🤝｜tech-support

[ornate elk]

Note that sdxl models won't work with that. And some extensions might have errors. But most of them are fine

[bleak prism]

ok thx for your help, i will try to install it

[ornate elk]

The speed increase for 1.5 models is immense. I can generate a 512x768 image, 30 steps, with hires fix upscaled by 2 + using controlnet openpose. And it only takes 20 seconds instead of 1:20-2mins

[ornate elk]

Np, if you have any question feel free to ask.

[bleak prism]

1:20-2min => 20s, u got some realy nice boost, i need atleast 5 min with hires fix

[ornate elk]

You need to make sure to use 10 hires steps

[bleak prism]

noted

[cosmic furnace]

found only install of zluda on alternative webui... vladmandic one =\

but he write it needs from 10-15 to even sometimes 60 mins for first laucnh- generation -_-

and anyway it got not as good UI as automatic directml y?

only thing i didnt tried is using 3.2 zluda but i doubt its that =\

[ornate elk]

Yea I dont like the sd.next ui and preimplemented stuff too

[ornate elk]

I'm not at home rn. But yea its not Zluda. Maybe try edit the share_init.py again and make a space (hit enter before it) so that the new added line is not connected to the above code

[fair oxide]

SD 3.0 https://x.com/StabilityAI/status/1760656767237656820?s=20

Stability AI (@StabilityAI) on X

Announcing Stable Diffusion 3, our most capable text-to-image model, utilizing a diffusion transformer architecture for greatly improved performance in multi-subject prompts, image quality, and spelling abilities.

Today, we are opening the waitlist for early preview. This phase…

Twitter · Feb 22, 2024, 1:24 PM

[wraith bronze]

I randomly started to get this error in img2img, I've generated 75 images so far without issue, what could be the cause of this? txt2img works fine. I've restarted SD several times.

I use SDXLFaetastic_v24 [07b985d12f] and sdxl_vae. I'll try SD1.5 and see if the same thing happens. Could it be the model or vae that decided to go corrupt on me?

[ornate elk]

Make sure your webui is whitelisted in any browser adblocker

Sometimes adblocker can mess up img2img

[wraith bronze]

SD1.5 works fine in img2img. I'll try SDXL again with my adblockers off

OK now it works. Thanks!

Well at least it doesn't crash, but the results are a different story xD

[ornate elk]

What's your GPU?

[wraith bronze]

I just bought a 3090 to have a NVIDIA GPU ^^ it isn't slow and normally works perfect.

[ornate elk]

Ah nice, make sure to use --xformers OR --opt-sdp-attention in your webui-user.bat to get the best performance

[wraith bronze]

I was using AMD, but then 512x512 took 15min to complete xD
Oh I will check if I forgot to add that back after swapping to forge

Oh I think I see why it looked messed up, after restarting I forgot to set the resolution to the same size so it was downscaling it ^^

[ornate elk]

Oh well

xD

[wraith bronze]

I still look ugly, but not as ugly as a downscaled me! xD

[valid gorge]

How do I enable the gallery view in Comfy? I had in on Mac, I think it was there without any extension (not sure). EDIT: JNodes, Image Drawer

it simply showed the generated images on the bottom / left / right / top as a simple row

[paper crag]

say guys, did you controlnet ever disappear from you page?
I had it there a few days ago, now its gone.

[valid gorge]

Hmm, I wonder what happened. I got earlier ~17it/s, now I'm getting ~13it/s with the same settings etc

[old parrot]

Anyone know why 2.5x 0.3 Denoise, Hires. Fix in the Txt2img tab takes 82 minutes and the img2img SD upscaler only takes 4 minutes? Same settings on both

[placid torrent]

anyway to get fast access after sign up

sd3

[ornate elk]

Check if its shown in the Extensions tab

[ornate elk]

Its slightly different to upscale with hires fix than with SD upscale script.
Make sure to use 10 hires steps with hires fix

[ornate elk]

Did you get it to work?

[paper crag]

its there. I even updated it and restarted my PC. Still not appearing

[ornate elk]

Make sure your webui is whitelisted in any browser adblocker

Also update the Extensions

[paper crag]

i've disabled the adblock on the page, and update the extensions, still nothing

[valid gorge]

Looks like it's both A1111 and Comfy. So I'm guessing it's not SD related

[bleak prism]

i tried but got error, so i will test out tomorow

[wraith bronze]

What do your extensions say? Is it in that list? enabled?

[dire blade]

ah ok here is the help area

can anyone help me? When i clone a repository it only clones the frame not the files inside it

what am I doing wrong?

[wraith bronze]

I guess you could try and disable it, restart, enable it and restart again. I get "- ControlNet - INFO - ControlNet UI callback registered." if it is loaded.

[dire blade]

for instance

https://huggingface.co/h94/IP-Adapter/tree/main

if I clone here it clones the folders but not the contents

I have to manually download all the contents and place them in the folder from my download directory, I imagine this isnt by design right?

[wraith bronze]

If that doesn't work I guess you need to reinstall controlnet

[wraith bronze]

In the correct folder?

[dire blade]

what do you mean?

I have to click and download each and every file as they are not in the clones folders when i clone a repo

afaik when you clone the folders and the internal files should be clones

not just the folders

[wraith bronze]

But you need to place the files in the appropriate place as well not just plopp them down whereever

[dire blade]

so as per the link above when I clone that it clones the models and sdxl folder on my pc but none of the files

[dire blade]

yes, I am aware

nevermind problem fixed

thanks anyways

[old parrot]

Thanks! How many Hires steps do you suggest for img2img SD upscale?

[cosmic furnace]

@ornate elk but what choice do i got now? Except trying install based on this vladmantic? https://github.com/vladmandic/automatic/wiki/ZLUDA . At least to make sure its smthing broke with automatic install not other one dunno .

[drifting aspen]

Hello which vae do you recommend for enhancing coloring page type images? Because I've been testing all day and I don't see any difference with vae and without. Unless I am doing something wrong. Anything you guys can advise?

[scenic saddle]

Hey y'all I'm trying to run stable diffusion through colab but keep getting this error - python3: can't open file '/content/gdrive//sd/stable-diffusion-w/webui.py': [Errno 2] No such file or directory - I've tried deleting the sd files but I keep getting the same error. dose anyone know how to fix this issue

[thorn wraith]

Check the information on the model you're using. It's been my experience that some work better with certain vae's.

[drifting aspen]

And where this information is included in info civitai does not write too much

[thorn wraith]

Sometimes they do, sometimes they don't. You can try asking in the comments of the specific model.

[drifting aspen]

That's what I think I'll do

[thorn wraith]

Good luck finding an answer.

[upbeat atlas]

Yeah on every browser, and yes both C and D drive are on SSD. At first I thought it was a probelm only in ZLUDA but directml also was acting like that. I'm currently running the HIP version for the ZLUDA, so I'm now re downloading Adrenaline 24.1.1 and installing it to see if it fixes it, ima run the windows fix command after if it doesn't work. thanks again.

So I did both the adrenaline 24.1.1 re install and the fix command, still having the same issue. Any other ideas on how to fix it?
The program is running correctly, it's the browser interface that's laggy and doesn't work right

[ornate elk]

did you tried the step i mentioned with editing the shared_init.py`?

[ornate elk]

there are no hires steps in general when using the upscale script in img2img, because in img2img it tiles the input image into multiple tiles and will generate multiple images with the steps set at the top, like generating a normal image, then they get merged to the final output

so 20-30 is enough

and denois should be below 0.3

[cosmic furnace]

Yes 😦

and not even sure how to try this vladmandic because i cant launch it even once... and nothing mentioned about it in guide

and webuser.bad got alot of strings there >_> and not recomended to edit it on top of it 😒

[ornate elk]

Yea vladimanic didn't worked for me either

[pliant flame]

Helllo!!! Super new to SD trying to install Reactor and I am facing a error

[upbeat atlas]

Any idea what might be causing this?

[solar sundial]

hey guys, not sure anyone saw my message from yesterday. But I'm trying to run StableSwarmUI behind a reverse proxy. but I'm getting API error messages. I've tried nginx and apache, but same issue. Has anyone here done the reverse proxy thing before?

[ornate elk]

hey, did you followed the steps on the github of reactor?

[ornate elk]

can you test this in firefox?

[pliant flame]

Yes sir using Firefox

And also followed several troubleshooting methods - installing a prebuilt insightface package, MICROSOFT visual studio dev and C++ tools etc

[upbeat atlas]

It's the same in Firefox, edge, opera gx and brave

[pliant flame]

No it still says that it cannot find insightface and it is necessary to run the application

[ornate elk]

do you have any third party antivirus software installed?

[upbeat atlas]

No nothing. It was working just fine until yesterday, I think I messed sth up while installing ZLUDA.

[ornate elk]

hmm normaly nothing affects the directml webui when you installed the zluda webui in a different folder

[upbeat atlas]

I'm not even sure if this has anything to do with it, but is the path files correct?

[upbeat atlas]

The only thing that comes to mind is either messing up some path files or messing up files while installing HIP to replace Adrenaline 24.1.1

But I did reinstall adrenaline so at this point I'm just lost 🧍‍♂️

[pliant flame]

No, ser

[cosmic furnace]

Mb theres detailed log file telling whats wrong exactly? 🤔

[solar sundial]

it's probably easier to run as a VM on linux versus windows... i was up in running in a few minutes

[upbeat atlas]

I have absolutely no experience and knowledge with linux, is it possible to run it with windows as well or do I have to completely change to linux?

