🤝｜tech-support

[vocal burrow]

scam bot

[sonic pine]

yeah I wouldve clicked it if it sent me to another channel and not another discord

I keep getting errors like " File "C:\Users\User\AppData\Local\Programs\Python\Python313\Lib\gzip.py", line 359, in close
fileobj.write(self.compress.flush())
ValueError: I/O operation on closed file."

lastly, how do I uninstall extensions?

cause I think the "latent couples" extension is stuck on

[mystic basin]

i have an issue, confyui get stuck at 50% with flux, it says CLIP Text Encode (Prompt), i don't know why it get stuck, also it says clip missing even though i have it

i'm pretty sure it use my gpu and not igpu (ZLUDA)

[mystic basin]

from what i understand, it says clip is missing even though it's not, it says it use my gpu and it does, i checked in task manager and my gpu vram is being used but other than that, i don't understand anything else

[sonic pine]

yeah it detects ur gpu but its not using it

i mean

I dont think it is

Im guessing

[mystic basin]

task manager says the vram of my gpu is being used

but other than that my gpu stays at 0% usage

[sonic pine]

have you tried other models?

[mystic basin]

nope

[sonic pine]

try that out

[mystic basin]

like other type of model other than flux like pony or another flux model is fine?

[sonic pine]

try both

[mystic basin]

alright

[sonic pine]

maybe it's flux, maybe it's the specific model

[mystic basin]

i'll try

[sonic pine]

try to set the cuda device to 1 also

what gpu do you have?

[mystic basin]

an rx 6850mxt

[sonic pine]

oh idk about amd gpu's

Have you used comfyui before?

[mystic basin]

confy was working before but now it doesn't

[sonic pine]

gotcha

[mystic basin]

yeah

it doesn't work with another flux model, i just need to download another type of model to try

[sonic pine]

Arent CUDA cores only Nvidia?

or am i trippin

[mystic basin]

yeah but i'm using ZLUDA, it emulate cuda

that's why it's saying cuda

[sonic pine]

gotcha

[mystic basin]

currently downloading an illustrious model

to test if it's only with flux or not

i forgot i already have a pony model i never used, i don't need to download any model

[safe inlet]

how long did you wait, it looked like it was working

[mystic basin]

a few minutes, the first time i was waiting, it crashed

and i retried but with a different flux model

it seem to get stuck with a pony model too

my gpu vram is used but other than that i don't think my gpu is being used

the last time i used confy was 15 days ago

[misty marten]

my Controlnet is installed, but its tab does not appear in the webui

I tried using the regional prompt, but it's not working, is there a guide for this?

[misty marten]

I've tried generating hundreds of images, it doesn't work, watching videos and seeing guides on reddit, it doesn't work

[sonic pine]

any reason why I keep generating black squares in reforge?

I feel like it's because of vpred, but idk

It works on the old forge

whats this mean?

    Traceback (most recent call last):
      File "Z:\Other\stable-diffusion-webui-reForge\modules\scripts.py", line 533, in load_scripts
        script_module = script_loading.load_module(scriptfile.path)
      File "Z:\Other\stable-diffusion-webui-reForge\modules\script_loading.py", line 13, in load_module
        module_spec.loader.exec_module(module)
        ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^
      File "<frozen importlib._bootstrap_external>", line 1022, in exec_module
      File "<frozen importlib._bootstrap>", line 488, in _call_with_frames_removed
      File "Z:\Other\webui-forge\webui\extensions\adetailer\scripts\!adetailer.py", line 15, in <module>
        from rich import print  # noqa: A004  Shadowing built-in 'print'
        ^^^^^^^^^^^^^^^^^^^^^^
    ModuleNotFoundError: No module named 'rich'

---```

[sonic pine]

how do I install soft inpaint with reforge?

[gilded laurel]

Has anyone used n00mkrads sd gui? Please message me

[ornate elk]

Looks like adetailer has an issue.
Make sure the extensions are updated in the extensions tab

[ornate elk]

Can you show the whole txt2img site with your settings etc

[sonic pine]

yall know a reliable way to add text in an image without photoshop?

[misty marten]

and this settings in Regional Prompter

I unchecked the BREAK change chunks option, but nothing changed either

[ornate elk]

You have to use 2 breaks

Because you have base prompt active

So its
Base prompt BREAK
1 character BREAK
2 character

[misty marten]

i will test here

didn't work :/

I tried using Latent Couple, but it didn't work either

[ornate elk]

You have to prompt for 2girls in the top one

I'm not at the PC rn to share an example

[misty marten]

same result

At least the character that is being generated is cute hahaha

[vocal burrow]

Some examples from a long time ago.
(Using ADDCOM, ADDBASE to share parts of the prompt betweeb subprompts you don't have to, it's just an additional thing for you to research and learn about)

cf https://github.com/hako-mikan/sd-webui-regional-prompter?tab=readme-ov-file#usage

[misty marten]

I'm leaving work now, as soon as I get home I'll test

[misty marten]

It looks like it finally worked, thank you very much friend

[cerulean jewel]

hey im very new to this and looking for a little guidance. i can't seem to figure out why i keep generating similar images. for instance, i am using a model and i generated her poolside. i deleted and changed the prompt, and sd is still outputting her by the pool. i haven't encountered this. there is no language saying to put her there. is there some sort of cache i need to clear?

[ruby fjord]

I want to input text and have it output 4 images.

The endpoint is this. (stable diffusion 3.0 midium)
https://api.stability.ai/v2beta/stable-image/generate/sd3

However, there is no SAMPLE option to specify the number of images.　What should I do?

[formal ridge]

IIRC A1111 has some sort of caching issue if you're using that WebUI. I'd recommend ComfyUI over everything else even though it has a bit of a learning curve

[half island]

did latest comfyui cause workflows to suddenly be different than the workflow should have outputted? Even my own gens from 2 days ago now looks different using the exact same workflow

[cerulean jewel]

yeah, thats exactly what i'm using. that's a bummer. i'll look into comfy i suppose.

thank you

[lucid yoke]

SwarmUI might be a bit easier to get into while still having the benefits of Comfy

in the pinned messages theres a guide but the swarmUI install is as easy as it gets

[cerulean jewel]

thanks, ill check it out

[dense tapir]

Is it possible to have a setup in Forge where after your image is done in Txt2Img, that image automatically gets put into Img2Img and hits generate?

[ornate elk]

Hey make sure the seed is set to -1

So its always random

If your new, Auto1111 is totally fine to use. Dont start with comfyui

[cerulean jewel]

Yep, it's -1. I only messed with guidance and for about a week my settings were outputting great images. Now it seems to be stuck in a loop. Comfyui looks a little complicated for what I want it for since I have limited time to learn so I might try something else first.

[ornate elk]

Can you show your txt2img settings?

Its mostly a simple setting that might be wrong

[mystic basin]

yeah i haven't manage to fix this, it get stuck at the prompt area with flux and pony, i haven't tested other models but i'm pretty sure it happen with every models

it keep getting stuck

[mystic basin]

yeah, my ram usage instanly goes down and vram usage instantly get stuck at 6gb when arriving at the prompt area

my gpu usage become 0% and my igpu is certainly not the one being used for comfy

[scarlet igloo]

Is there anyone familiar with FluxGym, the software for training LoRA? I have some usage issues and need help.

[formal ridge]

Why FluxGym and not OneTrainer? I ask because I've never heard of FluxGym

[ornate elk]

What's your GPU and vram?

[mystic basin]

rx6850mxt, 12gb of vram on it

i use the ZLUDA version of comfy

[scarlet igloo]

It might be simple for me, and it supports training Flux LoRA models with 12GB of VRAM.

[ornate elk]

Make sure its updated by using git pull

[mystic basin]

can't the the comfy manager do that?

[ornate elk]

Pony models should work on your GPU

[mystic basin]

i've tried with pony, it did not work

but i can try again

[ornate elk]

Can you show the error you get and maybe the workflow too?

[mystic basin]

sure

the workflow

i can zoom in somewhere if needed

CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16
clip missing: ['text_projection.weight']
model weight dtype torch.float8_e4m3fn, manual cast: torch.bfloat16

i have the clip though

but even when trying with a pony model, it get stuck at the same place

[ornate elk]

Thats a flux workflow

It won't work with just changing to pony model

[mystic basin]

i know

i changed the workflow for pony

[ornate elk]

Can you show me the pony one

[ornate elk]

What does the cmd shows?

[ornate elk]

Looks normal but you only have 16gb of RAM

That could be the issue when it tries to load the clip model

You have to adjust your Windows Pagefile

[mystic basin]

it used to work the last time i used it 16 days ago

[mystic basin]

how do i do that?

[ornate elk]

https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Set it to
16000min and 24000max for the C drive.
And make sure its only enabled for C.
Restart the PC to apply the changes.

[mystic basin]

alright

i'll do that

thanks

[ornate elk]

Np let me know if that helped

[mystic basin]

0%| | 0/20 [00:00<?, ?it/s]

for the pony, it got further, it's stuck at the KSampler for a while a but the ram usage went up a little

i haven't tested for flux yet

it worked but it took longer than usual

i'm pretty sure what happened is that the last time i used it was 16 days ago so what happened it since then, it got updated and all + i had to send my laptop for repair for a cpu issue

so it probably acted as if it was my first time using comfy which is supposed to take more time for the first image generation

(yeah sorry for my english)

[ornate elk]

Ah that explains it. It mostly did compiled zluda again

[hard mountain]

I'm trying to do image-to-3d and it converts the image to a flat 3d object every time with both models...

[ruby yoke]

what are you using and what steps are you taking?

[hard mountain]

https://platform.stability.ai/docs/api-reference#tag/3D/paths/~1v2beta~13d~1stable-fast-3d/post just hit the curl with an image of a hand drawn ghost and it outputs a glb as a flat 3d mesh

(this is both for the fast 3d one and the new preview model)

[ruby yoke]

that's what i'd expect. that's a plane

you can extrude it

[hard mountain]

I mean how can I turn a drawing into a 3D glb? Isn't that the whole point of these models?

Can you feed it multiple images from front/back/side so it has some concept of depth?

[lunar kite]

Does anybody else have trouble with numpy on stable diffusion deforum ~ collab notebook?

[tribal anvil]

Is adetailer guaranteed to fix eyes?

[crystal nacelle]

I busy updating now , ofcourse HIP SDK Core is required, but do i need all the other optional installs for Stable D ZLUDA Auto1111? Namely the HIP SDK Runtime, Libraries, Ray Tracing and Visual Studio?

[tribal anvil]

Tried it and it didn't work so I hope I'm using it wrong

[lunar kite]

Is it still possible to use SD Deforum with Notebook Collab?

