🤝｜tech-support

[frank wasp]

and now i add the new line

to COMMAND

[ornate elk]

wait

first delete the venv folder
then open up a cmd and type
pip cache purge
then edit the webui-user.bat and add the commands there
then save and launch it

[frank wasp]

ok

one sec

i alr did this pip cache purge

i need to do it again? afyer vnv delete

@ornate elk

[ornate elk]

yes

i thought your using it on gpu?

[frank wasp]

ok

[ornate elk]

oh wait

dont launch the cmd inside that folder

launch it normaly

@frank wasp then do pip cache purge

[frank wasp]

on the finder ?

[ornate elk]

yes

[frank wasp]

this is normal?

[ornate elk]

okay looks good

[frank wasp]

aight

now il add to the COMMAnd

now launch

[ornate elk]

that doesnt look good

[frank wasp]

😦

[ornate elk]

did you had an nvidia gpu before?

[frank wasp]

yeah

[ornate elk]

ahhhh

[frank wasp]

i just upgraded yesterday

[ornate elk]

xD

[frank wasp]

😮

[ornate elk]

the SD installer found rests from the nvidia driver

[frank wasp]

do you may now how to delete that

maybe in services

[ornate elk]

you need to uninstall the driver in the system settings

but i would advise the DDU uninstaller tool

that should make a clean uninstall of all nvidia stuff

[frank wasp]

ok il try that

[tame yoke]

no the gpu is at 0%

[ornate elk]

whats in your webui-user.bat?

[tame yoke]

in task manager

[ornate elk]

remove --skip-torch-cuda-test --precision full

[tame yoke]

but it gives me a error then

[ornate elk]

also add --opt-sub-quad-attention

[frank wasp]

do i select here the one i want to keep so in my case amd?

or do i select the one i want to uninstall

nah nvm seems obvious xd

nvidia

[ornate elk]

you select what you want to uninstall

[frank wasp]

yeah stupid question fr

xd

[ornate elk]

yea torch doesnt support your old gpu

so it wont work :/

[tame yoke]

awesome

is there a huge difference between gpu and cpu

[ornate elk]

yes

a very big difference

while cpu takes minutes for one single image a gpu can make it in seconds

[tame yoke]

is the result the same

[ornate elk]

yes

the quality isnt affectet by the hardware

but larger resolution will take like ages on cpu

also some stuff wont work if it requires a gpu

[tame yoke]

[tame yoke]

do yk whats the oldest it supports

[frank wasp]

i rebooted

[ornate elk]

any Nvidia GTX gpu or AMD RX 5 series

[ornate elk]

You need to do the pip cache purge and the venv delete again

Before launching

[late tree]

Supposedly someone managed to run SD on a Raspberry Pi Zero 2, but it apparently takes over an hour to generate

[frank wasp]

its loading xd

i need to wait bit

[ornate elk]

Alright

Hopefully it downloads the stuff amd needs xD

[frank wasp]

IT WORKKSSSS

THANKYOU ALOT MAN

YOU :

[ornate elk]

Perfect no problem 🙂

If you have any questions about resolution or upscaling with amd. Let me know

[tepid crypt]

how can i do to generate my character with a specific costume? i want to generate her with this costume:

[ornate elk]

You would need controlnet for that and then maybe the style or a ti adapter model

[final vector]

Just out of curiosity, would utilizing dual-graphics cards boost speeds for image generation? For example, if utilizing dual-4090s, would this allow for quicker generations than a single-4090?

[frank wasp]

i will thanks alot

[tepid crypt]

what do you mean with style/ti adapter?

[ornate elk]

Nope you can't use two GPUs for a single image generation. But each card can use its own SD instance to create images at the same time

[ornate elk]

These are models for controlnet.
Style Adapter can take a style and put it into the output image

And then there are experimental ti Adapter for ControlNet

Some of them could work too for that purpose

[tepid crypt]

k going to search it

[frank wasp]

how do i get good quality and normal faces xd?

[final vector]

I see, would I also not be able to do this with SD Video?

[ornate elk]

Yes with any Ai Programm

[ornate elk]

Upscaling ^^

[final vector]

If you're using standard 512x512, the main way I found is upscaling if the subject is further away, but you can technically get decent results through portrait or up-close photographs

[frank wasp]

is this normal ?

RuntimeError: Could not allocate tensor with 7859200 bytes. There is not enough GPU video memory available!

[ornate elk]

Also use some good negative prompts and quality prompts

[frank wasp]

of i upscale too much

[ornate elk]

Yes

[frank wasp]

ok

[ornate elk]

The directml version for amd is Vram hungry.
On Linux you would get a much better performance btw. But in a few months AMD will get the same speed on Windows too

@frank wasp for upscaling you need an extension called Tiled Diffusion

That will safe you from the out of vram error

[frank wasp]

i will try it

thanks \

[vocal burrow]

multidiffusion, https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111

[ornate elk]

Also every time you get an that out of vram error you need to restart SD to clear the vram.

[ornate elk]

Thats the old name ^^ in the extensions tab its called tiled diffusion

But he couldn't changed the github name

[limpid tiger]

using the stability ai API for image-to-image/masking, how do i set denoising and blur? I don't see it in the docs: https://platform.stability.ai/docs/api-reference#tag/v1generation/operation/masking

[ornate elk]

@frank wasp
You can try following:
Resolution: 512x768
Steps: 30
Highres Fix:
Hires steps: 10
Denois: 0.5
Upscale by 2
Upscaler: Esrgan4x

[frank wasp]

thanks

where to put the negative promps

[vocal burrow]

complete list of parameters https://github.com/Stability-AI/api-interfaces/blob/main/src/proto/generation.proto

[ornate elk]

In the second text field

[vocal burrow]

Also #1042896447311454361 might be better suited for official Rest API related questions

[limpid tiger]

Hm I don't see denoise on there 😮

oh, I will try platform I guess

[limpid tiger]

hmm I'm not even sure how to use these on the official rest api

[ornate elk]

@frank wasp im off for now.
Try the settings. If they dont work install the tiled diffusion extension but only enable the tiled VAE option.
Then set decoder to 1024 and encoder to 128

That will work

[frank wasp]

your a g

i get perfect images now

[frank wasp]

thanks again have a nice day

[vocal burrow]

oh yeah my bad that's for the gRPC API

anyways #1042896447311454361

[ornate elk]

Nice have fun! 🙂

[hasty orbit]

Does Stable Diffusion's images leave behind any sensitive metadata on the generated images?

[glossy lynx]

Thanks CS1o, its noticably better now - cheers!

[subtle quarry]

So does anyone know why comfyui is doing this when using Ultimate SD Upscale?

ERROR diffusion_model.output_blocks.5.1.proj_in.weight shape '[640, 640]' is invalid for input of size 1638400
ERROR diffusion_model.output_blocks.5.1.transformer_blocks.0.attn2.to_v.weight shape '[640, 2048]' is invalid for input of size 983040
ERROR diffusion_model.output_blocks.5.1.transformer_blocks.0.attn1.to_out.0.weight shape '[640, 640]' is invalid for input of size 1638400
ERROR diffusion_model.output_blocks.5.1.transformer_blocks.0.attn1.to_q.weight shape '[640, 640]' is invalid for input of size 1638400```

It spits it out hundreds of times

but completes the image anyway. This is in comfyui.

This should provide the workflow

(minor nsfw? I guess? I'll delete it if it's too bad)

Using an SDXL model as well.

[languid cobalt]

Tried to train LoRA on Kohya, but got this error ModuleNotFoundError: No module named 'bitsandbytes.cuda_setup.paths' , couldn't figure it out but read that Kohya portable lora had better success, but I get this problem OSError: [WinError 126] The specified module could not be found. Error loading "C:\Users**\Kohya_ss-GUI-LoRA-Portable-main\Kohya_ss-GUI-LoRA-Portable-main\venv\lib\site-packages\torch\lib\caffe2_nvrtc.dll" or one of its dependencies.

[stuck glade]

can you use --share in comyui?

[hollow drum]

I need help with additional networks

It is refusing both input lora's. Did I do my mask incorrectly?

I mean since the lora's are not even effecting the people I would say its just not working.