[solar sundial]

one moment.. let me link you

[ornate elk]

python 3.11 shouldnt be installed if your using auto1111

but if you need it for something else its okay

but the python 3.10 scripts folder is missing 😮

[solar sundial]

https://github.com/AbdBarho/stable-diffusion-webui-docker

[ornate elk]

do you have a models/insightface folder?

[ornate elk]

can start the webui and show me the full cmd output?

[solar sundial]

This feels like a dumb question to me. But what is the difference between stableswarmui vs auto1111? I understand thats like... base stable difusion?

[upbeat atlas]

Thanks, ima check it out ❤️

[pliant flame]

Yes, also it has onxx file if I’m not wrong

Approx 500mb

[ornate elk]

stableswarm ui is a different ui, it has a similar ui to auto1111, the big difference is that its using Comfyui as backend

[upbeat atlas]

I added the scripts, still the same ):

[ornate elk]

do you need python 3.11?

you should uninstall it if not needed
and then delete the venv folder

[solar sundial]

whats comfyui?

[upbeat atlas]

No I don't even know what it's for xD

[ornate elk]

another webui to use stable diffusion

[ornate elk]

then pls uninstall it and then delete the venv folder, (for zluda you then need to do step 4 again)

[solar sundial]

ok... so... what is the 'generative-model' thing then? Im confused how to use that one specifically

[ornate elk]

which one?

[solar sundial]

https://github.com/Stability-AI/generative-models

this thing

[upbeat atlas]

Ok, but just to make sure, it's the one on the bottom, correct?

[ornate elk]

can you show me the vc++ when you open it?

[ornate elk]

oh you have miniconda

thats the 3.11

if you dont need it, uninstall it

best would be to uninstall evrything from python

then install python 3.10.11 64bit

then everything should work

[solar sundial]

I concur with all things here...

[ornate elk]

this is a collection of multiple SD models, you can download SDXL, turbo and 1.5 from huggingface or civitai and use them in any Webui (auto111, stableswarm, comfyui)

[solar sundial]

ok.. so.. I do have it downloaded. is it as simple as just moving the folder to the stableswarmui directory?

like just dump it in the Models directory or something?

[ornate elk]

you need a .safetensor file and then you can move it into the models folder

[solar sundial]

where do you get that?

[ornate elk]

SDXL: (sd_xl_base_1.0_0.9vae.safetensors) https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main
SDXL Turbo: https://huggingface.co/stabilityai/sdxl-turbo/tree/main

[solar sundial]

ah. huggingface. are they the... i dunno? gatekeepers for some stuff so we can run locally?

[ornate elk]

there are no gatekeepers, you can run any model localy

[solar sundial]

i mean, for this file you're mentioning?

[ornate elk]

i dont understand the question sry

[solar sundial]

its ok. I'm still figuring this stuff out. not sure what questions or how to ask some of these details

[ornate elk]

the only thing that prevent you from using it if your graphics card has 4gb or less vram

[solar sundial]

yeah I have a rtx 2060 (the old 6GB version)

[ornate elk]

wouldnt recommend using sdxl with that, but it can work

it will be slower than 1.5

[solar sundial]

it's slow at times i'll say that

not too worried about speed. This is an intern project I'm doing

i need to milk the hours

[ornate elk]

😄

[solar sundial]

so either of those safetensor files are ok to use? I'm not sure how to use these files with the generative-models i downloaded

is there a guide somewhere I just havn't found yet?

[upbeat atlas]

I did that just now and still no luck, made sure it's added to the path too, I'm beat so I guess I just take the L for now, if you happened to find anything that helps pls @ me, thanks again 🙏

[cosmic furnace]

@ornate elk Any idea what to do with this ZLUDa - torch error? 😒

[ornate elk]

have you deletet the venv folder?

[solar sundial]

go install pop_os! as a virtual machine. you might have some better luck with it

[upbeat atlas]

Yeah I did

[upbeat atlas]

Yeah might as well try that, automatic1111 webui just isn't working

[solar sundial]

the install instructions are simple enough, you will be up and running pretty quick I bet

[ornate elk]

you can share me screenshots of your zluda folder, of your environment path with both added, a screenshot of the editet share_init.py and we can compare with mine

[solar sundial]

@ornate elk hey what do I do with the safetensor file? how do i use this with the generative-models folder? am i just moving files together as one or something?

[ornate elk]

are you trying to use sdxl or svd ?

[solar sundial]

i'm trying to use anything in general. i have no preference

[flat olive]

how i can get the onnx and olive tabs on my stable diffusion?

[solar sundial]

in my stableswarmui install I do have the SDXL model available (used it a couple of times). am I trying to do something that is already installed/available to use?

[ornate elk]

yes

xD

if you already have sdxl running you dont need the generativ-models

you find more models on Civitai.com

these are .safetensor files

these go to the same folder as the sdxl model

[ornate elk]

trust me you dont want them xD, whats your gpu?

[solar sundial]

ah, ok. So, the generative-models stuff i downloaded is if I'm trying to make some custom stuff?

[flat olive]

rx 580 8gb

i see that using some option that optimize the it/s

[ornate elk]

yes but its pretty restrictet, extensions wont work, loras are not supported etc, each model needs to be converted and is fixed on a resolution

not worth the speed you get

better try out Shark, its the same but better

[flat olive]

i have a issue with shark

i posted it on the github but im waiting for some response and on the discord server no one says nothing about so

my only option its this

[solar sundial]

@ornate elk One thing I'm noticing in my stableswarmui is I'm missing the video models to select from. is that not built in with SDXL?

[flat olive]

about this, vulkan shaders compilation never ends if u wanna know

[ornate elk]

oh okay, have you tried an older version?

also if you want to get Onnx/olive running, follow the devs guide here:
https://github.com/lshqqytiger/stable-diffusion-webui-directml/discussions/149#discussioncomment-8392257

[flat olive]

ty

[flat olive]

yep i tried one from 2023 and the last one from this year from january

[solar sundial]

@ornate elk thanks for your help by the way

[cosmic furnace]

is there a way to check if hip-sdk & zluda installed correctly & working fine? dunno 🤔

well with hip sdk nothing cud go wrong in theory y? 🤨

[ornate elk]

you can try open up a cmd and type
zluda --help

and it should show stuff

[ornate elk]

did you moved the files in the correct folder?

[cosmic furnace]

seems like yeah

considering ive done full steps install twice

[ornate elk]

looks good, but you dont need the 64_10

only the 64_11

[cosmic furnace]

well that was a try what if win 10 🙂

open cmd inside zluda folder?

i guess its working -_-

[ornate elk]

can you show me the torch folders in your venv ?

[cosmic furnace]

?

this?

or the torch folder itself?)

102 files

[solar sundial]

should I set my 6GB rtx 2060 to lowVRAM mode? I'm getting lots of errors/failures on the backend with cumfyUI when doing repeated image generations

[ornate elk]

compare it with mine

hmm looks the same

[ornate elk]

would recommend you to use automatic1111 instead of swarm

[solar sundial]

Yeah i'm thinking the same....

[ornate elk]

there you can adjust the performance args better and get better help

[solar sundial]

its all the same in the end? SDXL model etc?

[ornate elk]

yes

[solar sundial]

sweet

[ornate elk]

the models can be used with any webui

[solar sundial]

youve been a big help. sorry for asking very noob questions

[cosmic furnace]

theese 11 torch folders got 12937 files & 786 subfolders?) & 4.57 gb size? to make more exact comparing

[ornate elk]

12934 and 4.56gb

for me

[cosmic furnace]

ah yeah its those "spare" _10 files ...

[ornate elk]

np, if you want to install Automatic1111, please follow my install guide, its in the pinned messages of this channel

[solar sundial]

yeah i found it

[cosmic furnace]

😵‍💫

[winged stone]

i see lots of prompts with < > and "lora" text inside with numbers. what do they mean?

[cosmic furnace]

it shudnt launch with directml command y ?)

[ornate elk]

what did you changed?

[cosmic furnace]

just added directml back into com args)

to see if it run >_>

[ornate elk]

you need to delete the venv folder then too

to remove nvidia cuda stuff from torch

[cosmic furnace]

oh)

so theres no other way to test this "torch" huh? -_-

but just laucnh webui?

[solar sundial]

@ornate elk is it imperative i install tcmalloc?

or ignore...

[ornate elk]

are you on linux or windows?

[solar sundial]

pop_os!

linux

[ornate elk]

ah okay, can be ignored

[solar sundial]

i figured.. i just see this red letters 'improves cpu memory usage' in ()

[ornate elk]

if you know how, you can install it xD

[solar sundial]

i was about to then i began to feel like... i just dont want to lol

so will THIS run behind a reverse proxy without anything really special to do?

[ornate elk]

It runs localy so yes, you can also specify the --port

[solar sundial]

omg these images from automatic1111 suck in comparison to stableswarmui....

[ornate elk]

Why? Did you used the same model ?

Auto1111 comes with 1.5 base. Its not good

[solar sundial]

i tried it once,... giving me a lot of errors in the UI. i'm just running it on a KVM kind of console via proxmox.