[fiery ravine]

very intrested to start learning about text to image and video/ animations. need help as to where to start and what to download. can anyone please help

[ornate elk]

Yes it should nearly always work

[ornate elk]

Yes you need all of that when Installing HIP SDK. But dont check the last thing where it says "Pro driver"

[crystal nacelle]

thank you

[tribal anvil]

I just have to click it right? That's what I did and it didn't work

[ornate elk]

Can you show the cmd log?

[tribal anvil]

How do I find that?

[ornate elk]

Just copy the full cmd text in here

[tribal anvil]

1 sec

Very bad lag

@ornate elk

Could it be because I only have 3gb of vram

[ornate elk]

move the forge webui out of the downloads folder into a new folder you created. not desktop, downloads, documents

[tribal anvil]

Can I move it to my SSD?

Aight thank you

[ornate elk]

then try again

also your using an big model (6gb) also with 2 loras on your 3gb gpu

that slows down a lot

[last tide]

Anyone knows a tutorial for creating your own loras?
Im currently failing to achieve this tutorial since it gives me char encoding error, appears to be a 🎉 but cant seem to understand where. Since its in a custom_node.

https://www.youtube.com/watch?v=hZCTiZi-EPw
Custom node in question: https://github.com/rgthree/rgthree-comfy

[ornate elk]

hey, use seperate tools for creating loras like OneTrainer or Khoya_ss

[lucid yoke]

All of that depends on what gpu you have

[tribal anvil]

Just slows it down or its also why adetailer doesn't work?

[ornate elk]

slows down. adetailer uses cpu so it could also fail if the pc is at max load

[tribal anvil]

I see

[last tide]

Oh, I'll look into this some more. Thank you!

[mystic basin]

wait for this free memory custom node, do i keep it like that or i need to separate the path, one path that goes to save image from the vae decode and the other going from vae to the free memory (image) node? I'm having issue where sometime after the flux image is generated, my pc shutdown due to ram issue, i've got all the other free memory node placed in my workflow but i'm not sure where to put this one.

i have a ram issue and using custom node to help is easier than modifying certain file to get rid of the issue i'm encountering, it's annoying to reopen my pc and reopen comfy completely each time my pc shutdown because of ram issue which does not always happen

[ornate elk]

Have you increased the windows pagefile?

[mystic basin]

yeah, i did like you told me yesterday

well i've only had this issue once since i added the free memory custom node, like about everywhere i could in my workflow but i still don't want it to happen again

so that's why i'm looking for a way so that never happen again

[ornate elk]

Use tiled VAE decode instead of VAE decode

That reduces vram usage alot

[mystic basin]

this?

or this?

[ornate elk]

both work, but test VAE Decode (Tiled) first

[mystic basin]

alright i will, thank you

[chrome rampart]

Hello all. Anybody know more info when controlnet options like open pose/reference might come to SD Forge with flux?

[lunar kite]

Is it still possible to use SD Deforum with Notebook Collab?

[knotty turret]

Illustrious-XL need vae?

[ornate elk]

nope but you can use one

[misty marten]

To use a positional LORA with Regional prompt, do I place the lora at the base, or at the characters' prompt? I want the lora to be applied to both characters

[lunar kite]

Anybody?

[median harness]

can someone help, i updated my webui stuff and now im getting this error

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

[zenith sail]

Did you select low vram / cpu mode?

[zenith sail]

AFAIR collab removed all high gpu notbooks you have to rent a gpu to use, it's like $9 a month for x min.

[median harness]

I'm using forge and it's been a while since I had to change settings, would you know where it is?

[zenith sail]

No idea it should be in your startup args.

https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1327

That explains what the arg is

[flat olive]

which ui is the best for pony models?

[lunar kite]

Ty for this response! Really helpful ~ how do I do this?

[lunar kite]

Does this relate to Google Collab?

[lucid yoke]

? But currently the best ui's (im my opinion) is comfyUI, swarmUI or Forge

Though forge has some limitations iirc

[flat olive]

would you recommend forge? i have a 1660 super

[lucid yoke]

Oof hmm with 6gb vram its gonna take a while for sure

Its been optimized a lot but i think you could give it a try but imo swarmUI is the easiest to install

[hoary gazelle]

Excuse me, is there a plugin that can queue tasks for the Forge version, such as agent - scheduler? But it seems that it can't be used in Forge(′；ω；`).

[formal ridge]

ComfyUI simply because of "Note" Nodes.

[sullen hawk]

I've heard that a new NFT coin called X Whales is launching. Is that true? Can someone tell me more about it? I've heard there's a lot of profit in it. Kindly guide me.

[crystal nacelle]

im busy installing SD Next now. got the HIP SDKs installed. im being told a new release of PIP is available, should i install it?

im also getting some warnings about version miss-matches, is this okay?

[lucid yoke]

Scam, hope this helps

[ornate elk]

Why sd next?
IMO its the worst to use xD

[formal ridge]

SD.Next was nothing but issues here too. I was able to generate a single image. After that it refused to ever work again.

[flat olive]

Whatas note nodes?

[ornate elk]

What's your GPU?
Pony models can be used in any webui but some have better optimisations for lower vram

[formal ridge]

Note nodes are basically little notepad txt files in a box. That way you don't have to remember all the weird stuff pony does, like rating_safe, rating_explicit, score_, etc.

[flat olive]

1660 super

[ornate elk]

Then you should go with Forge webui

Or ReForge

[flat olive]

You never mentioned ReForge, what is it?

[ornate elk]

Its a mix between Auto1111 and Forge with more frequent updates than Forge

[flat olive]

it is in the general installation guides?

or is it not a guide from yours

[ornate elk]

Its only linked there but no guide.
But you would only need to git clone it like Automatic1111

[flat olive]

okey will try that, ty

[ornate elk]

Np

[fair oxide]

Is rocm faster on rx 6600 or is zluda faster?

[ornate elk]

Linux with ROCm is faster

[fair oxide]

Oh

[ornate elk]

But zluda is easier to setup if your already on windows

[fair oxide]

Can i use it with wsl 2? Idk if my gpu is supported

[fair oxide]

True, im on windows and i use it

[ornate elk]

WSL2 should work but I think only the 7000 series are supported

[fair oxide]

Yeah could be cause there is no mention of 6000 series card being supported yet

I currently use comfy-ui zluda using stability matrix and its runs good

But is it true that, onnx and olive is technically faster than zluda?

[crystal nacelle]

zluda a1111 gave me an error so i figured id try the SD Next it recommends

i get this error with zluda after upgrading my HIP SDKs

[ornate elk]

Technically yes, but worse in compatibility and usability

[ornate elk]

I hope you dont use rocm5.7 with it and upgraded to 6.2 instead.
Its better for 6000 series to use the newer hip SDK and zluda versions

[ornate elk]

Delete the .zluda folder and relaunch

[crystal nacelle]

I tried deleting the ZLUDA folder in my C drive (that one that has rocm6 unzipped files) and i still get the same CMD error when i run the webui for A1111

yes i got the latest rocm aswell

[fair oxide]

i get an error using 6.2

[fair oxide]

It says cuda not detected

so i am still using 5.7.1

[fair oxide]

Is it faster?

[ornate elk]

Its more stable, could be faster, but not by much

[ornate elk]

No not that folder.
The .zluda folder is in the webui folder

Where the venv is

[spiral tinsel]

hello there, i want to ask how to assign specific gpu for SD run?

[ornate elk]

Hey which webui do you use?

[spiral tinsel]

Comfy UI

[ornate elk]

Okay there you need the launch arg --cuda-device 1 or 0 depending on the GPU

[random night]

i really wish reforge has zluda support

[ruby yoke]

wrong discord

[violet swift]

Is there a ROPE deepfake based repository that can work in bulk? That tool is incredible, but I have to do everything manually.

[ornate elk]

Checkout FaceFusion
https://github.com/facefusion/facefusion

[violet swift]

It's bad, I've used it for a long time, I'm looking for something based on rope

[ornate elk]

then this?:
https://github.com/Hillobar/Rope

or maybe Roop-unleashed

[analog cosmos]

which folder do i put workflow in? or is it extension?
(using automatic1111 SD)

[ornate elk]

Workflows are only made und usable for ComfyUI/SwarmUI

[analog cosmos]

ahh ok, i see. thx

[warped canyon]

I want to train a lora based on my own ai generated pictures. For this, should I use the original outputs (832x1216 / 896x1152 / 1024x1024, etc) or should I use the 2x upscaled versions of them? (i usually always upscale them using img2img 0.15 denoise with sd upscaler ultrasharp)

I think they say that kohyaa automatically downscaled images of higher resulotions to the normal 1024 resolutions. So I'm not even sure what resolution i should use

[analog cosmos]

720x720 is a good stable resolution where you get some quality but also not sacrifice too much of the speed. and then upscale it 2x. then you will produce very good pictures, for 16:9 and 21:9 i use a size that will be approximatly in the same area of res, 1280x720 divided by 1.3-1.5 and then upscale it to be at desired size (1080p or 1440p)

for training lora purpose it depends what your training for, gold detail ultra high detailed things, high resolution, for just a character or style you can use original.

i would much rather have the lora to be trained on big database to give accurate results instead of focusing on resolution.
if you got beefy enough pc then max out your capabilities, quality over quantity

[warped canyon]

the lora would be of the style of the images i mage with AI itself

[analog cosmos]

then resolution wont have much to say unless its a detailed style, the more details the higher resolution. if its cartoon then its not alot of detail and you dont need high resolution.
but dont train a lora on 15 images, make it 200-300 pictures atleast, <

gives user of lora more diversity of usage when having bigger database and it will be more correct in other scenarios than you using

also make sure you cover different scenarios in that style your training

[warped canyon]

Yeah I have like 400

[analog cosmos]

i have way too many loras that only work in 1 way cause they have been trained on 15 pictures of 1 position, its annoying

[analog cosmos]

Hunyuan video. Is that a standalone thing or extension or what? I have the ltx video and animatediff and have made txt2vid

But i want ti have img2vid. I dont need large videos, just i can loop them and make good abstract move in a good controlling way. I have hunyan models and everything but havent quite figured how this hangs together,

Very poor information on how to install or what folder its going in or whatever help there is from them. This is the only thing o havent figured out yet cause no information on how to actually run it. Img2vid

[tardy peak]

was anyone able to run rocm and web ui (any) for stable diffusion on rx 6600?

for some reason rocm uses cpu, I looked for a solution on Google but there are only 3-5 pages with the same problem, and the solution did not help.