[rich zodiac]

I see tons of LoRA's on civitai recommend a specific weight for the LoRA, how do i apply those? I use sd.webui; Automatic1111 WebUI

I have an extension called "LoRA Block Weight" but i have no clue if that's what i need or how to use it

[hollow drum]

its the number before the final >, it is by default 1, you can change it from anywhere between 0.1 and 1

[rich zodiac]

thanks

[languid cobalt]

You can also use the extension Additional Networks for multiple Lora and it has a weight slider

[dire raptor]

guys, im trying to generate my bedroom design using the workflow above, what went wrong? please help me

[dire raptor]

would like to also mask the wardrobe on the left and inpaint not masked. can anyone teach me how to do it?

[delicate urchin]

I get following error when running simple refiner workflow with sdxl:

RuntimeError: mat1 and mat2 shapes cannot be multiplied (154x2048 and 1280x768)

Anyone knows what the problem is?

[thorny wyvern]

This is maybe a better channel to ask this in. Im trying turbo out in comfyui and it generates images in about 0.3 sec but as soon as i change the prompt it 10x the generation time. What i can find it seems like its the encoding of prompt to embedding that seems to slow it down. Anyone have any suggestion what i can do to fix this?

[fair oxide]

Hi everyone, a tech question more related to speed and costs. I wanted to generate a 3min video, probably using disco diffusion or warp fussion (similar to https://www.instagram.com/gerdegotit/ content) and I will need to render pretty fast. For this tools is it Google Colab the only cloud option?? anyone has used https://rundiffusion.com/ ? Thanks!! 🙂

[formal trout]

Hi, I need help with Hi-Res fix, I wanted to try it but I get not enough memory. I have RX 6700 (12 GB VRAM). I'm generating at 832x1152, maybe that's the issue. I tried by 2 and by 1.5. Also I don't see the option of using models from my model directory.

[ornate elk]

You can't use such a high native resolution

Auto1111 or comfy ?

[formal trout]

ahhh that's what I thought

Auto1111

so what's the deal them, I just wanted to try to see if I was missing something, I usually just use SD XL and stay away from upsacallers and SD 1.5 xD

Also why there isn't a tile model for up scaling in SD XL 😒

[ornate elk]

You can use 512x768 or 540x960 upscale by 2
The later one only works with tiled VAE enabled

For example
You can try following:
Resolution: 512x768
Steps: 30
Highres Fix:
Hires steps: 10
Denois: 0.5
Upscale by 2
Upscaler: Esrgan4x

[formal trout]

let's see

oohh tiles?!?

[ornate elk]

Yes

Upscaling with hires fix or upscale scripts always works by tiling

[formal trout]

Cool I though it was a simple img2img

[formal trout]

hey ty, cool, I like the quality

[ornate elk]

Looks good!

You can play around with the denois and upscalers
For denois dont go over 0.5

[formal trout]

kind of surprised, I didn't expect it. Many come out nice.

I was dodgy about using upscaller and not generating in full SD XL resolution

[ornate elk]

Retro sci fi looks cool

[ornate elk]

I still use 1.5 with hires fix more than sdxl. I think most people still do

But you also can upscale sdxl images. But with your GPU you have to do that in img2img with the SD upscale script

[formal trout]

this is same prompt but with native SD XL resolution. I don't know, they have another kind of detail? But both seems to be able to generate good images.

[ornate elk]

Nice ones. Upscalers can increase the details depending on the upscaler used. There are also upscalers to get sharper images or softer/smoother

You can get additional details by using Latent upscalers

The sdxl benefit is the 1024x1024 resolution as the more pixels you have the more stuff you have in the image

More place for details

[formal trout]

Haven't try that

[ornate elk]

Me neither but it should work. xD

Can't test it rn

[formal trout]

I remember some time ago trying to upscale and image with tile diffusion and upscalers and I always found they add fake/bad detail, it was good in some aspects and bad in others, so when SD XL came I was super glad, it generated in big resolution with real detail.

And yes sometimes they bad (or low quality) detail is there, but kind of good or it may pass some kind of style

But can be fixed with inpainting

the whole image and big details looks good

[ornate elk]

Yea sdxl was a huge step forward for highres images

I will test some sdxl upscaling in img2img maybe later today

[formal trout]

I'm trying 540x960 with Tiled VAE, it works like a charm, thank you

[ornate elk]

Perfect 🙂 if you get any weird Gray squares then set the encoder to 1024 and decoder to 128

[full knoll]

How do I disable a refiner?

[fair oxide]

what are the requirements to using t2i-adapter_diffusers_xl_ controlnet models in this list https://huggingface.co/lllyasviel/sd_control_collection/tree/main

i see t2i adapters are lot smaller controlnet models

but would they work by selecting controlnet options or are there any other required files?

[ornate elk]

In auto1111 you have to click the small arrow to collapse it

[tame yoke]

do i put a LoCons into a LoRa folder?

[subtle quarry]

yes, you put them in the lora folder

[pure wedge]

Using ddpm_2M_sde + Karras breaks my characters faces?

Any reason why??

[fair oxide]

I used to have amazing results with DDIM (1 Year Ago) on img2img but now the results are amazingly bad. Any ideas? I am pretty sure every setting I use is same but sampler.

[full knoll]

That makes sense ig. Thanks!

[hollow drum]

mine dont work

[crystal urchin]

Hello, just started using Stable Diffusion and trying to create consistent characters. I have a close up face picture and am trying to put that same face on a full body picture using img2img, but it keeps giving me more portrait photos. Among my other prompts used to create the original face, I have "full body image, cute black outfit, streets of brooklyn" Any recommendations?

[terse crane]

Hello!
I'm using comfyui and have a dedicated path to store my models but comfyui isn't able to find the path to my GFPGAN file (D:\models\GFPGAN\GFPGANv1.4.pth) while it can see and use my models and loras properly (D:\models\Lora and D:\models\Stable-diffusion), I'm looking for the right keyword to use in the extra_model_paths.yaml file

If there is a file listing all these "keywords" I'm very interested, like checkpoints, clip_vision, etc.

[elder turtle]

You can use Reactor for that for example

[elder turtle]

mine looks like this and they work

[terse crane]

Where is your GFPGAN model file on the disk?

Because I inherited a models/GFPGAN/ folder (I don't know which GUI created it, I've played with so many of them) and it comfyui can't find it there

[elder turtle]

You probably just have to add it below ESRGAN

[terse crane]

Thanks, I'll try!

[elder turtle]

with the same syntax

I didnt try it tho, just guessing

[delicate urchin]

Is it possible to do img2img workflow with sdxl turbo?

[terse crane]

Just tested, it still can't find it in the ESRGAN folder (my extra_model_paths.yaml lists this folder iin particular under upscale_models:)

[crystal urchin]

what is Reactor?

[elder turtle]

https://github.com/Gourieff/comfyui-reactor-node

[elder turtle]

Thats unfortunate

[crystal urchin]

thank you

oh boy, this looks confusing 😂

[terse crane]

yup ;D

[elder turtle]

Its pretty much face swapping, if you can get behind the UI its very simple to use

Or you try it for automatic

they have an extension as well

[crystal urchin]

automatic1111?

[elder turtle]

https://github.com/Gourieff/sd-webui-reactor

Yep

[crystal urchin]

I've heard of that but haven't tried it myself

[elder turtle]

Whichever UI you prefer

[crystal urchin]

I'm currently using this so I might just stick to this and learn one instead of trying to learn two UIs at once

[elder turtle]

Thats the Automatic1111 UI isnt it

[elder turtle]

This extension should work there

[crystal urchin]

haha no clue

[elder turtle]

Go to the extensions tab and install it either via the link or by searching for it in the list

[crystal urchin]

oohhh I see. Thank you so much for your patience 😁

[elder turtle]

No worries

[rich zodiac]

Im using Automatic1111 UI rn and i cant seem to use the img2img tab anymore, i keep getting the same error message on inpaint, and regular img2img;
NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.
Iv enabled the setting, and have no clue how to do the commandline thing, but the weird thing iws this all worked yesterday

[terse crane]

I found an explanation, for some reason it wasn't seeing it when it was in ESRGAN (even though I declared the path in the extra path file) but I also had a models folder inside my Comfyui one and moving the file to models/facerestore_cf (as per the readme for that node) worked, thanks!