I have no idea how to select the model to be honest

[solar sundial]

i see v1.5 pruned emaonly safetensors

as a SD checkpoint...

[ornate elk]

You need to download a model from civitai.com
For example Dreamshaper v8 or Dreamshaper xl
Then put them into the models/Stable-diffusion folder

Checkpoints = models

[solar sundial]

ah. ok

word

dreamshaper xl is just SDXL?

[ornate elk]

is based on sdxl

and finetuned for better outputs

[solar sundial]

word

any suggestions here?

CUDA out of memory. Tried to allocate 14.00 MiB (GPU 0; 5.79 GiB total capacity; 5.46 GiB already allocated; 13.69 MiB free; 5.65 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[ornate elk]

You need to made changes to your webui-user.sh

Out comment the set COMMANDLINE args

There add: --xformers --medvram --no-half-vae
Then save and relaunch

[solar sundial]

you're talking this section?

Commandline arguments for webui.py, for example: export COMMANDLINE_ARGS="--medvram --opt-split-attention"

#export COMMANDLINE_ARGS=""

[tired copper]

Will we see SD3 in ClipDrop?

[ornate elk]

Remove the second #

Before the Export

Then add the commands I said

[solar sundial]

export COMMANDLINE_ARGS="--xformers --medvram --no-half-vae"

this?

[ornate elk]

Yes

[solar sundial]

it was still running out of vram.. i think its because of the desktop running. i'm trying to get the reverse proxy configured correctly right now

[ornate elk]

What settings did you tried?

Is xformers installed correctly?

[solar sundial]

i think so.... i had no errors with install

is xformers a python thing?

[ornate elk]

No its for pytorch and cuda

[solar sundial]

i dont think it was installed... i did pip install xformers... install abunch of stuff

it's installing i should say

@ornate elk still running out of memory...

[ornate elk]

What settings do you try?

Check at the bottom if xformers is installed

[solar sundial]

its the one from earlier... # Commandline arguments for webui.py, for example: export COMMANDLINE_ARGS="--medvram --opt-split-attention"
export COMMANDLINE_ARGS="--xformers --medvram --no-half-vae"

check bottom of what?

[ornate elk]

In the browser

At the bottom it should show the xformers version

[solar sundial]

oh.. i see. yeah. n/a it says. i installed it after i did the 'install' for a1111

[ornate elk]

Na means its not installed correctly

[solar sundial]

yeah i figured

[ornate elk]

Try that:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers#building-xformers-on-linux-from-anonymous-user

[narrow swift]

Anyone knows what’s the VRAM requirement for the new SD3?

[solar sundial]

this is annoying... wont install torch now

even though its installed

[umbral storm]

Anyone here who has experience with sending files to the instance from google drive on vast.ai? I am a newb to this and I have no idea how to do it.

[solar sundial]

can you not download locally and just do a scp

[upbeat atlas]

Hi, so I tried shark and comfyui and they both work just absolutely fine. Automatic1111 still has the problem. I was wondering if anyone knew how to increase/optimize comfyui for AMD windows. Like does it have arguments like in automatic1111? (Such as medvram, no half and so on)

[unique bear]

How do I fix this error? Error occurred when executing VHS_VideoCombine:

Could not find ffmpeg

File "/workspace/ComfyUI/execution.py", line 152, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "/workspace/ComfyUI/execution.py", line 82, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "/workspace/ComfyUI/execution.py", line 75, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "/workspace/ComfyUI/custom_nodes/ComfyUI-VideoHelperSuite/videohelpersuite/nodes.py", line 298, in combine_video
raise ProcessLookupError("Could not find ffmpeg")

[nocturne topaz]

do you have ffmpeg installed on your system?

[solar sundial]

random question to that... do we need ffmpeg installed for the video function to work?

[nocturne topaz]

depends on the function

that one, yes

[solar sundial]

im using stableswarmui. i noticed i have no video module in the drop down menu for that function

[unique bear]

solved it using apt-get

On ComfyUI i'm trying to make animatediff longer than 16fps, it's not reading my context_length properly

[quick forge]

can anyone suggest any good sensor monitors that will show live temperatures of all sensors?

also, how do i unstall something that i installed using wget (ubuntu)?

[fair oxide]

For Ubuntu/Debian: sudo apt remove <program_name>

[quick forge]

i have this LORA, "Sarah.safetensors"
Anybody know how I can find out where I downloaded this from because it doesn't seem to be Civitai although I thought that's where I got it.
I know it has a picture of a woman sitting in front of a shop so it might be civitai still.

[nocturne topaz]

https://civitai.com/models/162836?modelVersionId=310277

[quick forge]

hmmm....not the image i remember..... but i think thats the right one.
how did you find that so fast?!

[nocturne topaz]

googled sarah.safetensors, was the third result

[quick forge]

i did the same thing

today has not been a good day :/

[unique bear]

Anyone have issues with batch prompt schedule only looking at the first line for prompts?

[upbeat atlas]

Quick question, is it possible to run forge on amd windows?

[cosmic furnace]

Anyone knows why SD next uses CPU instead of GPU ?

[nocturne topaz]

Torch not installed correctly, gpu not compatible

[cosmic furnace]

y seems like it =\

all atempts to run zluda & install torch fail by some reason -_-

[topaz canyon]

How can I get sdxl lightning loras to work on comfyui?

I add them as a lora through the CR Lora Stack node and they don't seem to make good generations in 2/4/8 steps on euler/sgm_uniform

Also CFG is set to 1.

Ah, I see where I went wrong. It doesn't work with clipskip based models

[harsh egret]

hello
I don't have InterrogateCLIP.

How can I get it?

[formal trout]

it is the clip icon

[formal trout]

I'm trying to find where this model is, from stable cascade extension for Automatic11111, the 7 GB mostly

[nocturne topaz]

Google huggingface cache location

I dont remember where, but thats what it is hiding in

[formal trout]

oh yes it seems to be there, ty

[dense obsidian]

trying to get stable diffusion to work for AMD, there's some tutorials and things I'm working through, but currently stuck on permissions to update a file. I tried figuring it out for a few hours, but no luck

stderr: ERROR: Could not install packages due to an OSError: [WinError 5] Access is denied:

Can anyone help?

I am running as the admin, I should have full access to update any files

[karmic crown]

you probably have mcafee or some other 3rd party antivirus that makes a mess

the other posibility is that you are trying to install stable diffusion in a wrong place, like onedrive or some other shared folder

#🤝｜tech-support message <- AMD guide is here

[dense obsidian]

just windows defender, no other antivirus.

I'm trying to do all the cmd commands in the folder, the lib folder, and running the same webui-user.bat that is above it

I'll take a look at the amd guide, maybe I'm missing something there

I'll start over from the beginning using the guide

[karmic crown]

be sure to follow step 1, and do not use mentioned folders

what is your file system btw?

ntfs?

[dense obsidian]

yeah

I appreciate you taking the time to help :D, I'm determined to get sd working tonight

new error haven't come across yet.

ERROR: Could not find a version that satisfies the requirement torch-directml (from versions: none)

[winged stone]

is it bad to merge similar models with each other? like 3 separate models trained for detailed faces and merging them into 1 to not switch between them with each generation

[karmic crown]

have you executed pip cache purge as instructed before beginning?

there is a part like "if you failed to install...."

[dense obsidian]

I did on my previous attempts, but didn't try this time XD, whoops, let'se if it fixes

same error

RuntimeError: Couldn't install torch.
Command: "C:\ai\stable-diffusion-webui-directml\venv\Scripts\python.exe" -m pip install torch==2.0.0 torchvision==0.15.1 torch-directml

[nocturne topaz]

nope.. experimenting with that sort of stuff is how some progress gets made

[winged stone]

oh thank god.. ill make one of each contextual models by merging stuff then

like architectural models into 1, detailed faces into 1 etc

[dense obsidian]

Just to make sure I'm understanding what the the error is telling me, go down to the scripts folder, open python, then paste the

-m pip install torch==2.0.0 torchvision==0.15.1 torch-directml

right?

I'm reading people need to delete the venv folder after reinstalling python, no where did I read in the guides I needed to delete python and reinstall it.

[karmic crown]

be sure to execute step 3 as one line (all at once)

[dense obsidian]

yup, it is

[karmic crown]

for step 5, find webui-user.bat and open with edit

find this like

and make it look like this

set COMMANDLINE_ARGS=--use-directml --medvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling

[dense obsidian]

ohh I think I found the issue, all the tutorials I watched said to install 3.10.6, the above guide says 3.10.11

[karmic crown]

3.10.6 should work

wait

i mixed tutorials

sorry

there are 2 amd tutorials now

[dense obsidian]

ah, I had copied the above tutorial

[karmic crown]

just be sure to follow one of them step by stepž

do not mix steps

come to the dark side (mac) we have a script for this 🤣 (i made it)

[dense obsidian]

lol, if I ever make it big as youtuber, who knows I might consider it

[karmic crown]

now serously, copy paste tutorial to some txt file, and follow it step by step

[dense obsidian]

going to do a full delete and try going through the steps again. Do you have any recommendations on if I need to clear any other caches or things before I restart the process?