(Linux)

[tame yoke]

does this mean i have conflicts, or is it just a list of nodes it can conflict with

[vocal comet]

I am using version v0.3.14-34 of comfy ui and the list that comes up when you click on the checkpoint loader or any other else like loras and any of the nodes where there are choices has stopped working. i have tried different browsers and different computers is there a setting that i need to change?

[formal ridge]

I bet you're using rgthree...

The new ComfyUI (with the sidebar and topbar) breaks a lot of crap right now

[vocal comet]

yes i am using rgthree should i uninstall it?

[formal ridge]

I had to get rid of mine for the time being due to the problem you're having.

Everything went back to normal after that.

[vocal comet]

thanks that worked for me too👍

[upbeat atlas]

Hi, so I deleted venv and I'm still getting this error after I run the webui. I'm on windows with AMD GPU.

[ornate elk]

hey, show the full cmd log pls

[fickle zealot]

hi i might be stupid but i dont understand this
if i had webui folder on disk s i should put this into this or where

S:\SD-Zluda\stable-diffusion-webui-amdgpu

and where i should put this

[ornate elk]

hey, it doesnt matter where the webui is located as long as its a new created folder

same goes for zluda

[ornate elk]

these are the wrong files

[ornate elk]

you only need this

[fickle zealot]

yep i done that

also

[ornate elk]

then you can proceed with the next steps

[fickle zealot]

if i done all of that

what should i next

i have rx5700xt

[ornate elk]

have you restarted the PC?

then follow the guide

[fickle zealot]

yes

[ornate elk]

then your done

[upbeat atlas]

here's the full cmd log

[ornate elk]

did you downloaed the new zluda files?

[upbeat atlas]

no i didnt download anything just the autoupdate for webui

[ornate elk]

oh your still on the old rocm version

whats your gpu again?

[upbeat atlas]

amd radeon 6900 xt on windows

[ornate elk]

try delete the .zluda folder in the stable-diffusion-webui folder

then relaunch

[upbeat atlas]

it was working fine but a week or two ago it did an autp update for webui and it stopped working

[upbeat atlas]

okk

[ornate elk]

okay try to delete the venv folder and relaunch

[upbeat atlas]

full log after deleting venv

[ornate elk]

relaunch

it will go further

[upbeat atlas]

you mean rerun the webui?

this is what i get after i rerun the webui

[ornate elk]

ok you have to upgrade your rocm it seems

your on old 5.7 and the new one is 6.2

you have to uninstall all of HIP SDK in the control panel, then install hip sdk 6.2.

[upbeat atlas]

do you have a link for it? i can't find the 6.2 windows version only linux

[ornate elk]

first uninstall all of 5.7, here is the link to 6.2

https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html

[upbeat atlas]

ok thanks

[upbeat atlas]

do i uninstall ALL of them?

[ornate elk]

yes

[upbeat atlas]

ok so I deleted the old rocm, installed the 6.2 version, added it to path, restarted my pc, deleted old zluda files in drive c, put in the new zluda files and did the copy paste rename thing, deleted venv and .zluda in the stable diffusion and ran the webui I still get this error:

[ornate elk]

damn

okay, uninstall (in the control panel), conda, anaconda and every python version you see.
Then install python 3.10.11 64bit from here:
https://www.python.org/ftp/python/3.10.11/python-3.10.11-amd64.exe
Then delete the venv folder and relaunch the webui-user.bat

[upbeat atlas]

ok

[ornate elk]

also show me the extensions of the extensions folder

[ornate elk]

looks okay

[crystal nacelle]

greetings again. im still having the same problem as before: i have updated my HIP SDKs and downloaded updated ROCm Zluda files :(. Can someone help please

i tried deleting the ZLUDA folder aswell, didnt make a differnece

[ornate elk]

Hey, delete the .zluda folder and not the C:\zluda folder

Then relaunch

[crystal nacelle]

what is the .zluda folder?

[ornate elk]

A folder like venv

[crystal nacelle]

is it the one that has my webui and loras?

[ornate elk]

No

[crystal nacelle]

ah i found it!

this one jea?

[ornate elk]

Yes

[crystal nacelle]

also, do i need pytorch? i think i uninstalled it along with HIPSDK but never reinstalled it

[ornate elk]

You cant uninstall pytorch it will get reinstalled when launching

[crystal nacelle]

same issue with deleted .zluda

[ornate elk]

No its a different error now

[crystal nacelle]

cublas64 why you....

[ornate elk]

Did you restored the C:\Zluda folder ?

Because you told me yesterday you deleted it...

[crystal nacelle]

yes i did restore it today. i see it doesnt have cublas64 though

i got these contents from ZLUDA nightly rocm64.zip

[ornate elk]

Thats the issue, dont use nightly (if you dont know how)

Use the normal one zluda-rocm6

[crystal nacelle]

the normal one doesnt have cublas 64 either so im getting the same error

[ornate elk]

No you shouldn't

[crystal nacelle]

i got the 13MB file

rocm6

[ornate elk]

Have you replaced that in the C zluda folder?

Then delete the .zluda folder in the webui again and relaunch

[crystal nacelle]

jea i have. ill try again one more time. this is so weird, coz the .zluda folder clearly has the cublas64.dll

[ornate elk]

What's in your webui-user.bat?

Did you by any chance downloaded the hip-sdk extension zip from the zluda github?

[ornate elk]

Looks normal

[crystal nacelle]

yes i used the link you pinned at the top, followed A1111 ZLUDA and then got the HIP DSK 6.2.4 link, the top one for win 10&11 6.2.4

[ornate elk]

Yea thats okay

I just meant the additional file on the zluda repo

That shouldn't be used

[crystal nacelle]

i dont remember using an additional file. i see the RX6800 is technically stronger than my RX7600 according to benchmarks. Should i try a different rocm version? altho this wasnt a problem before

[ornate elk]

Also what error do you get now after deleting the .zluda folder again?

[ornate elk]

Nope, yours is newer so thats fine

[crystal nacelle]

same error

i updated python aswell, maybe i need to change the path enviornment variables again?

[ornate elk]

Delete the venv folder and relaunch

That should fix it

The error searches cublas inside the venv so if it gets rebuild it could work again.

The nightly zluda broke the venv

[crystal nacelle]

deleting now. will relaunch aswell. thank you.

[crystal nacelle]

nightly! why you...

[ornate elk]

NP, let me know if it worked.
Also if you get a numpy error you just have to relaunch again.

I'm off now its very late. Gn

[crystal nacelle]

thank you, goodbye for now

[upbeat atlas]

so it did a bunch of ddownloads and a lot of compilation and now it is fixed, thanks.

[tiny gorge]

anyone got an idea why my sd is doing this

[lucid yoke]

I can't help but notice time taken 5min

[tiny gorge]

im new to sd so im not even sure if thats bad or good 😭

im using a 6700xt

[lucid yoke]

Also resolution is uhh too smal

[formal ridge]

It's very bad. Are you using A1111? It's the slowest of them all, except Easy Diffusion

[lucid yoke]

I recommend using zluda & forge (guide by cs1o in pinned messages)

And try using 1024x1024 images or similar aspect ratioa

[tiny gorge]

i just followed this https://www.youtube.com/watch?v=YazUwPNsdzE

[lucid yoke]

Id still recommend following cs1os guide

Youtube tends to be outdated

[tiny gorge]

recommend me reinstalling everything?

[lucid yoke]

Hmm maybe the venv folder

But a1111 is very outdated if your using that Webui

Cant see from the screenshot

[tiny gorge]

i feel like i did something really wrong xd

[lucid yoke]

Since its similar to forge

[lucid yoke]

Nah it's common amd issues

What's the folder called that its in

[tiny gorge]

stable-diffusion-webui-amdgpu-master

i think i should redo it following the updated guide thats pinned rather than a youtube video

[lucid yoke]

Ah yeah thats a1111 isn't it. I honestly recommend using the forge guide

Forge is also optimized for sdxl better

[tiny gorge]

whats forge

[lucid yoke]

It's a different UI

[lucid yoke]

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-install-guides

[tiny gorge]

thank you

[lucid yoke]

Personally i use swarm but forge is beginner friendly

People tend to get confused with swarmUI if it's their first ui

[tiny gorge]

hmm i think something went boom again

i followed the tutorial/guide exactly though

[lucid yoke]

Id wait for cs1o to wake up ngl

Hes the amd expert

[ornate elk]

Can you show me the full cmd log?

[elder mauve]

@ornate elk

[warm crane]

hellp

can anyone help me

[ornate elk]

Looks normal

[ornate elk]

Depends on the issue

[elder mauve]

👀 what can be the issue then

[warm crane]

its my first time so i want to use stable diffusion and models from civtal but my pc is not that strong so i used collab but i only worked like one basic notebook i made by watching a youtube video, so the quality of images is not good and i cant add lora so how do i get better with it

and find notebooks

or vids that r lastest that can help me

[ornate elk]

whats your GPU in your PC ?

[warm crane]

AMD redeon(TM) RX6550M 8gb gone upgrade to 16 up or more in summer

[ornate elk]

ah okay

yea with 4gb vram and 8gb ram its a bit hard

[warm crane]

then what about running it on collab

[maiden frost]

Guys I'm kinda new and would like some help with consistent faces.

Basically I generated a bunch of fantastical characters (e.g. mage, hunter etc.) and I want to be able to use one face to show him in every class— i.e. show what this guy would look like if you chose him as different classes in a game. This is for concept art for a personal project.

Can anybody point me in the right direction?

[fickle zealot]

anybody can help?

[ornate elk]

Try to use an other model. Also make sure the webui is whitelisted in any browser adblocker

[ornate elk]

Do you have anything open in the background? Like Wallpaper engine or multiple Browser tabs?

[random night]

does anyone happen to have your automatic 1111 amd can't detect zluda after the update?

Already up to date.

C:\AI\stable-diffusion-webui-amdgpu>webui-user.bat --opt-split-attention --force-enable-xformers --lowvram
venv "C:\AI\stable-diffusion-webui-amdgpu\venv\Scripts\Python.exe"
WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next.
Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr  5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]
Version: v1.10.1-amd-27-g1c3d2aae
Commit hash: 1c3d2aae8cdedbaea43c7ae333df42b471a6bba1
ROCm: agents=['gfx1103']
ROCm: version=6.2, using agent gfx1103
ZLUDA support: experimental
Failed to load ZLUDA: function 'zluda_get_nightly_flag' not found
Using CPU-only torch
Installing requirements
Collecting protobuf<=4.9999,>=4.25.3
  Using cached protobuf-4.25.6-cp310-abi3-win_amd64.whl.metadata (541 bytes)
Using cached protobuf-4.25.6-cp310-abi3-win_amd64.whl (413 kB)
Installing collected packages: protobuf
  Attempting uninstall: protobuf
    Found existing installation: protobuf 3.20.2
    Uninstalling protobuf-3.20.2:
      Successfully uninstalled protobuf-3.20.2
Successfully installed protobuf-4.25.6
Installing sd-webui-controlnet requirement: changing opencv-python version from 4.11.0.86 to 4.8.0```

like this

oddly enough, when i use forge, it's run pretty much normal, while only automatic1111 has that problems

[ornate elk]

Delete the venv and the .zluda folder and relaunch

[random night]

well, good start, it can detect zluda now

[tiny gorge]

sure let me replicate and i will send the cmd output, here is before generating and normal logs:

here we go

[ornate elk]

can you share the full cmd log after you generated the image?