[ornate elk]

Hey what was your GPU?

[rich zodiac]

rtx 4060 ti overclocked edition, 16 gb of vram

[rich zodiac]

but whats confusing is it worked yesterday, the only thing i did since my last batch of generations and today when i tried was sleep

[clear magnet]

Does anyone know if the stability foundation will develop its own training tool for training models as opposed to all of us using Kohya_ss? im pretty tired of Kohya breaking every other month.

[cinder ermine]

decided to try fooocus with amd on windows, im getting "out of memory" for pretty low res generations

i have an rx 5700 xt

[ornate elk]

Make sure the model size is larger than 1.98gb

[rich zodiac]

it was, i actually got it working somehow? i think the image itself was the issue somehow

[cinder ermine]

DefaultCPUAllocator: not enough memory: you tried to allocate 6553600 bytes. - output from the cmdline using fooocus

what settings should i change

[unborn hazel]

Hello
https://civitai.com/images/3332548
I'm trying to replicate the third example here but i'm finding some difficulties
I can notice 2 things from my metadata, the version and the vae, the vae shouldnt make this much of a difference so I assume the version is the reason? or am I missing something(more likely)

Hey thanks 🙂

[hollow schooner]

What’s the latest version of sd that amd can run cause mine is at 1.6

[ornate elk]

1.6 is the webui not SD

Auto1111s version in fact

It works with any SD model

[languid cobalt]

there's and SD2.1 model and there's also SDXL

[hollow schooner]

Ooooh okay so is the latest of webui 1.6?

Also should i switch to low vram if i keep getting “could not allocate tensor”

[tame yoke]

whats the maximum number i should decrease this to>

for upscaler

[languid cobalt]

That's the latest. You can try the low VRAM.

How much VRAM do you have?

[hollow schooner]

8gb

[languid cobalt]

Do you have NVIDIA or AMD? Or Intel 👀

[hollow schooner]

Amd

The drawing gets really close to being done then just says could not allocate tensor

[languid cobalt]

Just one image, no batches?

[hollow schooner]

Yeah

I am using a hyper net though too

[languid cobalt]

How about the resolution and Hires settings?

I have 8gb too.

but nvidia and xformers

[hollow schooner]

There’s nothing in refiners

[ornate elk]

You need to always set hires steps. 10-15

And denois 0.5

What happens when you click on it ?
What does the cmd shows?

[languid cobalt]

Yeah I think hires 0 steps does 50 by default

[ornate elk]

0 does use the same amount as steps set

[languid cobalt]

Oh make sense, I put it on 50 if I use hires, otherwise I usually do 100

[ornate elk]

it doesnt make sense to use 50 hires steps, or 100 normal steps,
mostly your fine with 30 steps + 15 hires

[languid cobalt]

I do 10 hires usually, and 50 regular steps. yeah 100 steps seems to be pointless, but 30 seems to low with some models.

[ornate elk]

can you show me the whole cmd text?

[ornate elk]

yea some models need more, 30 is the average

[languid cobalt]

@hollow schooner I wonder if using --opt-split-attention would help, there's some more args you could use probaby.

[ornate elk]

yea --opt-split-attention is usefull for amd cards

but he uses it already

that model (superhero diffusion) isnt compatible with the new auto1111 it seems as its over a year old.
Can you try an other model like Dreamshaper v8 for example?

that should work

[languid cobalt]

thats the base sd model so it should work

[ornate elk]

can you try to generate something?

[languid cobalt]

I think he's saying he can't switch, like the textbox remains blank.

[ornate elk]

whats your GPU?

okay, if your using an browser adblocker, make sure to whitelist the webui in it

also you moved your webui into a new folder. that prevents future updates when not fixed

[languid cobalt]

Some extensions did cause problems for me

Like not being able to use controlnet at all..

[ornate elk]

oh okay, hopefully not multiple adblockers because ublock origin is the only one needed

also script blockers can cause issues for auto

okay, did you git cloned it into a folder thats not documents/desktop/onedrive etc?

you would need to git clone it again directly in the folder

i also suggest you install python 3.10.9 or 3.10.11 (you can overwrite 3.10.6)

less bugs

because you can get permission errors if the folder was initialy created by windows instead of you

[languid cobalt]

run it with no args, too.

[ornate elk]

you shouldnt run it as administrator

what guide did you watch?

[tame yoke]

can u force torch to use ur gpu

[ornate elk]

first one is for nvidia,
Second one is not usefull for most amd users
third one is "okay" but has a few bad tips

but yea install the new python version i mentioned and git clone the repo into a folder again

[hollow schooner]

So i restarted my pc and tried it without hires steps and it worked but after trying hires (13) it said could not allocate tensor again

[ornate elk]

try 10.9 first

[hollow schooner]

The hi res did look better though it just won’t load for some reason

[ornate elk]

i have a guide for nvidia gpus, i should write on for amd too

[hollow schooner]

Any way to fix the hires thing?

[ornate elk]

you should try upscale by 1.5

also restart before trying

every time you get an out of vram error you have to restart SD

if it still dont work you need to upscale it in img2img. i can guide you

[hollow schooner]

Where do i restart SD

[ornate elk]

close the cmd and relaunch the webui-user.bat

[hollow schooner]

Is out the vram error the (could not allocate tensor)?

[tame yoke]

with stable diffusion colab, can u use models and loras from ur pc?

[ornate elk]

if you done it before with 3.10.6 then it wont be a problem

[ornate elk]

yes should be possbile as it uses your google gdrive

[tame yoke]

would it be faster than using my cpu

[languid cobalt]

yes.. lol

[tame yoke]

ok ok thx ❤️

[languid cobalt]

Not big performance on free, but you can get some decent horespower with paid colab. I think free is on tesla k80, which is pretty old, but still should be better than running on modern cpu

If nothing changed, then yeah

[tame yoke]

how much $$

[ornate elk]

add them after you tried a 512x512

[languid cobalt]

$10 a month for the lowest plan, but you will need to buy "compute units", which exhaust from computation, could get expensive if you're doing a lot of work

[ornate elk]

yes these should work for your gpu. but you may change --medvram to --lowvram if you encounter out of memory issues

[tame yoke]

and do yk where i can find the free collab

[languid cobalt]

just google colab and you'll find it 😛

[tame yoke]

https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_diffusion.ipynb this?

[ornate elk]

did you updated python ?

do you have the full cmd log again ?

[languid cobalt]

https://colab.google/

[ornate elk]

looks good so far

[languid cobalt]

What gpu do you have i forgot?

hmm, is it 8gb?

I would also recommend increasing your pagefile to like 30gb

windows key > type system(and select the app) >advanced system settings > (under advanced tab select settings on the performance section> advanced tab> virtual memory (change)

I guess so, looks different on mine

[ornate elk]

do you have any other programms open beside SD ?
did you restarted the PC in the last few hours?
How much free space do you have on C ?

yes do that. it needs 15gb free
i also found maybe the cause of the error. The installation files could be corrupted.
So do this:
open up a cmd and run
pip cache purge
then git clone the repo again into a new folder.
i think that will work

[languid cobalt]

I've ran a1111 SD with as little as 500mb left on my C

but now i can't recall if I had it already running

[ornate elk]

yes but it always uses the cached files, these are stored on C cache
if they are corrupt you can try 100 installs and it always gives that error

from the searchbar

[fair oxide]

ive downloaded a stable difusion model, put it inside the modelfolder, and when I try to run stable difusion it still wants to install some other model cuz it probably doesnt find the model I installed. IDK WHy and ive tried reinstalling stable difusion and using diffrent models like 4 times now 😭

mhm

oh

he went poof

😛

it also says this:

Launching Web UI with arguments: --skip-torch-cuda-test --upcast-sampling --no-half-vae --use-cpu interrogate
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
Warning: caught exception 'Torch not compiled with CUDA enabled', memory monitor disabled

Is the prossessing thing in the corner usual?

I might just start using https://github.com/divamgupta/diffusionbee-stable-diffusion-ui if this doesn't work out. This seems super sketchy tho 🤷‍♂️

cuz ive wasted hours trying to figure this out

[ornate elk]

Hey what's your GPU?

[fair oxide]

Im using a macbook m1

[ornate elk]

Ah okay

And is the model file bigger than 1.98gb ?