Wooo! It works finally, and using the gpu now, 6.9 secs for a dog in a hat

[formal trout]

I did one generation with cascade in Comfy and now I'm getting

made this image without changing the prompt, and setting the models. Then it doesn't work.

[ornate elk]

Yea one is for experimental Zluda xD sry for the confusion

[ornate elk]

Nice! What's your GPU?

[dense obsidian]

7900 xtx

[ornate elk]

Epic, I have the same

[dense obsidian]

took me 13 hours to get it to work, now I can start adding all the extensions and models

[ornate elk]

Did you followed my directml or my Zluda guide?

[dense obsidian]

directml

[ornate elk]

Did you watched broken YouTube amd guides? xD there are a lot

[dense obsidian]

oh so very very many

[ornate elk]

If you have any questions about anything AMD related in SD, let me know 🙂

[dense obsidian]

thank you so much for the guide. I'll most assuredly be back once I break something, which will be sooner rather than later

[ornate elk]

Sure np ^^

The Directml version should work with any extension. But you can't Train on AMD rn. So dont install Dreambooth

[dense obsidian]

ahh I don't think I'll be training, just converting videos to ai generated art for videos.

But I did look up a video saying that it may be a wise move to install a pickle scanner, and scan all the models that I download.

Do you think it's worth it to do?

[ornate elk]

That was a thing when the models where .ckpt (pickled files)
Now most models are .safetensor which is much safer, so you dont need to scan them.
If you download your models from Civitai.com your safe as well as they scan every uploaded file automatically

[dense obsidian]

oh good, I was primarily looking at their website, that saves me time

[ornate elk]

Make sure to follow videos/guides that are newer than 6 months ago. In ai time stuff is changing fast

What worked a year ago might works differently or easier now

[dense obsidian]

I read something also about errors coming up with the fp16, gets fixed if you change it to fp32, is that something I should update?

There's so many videos that pop up that are roughly around 8 months ago

[ornate elk]

There is no issue with fp16 or fp32.
There are some GPUs, like the nvidia GTX 1660 or the AMD RX 580 which can't process fp16 models. So they need --no-half to load them as fp32.
The webui will check and load the models in the type they are. So you dont get any issues with your GPU.

[dense obsidian]

ahh I see. Well that concern is taken care of.

now I'm just working figuring out temporalnet

[valid gorge]

Can SD1.5 embeddings be used with SDXL checkpoints?

[light glacier]

Yes, you need to use the smaller models with your GPU, and make sure the resolution is in multiples of 128.

[light glacier]

They can, but don't have much impact. You need to turn teh weight up high to see it.

[fair oxide]

ive' seen dreamshaper's author use his sd 1.5 embddinggs with sdxl, given he developed those himself, but generally you can't have sd 1.5 embdding work with sdxl, but there is x-adapter plugin that you can use sd 1.5 embeddings, lora, controlnet work with sdxl, so keep a look out for that public release

[formal trout]

does your workflow work with those?

[light glacier]

Which version?

[formal trout]

It is weird that it worked one time, then it didn't

[formal trout]

v4

[valid gorge]

I'll need to test with it thanks guys! @light glacier @fair oxide

[light glacier]

That can only use the ComfyUI models, which are smaller versions already. Make sure your comfy is the latest version (run the update) and don't use any brackets in the prompt.

[fair oxide]

np, we should have this integrated with common webuis soon hopefully https://showlab.github.io/X-Adapter/

[light glacier]

The problem with brackets should be fixed if you're up-to-date, but try without if you're getting problems and using them

[formal trout]

@light glacier should I worry about this ?

[valid gorge]

Perfect!

[light glacier]

Yes, that means you're not using the comfy specific models

[formal trout]

@light glacier also the first model uses 120 steps, I don't see that number in any variable

[light glacier]

All of the others don't include the CLIP and VAE in them.

[formal trout]

I download the ones from the link you shared

[light glacier]

120?!?!?

try 20 steps, in the sampler

[light glacier]

That's not the default

[formal trout]

I loaded the workflow, and changed nothing, and it generates like that

is there a stage a comfy model?

what are the correct models exactly ?

[agile sage]

No. It is embedded in the stage B model

[formal trout]

oh, and here what should I put?

[light glacier]

What are you doing?! This is from the flow that I uploaded, nowhere does it have 120 steps set.

[formal trout]

yes I know, I don't why it is doing that

[light glacier]

This is not my v4 workflow

[formal trout]

I'll check it

[agile sage]

If you are using the comfyUI checkpoints you don't need to worry about stage a or the clip models, B and C do everything. No extr VAE required.

[light glacier]

Ah wait, yes it is sorry

That's the stage C decoder. You need these 2 models in comfyui/models/checkpoints/
https://huggingface.co/stabilityai/stable-cascade/tree/main/comfyui_checkpoints

Then the preview.safetensor from here goes in the VAE folder
https://huggingface.co/stabilityai/stable-cascade/tree/main

[formal trout]

yes it is, I just re download, load it in Comfy, change resolution to 1024x1024, hit Queue, and it starts with 120 steps

[light glacier]

That last one goes in the node you showed above

[light glacier]

Are you loading the json or the png image?

[formal trout]

png

[light glacier]

I don't understand then, I never used 120 steps, and don't get it when I load the png.

[formal trout]

yes I'm using thossee

[formal trout]

oh I think that what was misssing

I was loading stage a instead

[light glacier]

I think there may be a bug in that node. I just downloaded the file from Civit and dropped it into Comfy. The nodes does say 40 steps, but when you run it there are 120 steps.

[formal trout]

well it generated again, with stage a in VAE

[light glacier]

finally! 😄

[light glacier]

ok, because of the way the node works (the restart node) it multiples the number of steps. Even setting at 10 steps makes it run 90.

[formal trout]

yes, I'm trying again, but it keeps doing 120 steps, I didn't anything from the workflow, loaded the json

Now trying with previewer.safetensors instead of stage a

still getting

[light glacier]

Yes, it's the node behaviour. Set the Stage C1 Sampler node to 5 steps, instead of 40

[formal trout]

ok!

[light glacier]

This looks like you have the wrong models in stage B and/or C loader

Like you've selected a bf16 or lite model, but not the comfy ones.

[formal trout]

oh great now I have a preview

[formal trout]

mm I'm almost sure they are the ones from the link

[light glacier]

It's reporting that the vae and clip are missing, which would be the case for any other models that aren't the comfy ones

[formal trout]

well it works nonetheless !(?) thanks!!!!

[light glacier]

Even with the error?

[formal trout]

yes

I tell you I think I got the right models

updated comfy

[light glacier]

That error should stop it working altogether. You're sure it's not just an old error, or does it appear every run?

Should look something like this:

[rgthree] Using rgthree's optimized recursive execution.
Requested to load StableCascadeClipModel
Loading 1 new model
Requested to load StableCascade_C
Loading 1 new model
100%|██████████████████████████████████████████████████████████████████████████████████| 35/35 [00:05<00:00,  6.63it/s]
Requested to load StageC_coder
Loading 1 new model
100%|██████████████████████████████████████████████████████████████████████████████████| 10/10 [00:01<00:00,  6.57it/s]
Requested to load StableCascade_B
Loading 1 new model
100%|██████████████████████████████████████████████████████████████████████████████████| 15/15 [00:05<00:00,  2.63it/s]
Requested to load StageA
Loading 1 new model
WAS Node Suite: Image file saved to:

[formal trout]

It appears in every run

I used 5 steps and now the it is 35 steps, nice

[light glacier]

Did you change the segments?

[formal trout]

no, didn't touch anything there

well less steps affects quality a bit, but still good

is there a prompt length limitation?

[light glacier]

You do need to change the compression along with resolution. There's no formula, but lower res needs lower compression

[light glacier]

Not that I know of

[formal trout]

because I got this

well fantastic @light glacier thanks for your help, it is working good

[light glacier]

Your'e welcome! Do you get the same error with a short prompt? You shouldn't need long stories with this model 🙂

[formal trout]

no, no with short prompt

[light glacier]

The way that C1 node works is strange. I got 35 steps when I set it to 5 and 26 steps when I set it to 8 🤷🏻‍♂️

78 steps when I set it to 10

https://tenor.com/view/krule-john-c-reilly-what-face-reaction-gif-16881707

[quick forge]

is there a stable diffusion 3 beta that people are able to sign up for?
I saw someone post that they signed up and was curious if it was true or not.

[light glacier]

Yes

[light glacier]

https://stability.ai/stablediffusion3

[formal trout]

oh now it is not showing

[quick forge]

Sweet.

I stil have no idea what version of SD im running

[light glacier]

Which model are you loading?

[dense obsidian]

does forge work for the amd version of stable diffusion?