[tiny gorge]

Using online LoRAs in FP16: False
Exception in thread MemMon:
Traceback (most recent call last):
  File "C:\Users\kyu\AppData\Local\Programs\Python\Python310\lib\threading.py", line 1016, in _bootstrap_inner
    self.run()
  File "E:\SD-Zluda\stable-diffusion-webui-amdgpu-forge\modules\memmon.py", line 43, in run
Loading Model: {'checkpoint_info': {'filename': 'E:\\SD-Zluda\\stable-diffusion-webui-amdgpu-forge\\models\\Stable-diffusion\\animagineXL40_v4Opt.safetensors', 'hash': '06de8aee'}, 'additional_modules': [], 'unet_storage_dtype': None}
[Unload] Trying to free all memory for cuda:0 with 0 models keep loaded ... Done.
    torch.cuda.reset_peak_memory_stats()
  File "E:\SD-Zluda\stable-diffusion-webui-amdgpu-forge\venv\lib\site-packages\torch\cuda\memory.py", line 370, in reset_peak_memory_stats
    return torch._C._cuda_resetPeakMemoryStats(device)
RuntimeError: invalid argument to reset_peak_memory_stats
StateDict Keys: {'unet': 1680, 'vae': 248, 'text_encoder': 196, 'text_encoder_2': 518, 'ignore': 0}
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
K-Model Created: {'storage_dtype': torch.float16, 'computation_dtype': torch.float16}
Model loaded in 46.3s (unload existing model: 0.2s, forge model load: 46.1s).
[Unload] Trying to free 3051.58 MB for cuda:0 with 0 models keep loaded ... Done.
[Memory Management] Target: JointTextEncoder, Free GPU: 11900.29 MB, Model Require: 1559.68 MB, Previously Loaded: 0.00 MB, Inference Require: 1024.00 MB, Remaining: 9316.61 MB, All loaded to GPU.
Moving model(s) has taken 17.17 seconds
[Unload] Trying to free 1024.00 MB for cuda:0 with 1 models keep loaded ... Current free memory is 10137.31 MB ... Done.
[Unload] Trying to free 7656.40 MB for cuda:0 with 0 models keep loaded ... Current free memory is 10138.46 MB ... Done.
[Memory Management] Target: KModel, Free GPU: 10138.46 MB, Model Require: 4897.05 MB, Previously Loaded: 0.00 MB, Inference Require: 1024.00 MB, Remaining: 4217.41 MB, All loaded to GPU.
Moving model(s) has taken 51.97 seconds
  0%|                                                                         | 0/20 [00:00<?, ?it/s]Compilation is in progress. Please wait...
 60%|██████████████████████████████████████▍                         | 12/20 [01:17<00:50,  6.31s/it]E:\SD-Zluda\stable-diffusion-webui-amdgpu-forge\modules\sd_samplers_common.py:75: RuntimeWarning: invalid value encountered in cast
  x_sample = x_sample.astype(np.uint8)
100%|████████████████████████████████████████████████████████████████| 20/20 [02:06<00:00,  6.31s/it]
[Unload] Trying to free 4495.36 MB for cuda:0 with 0 models keep loaded ... Current free memory is 5064.89 MB ... Done.
[Memory Management] Target: IntegratedAutoencoderKL, Free GPU: 5064.89 MB, Model Require: 159.56 MB, Previously Loaded: 0.00 MB, Inference Require: 1024.00 MB, Remaining: 3881.33 MB, All loaded to GPU.
Moving model(s) has taken 0.05 seconds
Total progress: 100%|████████████████████████████████████████████████| 20/20 [02:01<00:00,  6.09s/it]
Total progress: 100%|████████████████████████████████████████████████| 20/20 [02:01<00:00,  6.07s/it]

this is whatever was logged after i pressed generate

[ornate elk]

hmm make sure the webui is whitelisted in any browser adblocker

and then try a different model

[tiny gorge]

alright i was using opera gx integrated adblocker maybe that was it let me try

[ornate elk]

make sure operas vpn is disabled too

[tiny gorge]

dont use it, but turned everything off and its stuck on this now

Using online LoRAs in FP16: False
Loading Model: {'checkpoint_info': {'filename': 'E:\\SD-Zluda\\stable-diffusion-webui-amdgpu-forge\\models\\Stable-diffusion\\noobaiXLNAIXL_vPred10Version.safetensors', 'hash': '25dc06a8'}, 'additional_modules': [], 'unet_storage_dtype': None}
[Unload] Trying to free all memory for cuda:0 with 0 models keep loaded ... Current free memory is 10010.98 MB ... Unload model JointTextEncoder Current free memory is 11770.84 MB ... Unload model KModel Current free memory is 16762.51 MB ... Unload model IntegratedAutoencoderKL Done.
StateDict Keys: {'unet': 1680, 'vae': 248, 'text_encoder': 196, 'text_encoder_2': 518, 'ignore': 0}

not moving since a couple min

[ornate elk]

how much RAM do you have?

[tiny gorge]

32GB

[ornate elk]

is the generate button grayed out?

[tiny gorge]

its on interrupt/skip now it started loading finally

[tiny gorge]

heres the output:

Using online LoRAs in FP16: False
Loading Model: {'checkpoint_info': {'filename': 'E:\\SD-Zluda\\stable-diffusion-webui-amdgpu-forge\\models\\Stable-diffusion\\noobaiXLNAIXL_vPred10Version.safetensors', 'hash': '25dc06a8'}, 'additional_modules': [], 'unet_storage_dtype': None}
[Unload] Trying to free all memory for cuda:0 with 0 models keep loaded ... Current free memory is 10010.98 MB ... Unload model JointTextEncoder Current free memory is 11770.84 MB ... Unload model KModel Current free memory is 16762.51 MB ... Unload model IntegratedAutoencoderKL Done.
StateDict Keys: {'unet': 1680, 'vae': 248, 'text_encoder': 196, 'text_encoder_2': 518, 'ignore': 0}
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
K-Model Created: {'storage_dtype': torch.float16, 'computation_dtype': torch.float16}
Model loaded in 373.8s (unload existing model: 3.2s, forge model load: 370.6s).
[Unload] Trying to free 3051.58 MB for cuda:0 with 0 models keep loaded ... Done.
[Memory Management] Target: JointTextEncoder, Free GPU: 11928.81 MB, Model Require: 1559.68 MB, Previously Loaded: 0.00 MB, Inference Require: 1024.00 MB, Remaining: 9345.14 MB, All loaded to GPU.
Moving model(s) has taken 0.25 seconds
[Unload] Trying to free 1024.00 MB for cuda:0 with 1 models keep loaded ... Current free memory is 10163.77 MB ... Done.
[Unload] Trying to free 2881.78 MB for cuda:0 with 0 models keep loaded ... Current free memory is 10162.91 MB ... Done.
[Memory Management] Target: KModel, Free GPU: 10162.91 MB, Model Require: 0.00 MB, Previously Loaded: 4897.05 MB, Inference Require: 1024.00 MB, Remaining: 9138.91 MB, All loaded to GPU.
Moving model(s) has taken 0.02 seconds
100%|████████████████████████████████████████████████████████████████| 20/20 [02:18<00:00,  6.94s/it]
[Unload] Trying to free 4495.36 MB for cuda:0 with 0 models keep loaded ... Current free memory is 10162.41 MB ... Done.
[Memory Management] Target: IntegratedAutoencoderKL, Free GPU: 10162.41 MB, Model Require: 159.56 MB, Previously Loaded: 0.00 MB, Inference Require: 1024.00 MB, Remaining: 8978.86 MB, All loaded to GPU.
Moving model(s) has taken 0.05 seconds
Total progress: 100%|████████████████████████████████████████████████| 20/20 [02:14<00:00,  6.71s/it]
Total progress: 100%|████████████████████████████████████████████████| 20/20 [02:14<00:00,  6.80s/it]```

[ornate elk]

can you try a different model and maybe a different browser?

[tiny gorge]

i have 3 models installed which one would you like me to use

and yeah i can try chrome

[ornate elk]

sdxl base

[tiny gorge]

Using online LoRAs in FP16: False
Loading Model: {'checkpoint_info': {'filename': 'E:\\SD-Zluda\\stable-diffusion-webui-amdgpu-forge\\models\\Stable-diffusion\\sd_xl_base_1.0.safetensors', 'hash': 'be9edd61'}, 'additional_modules': [], 'unet_storage_dtype': None}
[Unload] Trying to free all memory for cuda:0 with 0 models keep loaded ... Done.
StateDict Keys: {'unet': 1680, 'vae': 248, 'text_encoder': 197, 'text_encoder_2': 518, 'ignore': 0}
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
K-Model Created: {'storage_dtype': torch.float16, 'computation_dtype': torch.float16}
Calculating sha256 for E:\SD-Zluda\stable-diffusion-webui-amdgpu-forge\models\Stable-diffusion\sd_xl_base_1.0.safetensors: 31e35c80fc4829d14f90153f4c74cd59c90b779f6afe05a74cd6120b893f7e5b
Model loaded in 103.7s (unload existing model: 0.2s, forge model load: 84.3s, calculate hash: 19.2s).
[Unload] Trying to free 3051.58 MB for cuda:0 with 0 models keep loaded ... Done.
[Memory Management] Target: JointTextEncoder, Free GPU: 11928.81 MB, Model Require: 1559.68 MB, Previously Loaded: 0.00 MB, Inference Require: 1024.00 MB, Remaining: 9345.14 MB, All loaded to GPU.
Moving model(s) has taken 0.99 seconds
[Unload] Trying to free 1024.00 MB for cuda:0 with 1 models keep loaded ... Current free memory is 10163.77 MB ... Done.
[Unload] Trying to free 2881.78 MB for cuda:0 with 0 models keep loaded ... Current free memory is 10162.91 MB ... Done.
[Memory Management] Target: KModel, Free GPU: 10162.91 MB, Model Require: 0.00 MB, Previously Loaded: 4897.05 MB, Inference Require: 1024.00 MB, Remaining: 9138.91 MB, All loaded to GPU.
Moving model(s) has taken 0.02 seconds
100%|████████████████████████████████████████████████████████████████| 20/20 [03:19<00:00,  9.96s/it]
[Unload] Trying to free 4495.36 MB for cuda:0 with 0 models keep loaded ... Current free memory is 10162.41 MB ... Done.
[Memory Management] Target: IntegratedAutoencoderKL, Free GPU: 10162.41 MB, Model Require: 159.56 MB, Previously Loaded: 0.00 MB, Inference Require: 1024.00 MB, Remaining: 8978.85 MB, All loaded to GPU.
Moving model(s) has taken 0.05 seconds
Total progress: 100%|████████████████████████████████████████████████| 20/20 [03:10<00:00,  9.52s/it]
Total progress: 100%|████████████████████████████████████████████████| 20/20 [03:10<00:00, 10.28s/it]```

[ornate elk]

okay something is wrong

5 minutes is bad too, it should take seconds

[tiny gorge]

yeah ive heard 😭

[ornate elk]

so your zluda setup is maybe not correctly setup

[tiny gorge]

i mean i followed the guide, everything in relation to zluda should be right have it in path etc

[ornate elk]

can you show me the path settings?

and your zluda folder with the files in it

[ornate elk]

looks okay

[tiny gorge]

seems like its fine

heres the ROCm if that matters

[ornate elk]

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

have you also restarted the PC ?