[fair oxide]

what file

[fair oxide]

?

[ornate elk]

You said you downloaded a stable-diffusion model

[fair oxide]

yeah

5 723 573 918 byte (5,84 GB)

For the whole folder

oh

u mean

the file

1 sec

4.27

gb

[fair oxide]

I got it from here: https://huggingface.co/runwayml/stable-diffusion-v1-5

ive tried the XL thingy too

It managed to load this file this time

but

it still gave me errors when making images

[ornate elk]

What errors?

[fair oxide]

like it didn't automaticly download a model, which it did earlier

[fair oxide]

To create a public link, set share=True in launch().
Creating model from config: /Users/kaare/ai/stable-diffusion-webui/configs/v1-inference.yaml
Startup time: 12.3s (prepare environment: 0.1s, import torch: 2.0s, import gradio: 0.6s, setup paths: 0.6s, initialize shared: 0.2s, other imports: 0.7s, load scripts: 0.3s, create ui: 0.2s, gradio launch: 7.4s).
-[IOGPUMetalCommandBuffer validate]:215: failed assertion `commit an already committed command buffer'
./webui.sh: line 255: 4916 Abort trap: 6 "${python_cmd}" -u "${LAUNCH_SCRIPT}" "$@"

thats only some

its the latest

but I got like a timed out error on the port too

like the interface thingy

[calm cairn]

hi all--not sure if you wanna help me because images are NSFW, but I am trying a prompt I found on CivitAI, and on one model (unrealityv30) I keep getting weird artifacts in the images

i was wondering if anyone can tell me what is wrong--im not using any LORA, and I tried running the positive and negative while removing sections to see if it was one particular term (it was not)

(it is behaving as if you put the value too high on a lora)

[fair oxide]

scam lol

[calm cairn]

haha thanks

[rich zodiac]

have you tried it in a different model to see if the issue persists?

[calm cairn]

yes, I put it through around 20 different models, and it is only creating this problem on the unreality model

[fair oxide]

would u guys trust this?

[fair oxide]

its open src but idk how to compile it, cuz the downloader on their website could be something completely diffrent

and their offical discord seems hella botted

1k+ members with -- lots with chinese character names

[dreamy lodge]

hey guys im getting an error for tdqm.auto when i installed the text2video extension. Any way to fix?

[languid cobalt]

someone on the comments recommended clip skip 2

[dreamy lodge]

r__
submod = importlib.import_module(submod_path)
File "C:\Users\tanne\AppData\Local\Programs\Python\Python310\lib\importlib_init_.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "C:\Users\tanne\stable-diffusion-webui\venv\lib\site-packages\huggingface_hub\hf_api.py", line 47, in <module>
from tqdm.auto import tqdm as base_tqdm
ModuleNotFoundError: No module named 'tqdm.auto'

[calm cairn]

ooh good idea, thanks, will try immediately

[dreamy lodge]

i tried uninstalling the extension manually but that didnt work

[calm cairn]

that seemed promising--original artifacts created on clip skip 2, so I ran prompt again at clip skip 1, and again at 3, but sadly got the same error, so probably something else

[ornate elk]

Would need to see the error but you should try to download the Dreamshaper v8 model to test

[fair oxide]

im just compiling diffusionvee rn

I think ill use that if it works instead

well see

[ornate elk]

Do you have an example image?

[ornate elk]

You need to delete the text2video extension and then the venv folder too

Than it will work again

[dreamy lodge]

the entire venv folder?

[ornate elk]

It will be recreated

[dreamy lodge]

k. but is there a way to get the text2video extension to work?

or is it broken

[calm cairn]

i do but its nsfw, let me run it again without the "topless" keyword and I imagine the same problem will happen, only that I can then post it here 🙂

[ornate elk]

Its not updated anymore but it seems they will work on it again

A group named deforum-art will work on it. Thats written on the github page

So I think you have to wait

[dreamy lodge]

The collecting torch command lagged the fuck out of my pc lmao

Swear I was running at a frame a second

[ornate elk]

Diffusion bee has not many features and dont get updates

[ornate elk]

Your on 8gb ram ?

[dreamy lodge]

16

[ornate elk]

Ah okay

Yea that lag can happen with 16 too

If the pagefile is small

[calm cairn]

ok I got my images, should I upload directly here, or do catbox etc?

[ornate elk]

Here is fine

[dreamy lodge]

ye my pc isnt the best in the world. A simple msi gf63 but this thing can kick ass

[calm cairn]

says blocked by server, too large? only 512x512, and no nudity

what is the problem? I'll try 1 at a time

yeah wont take either of the 2

[ornate elk]

There is only an 8mb file limit from discord

[calm cairn]

400kb

[ornate elk]

Sometimes discord blocks images because of the prompts in the negative box

You can make a screenshot of the image

That should work

[calm cairn]

yeah if I save as different image file, that also scrubs exif iirc

[ornate elk]

Maybe

[calm cairn]

positive
1girl, 1 girl, full body, masterpiece, intricate, public, supporters, tank top, boxing gloves, in a boxing ring, (boxing:1.2), incredible detail, confident, eye contact, dyke, butch, (short hair:1.2), perfect face, very slim, (beautiful face), athletic, fit, petite, (30 year old:1.3), (sweating), (shiny skin), detailed face, poor, lower class, dirty

negative
asian, deformed, malformed, deformed face, bad teeth, bad hands, bad fingers, bad eyes, long body, blurry, duplicate, cloned, duplicate body parts, disfigured, extra limbs, fused fingers, extra fingers, twisted, distorted, malformed hands, mutated hands and fingers, conjoined, missing limbs, (facial hair), moustache, mustache, bad anatomy, bad proportions, lowres, mutated, mutilated, artifacts, gross, ugly, badhandv4

yeah convert to JPEG worked

[ornate elk]

Okay and what are the txt2img settings?

[calm cairn]

sampline Euler a, steps 20, cfg 7, seed -1, but I can pull up the seed if you want that too

[ornate elk]

Any VAE file ?

Or embeddings used?

[calm cairn]

ah, forgot to check the top--clip skip 2, vae ft mse 840000 ema pruned

[ornate elk]

Did you tried DPM++ SDE Karras ?

[calm cairn]

again, model is unrealityv30

[ornate elk]

Can you try an other sampler and without a VAE ?

[calm cairn]

sorry graphics card isn't super good, going slowly, but I'm doing it

[ornate elk]

What's your GPU ? Maybe we can speed it up

[calm cairn]

different sampler did the trick, going to check VAE + card name/specs

ah right, i remember this one--got it second hand from store
NVIDIA GeForceGTX 1660 Ti

also slightly interesting--running prompt without VAE did make it better, but not totally fixed

(so yeah it was sampler)

any idea why sampler would cause artifacts?

[ornate elk]

Maybe the model got trained on it specifically

On DPM samplers

Euler a is the oldest sampler but still a good one

[ornate elk]

Do you use --xformers ?

[calm cairn]

i remember looking at it before but it was back when I first started, how do I check?

[ornate elk]

Right click the webui-user.bat and click edit.
Then check what's in the commandline_args= line

[calm cairn]

lol i forgot it was a batch file, checking now

[ornate elk]

For your card there should be
--xformers --medvram --no-half

That gives you the best performance

[calm cairn]

find xformers "cannot find xformers"

[ornate elk]

Okay make sure your in the webui-user.bat and not webui.bat

[calm cairn]

ohhh

set COMMANDLINE_ARGS=--xformers --medvram --no-half

thanks in advance cant wait to see the difference

and thanks for figuring out that my problem was the sampler, i will remember that in future

seriously you guys are great

[ornate elk]

Thx, No problem 🙂

[calm cairn]

next time someone requests a wizard to change their picture, I can say I've been to more wizard school

[ornate elk]

Also always launch the webui-user.bat and not the webui.bat

To start SD

[crystal urchin]

are there any generic phrases for prompts and negative prompts I can enter to get better hands?

[ornate elk]

Hey, thats a question for #📝｜prompting-help
Yes there are some prompts and Lora's and embeddings to help

[crystal urchin]

thank you! Love the pfp btw

[ornate elk]

Np and thanks! OP best 😄

[lone wolf]

Idk, I did the SD yesterday, everything was fine

same prompts today, and faces are weird af

This is me making a blood elf today with same prompts as yesterday

This was yesterday

What happened?