[ornate elk]

nope, i tried it, and it was bugged,

[dense obsidian]

aww that is a shame, would have liked a little bit of a speed boost

[ornate elk]

do you need more speed?

there is the zluda backend for directml, its experimental (and some stuff like sdxl wont work) buts its blazing fast

[dense obsidian]

not yet, I have to figure out how I'm going to get the videos to image format

and from there, the tutorials for how to set up temporalnet are terrible, might have to explore that down the road

[ornate elk]

for that you need ffmpeg or a tool which uses ffmpeg

didnt looked into temporalnet

[dense obsidian]

the video sample of the difference between the standard and temporalnet looked impressive enough, just a matter of trying to figure it out what the tutorials mean by "install here" and they don't show where they want it to go

[dense obsidian]

found the tool for frame extraction. If you happen to look into temporal net and figure it out, would love to hear how you got it to work.

going to call it a night for now, I'll figure out what models from civitai later today

[ornate elk]

ffmpeg or an other tool?
im gonna try out temporal net and let you know if it works

[dense obsidian]

vlc media player

[valid gorge]

Maybe a stupid question but does LoRAs have weight like normal words in prompts, depending on which part of the prompt they're introduced? (A1111)

[ornate elk]

any word thats nearer to the beginning has more attention

loras have their own weight by adjusting the number after its name

[valid gorge]

So it I assume it does not matter if I just place them in the end of prompt, they don't have more weight if they're place earlier in the prompt

[ornate elk]

yea and you also can adjust the lora weight by chaning its number

doesnt make that difference to add them first or in the middle of a prompt

[valid gorge]

Yea ok. So basically this,
Massive three headed dog <lora:flaming eyes dog:0.8> in a huge cave filled with monsters, illustration, 3D, realistic lora:MJ52:0.3

Is same as this,
Massive three headed dog in a huge cave filled with monsters, illustration, 3D, realistic lora:MJ52:0.3 <lora:flaming eyes dog:0.8>

[ornate elk]

you wont get the same output with it, but yea

also make sure to seperate loras with ,

[valid gorge]

Ok, I'll just keep adding the loras in the end of a promt. Yep, sorry forgot the comma when typing the example. Thanks again for the help!

[ornate elk]

no problem!

[thorn yarrow]

What’s the best way to get in touch with Stability.ai regarding commercial licensing of the recent models? The website forms are fruitless as I never got a response over months now

[ornate elk]

Here:
https://stability.ai/contact

[cosmic furnace]

@ornate elk Installed sd next and ...and... uses cpu only -__- . As if i dont have zluda-cuda-torch w/e

sd next discord in the end also told step by step reinstall 😵‍💫

https://github.com/vladmandic/automatic/wiki/ZLUDA have u tried this?

im beginning to think of reinstalling hip sdk - zluda indeed 😒

zluda 3.5 is out btw

he says this needed for zluda o.0

[chilly knot]

on some models i have black images

i have 4 gb vram

i can fix it with bat arguments?

[ornate elk]

now way xD i tested 3.4 yesterday

time for 3.5 then

[ornate elk]

this isnt needed for zluda to work

[ornate elk]

yes, whats your gpu ?

[chilly knot]

1050 ti

[ornate elk]

and what do you currently have in your webui-user.bat?

[chilly knot]

--disable-nan-check --xformers --upcast-sampling --no-hashing --always-batch-cond-uncond --medvram

[ornate elk]

remove --disable-nan-check --upscast-sampling --no-hashing --always-batch-cond-uncond
and add: --no-half-vae
so in the end you only should have: --xformers --medvram --no-half-vae

next step is control your models if they are 2gb or larger.

[chilly knot]

thx i will try

but how i control model

[ornate elk]

go in the models/stable-diffusion folder and look at the file sizes of the models

[chilly knot]

2 gb and 4

[ornate elk]

ok good

[chilly knot]

bad if more 2 gb?

[ornate elk]

nope but slower, if you can take 2gb models

most of the new models are 2gb anyway

[cosmic furnace]

Sdnext behaves as if i launch automatic non directml... Uses cpu. Not showing device gpu. Unlike other ppl seems to have shown exact model used there 😒

Also used that guide step by step... as if something wrong with hip -zluda or that path. 😑 . As if i didnt installed at all that =\

[thorn yarrow]

But that form gets no response at all so far

[tranquil vigil]

Hi, in the end, I faced the same error of Runtime

No CUDA runtime is found, using CUDA_HOME='C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8'
Traceback (most recent call last):
  File "C:\AI\DragGAN\visualizer_drag.py", line 403, in <module>
    main()
  File "C:\Users\Chandrajit Rout\AppData\Local\Programs\Python\Python310\lib\site-packages\click\core.py", line 1157, in __call__
    return self.main(*args, **kwargs)
  File "C:\Users\Chandrajit Rout\AppData\Local\Programs\Python\Python310\lib\site-packages\click\core.py", line 1078, in main
    rv = self.invoke(ctx)
  File "C:\Users\Chandrajit Rout\AppData\Local\Programs\Python\Python310\lib\site-packages\click\core.py", line 1434, in invoke
    return ctx.invoke(self.callback, **ctx.params)
  File "C:\Users\Chandrajit Rout\AppData\Local\Programs\Python\Python310\lib\site-packages\click\core.py", line 783, in invoke
    return __callback(*args, **kwargs)
  File "C:\AI\DragGAN\visualizer_drag.py", line 397, in main
    viz.draw_frame()
  File "C:\AI\DragGAN\visualizer_drag.py", line 168, in draw_frame
    self._async_renderer.set_args(**self.args)
  File "C:\AI\DragGAN\visualizer_drag.py", line 295, in set_args
    self._set_args_sync(**args)
  File "C:\AI\DragGAN\visualizer_drag.py", line 312, in _set_args_sync
    self._renderer_obj = renderer.Renderer()
  File "C:\AI\DragGAN\viz\renderer.py", line 80, in __init__
    self._start_event   = torch.cuda.Event(enable_timing=True)
  File "C:\Users\Chandrajit Rout\AppData\Local\Programs\Python\Python310\lib\site-packages\torch\cuda\streams.py", line 163, in __new__
    return super().__new__(
  File "C:\Users\Chandrajit Rout\AppData\Local\Programs\Python\Python310\lib\site-packages\torch\cuda\_utils.py", line 49, in err_fn
    raise RuntimeError(f"Tried to instantiate dummy base class {class_name}")
RuntimeError: Tried to instantiate dummy base class Event

[ornate elk]

i dont know where the problem is with your installations :/

[cosmic furnace]

Have u tried zluda on sdnext ?) or u dont like the UI

[ornate elk]

i tried to install sd.next but it didnt worked, also i dont like its ui xD

but some on the shark discord got it installed with zluda

[cosmic furnace]

theres vladmantic install & sd next instal

the vladmantic doesnt suit amd kinda

but sdnext feels about right

• i can see some ppl had issues too, but kinda solved them in this way. shud do another type of install for it tho it seems

[ornate elk]

vladiamnics is sd.next

[cosmic furnace]

o.0 but theres 2 kinds of guide for it) about zluda

the one mentioning using directml for at least single launch and the one dont

sec

or theyr seem to just updated guide 😄

-___-

[ornate elk]

there should be only the sd.next guide and my guide for directml

[cosmic furnace]

y seems about right -_-

so u cudnt launch it on sd.next huh? intresting 🤔

[ornate elk]

yea but i dont really tried to fix it

i gave it one chance xD

for me the directml webui (with and without zluda) does everything i need while beeing as close to the original auto1111

[tight sinew]

anyone know how to use ReActor before ADetailer in a1111?

[vocal burrow]

I m off grid for a few days. Try opening a ticket on their github

[old parrot]

Is 193 minutes using 2.5x SD upscale normal with these settings on SD Forge? 6gb VRAM

[valid gorge]

Going to try training a lora tomorrow, let’s see how that goes. Anything I should watch or read on top of the Lora videos from Olivio Sarikas and All Your Tech Ai YouTube?

[ornate elk]

No

Did you used img2img or txt2img hires fix?

[old parrot]

Img2img SD upscale

Do I need to lower this?

[ornate elk]

no thats okay, but when using this you need to set the resolution at the top to 512x512

because its tiling images in multiple tiles

these need to be squares

the output wont be a sqaure and also not 512x512, its just for the upscaling

[old parrot]

Makes sense, thank you! 512x512 in the "resize to" section?

I don't see anywhere else to change

[ornate elk]

yes

[old parrot]

Thanks, you're amazing!

[twin shale]

A frog climbing a ladder

[old parrot]

Easy fix! What sampling steps do you suggest please? These tutorials are all over the place

[ornate elk]

30 is enough

also make sure to select an upscaler at the bottom

and set the denois below 0.3

[old parrot]

Interesting, last video said 150 steps lmao

Thanks again! Went from 193 minutes to 6 minutes... 🧠

[ornate elk]

lol 150 steps makes no sense at all xD

[gray oyster]

When generating 640x640 images it generates at about 5 - 6 it/s with Nvidia memory fallback disabled, but slows down quickly as it runs out of memory, however 1024x1024 generates at 2.5 it/s but barely uses any vram?

correction: hovering around 2.3 and 2.4 it/s

for the 1024x1024

[umbral plover]

Im trying to use MeshGraphomer Hands Refiner with sdxl lightning but im getting blurry results - is it something to do with settings or it's just that this model isn't good for this type of refinment, face detailer works good

[ornate elk]

whats your GPU btw?

you may need some performance args in the webui-user.bat

[old parrot]

I have 1660 Super 6gb

[ornate elk]

then you need to add: --xformers --medvram --no-half
in your webui-user.bat at the line COMMANDLINE_ARGS=

that will give you the max performance on that gpu

[frosty pivot]

I'm using a Mac and am getting

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.