[tiny gorge]

yeah

its just reinstalling all the modules etc

[ornate elk]

if it then still doesnt work, you should try setup auto1111, its easier to troubleshoot for me where the error is

[tiny gorge]

took ages to install the modules it just finished, and launched the webui if it doesnt work ill try auto1111

i wouldnt need to reinstall ROCm and zluda etc right?

[ornate elk]

yep, you only need to setup auto and skip the zluda parts

[tiny gorge]

i think my pc is still having a stroke is this normal after deleting venv and .zluda?

[ornate elk]

that looks perfect

its compiling zluda

so after thats done your ready

it takes a bit

[tiny gorge]

yippie?

24min i almost fell asleep (i didnt)

can i use any other model now again or?

[ornate elk]

yep you can now try the others

[tiny gorge]

this looks a bit scuffed idk if thats cos of the bad prompt etc

[ornate elk]

its because sdxl base model needs the sdxl vae

for color correction

you can download it here:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

[tiny gorge]

for the other models i got from civitai do i need the vae too?

[ornate elk]

some models need one yes

it depends on the model version

with the vae i linked you can use any sdxl / pony / illustrious model

[tiny gorge]

theres two vae files from the link does it matter which?

[ornate elk]

take this:

[tiny gorge]

and will it always take this long henever i use a new model?

[ornate elk]

nope

the compiling will only appear when torch or your gpu driver gets an update

or you upgrade zludas version

[tiny gorge]

i switched models to animagine and it took like a minute for it to just move the model rendering the image now says eta 13min

[ornate elk]

what does the cmd shows?

[tiny gorge]

it is taking shape though just taking a while

[ornate elk]

for the next image try a resolution of 768x1024

[tiny gorge]

ehm

lmao

also took 5min

[ornate elk]

do you have anything running in the background like wallpaper engine or multiple browser tabs?

games etc

[tiny gorge]

i do have wallpaper engine running as well as multiple tabs

this did a lot better

i mean i do have 3 displays plugged into my gpu...

[ornate elk]

close and exit wallpaper engine (not pause), it slows down massivly

also close browser tabs

[tiny gorge]

okay

[ornate elk]

gen time should be under 1 minute

if everything else is closed

[tiny gorge]

left 2 tabs open the webui and civitai for an image reference, closed wallpaper engine and rainmeter took 2min 43seconds

[ornate elk]

close civitai too, its crazy resource intensive xD

worst optimised website

[tiny gorge]

closed it took 2 min. 31.4 sec.

[ornate elk]

ah okay

[tiny gorge]

is that normal for my specs? 6700xt with 32gb ram

[ornate elk]

it should take less for a 768x1024

normaly

but you said you have 3 monitors plugged in too

[tiny gorge]

yeah 1 ultrawide 1 1080p and one tiny pc monitor

[ornate elk]

also check your task manager and look at the vram gpu usage which process takes a lot

[tiny gorge]

for some reason i dont have my gpu here lol

[ornate elk]

right click on cpu or mem

then select gpu

[tiny gorge]

yeah its mainly just opera when i do have it open

thats about it

rest is less than 1%

[ornate elk]

hmm

[ornate elk]

you have two browser open

opera with 21 tasks lol

[tiny gorge]

yeah that was after generation, also is there a reason my generations have way harsher outlines than the people i copied the prompts from civitai?

[left gorge]

noob question what is the reason for my lora not being visible in stable diffusion gui? I installed the most popular lora for logo creation yet I cannot choose it

[ornate elk]

different settings

maybe they used an upscaler or an other resolution

[ornate elk]

mostly its not compatible with the selected model, then it wont show up

[left gorge]

thanks I now see I need SD XL checkpoint. I should put in models folder after downloading it right?

[ornate elk]

yep

best is to make subfolders for models and loras in their folders

like models/lora/sdxl
models/stable-diffusion/sdxl

[brave frigate]

can anyone help me with my ai? for some reason it won't start generating... Its just stuck on this. I have AMD

these are the same setting i have used a few days before for this image https://civitai.com/images/64039896

[ornate elk]

Your rocm hip SDK version is to old

Uninstall everything from HIP SDK in the control panel.
Then install HIP SDK 6.2.4

Then reboot the PC.
Then delete the venv and the .zluda folder in the stable-diffusion-webui folder.

Then remove --medvram from the webui-user.bat
And then relaunch it.

Also move your Lycoris models to the lora folder and remove that old extension. Its outdated

[brave frigate]

can i also just remove lycoris, i never use it tbh

[ornate elk]

Sure

They work without a extension now

[brave frigate]

i see

[tiny gorge]

is video generation avaliable for amd?

[mystic basin]

i'm pretty sure it is, comfyui support video generation

and it support amd

but i'm not sure though if comfy only support video generation for nvidia or if it also work with amd

and as for the other ui, i don't know

[tiny gorge]

huh im pretty sure i saw comfy only works with nvidia

[mystic basin]

you mean the ui or the video generation?

[tiny gorge]

ui

[mystic basin]

it work with amd

i have an amd graphic card and it work

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides (you can find a guide to download for amd here)

by the way the zluda version use less vram, less ram and have better performance than the direcmlt one

[ornate elk]

Yes in comfyui I used ltx video

[tiny gorge]

if i want to install comfyui do i just need to install that alone without rocm or zluda?

[ornate elk]

You dont need to install hip SDK again.
You only need to setup Comfyui and then you have to copy and rename some files in the zluda folder to replace them in comfyui

The guide covers all steps

But with a 6700xt you shouldn't focus on video generation

I will test the new ltx video models tomorrow, then I see how much vram they need

But normaly Video gen is very resource intensive

[formal ridge]

I can use Wan2.1 on an 8GB card. It's not that intensive.

[timber oak]

hello

i just updated my comfy

i got this error

i cant start it up anymore

[tiny gorge]

how do i use a reference image to generate images?

[mystic basin]

in comfy?

[tiny gorge]

in forge

havent installed comfy yet will do soon

[mystic basin]

i can't help you then

sorry

[tiny gorge]

allg np

[tiny gorge]

why when i select the webui controlnet it keeps automatically deselecting

[flat olive]

which extension i can use to make img2vid?

like animate a image?

[random night]

forge block controlnet extension, can only use intergrated onne

[crystal nacelle]

i love your theme! how do i get it?

[crystal nacelle]

should i also remove --medvram? you helped me do all this yesterday except we didnt mention the medram removal

[lucid yoke]

You use a different model for that, not extension. No idea if forge supports wan/hunyuan though

[crystal nacelle]

also...ive been using SD1.5 for like a year now. it has terribadprompt adherence and i gotta type the same thing like 20 times to get decent results... which version do you recommend I upgrade to for a 16GB ram i5 9th gen AMD GPU pc? could I go to SDXL, pony or illustrious?

[lucid yoke]

I'm gonna be honest chief. I recommend illustrious for anime style images but if you can pick a gpu and you wanna use it a lot more for ai i do recommend Nvidia for less headaches

[formal ridge]

And I'm not going to lie, no matter what you choose that 16GB of RAM is going to hurt.

I wouldn't run AI anything on less than 32GB

[ruby yoke]

i have 16 gig, and the only thing i don't do is videos

[formal ridge]

Do you have any issues on SDXL at all?

I ask because I have lots of old hardware begging for use, even if it's CPU only gens.

[ornate elk]

If you have 8gb vram change it to --medvram-sdxl
If you have the 16gb variant then remove it

[ornate elk]

Ah, need to test wan then too

[formal ridge]

I used the exact files that @old sky recommended at https://comfyanonymous.github.io/ComfyUI_examples/wan/. And it worked for me.

[ruby yoke]

Nope. I just dont run any of the open source video models

[formal ridge]

Alright, thanks for the reply. I'll just have to try it then on my older hardware.

[frigid timber]

any tech support online?

[zinc comet]

I am trying to us this model with diffusers
pipe = StableDiffusionXLPipeline.from_pretrained(
"recoilme/ColorfulXL-Lightning",
torch_dtype=torch.float16,
variant="fp16",
use_safetensors=True,
custom_pipeline="lpw_stable_diffusion_xl",
)
but I got this error
RuntimeError: mat1 and mat2 must have the same dtype, but got Float and Half
anyone knows what to do ?

[ornate elk]

That means something isnt compatible

Like an sdxl model and a 1.5 lora or embedding

Or controlnet for 1.5 with an sdxl model

[zinc comet]

mmhhh ok, it is all sdxl and it fails before loading the lora weights, will investigate

[zinc comet]

found the fix:
https://github.com/huggingface/transformers/issues/30914
with torch.cuda.amp.autocast():

[brave fiber]

I am trying controlnet-inpainting but get a strange error:
def stable_diffusion_controlnetinpainting_rendering(sd: StableDiffusion, images: Images, controlnet_images: list[Image]) -> Image:
print(f"type: {type(images.rgb)}")
return sd.pipe(
prompt=sd.positive_prompt,
negative_prompt=sd.negative_prompt,
num_inference_steps=sd.num_inference_steps,
controlnet_conditioning_scale=sd._controlnet_scales,
controlnet_conditioning_image=controlnet_images,
generator=torch.Generator(device=sd.device).manual_seed(int(time.time())),
strength=sd.strength,
output_type="pil",
image=images.rgb,
mask_image=images.mask,
guidance_scale=sd.guidance_scale,
padding_mask_crop=5
).images[0]

output:
type: <class 'PIL.Image.Image'>
ValueError: The image should be a PIL image when inpainting mask crop, but is of type <class 'NoneType'>.