Why is it so weird?

[young bridge]

Hey guys, does anyone know how to update fooocus, I think there is an update but it’s been a while since I used it

[languid cobalt]

100% sure it's same prompts, same TI's and Loras? Use the seed from the img name.

[calm cairn]

ty for the tip, but the one you specified is foolproof--made a shortcut to it on start

[fair oxide]

Hey can some help me on how to add an image to change it with prompt? Reply this msg if yes

[languid cobalt]

like img2img inpaint ? you can also use controlnet's inpaint

[radiant stratus]

it worked pretty well, though do you know how to get rid of these noisy

[late tree]

You'd need to do that before upscaling it. Maybe try inpainting

[radiant stratus]

I'm using playground, would it be better to inpaint before upscaling or after?

@late tree

[late tree]

Are you using Playground to generate or just to upscale? If just upscaling then do it before because it would take more vram to inpaint a larger size image than a smaller one.

And it would take longer

[radiant stratus]

No real generation

just upscaling an image that was an old wallpaper

though I am also considering doing some generation to extend it into a 16:9 wallpaper

[late tree]

Then inpaint it first and then upscale, will be faster that way.

[radiant stratus]

with playground are you able to provide masks so only certain areas regenerate?

without totally erasing an area?

[late tree]

I'm not sure since I only use it to upscale

[radiant stratus]

I guess I should actually download stable and learn to use it then do upscaling with playground

[late tree]

Playground has a limit of 1000 images (I think in a 24 hour period), so I prefer to use it just for upscaling. I think there is a way to inpaint with it, but I just have never tried.

[elder turtle]

What do you need?

[lone wolf]

Yes, idk why it happened....

[uncut jewel]

Hey guys! Wassup ? I'm trying to find LCM Sampler on A1111 but can't find it ... Meanwhile i got it on Comfyui, but nodes are different from the Webui

Any idea how to get it ?

Same for Euler Ancestral

I got Euler A, but not Ancestral on A1111

[chrome rampart]

morning all. question > what commands (if any) do i put in my comfyui startup bat file when having nvidia 1660 graphics card with 6Gb vram?

[elder turtle]

Euler A stands for Ancestral

[uncut jewel]

Dope i was thinking about that, but wasnt sure aha

For LCM Sampler, it's an extension to add or something manual to do ?

[jaunty root]

Hello! I would like to use impaint mode on SD, while put Disclosure Face Effect on an original pic. I'm using Euler A but the result is very awful, glitched.
How can I solve it?

[karmic crown]

install sd-webui-animatediff extension and restart it will give you LCM sampler, but do not forget to use the LCM lora too

[sand trail]

is this right?

[uncut jewel]

Well well, i tried many times to install AnimateDiff on Webui but it never worked unfortunatly, cause of some wrong path related issue... I'll try it out again anyway, thanks!

[karmic crown]

go to extensions tab

then find available

click on button to get a list

search for it

click install

wait a few seconds (to download)

and then go to aplly and restart

but i am not sure if it is enough for lcm to show, or you need to restart the whole a1111

[uncut jewel]

Yea that's what i made many times in the past, but i got an issue when webui is starting with a path going to the wrong direction from AnimateDiff then it makes the whole plugin break even tho i tried different installation methods overtime and i got different AnimateDiff on ComfyUi that works perfectly 😒

[tame yoke]

the heck is this someone help

[strange thunder]

Is there any extension that will allow me to search images with and without certain meta data prompts?

[karmic crown]

infinite image browser

works beter thn the other one

[tame yoke]

does xformers affect generating with cpu?

[karmic crown]

xformers is for nvidia

[tame yoke]

oh

common sense went poof

thanks

[sand trail]

is there a video on how to install xformers bc i seem lost

i own an rtx3060 6gb

installed cuda 11.3

vs builder tool

but then i lost

[tame yoke]

it installs automatically when u type --xformers into webui.user.bat
if u mean manually idk

[karmic crown]

give me a second

[sand trail]

its been like this for the last 30mins

[karmic crown]

i am trying to find a simple guide

give me a second

[sand trail]

alright

[karmic crown]

here it is #🤝｜tech-support message

his guide is simple and easy to follow

if you already finish everything

[karmic crown]

ok

[sand trail]

never miin its working now

[karmic crown]

add this in command line params

[sand trail]

thanks man

[karmic crown]

what is your gpu?

[sand trail]

3060 6gb

[karmic crown]

ok

since you have 6GB

use this

--xformers --medvram-sdxl --no-half-vae

sorry

[sand trail]

this way?

[karmic crown]

remove the "

or add it after vae

i think that on windows you should not use any

[karmic crown]

mine (on Mac) looks like this export COMMANDLINE_ARGS="--skip-torch-cuda-test --opt-sub-quad-attention --upcast-sampling --no-half-vae --medvram-sdxl"

[karmic crown]

this looks good

[sand trail]

didn't work

fixed it

[karmic crown]

no quotes on windows? right?

[sand trail]

for xformers only

[karmic crown]

?

remove them compleatly

[sand trail]

"--xformers" is just left it this way

[karmic crown]

you do not need it

set COMMANDLINE_ARGS=--xformers --medvram-sdxl --no-half-vae

[sand trail]

allright

[karmic crown]

you should be able to use SDXL too now without a problem

[strange thunder]

very dumb question: With privacy and security, what exactly do I need to be concerned with here? Like, is the search function connecting to anything? I run local, but I guess now I'm wondering if that isn't enough on its own. lol.

[karmic crown]

Infinite image browser doesn't work if you enable remote unless you add --enable-insecure-extension-access parameter

you will get something like "your settings doesn't allow bla bla"

you can always disable your network adapter and try 🙂

[strange thunder]

Is there a best way to go remote with SD? I've looked into it, but, tbh, haven't really committed to trying anything/just feel more comfortable doing everything local.

[karmic crown]

if you know how to use ngnix or apache, than i would run it behind reverse proxy but with --no-gradio-queue, since websocket proxing can be little tricky on apache

[ornate elk]

You should use --medvram instead of --medvram-sdxl

Oh wrong person

[karmic crown]

oops

[ornate elk]

You should use --medvram instead of --medvram-sdxl

[karmic crown]

i thought it would work with --medvram-sdxl

[ornate elk]

--medvram is for 6gb and lower.
--medvram-sdxl is for 8-10gb

[karmic crown]

thanks

[ornate elk]

There is the --share (global link) and --listen (local link) commandline arg.
Also you can connect to your PC via Rustdesk

[karmic crown]

share will create xxx.app.gradio link, while --listen gives you more control

[tame yoke]

do models that are a .safetensors file also go into stable diffusion folder?

[ornate elk]

If they are larger than 1.98gb then yes

[tame yoke]

2.08

[chrome rampart]

question > what commands (if any) do i put in my comfyui startup bat file when having nvidia 1660 graphics card with 6Gb vram?

[gray oyster]

How do I fix this?
RuntimeError: [enforce fail at …\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate 238551040 bytes.

I already have 16GB of ram where only 3.6gb is taken up by windows and a swapfile on the same drive as my webui of 8192mb

[karmic crown]

you can try this https://github.com/CompVis/stable-diffusion/issues/329#issuecomment-1361030238

[gray oyster]

i will look into it once my laptop has loaded it (I'm running the AI off of my main pc though)

my WiFi is just pretty crap

thing is that upgrading the RAM is not an option for me unless i want to rebuild my entire pc

because the mainboard that i have doesnt exist anymore other than retail and used and my current mainboard has broken ramslots so 2 out of 4 slots are broken entirely

so i cant even do dual channel

i do have the ram sticks though

just not the mainboard to do it unfortunately

i could buy 2x 16 though

but shouldnt a swapfile fix the problem?

i guess i will have to move my AI to my sata ssd and increase the swapfile

also im having this issue with oobabooga webui, however the error seems to be consistent with every other application with AI

its 31gb on C:/ 🗿

[gray oyster]

no matter the swapfile size it stays the same number of bytes

[ornate elk]

What's your GPU and what's inside your webui-user.bat?