[gray oyster]

Well from the error I would assume it wants you to change the command line args to what it says

Auto1111 did mention something similar in the AMD tutorial too

[frosty pivot]

Yeah i've tried it.

[gray oyster]

setting --no-half and --disable-nan-check?

[frosty pivot]

where is that tut?

yes I can check again.

[gray oyster]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs

I guess not just AMD has that problem

yea I would suggest you try adding cmd line flags until it works one by one

[frosty pivot]

Thank you i'll look into this

[gray oyster]

eg.
1st try: --no-half --disable-nan-check
2nd try: --no-half --disable-nan-check --upcast-sampling
3rd try: --no-half --disable-nan-check --upcast-sampling --precision full

[old parrot]

You're amazing, thank you sir! You need a tip jar

[ornate elk]

you can also use the MAC auto install script for SD, that will automaticly set the best flags for you

its linked in the pinned messages of this channel

[graceful python]

stderr: ERROR: Could not install packages due to an OSError: [WinError 5] Access is denied: 'C:\stable-diffusion\stable-diffusion-webui-directml\venv\Lib\site-packages\onnxruntime\capi\onnxruntime_providers_shared.dll'
Check the permissions.

help pls

RuntimeError: Couldn't install onnxruntime-directml.
Command: "C:\stable-diffusion\stable-diffusion-webui-directml\venv\Scripts\python.exe" -m pip install onnxruntime-directml --prefer-binary
Error code: 1

trying to install SD on an AMD GPU

[formal trout]

I'm still getting this while using cascade workflow

It is a long list

[nocturne topaz]

Have you updated comfy?

[formal trout]

from Manager, should I git pull?

[ornate elk]

hey, did you followed my install guide ?

[formal trout]

still getting it after git pull

[formal trout]

I'm running from Automatic1111 venv, may it be related?

At some point it started to doing totally unrelated generations to prompt

[nocturne topaz]

No idea sorry, my initial thought was nodes might not know how to deal with those models

[formal trout]

It is strange as it generated all right a bunch of times

I'm going to try to run it in its own venv (my poor disk)

[karmic crown]

which mac do you have?

[nocturne topaz]

Which mac, and are you using a sdxl model?

@frosty pivot

[formal trout]

I tried running Comfy without activating Automatic venv but it seems it is using it now anyway as it didn't install nothing

maybe I should install requirements?

And then after another try it follows the prompt

Still Missing VAE keys messages

working again it seems

I downloaded that workflow but it seems to ask for UNET models in the unet folder, not the Comfy ones

https://github.com/comfyanonymous/ComfyUI/issues/2785#issuecomment-1954866357

doesn't detect the models in the directories that work with the GTM workflow

[swift jacinth]

Hi

[chilly knot]

cs10 you there?

https://tenor.com/view/cat-кошка-sad-грустно-cry-gif-23938046

[formal trout]

same thing happening here https://github.com/Stability-AI/StableCascade/issues/86

[brittle falcon]

I haven't used Stable Diffusion in over 7 months now and I'm getting this error when launching A1111. I need help understanding what the problem is, please.

[karmic crown]

you will probably not the like the answer. most likely the launcher is a problem. better follow guides for your GPU from this channel

[brittle falcon]

Ah, I see.

[formal trout]

I have a RX 6700 with 12 GB, it is ok this speed for cascade? It takes really long

[boreal bison]

Stuck here

after downloading the stable video xt 1.1 and restarting comfy

DId I break my comfy UI?

@ornate elk Hi 🙂

Anyone?

[ornate elk]

hi

[boreal bison]

I broke my comfyUI @ornate elk

[ornate elk]

i see :/ cant help much with that

[boreal bison]

It’s some kind of pytorch error

[glossy spindle]

I am reading a guide on "NVIDIA GPU Accelerated Computing on WSL 2" and I see a lot of references to Windows 11.. I'm guessing there's some new tech on Windows 11 that is required for sharing a GPU with WSL2?

[boreal bison]

What file should I delete to reset settings in comfy?

[worldly comet]

Why can it not download torch when opening the web ui file?

[prisma quartz]

where to get like this ?

[worldly comet]

yo dawg you think you can help me?

[formal trout]

MM I'm starting to believe that the generations that has nothing to do with the prompt are related to this ImpactWildcarProcessor node, it seems like sometimes it doesn't autopopulate or something,@light glacier EDIT: yes that was it, messing with the second field

Now I set it to "Populate", I don't know what it does

[fair oxide]

anyone know the fix for this?
FutureWarning: LoRACompatibleLinear is deprecated and will be removed in version 1.0.0. Use of LoRACompatibleLinear is deprecated. Please switch to PEFT backend by installing PEFT: pip install peft.
deprecate("LoRACompatibleLinear", "1.0.0", deprecation_message)

[storm spruce]

stupid question, is there a prebuilt ui to download? If there is not, everything else seems to be github!

[spiral lotus]

Hello SAI.

You closed my ticket with no follow up not a solution. Nothing.

I am a paying customer and still waiting to know what it actually mean because for now all I do is paying 20$ à month.
I have no special role on the forum, not a priority on support, not a paper telling me I have the commercial rights on the core models, not a customer space to log to...

I am a day 1 membership subscriber. And I mean it when I say literally DAY ONE: go check out on your received payments.

Could you provide some answers? I really like you people at SAI and support you with your development, but please, if you launch a customer portal please have some SAV 😉 (which is French for customer after sale support)

@eternal elbow

[fiery shard]

MM I'm starting to believe that the generations that has nothing to do with the prompt are related to this ImpactWildcarProcessor node, it seems like sometimes it doesn't autopopulate or something,@light glacier EDIT: yes that was it, messing with the second field

[eternal elbow]

Hi, sorry for any confusion.
At times, we have issues with our mod-support bot. Will get back to you after weekend.
It's important to note what you're trying to get answers in is in the Discord moderator support system. You can contact SAI directly via the actual official website: https://stability.ai/
Thanks! 😉

[light glacier]

It's because the "fixed" option is selected. Change it to "populate". Fixed causes it to not change the the prompt,but populate then chooses the random options that you enter, which you're not using there anyway.

[formal trout]

yes in populate the problem it is gone, it was driving me crazy, just making totally random prompts

like

Any interesting usage of that node? Also if you don't mind, what about the styles, I see there are 3 nodes with styles

[light glacier]

Yes, it's a powerful tool to create wildcard prompts and you can create text file lists to use with it. The styles should be obvious, just try them.

[dense obsidian]

trying to use temporalnet, but it's not getting me the key frames or running after I put in my video. I heard there was issues with it last year, wondering if anyone has any luck with it recently

it creates the folders that are supposed to be filled with the frames, but all are empty and it just gives me a copy of the video I put in

[formal trout]

but in what combination, it acts in certain stages or it combines right away?

[light glacier]

What? It's part of the prompt.

[light glacier]

You either create the text files or use the format {one|two|three|four} and it will randomly pick one of those.

Eg.
A photo of a {man|woman|dragon|majestic castle|futuristic android}

It's part of the Impact Pack of custom nodes, so you can read about it there. https://github.com/ltdrdata/ComfyUI-Impact-Pack/tree/Main

[unique bear]

Error occurred when executing CLIPLoader:

Error(s) in loading state_dict for CLIPTextModel:
size mismatch for text_model.embeddings.token_embedding.weight: copying a param with shape torch.Size([49409, 768]) from checkpoint, the shape in current model is torch.Size([49408, 768]).

File "/workspace/ComfyUI/execution.py", line 152, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "/workspace/ComfyUI/execution.py", line 82, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "/workspace/ComfyUI/execution.py", line 75, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "/workspace/ComfyUI/nodes.py", line 858, in load_clip
clip = comfy.sd.load_clip(ckpt_paths=[clip_path], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type)
File "/workspace/ComfyUI/comfy/sd.py", line 388, in load_clip
m, u = clip.load_sd(c)
File "/workspace/ComfyUI/comfy/sd.py", line 145, in load_sd
return self.cond_stage_model.load_sd(sd)
File "/workspace/ComfyUI/comfy/sd1_clip.py", line 518, in load_sd
return getattr(self, self.clip).load_sd(sd)
File "/workspace/ComfyUI/comfy/sd1_clip.py", line 197, in load_sd
return self.transformer.load_state_dict(sd, strict=False)
File "/workspace/ComfyUI/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 2153, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(

Got this error from CLIP

[formal trout]

I mean these three nodes styles, they combine right away with same strenght?

[light glacier]

Yes, they just add additional text to the prompt, but you don't see it. If you add a text node to the output you will see it tagged onto the end. You can combine styles.

[dense obsidian]

anyone have trouble with ebsynth utility? I can't get it to extract frames from my video

[valid gorge]

I'm installing Kohya_ss, how long does the process usually take, not sure what's going on, the screen has stayed the same for some time now.

[manic snow]

can you check if it's still actively using system resources?

[valid gorge]

Good point, I will do that!

Ooh, something happened on the screen hehe. I wonder why there's no info of the progress what so ever.

[dense obsidian]

I wonder what a successful ffmpeg -version command is suppose to look like.

[valid gorge]

Kohya_ss installation done. Should I install the cudnn files?

[dense obsidian]

also, anyone know what this means in my stable diffusion?

[manic snow]

what webui is that?