[zinc comet]

as the message says looks like images.rgb is None, can you print it to check ?

[brave fiber]

i already have the line print(f"type: {type(images.rgb)}") in the function and the output is: type: <class 'PIL.Image.Image'. is there anywhere else you want to print?

[zinc comet]

ah my bad sorry, read too fast

[zinc comet]

strange effectively, what is the type of images ? it is not part of PIL no ? how do you load the images ?

[brave fiber]

images is a class i made:

from PIL.Image import Image

@component
class Images():
rgb: Image = None
mask: Image = None

[zinc comet]

and what is this @Component decorator ? how do you load the rgb/mask data from the file on the disk ?

[brave fiber]

@component was just an alias I made for @dataclass from Python's built-in dataclasses. It doesn't do anything special beyond providing convenient initialization and data storage.

images are never saved to disk—they’re generated and passed around purely in memory. They’re assigned directly as objects to the Images component earlier in my pipeline

image = np.asarray(visualiser.capture_screen_float_buffer(do_render=True))
img_np = (image * 255).astype(np.uint8)
images.rgb = Image.fromarray(img_np).convert("RGB")

This worked fine for Inpainting, but got the error trying to add Controlnet

[zinc comet]

weird, the convert provide a copy, it should not disappear, the StableDIffusion is what ? can't find it in the diffuser library, looks like the error would come from somewhere inside

[brave fiber]

StableDiffusion here is just a custom @component class I created around the Diffusers pipeline, containing some common configuration settings (prompts, inference steps, scales, etc.)

[zinc comet]

code please ? 🙂

[foggy sail]

hi, i need help. i think i need to update my huggingface to version 0.26.0 but i dont know how to

[rough blaze]

hey, Im currently trying to install Stable Diffusion using ChatGPT as a guide for errors, but it cant seem to find Python, eventho its installed and ChatGPT only goes in Circles and i dont know how to fix it.
Would anyone please help me getting Stable Diffusion to run?

[rough blaze]

help... anyone..

[formal ridge]

If you're on Windows, forget doing it manually and just give Stability Matrix a try. If you're on Linux, that's beyond my ability to help.

[rough blaze]

im on windows. what is stability matrix? wouldnt i still need python?

[formal ridge]

Stability Matrix is an executable that will set up Whatever UI you choose to run automatically. After that you just browse for models (it has a built in CivitAI browser), it downloads them and off you go.

[rough blaze]

this one?:
https://github.com/LykosAI/StabilityMatrix

[formal ridge]

Yes, that one. Just check releases and grab the latest windows zip.

It even has a built in inference generator, but I've not tested it.

[hot briar]

Exception training model: 'Using low_cpu_mem_usage=True or a device_map requires Accelerate: pip install 'accelerate>=0.26.0''. i got this issue, i had 1.5.2 when i got this. Uninstalled and installed 0.26.0 version but it didnt work either i got same error

[ornate elk]

hey, the rocm library for GFX1201 (9070 and 9070XT) has been released!
https://github.com/likelovewant/ROCmLibs-for-gfx1103-AMD780M-APU/releases/tag/v0.6.2.4
So the Zluda Guide could work on those too now.

@shadow path

[cursive hazel]

It worked for me, on a 9070!

On Automatic1111, using DirectML, I generated an image of 856x1024, 20 steps image in about 60 seconds.

On ForgeUI, using Zluda, I did the same thing at 35 seconds. So, nearly 2x the speed. Using CyberRealisticPony model on both.

[ornate elk]

Thanks for letting me know!
That means I can update the Guide

[shadow path]

ooooooooooooh

I'll give it a shot tomorrow

[ornate elk]

Let me know how it goes then

[flat egret]

guys am i doing something wrong here? this is the result im getting with img2img

from this picture:

here's my settings

[compact salmon]

waaait what?! how did u get comfy Forge working 9070xt?!?!

i got SDnext working but honestly i much prefer forge ive realized. i want to get forge and comfy up and running on my 9070xt on windows os

couldnt find a single person who has forge or comfy running on 9070xt windows os with zluda

[cursive hazel]

CS1o just mentioned that there's a new rocm library for GFX1201 (9070 and 9070XT), here #🤝｜tech-support message

I installed rocm again, used that patch ^ (I did the "No optimized", and have no idea if that matters or not..)
Then I just followed the pinned instructions "The latest Stable Diffusion install Guides for Nvidia and AMD", for ForgeUI. I also got ComfyUI working too, but I'm a complete noob at it

[compact salmon]

i did grab that and the 6.2 rocm and no optomized hting as well

[compact salmon]

hell yeah homie. i hope its ok if i maybe bother u once more trying to figure this out

thats awesome

[cursive hazel]

Sadly, I can't really help too much. I'm stumbling my way through all of this tbh! CS1o actually knows how to fix problems. All I can really say, is that I did a fresh install of rocm 6.2.4, patched that with the aforementioned rocm library, got a fresh copy of Zluda, and.. The rest was just following that guide. It just worked for me.

[compact salmon]

i trust it

thats really helpful

thank you!!!

maybe ill only bother to show you have it up n running 🫡 lol

[cursive hazel]

Oh, and at some point the guide says to restart your PC!.. That is actually mandatory apparently! I tried it without, thinking.. Ohhh, it'll be fine! But my images were taking like.. 5 minutes.. So I have a restart, annnd it's now 2x the speed of DirectML.

[compact salmon]

siiiick

tysm

[hot briar]

@ornate elk you know how to solve it?

[stiff hawk]

Hey, fam. I've been trying almost non-stop for days to install a webui with ZLUDA, ONNX, and Olive. I keep getting errors, I've done all the research I know to do and have studied the documentation the best I can, and I have run out of ideas. I'm going to cry ^^;

I have Windows 11 and a Radeon RX 7600 with 8gb VRAM.

I've tried stable-diffusion-webui-amdgpu, SD.Next, and I researched Fooocus and started to attempt stable-diffusion-webui-amdgpu-forge. I prefer forge if possible; but if not, then webui-amdgpu. I have successfully gotten ZLUDA working sometimes, have very rarely gotten ONNX working, and I've never gotten Olive. And nothing is consistent. I've successfully gotten it to choose the right GPU, too.

[formal ridge]

Top Guide under Pinned Messages in this channel. That's the absolute best one I've seen for AMD users.

[stiff hawk]

@formal ridge This one?

Should I use Automatic1111, the fork for AMDGPU, Forge, or Fooocus? I like the idea that Fooocus is simple, but I don't want it to be like Midjourney where it begins to ignore my prompt info

Forge seems like the second-best bet but not if it has so much less documentation/support that I struggle to even get it working

Huh, looks like I got Forge working with ZLUDA. Now to see if I can get it working with ONNX

[stiff hawk]

TypeError: 'StableDiffusion' object is not callable
'StableDiffusion' object is not callable

[formal ridge]

all of them will sometimes ignore your prompt info. There's no such thing as a perfect model unless you caption it yourself.

[stiff hawk]

What are your personal tastes? Also are you able to help with ONNX?

[formal ridge]

No, I've never bothered to try ONNX. I can't help with that at all.

Personal tastes? It depends on what architecture you can run and what you're looking for, Realistic, Anime, etc.

[stiff hawk]

Not anime, that's for sure

[formal ridge]

Lately I've been messing with the realisticVision series, and 5.1 seems to be the best in that specific line.

[stiff hawk]

I got that too! In fact it's my default

I meant the WebUI though, which do you like

[formal ridge]

For SDXL, I highly recommend Everclear PNY by Zovya v2 & v3

[formal ridge]

Oh that, I primarily use ComfyUI these days, as I can get a bunch of custom nodes / samplers / schedulers.

[stiff hawk]

I'm not sure my RX 7600 with 8gb VRAM can do SDXL, but I know basically nothing. ComfyUI seems way over my head though. I am not technical

[formal ridge]

My 3060TI is only an 8GB card, and I can generate (not upscale) up to 3840x2160 images on SDXL. But I have an nVidia card. Not sure how that will work on an AMD card.

[stiff hawk]

"TypeError: 'StableDiffusion' object is not callable"

This no matter what I do when ONNX is enabled. I haven't been able to get past this for like 15 hours of work

[formal ridge]

I wish you luck, I've never tried to get ONNX to work as it sounded like a hassle for so little value.

[stiff hawk]

I've been working for like a month trying to create characters to my specifications. Midjourney doesn't work, and ONNX doesn't work, and I don't trust other online tools, so I guess I'm screwed forever

All because I don't have an expensive Nvidia card

[formal ridge]

Naah, it might be the model you're using. Give me a sec.

[stiff hawk]

Hmm

[formal ridge]

Look up "DucHaiten-Journey-XL" and tell me what you think.

I recommended that one a month ago.

[stiff hawk]

What is it?

[formal ridge]

SDXL model

[stiff hawk]

Oh, that's a big pruned model. I wouldn't even be able to consider using it without ONNX and Olive

I have an 8gb card

[formal ridge]

I have an 8GB card too.

[stiff hawk]

I don't understand, are you suggesting I use this 6.5gb model without ONNX? I'm gonna run out of memory real damn fast

[formal ridge]

It works fine in ComfyUI and Forge.

A1111 is the only place you'll probably have issues

[stiff hawk]

I ran out of memory using a 2gb model properly

[formal ridge]

Again, what UI?

[stiff hawk]

I'm using Forge

[formal ridge]

Do you have Never OOm enabled at the bottom?

[stiff hawk]

realisticvision, recommended settings

Where's that?

[formal ridge]

Left hand side, scroll down

[stiff hawk]

My only find of "never"

What is it, what does it do?

[formal ridge]

Open the Forge UI, look at the left hand side and scroll down.

It helps with Out of Memory errors

[stiff hawk]

OH

It's still slow as hell, which is why I need ONNX and Olive

It's amazing that these tools exist, everyone else on Earth apparently gets them working because my error isn't on Google, just my luck

[formal ridge]

If it works, it works. Why complicate things?

[stiff hawk]

It's far too slow. How am I supposed to take the 100 hours to learn how to use Stable Diffusion when I get a new result every 5-10 minutes?

[formal ridge]

https://tenor.com/view/yoda-patience-you-must-have-patience-gif-15254127

[stiff hawk]

I have patience for generation if I'm getting something. I don't have patience to literally spend a year of doing nothing else but trying to learn how to prompt and use my toolset. That's what slow generation gets me. A lot of failure I have to wait to finish. I could have burned myself out hard trying to learn Midjourney, let alone something like this

[formal ridge]

The time to generate depends a lot on the resolution.