[gray oyster]

4060Ti 16GB

and im not using auto1111

but what i noticed is that my ram usage doesnt even show the slightest spike

it instantly fails

though my vram seems to be slightly filled before it fails

[ornate elk]

And what do you use?

[gray oyster]

oobabooga webui

[ornate elk]

The Pagefile increase to 32gb should be enough

[gray oyster]

like automatic1111 but for LLMs

its at 39.9gb

[ornate elk]

Okay and how big is the oobabooga model?

[gray oyster]

13.5gb

should fit into the gpu

https://huggingface.co/TheBloke/Python-Code-33B-GPTQ

its this one

this one

[ornate elk]

Can you try a smaller model?

[gray oyster]

a smaller model does infact work but that would be a 13b model

thats significantly smaller

or do you mean a smaller model of the same distribution?

in that case i can try that too

actually nvm its the smallest of this one already

[ornate elk]

Ah okay

[gray oyster]

i could try the GGUF

though its the same size

what if oobabooga just cant access the swapfile for some reason?

because right now 4gb of ram are already taken up by windows

[ornate elk]

I dont think it will use swap

[gray oyster]

i have 11.9 left

[gray oyster]

yea that would make sense but then again it should say its missing like 2gb and not 200mb or 80

[ornate elk]

My thought is that it tries to load the whole model into your Ram

[gray oyster]

yea

to load it into the vram

[ornate elk]

That error message doesn't say much

Its same on auto1111 if its out of ram

[gray oyster]

hmm

is there any workaround for that?

in python

[ornate elk]

Idk, maybe you should ask on a discord for text Ais

[gray oyster]

fair

if i know one

[ornate elk]

Dont know if ooba has one

[gray oyster]

lemme check

thanks again though

[ornate elk]

No problem, for any auto1111 related questions I can help more

[gray oyster]

maybe its just a memory leak that has been fixed already, i will try to install a newer version

[gray oyster]

ah yea, for me i cant install it anymore

its like a torch installation error or something

[ornate elk]

Oh do you you have a screenshot of that?

[gray oyster]

i gave up on auto1111 or stable diffusion for now until auto1111 has released a stable branch after 1.6.0

[ornate elk]

1.6.0 is pretty stable

[gray oyster]

unfortunately not but i can try to reproduce it if it happens

yea except for me 😅

[ornate elk]

I can give you my easy install guide if you want

[gray oyster]

sure

that would help

[ornate elk]

Here is a quick Guide to install Automatic1111 Webui (Stable Diffusion)

You need to install Git 64bit: https://git-scm.com/download/win
and Python 3.10.11, 64bit: (any python above or 3.10.6) No 3.11, 3.12 https://www.python.org/downloads/release/python-31011/
And check "add python to path" when installing Python.

1. Make a new Folder on your drive (not on Desktop, downloads, documents, Programms, Onedrive) and name it Ai for example:
   C:\Ai\
2. You go into the folder you created in this case Ai, then click in the File Explorer bar (not searchbar) and type cmd then press enter.
3. Then you copy and paste this command:
   git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
4. Press enter and after its done you can close the cmd and launch the webui-user.bat inside the stable-diffusion-webui folder.
   After the Installation that can take a while. Youll get an URL http://127.0.0.1:7860
   Thats the webui you open in Browser.

[gray oyster]

oh right, i always directly used the release version

i never thought of using git

that would explain the "Fatal: not a git repository"

[ornate elk]

Yep ^^

[gray oyster]

thanks!

[ornate elk]

Git clone is the right way to install

No problem 🙂 if you encounter any error feel free to ping me.
Also if your done so I can help you to improve the performance of the webui

[gray oyster]

also by any chance do you know how to reduce vram usage for AMD cards in Auto111?

[ornate elk]

Yes

I mean no

[gray oyster]

i have a friend who has 8gb of vram and he struggles with 512x512 even wheras i had no problem with that on my older 2060

[ornate elk]

On AMD windows your out of luck.
It will always use as much vram as it gets

[gray oyster]

his card is a RX 6600

[gray oyster]

ah

well that sucks

[ornate elk]

But he needs some commands to not get out of vram that fast

That should help

[gray oyster]

he is already using things like --medvram

went through Auto1111s cmd line flags for hours and tried every single one

[ornate elk]

He needs:
--medvram --opt-sub-quad-attention --opt-split-attention-v1 --upcast-sampling --no-half-vae

[gray oyster]

I will send that to him

[ornate elk]

If he still gets out of vram he need to use --lowvram

But thats slow

The commands above should work for 512x768 images

He then can upscale images in img2img

[gray oyster]

well he can generate like 512x512 for 3 times until he has to restart the ui

and he says it takes him like 10 minutes for a picture

so either way that should be faster

[ornate elk]

Ah okay. Then yes he should try my commands

That should increase the speed

[gray oyster]

yea

[ornate elk]

I have amd and nvidia to test everything in SD

[gray oyster]

im currently trying your method for installing auto1111

[gray oyster]

thats nice, unfortunately i dont have the budget for that 😅

the installation looking good so far, usually it already fails at pytorch

[ornate elk]

Yea I didn't bought it for SD, but old PC was nvidia and new one is AMD

[gray oyster]

by the way, do i profit from ReBAR for AI?

[ornate elk]

Nope

[gray oyster]

ah

damn

well atleast now i dont have the annyoingly bright leds which flicker at night of my 2060 anymore

[ornate elk]

But RGB gives 100% additional speed in SD 😛

[gray oyster]

oh yea right thats a proven fact

unfortunately my 2060 was the only rgb capable device in my pc

no wonder im only getting 2.5 it/s in SD on my 4060Ti

oh that was also another thing i was confused about

[ornate elk]

On which webui?

[gray oyster]

why do i have such slow speeds and why is xFormers so much slower than sdp?

[gray oyster]

auto

[ornate elk]

I thought you couldn't install it?

Because of errors

[gray oyster]

well before i installed dreambooth from the addon section and auto decided to completely brick itself

[ornate elk]

Ahh xD Yea dreambooth is the SD destroyer

But its fixable

But i would recommend to use the new installation

[ornate elk]

By using git you can later update it easily

[ornate elk]

Hahahah xD

[gray oyster]

honestly i want to train so many SDXL models but unfortunately i dont have the dedication to go through the process of dreambooth add-on installation

i tried kohya too but oh well

[ornate elk]

I would suggest you use Khoya_ss for Lora's and dreambooth

But idk how well it does dreambooth

[gray oyster]

im just sad that Auto1111 doesnt support native model training

i mean it does the textual inversion thing

but its nowhere near how efficient things like LoRA or other model types are

like 4bit training or whatever

and i would probably use comfyUI if it wasnt for the fact that i cant use it from a different device like i can with Auto1111

the --api and --listen aswell as --share features are just too useful to switch

but yea im kind of disappointed by the slow speeds in SDXL especially because so little vram is being used

its using like 50% of the available vram but its slow as hell, yea there is sdp and token merging ratio but token merging ratio makes things look like a blob

[ornate elk]

Yes they are very nice

[ornate elk]

We can fix your speed.
You should only use --xformers --no-half-vae

[gray oyster]

yea i did that but with that i only get around 1it/s

i will see how it works once i have everything set up

maybe my installation was simply scuffed

im also sad the TensorRT feature isnt ready for SDXL yet but the installation looks like a pain

also what SDXL base should i download? the vae or the one without the vae?

or both?

[young bridge]

Hey all, does anyone know how to update fooocus? Thanks

[gray oyster]

this is with SD 1.5

xformers and no half vae

sdp and xformers both have the same speed at 512x512 it seems

so what is the difference between the VAE version of SDXL 1.0 and the non VAE version?

[ornate elk]

Yes but xformers does use less vram so it's better

[ornate elk]

The one with VAE

[gray oyster]

i did

also it seems like xformers is even slower now

before i had 2.5it/s

now its 2.1

imma do a bit more testing and share the results then

though i had hoped for atleast 3it/s or more especially after i read "improved performance for stable diffusion" in the nvidia driver notes

[tame yoke]

there are 2 versions of a model
ckpt and safetensors, which should i pick?

[karmic crown]

safetensors

[tame yoke]

is it like better?