[dense obsidian]

not sure what that means, it just shows up when I run webui-user.bat

[manic snow]

can you send a full screenshot of the ui? that way i can probably identify it

[dense obsidian]

of course

[manic snow]

that's Automatic1111, as for the logos, i'm not sure what they mean. maybe you have a plugin that allows you to download models directly from civitai and the red sphere indicates that you don't have your account linked? idk, that's just some speculation. i have no idea what the second logo is either

[dense obsidian]

ohhh, that is true I don't have a account linked for civit.ai

looks a little worrisome, there's a green dot that blinks on that heart's upper right corner

[manic snow]

do you know what the heart logo is from? i don't recognise it

[dense obsidian]

I don't know, it just showed up while I was working on stable diffusion. There was a popup but I was typing something so I coudn't read to see what it meant

I've tried reloading it to no avail, whenever I start up, it's there

nope, that weird heart icon is still there

ohhh, I think it is an extension

[wind sun]

RuntimeError: min(): Expected reduction dim to be specified for input.numel() == 0. Specify the reduction dim with the 'dim' argument.

[dense obsidian]

can mp4 files not be able to have their frames extracted?

[ornate elk]

Hey, I got temporal net to work. Also mp4 files can be extractet

[dense obsidian]

ooo, I'll have to ask you how temporal net works and if you can guide me through that. But first I can't seem to get my mp4 to extract, here's what I get from the cmd

[ornate elk]

What tool are you using to extract it?

[drifting aspen]

Hello I often use sahastrakoti model. How to look for a similar model I would like to see the effects of other models. But similar to sahastrakoti

[dense obsidian]

ebsynth utility, I made sure it is updated

[torn lotus]

is loading models from pata feasible?

[ornate elk]

You can checkout Civitai.com for a lot of models. Dont know how shastrakoti looks like

[ornate elk]

Okay, I haven't used ebesynth

[drifting aspen]

Try manually extracting from 7z

[dense obsidian]

I can get the frames extracted from an outside tool, but from the tutorial I'm following on how to go from video to generated image it is necessary since it will piece together the tools without me having to load them in individually and hitting generate

[ornate elk]

I used batch Input in img2img

[dense obsidian]

ohhhhh, hmm maybe I can use that instead 😄

I gave up on ebsynth and tried temporalkit, seems to work well, got my frames extracted.

Only thing is that the video I'm trying to make is a dance, but when I'm at the img2img generation, I keep getting the generated image of just the face. Face is fine thankfully

[ornate elk]

this was my test yesterday with just controlnet (openpose + temportalnet) and afterwards frame interpolating

[dense obsidian]

that is perfect!

that's more or less exactly what I'm trying to do, though I would prefer to get the dress of the character in it instead of what the original dancer is doing

[ornate elk]

ah okay, i need to try, but its pretty hard to implement stuff that isnt already there

that would require higher denois, that could also change the background maybe

[dense obsidian]

oh I don't mind the background too much. As long as the face is there

[ornate elk]

oh i have to say i used Adetailer too

as the frames are only 512x768, to fix the face

[dense obsidian]

still however you were able to do that, I would love to learn it

[ornate elk]

ohhh i might found out how to give her a costume

will test and let you know

[dense obsidian]

feel free to friend and dm me, I'm determined to get this to work in the next 24 hours

[chilly knot]

how openpose work with video ,its for pose on images?

[ornate elk]

its to get the same pose as the input image

frame by frame

[chilly knot]

on video?

[ornate elk]

yes

[chilly knot]

remeber you give me bat arguments?

[ornate elk]

convert a video into frames, then you can use controlent with it

[chilly knot]

not work with error(

[ornate elk]

whats your error again?

[chilly knot]

NansException: A tensor with all NaNs was produced in Unet. Use --disable-nan-check commandline argument to disable this check

[ornate elk]

with the 1050 ?

[chilly knot]

if i use nan check =black screen

yes

[ornate elk]

dont use --diable-nan-check

can you show me all your txt2img settings when getting the error?

model, steps, resolution etc

[chilly knot]

512 512

default step and sampler

but with diffrent model its works

[ornate elk]

and with which model do you get the error?

[chilly knot]

URPM )

i try download it from civit

[ornate elk]

hmm how big is the model file?

[chilly knot]

3,97

[ornate elk]

hmm okay, should work

[chilly knot]

this say me i can change setting and set upcast cross attention

but it not help

maybe its error download

[ornate elk]

yea maybe, because the latest urpm legacy (v1.3) model is 1.99gb

on civitai

[chilly knot]

hm version 2 gb do same result

i try redownload

best my model epicphotogasm_lastUnicorn

very good results

i like how that do people photo

[ornate elk]

photon and realisticvision are good too

[chilly knot]

how its cant work if it on 1.5 base

its only train base model?

[ornate elk]

its the version of the own model, not sd version

[chilly knot]

oh

strange sd ignore promp and negative promt

example if you write a man holding a woman in his arms, she refuses, and the woman holds the man)

very often

90%

work after redownload)

[safe raptor]

What causes the images from Cascade to often look as if the VAE isn't working properly?

[valid gorge]

ok, finally got kohya to do something haha. I'm guessing this will take hours and hours.

ugh, says 11 hours lol

[cunning matrix]

anyone got the olive/directml setup working on windows+AMD navi3? The docs are a joke. I'm trying to use this guide and everything is broken, their fork of automatic1111 in particular
https://community.amd.com/t5/ai/updated-how-to-running-optimized-automatic1111-stable-diffusion/ba-p/630252

the args from the guide are not even recognized anymore in the recent versions of automatic1111-directml, is there an up-to-date guide somewhere?

I have the non-olive path (rocm I guess?) running fine in another installation on the same machine, it's just slow

[dense obsidian]

I also am using the non-olive path, I think?

I I know didn't get the olive tab either that was shown in the guide. But I can live without it, I think? as long as the version I have can do what I need

[cunning matrix]

they claim 10x speedup which sounds worth it

even 3x would be, honestly

[dense obsidian]

I suppose that does sound nice, but how fast does it take you now to render a 512 image?

[cunning matrix]

around 25 secs so not great. I think it's choosing suboptimal paths somewhere, I had to mess with cmdline args just so it launched at all on my older card

--lowvram and such

I just upgraded to RX 7800 so I'm trying to optimize my config

[dense obsidian]

ahh gotcha

if you figure out how to do it, let me know too 😄

[dense obsidian]

@ornate elk while you're testing out how the costume part works. Maybe I can practice on what you showed me earlier

[ornate elk]

Hey you have two options, dont use olive, the usage is not worth the speed.
Use directml for the best support and usage. Or use the experimental Zluda for best performance but with some bugs.

You find boath of the install guides in the Pinned Messages of this channel

I'm using an 7900xtx

Sdxl only works on the Directml version (for now)

[dense obsidian]

lol, I was trying to get the speed from sub 4 seconds down to 3 seconds or less

[ornate elk]

@cunning matrix also in the Directml guide you find the best COMMANDLINE_ARGS for AMD

So maybe you just need to switch those

[cunning matrix]

zluda actually works? that's surprising

I thought zluda's maintainer claimed it's largely a toy and couldn't run anything like sd or llama right now

[ornate elk]

Yes, going from 4it/s to 16 it/s

[cunning matrix]

wow nice

that's on windows?

[ornate elk]

Yes

[dense tapir]

How to make SD scan 2 folders for checkpoints (A1111)

[valid gorge]

Ideas, why I get errors?

[cunning matrix]

hmm yeah, on 7800 I'm gettting around 3it/s with directml

so good enough for realtime iteration, but I'll try zluda

[cunning matrix]

RuntimeError: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasGemmEx( handle, opa, opb, m, n, k, &falpha, a, CUDA_R_16F, lda, b, CUDA_R_16F, ldb, &fbeta, c, CUDA_R_16F, ldc, CUDA_R_32F, CUBLAS_GEMM_DEFAULT_TENSOR_OP)

looks like a version mismatch of some sort, but for a jury rigged solution like this it went pretty far

should there be cublas.dll in site-packages/torch/lib, originally? because there wasn't

there's cublas64_11.dll which I assume may be loaded directly instead

>>> import torch
>>> torch.__version__
'2.2.0+cu118'

[weary sky]

hello guys

just a dumb question

does this mean its downloading the interface?

the bottons are unavaileble

[undone locust]

hi , its possible to crate exact same picture with full prompt + step and seed, where i can pass this data to SD?

[ornate elk]

yes, did you followed all steps?
did you run the webui ones, then installed the correct torch and then replaced the files in torch/lib ?

[ornate elk]

it means its loading somethin, check the cmd for more info

[ornate elk]

yes, you can drag and drop an image into the PNG-Info Tab and hit Send to txt2img or directly into the txt2img prompt and then press the blue/white arrow

[weary sky]

thank guys, seem ,like a network error

its fixed now., thanks brother

[undone locust]

thank it's work 🙂 it's crazy

[cosmic furnace]

Hey, I managed to get it working on my RX 7800 XT with SD.next without needing to add any of those lines of code - I just set the enviroment variable set DISABLE_ADDMM_CUDA_LT=1 in the .bat file I use to launch it, as per the recommendations of the dev of ZLUDA, for pytorch compatibility: https://github.com/lshqqytiger/ZLUDA?tab=readme-ov-file#pytorch

hmgm o.0

[cosmic furnace]

and didnt help ... device cpu. nothing about zluda-torch used or my gpu model mentioning -___-

https://tenor.com/view/annoyed-disappointed-mad-upset-gif-26051038

[cosmic furnace]

Not a single word about rocm besides it = false . When try launching next.sd with zluda

no gpu detection no gpu model...

