike comfy and anything node based stuff like that

[low basin]

When you say folders is it just the lora and model folders or is it the whole sd.webui folder package?

[wet flame]

i guess most artists may like their nodes and spider web of wires connecting nodes

[gilded canyon]

the entire stable diffusion folder

[gilded canyon]

me too, thats why im trying so hard to make my a1111 work without breaking lol

[wet flame]

i hate blueprint in unreal too

or those any node based softwares or programs

i wish all those comfy features would be in A111

minus those nodes

[low basin]

And you made sure the file paths are correct pointing to the models and loras?

[gilded canyon]

yes, 100% correct

the lora appears in the ui and everything

I just made a generation, it finished in like 20 seconds at 1.38 it/s

now if i try with a lora, same prompt same settings, it will not work

[wet flame]

this is what it becomes very easily

[gilded canyon]

yup, out of memory instantly

[wet flame]

also programming is faster than visual scripting xD

[low basin]

What does your task manager show for dedicated gpu memory usage when this is happening

[gilded canyon]

ill do it again and take a screenshot

[wet flame]

btw i still get that cuda out of memory in A111

everything works good as long i make 512x512 images

[gilded canyon]

but to describe it, it shoots up to 11.8/12 for dedicated GPU memory and then dies

[wet flame]

i did all the steps cso said

[gilded canyon]

the generation doesnt even get a chance to start

[wet flame]

but idk why i can't use sdxl models on a 24 gb ram and 16 gb vram rig

i checked performance it don't even use half of the rams

[gilded canyon]

a1111 is so broken

it sucks

[wet flame]

but its the best for my asset development

[gilded canyon]

a 16gb vram rig shouldnt run into cuda out of memory

[wet flame]

forged don't have a seamless tiling

it simple don't make anything seamless

i can't run foocus idk why

and i refuse to use comfy

no way screw that

and forged is becoming obsolete

[gilded canyon]

@low basin

i wanna break my computer this is so frustrating

[wet flame]

same lol

[gilded canyon]

it works fine for 2 months and then breaks randomly

[wet flame]

i invested 500 dollars for the new gpu

so i can use AI in my art making

[gilded canyon]

me too, i bought a new gpu for $400 and i run into memory issues

so bs

[wet flame]

i thought i can make my assets to sale faster with AI

i make stuff based on my doodles 😄

i am doing a workaround atm

[low basin]

Yikes and you're sure high-res fix is off and no other controlnets or anything else

[wet flame]

using upscayler to upscale

[gilded canyon]

everything is off

[wet flame]

same