[karmic crown]

it is "safer"

ckpt can contain malicious code, ulike sefetensor. everything else is the same

if ckpt is from trusted source there is no difference

recommandation is to use safetensors version if it is available

[sullen fiber]

Hello all!
I've been using Stable Diffusion succesfully for the past week, in a Docker container.
But I now tried to add a couple of more checkpoints from civitai.com, but it fails to switch to them? Which is strange, because it works perfectly with others that I used before.
I get this error:

webui-docker-auto-1  |   File "/stable-diffusion-webui/repositories/generative-models/sgm/models/diffusion.py", line 61, in __init__
webui-docker-auto-1  |     self.conditioner = instantiate_from_config(
webui-docker-auto-1  |   File "/stable-diffusion-webui/repositories/generative-models/sgm/util.py", line 175, in instantiate_from_config
webui-docker-auto-1  |     return get_obj_from_str(config["target"])(**config.get("params", dict()))
webui-docker-auto-1  |   File "/stable-diffusion-webui/repositories/generative-models/sgm/modules/encoders/modules.py", line 88, in __init__
webui-docker-auto-1  |     embedder = instantiate_from_config(embconfig)
webui-docker-auto-1  |   File "/stable-diffusion-webui/repositories/generative-models/sgm/util.py", line 175, in instantiate_from_config
webui-docker-auto-1  |     return get_obj_from_str(config["target"])(**config.get("params", dict()))
webui-docker-auto-1  |   File "/stable-diffusion-webui/repositories/generative-models/sgm/modules/encoders/modules.py", line 428, in __init__
webui-docker-auto-1  |     model, _, _ = open_clip.create_model_and_transforms(
webui-docker-auto-1  |   File "/opt/conda/lib/python3.10/site-packages/open_clip/factory.py", line 201, in create_model_and_transforms
webui-docker-auto-1  |     model = create_model(
webui-docker-auto-1  |   File "/opt/conda/lib/python3.10/site-packages/open_clip/factory.py", line 131, in create_model
webui-docker-auto-1  |     raise RuntimeError(f'Model config for {model_name} not found.')
webui-docker-auto-1  | RuntimeError: Model config for ViT-bigG-14 not found.

Models that do this:

• Counterfeit V2.5
• Omnium

Some that do work:

• edgDreamyDresses
• deliberateForInvoke
• ral-lava

Anyone know what might cause this and how I can fix it?
(Actually, Counterfeit does appear to work now. It's just Omnium that's left. When clicking it in the GUI, it tries to switch for 6 seconds and then switches back to what was selected previously)

[ornate elk]

what are your settings?

[gray oyster]

which ones?

[ornate elk]

resolution, model, steps etc

[gray oyster]

ah

DPM++ 3M SDE Exponential SDXL base 1.0

1024x1024
7 cfg
50 steps

[ornate elk]

okay, then its normal for the low it/s

[gray oyster]

well euler A has the same speed

[ornate elk]

you cant compare it/s with other cards because its usually tested with 1.5 models on 512x512

[gray oyster]

maybe its like 0.1 it/s faster

[gray oyster]

ah

[ornate elk]

try it and then youll get much more it/s

[gray oyster]

well i have a private server in which we did some testing and the 6800XT on arch linux with rocm reaches 3 to 3.5 it/s at 50steps 1024x1024 and 7cfg with sdxl base

been a long time since we did the comparisons though but the numbers shouldnt have changed

i can overclock my gpu to reach near 4it/s iirc but its very unstable and the webui keeps crashing after a few generations

[ornate elk]

yea i dont recommend that

[gray oyster]

yea

[ornate elk]

okay thats interesting,

[gray oyster]

tbf it was a red devil overclocked one

sadly the guy left the server and i lost contact to him

[ornate elk]

ah damn, would liked him to test again ^^

[lone wolf]

Is this normal?

My SD isnt working as intended 😦

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --update-all-extensions --medvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half --upcast-sampling

call webui.bat

Text2images are not behaving right

[ornate elk]

whats the issue?

[lone wolf]

Prompts before, yesterday

Today

Exactly same promps, same seeds, everything

its kinda fcked up

I didnt do anything but turn off my PC and on again

[gray oyster]

maybe the RNG is different?

[lone wolf]

I mean the faces keeps showing like this

no matter what I do

No matter what model

[gray oyster]

try all different options

[lone wolf]

Yeah Im only on GPU,

7900 XTX

Did the AMD install, worked fine, but not after quitting and starting again...

[gray oyster]

are you using a different VAE perhaps?

or any LoRAs?

[ornate elk]

the faces are not the same size. on the second image is no vae used

[lone wolf]

What happened to the vae?

[ornate elk]

idk, do you have any vae?

[lone wolf]

No idea what vae are tbh, I just copyed from my last SD before the Turbo stuff came

Then it just broke somehow

[onyx grove]

what exactly is a checkpoint in stable diffusion?

[ornate elk]

can you show me a new generated image with settings like steps and resolution ?

[ornate elk]

its a little bit of the large trained database of txt2img that got trained with many gpus.
Its the core for image generation. the knowledge of SD comes from the model you choose

[lone wolf]

The images just get so weird, this is the same prompt I've used with many different seeds, but after today it just gets weird like this

[onyx grove]

is there like specific checkpoints for specific arts? where can I learn more?

currently I'm running on this

[ornate elk]

can you try to remove all the negatives and add only blurry, deformed, low quality as negative?

[ornate elk]

yes, the community trained many models and you can find the most on Civitai.com
there are models for everything

[dire vine]

Does Stable Animation require an NVidia GPU? I'm using Intel ARC A750. I'm also curious if I can use it with Automatic1111?

[ornate elk]

idk about the stable animation support, but there should be an fork of auto1111 made for intel arc gpus

[dire vine]

I'm currently using Auto1111 just fine for all other purposes, I'm mostly just curious about Stable Animation (I don't know how/where to start with it and Auto1111)

[lone wolf]

Still very weird :S

[ornate elk]

oh i know the problem

[onyx grove]

i'm just a beginner looking to create images in this style, is sdxl good enough for me already?

[ornate elk]

@lone wolf you have hyperrealism:15 instead of :1.5

everything above 1.5 can cause massive issues

[ornate elk]

sdxl should be really good for that yes

you can find other sdxl models on civitai too

[onyx grove]

models? how does that work on civitai?

also how many checkpoints are there

[ornate elk]

models = checkpoints

[onyx grove]

ah, how many are they?? seems like a lot

[ornate elk]

idk how many ^^

a lot for sure

[lone wolf]

Even removed it, still wonky af, damn I have no idea :S I've tried to reinstall loras and models , still weird, Idk why it worked perfectly yesterday

[onyx grove]

but the default one right now is sdxl?

[ornate elk]

can you try euler a as sampler

[ornate elk]

there are official models from stability.ai these are 1.5-ema-pruned, 2.1 and SDXL 1.0
the rest are community made models

[onyx grove]

the one i'm using right now, is sdxl 1.0 or is it 0.9

also i see a lot of civitai people on sd 1.5

[ornate elk]

its sdxl 1.0 with the 0.9 vae

[ornate elk]

1.5 is less resource hungry and more compatible with extensions. there are also more 1.5 based models than sdxl

[onyx grove]

what exactly is that?

[onyx grove]

as in for lower end pcs?

[ornate elk]

nope its the base model before sdxl was released

[lone wolf]

Think its broken :S

[ornate elk]

can you share in an original image where its not broken?

then we can compare the settings, as they get stored in the meta data

[onyx grove]

oh i see, what's the difference between sdxl 1.0 and sdxl 1.0 with the 0.9 vae?

[ornate elk]

every model has an included vae. but at the sdxl 1.0 release the 1.0 vae was broken. so they added the not broken 0.9 to it

[lone wolf]

This was the same big prompts higher up as the same image generated 2 days ago when I freshly installed AMD one

[onyx grove]

so when people are downloading sdxl 1.0 they are downloading it with the vae essentially as the original one is broken

[ornate elk]

but as you can see thats not 512x512

so your not using the same settings like back then

[ornate elk]

you can download it without the 0.9 if you want to use other vae files. but yes normaly 1.0+0.9vae is fine

[onyx grove]

as a beginner and my goal of creating that image, should I just stick to my current version and learn the settings etc or explore other models/checkpoints

because my concern atm is whether I'm using the correct model/checkpoint to create the images I'm looking for

[lone wolf]

Yeah I cant run on any other settings than 512 now, Always get mem error "RuntimeError: Could not allocate tensor with 1310720 bytes. There is not enough GPU video memory available!
Time taken: 33.2 sec.