[stiff hawk]

From my experience, low resolution means unfinished images. I don't know why

[formal ridge]

If it's taking 5 minutes for a 512^2 image, something is definitely wrong

[stiff hawk]

https://huggingface.co/SG161222/Realistic_Vision_V6.0_B1_noVAE

I tried following the instructions

[formal ridge]

That's the bad one, you want v5.1

[stiff hawk]

Really? O:

[formal ridge]

v5.1 is much better than v6.0 in all my testing

[stiff hawk]

I see v51 and noVAE

and v51VAE_cn

I have no idea how anyone can keep track of this stuff

https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE

This one is the most-downloaded

[formal ridge]

Most people use CivitAI, not huggingface

[stiff hawk]

To download models?

[formal ridge]

yes

[stiff hawk]

Can't even find it there

Sorted by most-downloaded, 6.0 is first, no other version on list

[formal ridge]

https://civitai.com/models/4201?modelVersionId=130072

[stiff hawk]

But you said 6.0 was bad

[formal ridge]

Do you see the numbers at the top?

[formal ridge]

CivitAI lets users upload many models under one model card, just like huggingface

[stiff hawk]

I see V6.0 B1

Okay, I found it

[formal ridge]

v5.1 (VAE)

[stiff hawk]

Very poor understanding of UX psychology, it just looked like noise up there

Didn't realize they were choices, or even see them at all because they looked like noise

[formal ridge]

That's why i sent a direct link...

[stiff hawk]

Nah, I appreciate it, and also it's good that you pointed those tabs out to me, I needed to know they were there

[formal ridge]

Because Huggingface is a massive mess compared to civitai.

[stiff hawk]

Hmm

I can understand and navigate Huggingface better, and it hurts me a whole lot less to look at, but I see advantages in CivitAI

Huggingface is built more like a repository, CivitAI is built more like a showcase

Not that I can use repository websites easier myself, but it's easier for me to visually process Huggingface

But for AI, showcase is better in principle, so CivitAI should be better

Some of the UX in CivitAI though is too busy and not grouped or labeled properly, so your eyes glaze over when you're trying to understand how to find things, hence me calling the tabs "noise" after I actually saw they existed

[formal ridge]

CivitAI also has filters based on what you're looking for, Lora, Checkpoints, Model Architecture, etc

[stiff hawk]

I see

Why is 5.1 better than 6.0?

Do you know if you can have presets for models? So I don't have to set negative prompts or CFG scales, etc. when I switch models

Also is there a reset button for all Generation settings?

https://civitai.com/models/4201?modelVersionId=130072

This page gives me like no information. I don't know how many sampling steps to do

I don't understand how anyone is supposed to learn anything with so little documentation about how things work or how to fill in missing information to get things working. It's like Linux all over again

[formal ridge]

I'm not sure, it's been so long since I've used Forge.

I've been on ComfyUI for the past almost year now.

Where I can have a workflow for every model or model type that I want.

[stiff hawk]

Out of curiosity, how long did it take you to learn it?

[formal ridge]

About a day, but even if you know things, a lot of it is simple experimentation and repetition.

The first week I was modifying A1111's code and fixing his screwed up math so I could generate 1080p images without issues.

[stiff hawk]

So people who aren't programmers have no hope. That's the Linux Experience for ya

[formal ridge]

I'm not a programmer by any means. I just have a lot of time on my hands and can throw shit at the wall until it sticks.

[stiff hawk]

That makes sense

I'm going crazy with how much work I've been doing to try to get things working, and my housemate doesn't understand is angry at me for working too hard

And I'm like, I need this working, I need my results, I can't just stop

Sure wish Midjourney wasn't being impossible, I'd have been done a week ago

My results with Realistic Vision are so malformed even with my negative prompts that I'm about to give up and try Fooocus even with poor AMD support. I don't have all of my life to try to figure out what I'm doing wrong, and I can only be negative around people I'm asking for help for so long before they're sick of me

[formal ridge]

I'm assuming you have Zluda, right? There's a Zluda version of ComfyUI if you feel like giving it a shot. https://github.com/patientx/ComfyUI-Zluda

[stiff hawk]

Doesn't that have an enormous learning curve? Like, you have to learn the technical details of exactly how generation works and then use that knowledge in custom setups

I am not that kind of person. I'm an artist

[formal ridge]

Not really, you just drag and drop spaghetti strings, like any other workflow software

Then it's just experimentation and repetition

[stiff hawk]

I read that people have to know exactly how generation works on a technical level

"I didn't find it difficult at all.

I made the same journey as you (easy diffusion, A1111, comfy), and at each stage have not gone back.

I think the challenge is more understanding how stable diffusion works, then understanding comfy per se. For example, in comfy you start with an "empty latent image". This isn't very intuitive, and A1111 hides this stage from you. But once you understand that:

a) a latent image is the type of image that SD works on b) you encode an normal image into a latent before using it, and decode a latent into an image after generating it c) the only difference between txt2img and img2img is whether you start with an empty latent, or a latent encoded from an existing, normal image

...it all makes sense. Then you never need to think about latents again, you just feed the model whatever starting point (empty latent, encoded image) you need for your purposes.

Of course, once you have this, you can then expand on it. You can start with an empty latent, generate an image, then pass that latent elsewhere. You can upscale it, mask it, replace parts of the image etc. This is where the real strength of comfy comes out, beyond just typing a prompt and making an image, but controlling the workflow to do exactly what you want."

[formal ridge]

It's somewhat user friendly. Just need to load your checkpoints, your vae, and it has a built in default workflow in case you do manage to completely screw yourself too badly.

[ornate elk]

Yes it can do easily use sdxl.
Also don't try to get onnx or olive running with Auto1111 or Forge.
Its not worth as olive/onnx needs model conversion to work and dont support a lot of extensions.
If you really want to use something with olive and onnx try out Amuse webui.

But the best for your GPU would be Auto1111 or Forge with Zluda.

[ornate elk]

Hey, you have to set the denois higher

[ornate elk]

To less information on what are you using, like tool and GPU.

[hot briar]

I am using automatic 1111 and my gpu is RX 7800 XT

[ornate elk]

Then pls show the full cmd log.
Also zluda or directml?

[hot briar]

Directml, how can i show i dont know

[ornate elk]

Copy the full cmd log and paste it here

Also directml is not recommended.
Its slower and has more bugs than the zluda variant

You definitly want to switch to the zluda version

[hot briar]

I just copied a guide that uses amd and download

If its work on mine i can switch it rn

@ornate elk i probably zluda

webui.bat --use-zluda I use this bat when i open sd

[ornate elk]

Thats good okay.
We will see when you send the cmd log

[hot briar]

The error's log right?

[ornate elk]

The complete log

When you launch the webui

• the error after you triggered it

[hot briar]

error.txt

@ornate elk

[ornate elk]

Ah okay now I see the problem

You want to train a lora in auto1111

But thats not how it works

You have 3 broken extensions installed.
Dreambooth, Roop, and Easy Lora Training Script

You have to delete those extensions from the extension folder.
Then delete the venv folder and relaunch the webui-user.bat

And for lora Training please then install OneTrainer with Zluda:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Lora-Trainer-Setup-Guides#amd-onetrainer-with-zluda

[hot briar]

Relauch it with --use-zluda or without it

[fiery hearth]

I've tried both fast and conservative upscaling via API but get 400 errors on both. I've successfully used the API generate interface without any issues. If anyone has used either fast or conservative upscale pls share

[hot briar]

Okay deleted

Deleted venv folder too

[ornate elk]

With --use-zluda --skip-ort and dont launch the webui.bat

Only the webui-user.bat

Where these two launch args should be added by you

[hot briar]

@echo off
cd C:\Users\niceb\a\stable-diffusion-webui-directml
webui.bat --use-zluda --skip-ort

Like that?

[ornate elk]

Nope, right click the webui-user.bat and edit it.
At the line commandline_args=
You add --use-zluda --skip-ort

And I mean webui-user.bat, not webui.bat

[hot briar]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--use-zluda --skip-ort

call webui.bat

[hot briar]

Like that then

[ornate elk]

Yep and always only launch the webui-user.bat and not the webui.bat

[hot briar]

Okay ty

[hot briar]

I am launching it and doing this and done

[ornate elk]

Alright. As you already setup zluda you can skip that part on the Guide

[stiff hawk]

Super thanks

[hot briar]

ImportError: cannot import name 'broadcast_to' from 'numpy.lib.stride_tricks' (C:\Users\niceb\a\stable-diffusion-webui-directml\venv\lib\site-packages\numpy\lib\stride_tricks.py) said this while lauching

[flat egret]

will try, thanks

[ornate elk]

Relaunch

That error is normal the first time after deleting the venv

[hot briar]

Which parts should i do

[stiff hawk]

I'm trying to create character concept art. But the more information I gave Midjourney, the more they would start distorting or ignoring. I spent a week or two of non-stop attempts, but there's so much I don't understand. At least I know to mention the rest of the body if I want it to not do only the face.

I'm at a point where I'm trying to use other tools. Trying to get Fooocus with ZLUDA up and running, as I'm not that technical and I need to move forward. So while I have Forge available and I'll use it if it's my best option, I don't know if I have it in me to spend the next bunch of months non-stop learning technical stuff to get results I want, especially since I only have an AMD Radeon RX 7600 with 8gb VRAM

But with that said, I will hear any advice people have about tools, setup, and prompting. I imagine inpainting is a lot more important than I've been treating it.

I need realistic body proportions and anatomy and detailed body shapes for a realistic look, but in the end I want inked lines and painted colors, so I don't need photorealism. Trying to make concept art for a smart adult novel that clearly shows it's (SFW, non-gritty) adult. Def not anime either.