[ornate elk]

zluda is not rocm so thats ok

[ornate elk]

thats a problem xD

[cosmic furnace]

Is there a way to test if torch-zluda or rocm work on their own? 🤔

[ornate elk]

rocm doesnt work on windows rn with sd

only on linux

[cosmic furnace]

the thing is...

other ppl HAVE this rocm mentioning

when they launch sdnext with zluda

also they have their model name shown there

[ornate elk]

hmm

[cosmic furnace]

mine

his

u can see the difference?)

[ornate elk]

yes

do you have an intel igpu ??

[cosmic furnace]

yeah already asked about integrated

no doesnt seem like. process shows only 1 and its my 7900

mb i got fake 7900 ? o.0

but everything else is fine 🙂

[ornate elk]

does your taskmanager lists only 1 gpu?

or 2 ?

under perfromance tab

[cosmic furnace]

here? 1

[ornate elk]

here for example:

[cosmic furnace]

y only gpu 0

-_-

[ornate elk]

okay good

cant help you much with SD.next its so different with everything it implements xD

did you treid that guide?
https://github.com/vladmandic/automatic/wiki/ZLUDA

@cosmic furnace

[ornate elk]

and here you used ---use-cuda

i would try the new instructions

[cosmic furnace]

was adviced not to use -cuda but use -zluda

indeed

but it errored this. did i write smthing wrong? new instructions where? o.o

[ornate elk]

you didnt updated the webui it seems

it doesnt recognize the new argument

you need to switch to dev version

like seen in the screenshot

[cosmic furnace]

y he says run - git pull .

[ornate elk]

git switch dev

before doing git pull

[cosmic furnace]

u both seem to make same conclusions its a good sign for now :))

[ornate elk]

now it should work

[cosmic furnace]

y he says same)

oh my frteaking gosh it seems to work indeeed ;OO

or not hm

Nah still uses cpu for some reason 😑

[ornate elk]

maybe you need to delete the venv folder again and relaunch

[cosmic furnace]

just remove venv or reinstall it with some cmd?)

[ornate elk]

i would just remove it, it will get recreated

[cosmic furnace]

ive read some ppl instal zluda not on directml but just autmatic111, have u tried it?)

[ornate elk]

me not, but yea someone tried it

[cosmic furnace]

what? o.0

[ornate elk]

like it says

it doesnt found it in path

[cosmic furnace]

but i got it?

no?

😵‍💫

erm

[ornate elk]

yea looks okay

[cosmic furnace]

ive got a crippled system or smthing? 🤔

[ornate elk]

i dont know, you can check if you run a cmd as admin and then use
sfc /scannow
it can fix some windows errors

[cosmic furnace]

;O

....

G:\MODELLING\stable-diffusion-webui-1.6.0\Python3106\Scripts;G:\MODELLING\stable-diffusion-webui-1.6.0\Python3106;C:\Program Files (x86)\Common Files\Oracle\Java\javapath;C:\Windows\system32;C:\Windows;C:\Windows\System32\Wbem;C:\Windows\System32\WindowsPowerShell\v1.0;C:\Windows\System32\OpenSSH;C:\Program Files (x86)\NVIDIA Corporation\PhysX\Common;C:\Program Files\dotnet;C:\Program Files\Git\cmd;С:\ZLUDA;C:\Program Files\AMD\ROCm\5.7\bin;C:\Program Files\AMD\ROCm\5.7\bin;C:\Users\AMD\AppData\Local\Microsoft\WindowsApps;C:\Users\AMD.dotnet\tools

[ornate elk]

path problem

[cosmic furnace]

echo path like this

rly?

erm o.0

[ornate elk]

wait it shows zluda

[cosmic furnace]

y -_- he says seems correct 😦

but he doesnt find this >_>

crippled system? xD

[ornate elk]

maybe

win 10 or 11?

[cosmic furnace]

Well ive got python installed not on c if that matters. But didnt had any problems with directml

10

[ornate elk]

hmmm

im on 10 too, have you tried the sfc scannow?

[cosmic furnace]

y its runs now

needs time)

maybe shud try another zluda location? 😄 ive got problems some times with admin rights... when just disk drive folder. not C and rarely...but...

[ornate elk]

yea you could try an other folder, but not desktop, documents, downloads, onedrive

[tranquil token]

I cant seem to get Stable cascade tab to appear in Automatic 1111's UI, I followed the guides exactly. Anything I'm missing?

[ornate elk]

@cosmic furnace maybe it didnt recognized HIP SDK too ?

[cosmic furnace]

dunno?

how to check?)

[ornate elk]

where hipinfo.exe

in cmd

[ornate elk]

hmm looks good

[cosmic furnace]

nice nice

system had error & damaged files btw ^^

thx)

[ornate elk]

ah okay ^^ yea it can happen after updates

[dapper frigate]

Hey guys, I'm trying to upscale an image and I'm passing blob because I do not have file stored in the path , but I am getting error that it can not decode the form data ``` export const sdxlUpscale = async ({
imageUrl,
width,
}: {
imageUrl: string
width: number
}) => {
const req = await fetch(imageUrl)

console.log(imageUrl)

const blob = await req.blob()
console.log(blob)
const formData = new FormData()
formData.append("image", blob)
formData.append("width", 1024)
const response = await fetch(
${apiHost}/v1/generation/stable-diffusion-x4-latent-upscaler/image-to-image/upscale,
{
method: "POST",
headers: {
"Content-Type": "multipart/form-data",
Authorization: Bearer ${apiKey},
},
body: formData,
}
)

if (!response.ok) {
throw new Error(Non-200 response: ${await response.text()})
}

const data = await response.json()

return data
}

[cosmic furnace]

@ornate elk I knew it i shudnt be putting just straight root drive folder . Cuz always had problems with them. Mb numerous windows reinstall on same drive(although i did format c last time o.0) & few systems running is at fault.... anyways i have new error when launching now. Yours guide build... Gonna try next.sd >_>

ah wait it was smthing i edited there hm

y bringing back original initialize.py seems to do the case

forgot to move over based 1.5 model it seems to download it

[ornate elk]

Dont edit the initialize.py

My guide is for the shared_init.py

[cosmic furnace]

Whats thre difference between next sd & yours? UI ?

[ornate elk]

Both are forks of automatic1111

[cosmic furnace]

i didnt see karras 3m ++ experimental on next sd ... hmgm

[ornate elk]

SD.next is a more advanced version with a lot of changes and pre enabled extensions. It focuses on "working for everyone"

The Directml fork is the closest to the original automatic1111

[cosmic furnace]

Y my friend whos for some long time now on usual automatic said UI on next feels more plausible with except of few things)

U dont need 1st time to compile thing on yours? o.0

like 1st generation takes from 5 to 15 mins or smthing like that?)

[ornate elk]

Thats normal for every Zluda webui.
And it will do that every time you update the Zluda version

[cosmic furnace]

oh my gosh it finally works... ive a had feeling it shudnt be just just C folder why didnt tried straight away 😑

[ornate elk]

So does SD.next works or my guide?

[cosmic furnace]

Well from one side its good that zluda updated so often )

yur guide launched ) gonna try sd.next

[ornate elk]

Perfect 🙂

[cosmic furnace]

next also works 🙂

i can see 7900 in devices)

finally -___-

This is ancient problem of my pc , had afraid it might be the problem from the very start & dunno why ignored

[old parrot]

Can I use Stable Cascade with 6gb VRAM?

[ornate elk]

you should consider, reinstalling windows in the future ^^

[cosmic furnace]

y ive got new& fresh on ssd...but for sd using my old one on hdd >_>

didnt felt like reinstalling everything 😑

• not sure they say new ssd is very durable & 5 years warranty but... i have to use this virtual mem file cuz 32 gb RAM IS NOT ENOUGH . So afraid it might wear out too fast... shud i? hmgm.... Or just consider buying additional ram 🤔

or maybe sdxl will get optimized soon ? to not demand 48gb ram? 😏

[ornate elk]

Your running Windows on HDD?! ^^ please use your SSD haha

If you have 32gb ram its fine

A pagefile won't destroy your SSD

Even if its 32gb or only 16

[cosmic furnace]

Well 50/50... but yeah usually hdd lately cuz SD on hdd windows ^^

maybe sdxl turbo require too much ram. also an good option. if you go sdxl

[dense obsidian]

@ornate elk hello, was wondering how the progress went on your testing

[ornate elk]

hey, i tried experimenting with a third controlnet (reference)

that was the output

[dense obsidian]

OOo, please teach me 😄

[ornate elk]

in this test i used a Lora for Makima from chainsaw man

so i get a consistent style

it was "okay"

i tested a lot, and using three controlnet models takes more time for 158 frames (reference mode is heavy)

[cosmic furnace]

dude refering to have problems in sd.next there and have this python version 3.11.5 can it work with it? o.o

[dense obsidian]

hmmm, the frames I extracted from videos is roughly about 300-400, so it seems like its going to take awhile