A: 20.24 GB, R: 22.72 GB, Sys: 22.7/22.7236 GB (100.0%)"

Which is weird since 7900 XTX has GIGA memory

[ornate elk]

if your new i would suggest trying different models and maybe start by using 1.5 models, play around with the resolution, steps, cfg, and see what they do.
sdxl behaves differently in terms of prompting. most guides and example images are based on the 1.5 prompting style.
but if its enough for you, you can also stay with sdxl. but i would recommend download a few models that you like and test them out

also whats your gpu? and whats inside your webui-user.bat?
maybe we can adjust the performance

[onyx grove]

Definitely not the best for stable diffusion

[ornate elk]

Remember: everytime you get this error (out of gpu memory) you need to restart the whole webui to clear the stuck vram.
if not you wont be able to generate anything with higher resolution again.
Then make sure you dont run wallpaper engine in the background as it uses vram too.

[lone wolf]

Not using WP engine, but ait ill try restart again, ill restart the whole pc

[ornate elk]

okay and whats inside your webui-user.bat?

[onyx grove]

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

call webui.bat

[ornate elk]

okay, you should add --medvram --no-half-vae

then your ready for every heavy extension or upscaling images

[onyx grove]

could you explain what do those do?

[ornate elk]

--medvram does use less vram per block size. if a model or a task is to heavy in vram load. it gets split to not run out of vram.
--no-half-vae is for compatibility for any vae files

[onyx grove]

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

call webui.bat```

so this is my final setting?

@ornate elk quick question, between this pc with is running off that GPU and my mac m2 which is better for sd

[ornate elk]

i would say the m2 could be a bit faster but im not sure.

[onyx grove]

alright thank you! as for settings what should i go with? (using the current pc not the m2)

[ornate elk]

1.5 models got trained on a 512x512 resolution. you would need to upscale them to get a better quality.
sdxl is trained on 1024x1024 but takes much longer to generate

if you want to use 1.5 models, and you want to make a portrait format go with 512x768 or for wide format 768x512

[onyx grove]

alright i'm going through this article right now https://civitai.com/articles/904/settings-recommendations-for-novices-a-guide-for-understanding-the-settings-of-txt2img and a lot of it seems too high res for my pc do i just use the default one sdxl is giving me?

[ornate elk]

you cant use highres fix for sdxl with your gpu

but for 1.5 models it will work very good

but you need the right settings

This is an example
512x768, 30 steps,
Hires Fix:
Hires Steps: 10
Denois: 0.5
Upscale by: 2
Upscaler: Esrgan4x or Latent bicubic

[onyx grove]

this for the sdxl i'm using?

[ornate elk]

no for 1.5

[onyx grove]

isn't it like built in?

[ornate elk]

the hires fix is built in

but you cant upscale it by2 with 6gb vram

you can try but i dont think it will work

[onyx grove]

i tried it and it generates image doe

also i used runpod to train a lora style and lora character which i now have the files

[ornate elk]

okay nice

[onyx grove]

can it still be used in like 1.5

[ornate elk]

that depends on what youve trained the lora

on 1.5 or sdxl

[onyx grove]

sdxl

[ornate elk]

then it will only work for sdxl

[ornate elk]

hows it going?

[lone wolf]

Still generating xD

[onyx grove]

ahh so do i just use the default settings?

also what's up with this section

[karmic crown]

m2

[onyx grove]

you using a m2?

idk if RAM plays a difference but my m2 is the base model

meaning it's running off 8gb of ram

this pc is running of 16

[karmic crown]

i have m1

i also have Intel iMac with Radeon

and my sons PC has 1660

m2 with 8GB?

air?

[onyx grove]

ye

the latest one i believe

[ornate elk]

do you generate multiple images?

[karmic crown]

8GB will probably not be enough

i have m1 MacBook Pro with 16GB

but you can try

https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/5461

[onyx grove]

alr that makes sense

[karmic crown]

start with the initial setup from the fist post (first 4 steps)

and then continue with my setup

https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/5461

[onyx grove]

@ornate elk do i have to touch these?

[ornate elk]

hey @karmic crown i didnt wanted to ping you to not disturb, but glad your here for the mac questions ^^

[ornate elk]

you need a refiner model to use that

[onyx grove]

i believe this is fine, however what does switch at do?

[ornate elk]

if you have 100 steps and you set it to 0.8 (80%) it will switch at 80 steps from the sdxl model to the refiner to enhance the quality

just as example

[onyx grove]

that sounds about right, should i use the refiner or just leave it empty

[ornate elk]

you can try it out, i mostly dont use it

[karmic crown]

please do not use over 50 steps (friendly advice)

🙂

[onyx grove]

gotcha

[karmic crown]

anything over 50 is pointless

some samplers will do nothing, and some others will probably just add mess to your image

[onyx grove]

sounds good

if i'm trying to describe this image what should i say

[karmic crown]

you will get the best answers at #📝｜prompting-help

[onyx grove]

alright perfect thank you

[ornate elk]

you can also put images into img2img and press the clip interrogate button, the ai tries to describe the image

[onyx grove]

this? Also thanks so much for your tips! my sdxl generation is so much faster now

[ornate elk]

yes, and no problem 🙂

[onyx grove]

decided to turn off the upscaler with your advice and it's definitely running better, but what does the upscaler do actually and is it essential?

[ornate elk]

its essentials for low resolution images. when using 1.5 models

[onyx grove]

image resolution seems fine to me

[fair oxide]

Somone plz help me, im getting this error when launching. It does successfullly launch but theres something odd and it wont gen images:

Python 3.10.13 (main, Aug 24 2023, 12:59:26) [Clang 15.0.0 (clang-1500.0.40.1)]
Version: v1.6.0-2-g4afaaf8a
Commit hash: 4afaaf8a020c1df457bcf7250cb1c7f609699fa7
Launching Web UI with arguments: --skip-torch-cuda-test --opt-sub-quad-attention --upcast-sampling --no-half-vae --medvram-sdxl
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
Warning: caught exception 'Torch not compiled with CUDA enabled', memory monitor disabled
Loading weights [62bb78983a] from /Users/kaare/stable-diffusion-webui/models/Stable-diffusion/epicphotogasm_lastUnicorn.safetensors
Creating model from config: /Users/kaare/stable-diffusion-webui/configs/v1-inference.yaml
Running on local URL: http://127.0.0.1:7868

To create a public link, set share=True in launch().
Startup time: 4.8s (prepare environment: 0.1s, import torch: 1.7s, import gradio: 0.6s, setup paths: 0.7s, initialize shared: 0.1s, other imports: 0.6s, load scripts: 0.3s, create ui: 0.3s, gradio launch: 0.3s).

I get this when I gen an image:
NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

Ive tried basicly everything here: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/5461#discussioncomment-7533743

[ornate elk]

whats your device ? m1 ?

[fair oxide]

Ye

I managed to get it to work like 2 times

and then like the third time I tried ealrlier it died

[ornate elk]

make sure the model you use is larger than 1.98gb

[fair oxide]

so

[fair oxide]

Yeah it is

[fair oxide]

I started changing the code to make it work always, which then it just wouldnt work

THe time I got it to work I used this: python_cmd="python3.10" ./webui.sh --opt-split-attention-v1 --medvram

instead of ./webui.sh

but none of those work anymore

:/

[fair oxide]

I tried that agian:

To create a public link, set share=True in launch().
Startup time: 7.2s (prepare environment: 0.1s, import torch: 2.0s, import gradio: 0.6s, setup paths: 0.7s, initialize shared: 1.9s, other imports: 0.7s, load scripts: 0.3s, create ui: 0.3s, gradio launch: 0.6s).
Applying attention optimization: sub-quadratic... done.
-[IOGPUMetalCommandBuffer validate]:216: failed assertion `commit command buffer with uncommitted encoder'
./webui.sh: line 255: 6446 Abort trap: 6 "${python_cmd}" -u "${LAUNCH_SCRIPT}" "$@"