I can upload a ref of one thing I was kinda good at for an example of that last bit

If my dern NAS would stop being borked

[ornate elk]

The first 6 steps of setting up OneTrainer.
Then the last step 6 at the bottom

That should be enough

If you have a C:Zluda folder you can also do the 4-6 steps at the end additionaly

[ornate elk]

Hey im not sure if this is now a technical question for here or if you better ask in #📝｜prompting-help

But you should try out Forge Webui with Zluda if it works. That gives you the most control over the image with your GPU.
Try out some models like illustrij or Dreamshaper or find some loras that match the style you like

Its not to hard to get anatomical correct images

Mostly stick to sdxl/illustrious models for that

[stiff hawk]

What's Illustrious

[lucid yoke]

A sdxl finetune

For anime style images

[formal ridge]

He's looking for semi realistic / realistic, that's why I told him about DHJ-XL

[stiff hawk]

Followed the Fooocus with ZLUDA guide: https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-fooocus-with-zluda

M:\AI\stable-diffusion-fooocus>.\python_embeded\python.exe -s Fooocus\entry_with_update.py
Already up-to-date
Update succeeded.
[System ARGV] ['Fooocus\entry_with_update.py']
Python 3.10.9 (tags/v3.10.9:1dd9be6, Dec 6 2022, 20:01:21) [MSC v.1934 64 bit (AMD64)]
Fooocus version: 2.5.5
[Cleanup] Attempting to delete content of temp dir C:\Users\orphi\AppData\Local\Temp\fooocus
[Cleanup] Cleanup successful
Traceback (most recent call last):
File "M:\AI\stable-diffusion-fooocus\Fooocus\entry_with_update.py", line 46, in <module>
from launch import *
File "M:\AI\stable-diffusion-fooocus\Fooocus\launch.py", line 152, in <module>
from webui import *
File "M:\AI\stable-diffusion-fooocus\Fooocus\webui.py", line 10, in <module>
import modules.async_worker as worker
File "M:\AI\stable-diffusion-fooocus\Fooocus\modules\async_worker.py", line 3, in <module>
from extras.inpaint_mask import generate_mask_from_image, SAMOptions
File "M:\AI\stable-diffusion-fooocus\Fooocus\extras\inpaint_mask.py", line 5, in <module>
import torch
File "M:\AI\stable-diffusion-fooocus\python_embeded\lib\site-packages\torch_init_.py", line 143, in <module>
raise err
OSError: [WinError 126] The specified module could not be found. Error loading "M:\AI\stable-diffusion-fooocus\python_embeded\lib\site-packages\torch\lib\cublas64_11.dll" or one of its dependencies.

M:\AI\stable-diffusion-fooocus>pause
Press any key to continue . . .

---

This is what my life has been for a week

The wording was odd, but it appeared it was asking me to copy the ZLUDA files, rename them, then paste them over the ones in Fooocus's directory, overwriting. I def copied/overwrote those three files

[ornate elk]

Its missing the copied zluda files

Sry for my wording xD feel free to let me know how to right it better

You have to make a copy of the 3 files in the zluda folder.
Then you rename the copies accordingly to the guide.
Them copy these 3 renamed files into the foocus torch lib folder and overwrite

[hot briar]

@ornate elk did Getting the OneTrainer AMD Fork:

Make a new Folder on your Drive (not on Desktop, Downloads, Documents, Programms, Onedrive) and name it SD-Zluda.
Dont use the same if you had installed the DirectML webui before.
You go into the folder you created in this case SD-Zluda, then click in the File Explorer bar (not searchbar) and type cmd then press enter.
Then you copy and paste this command:
git clone https://github.com/lshqqytiger/OneTrainer.git && cd OneTrainer && py scripts/install_zluda.py
Press enter and after its done you can close the cmd.
Launch the install.bat, and wait until its done. Then close the cmd.
Launch the update.bat, and wait until its done. Then close the cmd. this part

[stiff hawk]

Oh, you wrote it? I mean no offense, I can probably help with gap-filling and wording to make things clearer. I'm an autistic novelist, clarity is my specialty. Out of necessity xD

[ornate elk]

Sure ^^ always good if I can improve the Guide for easier understanding

[ornate elk]

Okay good

[stiff hawk]

I did copy them, I'm not sure what could be missing

[ornate elk]

Can you show me the zluda folder content?

[stiff hawk]

The one in the Fooocus directory, or the original ZLUDA download?

[ornate elk]

Original zluda folder

[stiff hawk]

Here's what I downloaded: https://github.com/lshqqytiger/ZLUDA/releases/

And here's what the folder looked like after I copied and renamed the three files

Then I copied those three files into uh

Copy these two three to python_embeded\Lib\site-packages\torch\lib and overwrite if asked.

[ornate elk]

Okay thats correct

[stiff hawk]

I wonder what went wrong

[ornate elk]

Can you copy the files over again and then relaunch?

[stiff hawk]

Sure

Exact same error

[ornate elk]

Can you check the file you had to edit again if the changes are still in there?
Fooocus sometimes resets that file because it updates

I'm AFK back in 15-20 mins

[stiff hawk]

I copied cublas.dll, cusparse.dll, and nvrtc.dll, renamed them to cublas64_11.dll, cusparse64_11.dll, and nvrtc64_112_0.dll, and put them in python_embeded\Lib\site-packages\torch\lib, overwriting the ones already in there

[stiff hawk]

Will do

Looked like it was still there, I repasted again from the guide and saved, ran run.bat. Same error

[hot briar]

what should i do with settings @ornate elk

[novel owl]

hello im trying now to test comfyui with vidéo workflow and the task lock at the negative prompt

[ornate elk]

hey, what does the cmd shows?

[ornate elk]

ive left these on default

[novel owl]

ah it go to KSampler but stuck here

[ornate elk]

hmm looks okay

[novel owl]

im on amd graphics card with zluda

[formal ridge]

It's probably not stuck, the video vae takes a while to complete. Open your task manager (ctrl + alt + del if on windows), and see if the Processor is still being used by Python.exe

[hot briar]

me too but wouldn't be problem for amd?

[ornate elk]

it could be compiling in the background

[ornate elk]

nope

cuda gets emulated by zluda

just make sure to disable xformers in the Training tab. Set the Attention to Default or SDP instead of xformers as xformers won't work on AMD and you'll get an Error.

[novel owl]

ah yeah is that !

[hot briar]

@ornate elk is concepts important

[ornate elk]

its the part where you add the folder of your images for training

[hot briar]

is it enough to add a concept and select the path

or it is something more of it

[ornate elk]

hmm i reinstalled foocus acording to my guide ant it works.
maybe you should redo the steps with the custom torch install command, then launching foocus, then editing the model_management.py again and then paste the zluda files over again

[ornate elk]

maybe this helps a bit:
https://civitai.com/articles/4789/lazy-lora-making-with-onetrainer-and-ai-generation

in concepts you can just add a folder to the images

but you also need captions

you can create captions with the caption tool in onetrainer under tools

[stiff hawk]

For troubleshooting reasons, I want to add that I didn't install Fooocus in the main/Windows partition, but in a secondary partition

[hot briar]

@ornate elk is it done?

[ornate elk]

Thats okay

[ornate elk]

How long did it take?

Could be done but you see that in the steps

[hot briar]

10 min

[hot briar]

yea it's done but bad result

is problem sources or step count

[ornate elk]

Mostly step or epochs

Also images should be like 20-40 good quality images

You have to test around.
Lora training is not easy

[hot briar]

okay ill try something

[novel owl]

Sory

[ornate elk]

np, how fast is wan?

[stiff hawk]

Can you send me your model_management.py so I can compare to mine?

[ornate elk]

sure

[hot briar]

can we save it

or just stopping training and it's been saving

[ornate elk]

it does backups while training

[hot briar]

ah true, i totally forgot that

[compact salmon]

does anyone know where i can change device_map='auto' to device_map='cuda',

[ornate elk]

where? for what? xD more info needed

[compact salmon]

i get this issue

and the ONLY info i can find that did something for someone was this

https://discuss.huggingface.co/t/runtimeerror-cuda-error-cublas-status-execution-failed-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/109488

"This issue sounds tricky. The workaround, if it can be worked around, is to change the CUDA version or save VRAM usage.
If there is no workaround, it may be an unresolved bug.

In your case, you have already specified device_map=“auto”, so as long as the accelerate library is properly installed with pip, you should be able to offload as much as possible.
The only thing left to do is to reduce the amount of data to be passed on somehow."

[ornate elk]

that means you have to paste the 3 renamed zluda files into the torch/lib of comfyui

[compact salmon]

i thought i did that but let me go ahead and doi that again

i have to find those 3 scripts

[ornate elk]

the 3 renamed dlls are in the C:Zluda if you followed my guide

[compact salmon]

maybe?

where is that

ill follow yours

[ornate elk]

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[compact salmon]

oh this can install forge?

oh shit

im tryna do comfyui rn let me not get ahead of myself

im on 9070xt so im having more issues than normal

using the modified 6.2rocm

[ornate elk]

yep forge, auto1111, fooocus, comfyui and OneTrainer for lora training work with zluda

[compact salmon]

so hype to hear this

just need to get it up n going lol

[ornate elk]

someone already did yesterday with a 9070 here, so it should work

let me know if you need any help

[compact salmon]

thank you i appreciate that im almost there haha.

where is the copying part?

oh here Go into the C:\ZLUDA\ Folder.
There make copy of the cublas.dll and the cusparse.dll and nvrtc.dll repaste them inside the folder.
Rename the copies to cublas64_11.dll and cusparse64_11.dll and nvrtc64_112_0.dll
Copy these three files into ComfyUI\venv\Lib\site-packages\torch\lib and overwrite if asked.

i started the guide from the patientx tutorial in his it had 3 codes to type or something instead this is where yours is slightly more manual and may make it woirk for me lol

so i followed option 1 all the way

[ornate elk]

yes my guide should work, also make sure you installed hip sdk 6.2 and not 5.7

[compact salmon]

yes i have that

can i delete the c:/zluda after or is that permanent?

after that info i should be good to try

[ornate elk]

leave it there if you added the path in the environment variables

you can use that for other tools from my guide

[compact salmon]

oh okay so ill put that in my ai drive instead

D:\Graphics\AI

followed the steps and still CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)

here is the error report

[ornate elk]

thats the patient x repo, comfyui version

cant help with that

[wary tulip]

Hi

I want to ask if you guys know what is the best PC I can get with a low budget to use comfy UI

[ornate elk]

hey, that depends on the budget

low budget means 500 for some people while other say 800 for example

[wary tulip]

Yeah I'm not from America so I'm not using dollars but it's around like this between 500 and 700

[compact salmon]

ill follow ur guide @ornate elk and get back to u homie

[wary tulip]

@ornate elk

...

[ornate elk]

A 3060 with 12gb vram would be a good start

Idk how the GPU prices are for you, but also used GPUs can be a good starting point.
Also get 16-32 GB RAM.
And for the CPU go for a AMD Ryzen

[wary tulip]

Intel Core i7-6700
GTX 1080 8GB
16GB DDR4 3000MHz (2×8GB)
What about this one ?

[wary tulip]

.

[ornate elk]

Thats okay but slow

[wary tulip]

What should I upgrade and remember I'm too broke

[ornate elk]

For image generation the vram amount is the most important.
After that 32gb RAM is important to not get slowed down by some larger models.