🤝｜tech-support

[ornate elk]

Better is to not use the refiner at all.

Its outdated and not needed since you can upscale images faster

[left junco]

SwarmUI.. i think i prefer using ComfyUI tbh.

[ornate elk]

Its only okay for base sdxl models but the vram requirement makes it not that great for smaller GPUs than using upscalers

Thats good

An nvidia GPU with a large amount of vram is the most important.

Because the whole ai stuff is based on "checkpoints" that get loaded into the vram. The more it can fit the bigger and better trained versions you can use

[left junco]

What about unified ramlike M3 macs and also the NVidia Digits (FOrgot the new name)

Waiting for benchmarks.

[wanton wagon]

Is anyone else having an issue with Automatic1111 not releasing the VRAM after done generating? the only solution i've found is just Ctrl+C on bash and resetting it and using the old window.

[left junco]

Isnt that an option in settings? Also, isnt that just the model so it doesn't have to reload it next generation?

[ornate elk]

What's your GPU?

And which model do you use?

Nope each GPU can only be used to Generate its own image/video

[wanton wagon]

Nvidia GeForce RTX 4060, 8 GB.

And sorry i'm new at SD, do you mean which Sampler i use? Sorry I dont quite understand your second question.

Oh! I'm using mistoonXLCopper_v20Fast

[ornate elk]

Ah 8gb vram and sdxl.
If you use Auto1111 webui you need to add --xformers --medvram-sdxl to the webui-user.bat at the line commandline_args=
That speeds up the gen time and lowers the vram usage. Resulting in not getting stuck at higher vram

[left junco]

forge too? @ornate elk

[ornate elk]

In forge you can't add these.
But you can add --cuda-stream --cuda-malloc

[wanton wagon]

Yeah I have the --medvram and --xformers, I have also tried PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:64,garbage_collection_threshold:0.6. (However i removed those lines it didnt seem to like that).

What is -sdxl? Is that something that goes onto the --medvram?

Just trying to get it working so i dont have to reset bash each time it finishes generation

[ornate elk]

Remove the pytorch cuda alloc, split size stuff. That doesnt help.
Also change --medvram to --medvram-sdxl
--medvram is only for GPUs 6gb or less useful

--medvram-sdxl loads sdxl models better on 8gb GPUs

[wanton wagon]

Oh, i might be mistaken, but i thought i was running on 1.5. i know that sdxl. I know that sdxl eats up a lot more juice.

[ornate elk]

Yep, the model you names has XL in its name. I assume it sdxl based

[wanton wagon]

That might be it then

[ornate elk]

You can easily verify that as sdxl models are not smaller than 6gb

[wanton wagon]

I'll add the extra part

Yeah 6.5gb

Thanks! Lots to learn

[ornate elk]

No problem!

[tribal anvil]

How do I generate bigger images? Crashes whenever I get to 1k+

[ornate elk]

What's your GPU?

And which model do you use?

[tribal anvil]

Uhhhhh

I have 4 GB of vram I think

Maybe 3

[left junco]

What card is that?

[tribal anvil]

GTX 970m

[left junco]

Can that even do CUDA?

[tribal anvil]

Takes a little bit but ye

[left junco]

I have one here.. on a desk. Gathering dust.

Ti o_O 😛

Okay my man, i will think of your hardship next time i am annoyed with speed

[ornate elk]

Upscaling helps

[tribal anvil]

Stupid question how do I do that?

[tribal anvil]

🫡

[left junco]

Are you sure you are not using CPU?

[tribal anvil]

I honestly dk

I'm very bad with computers

[left junco]

Windows then i guess? System monitor or something will show what's being used. It has little graphs showing who is busy

[tribal anvil]

Yep

[ornate elk]

Best would be to go into img2img tab,
move the image in there.
Set the resolution to 512x512
Then set the denois to 0.2
At the bottom at Scripts, enable SD upscale script.
There choose Resrgan4x+
Then hit Generate.

[left junco]

Pretty sure it's CPU because i was pretty sure you need 1080 or higher for cuda.

[tribal anvil]

I'll try that thanks

[tribal anvil]

Maybe I'm using something else then

I'm using forge and I remember seeing cuda on GitHub

[ornate elk]

It can work on a 970m

[left junco]

Yeah? I don't recall i had one and bought 2080. Maybe was abecause it was hard to get anything to work. (Back then things werent as optimized as they are now)

[ornate elk]

True

2022 I used the first SD models on my GTX1080
They were 7gb large and took minutes for a 512x512

[left junco]

Yeah, now it's seconds

BAM! BAM BAM BAM! 4 images

I use flux now though. takes longer.

But it can go into RAM 😛

[ornate elk]

Yep flux hits hard on lower vram

[left junco]

All new stuff is hard on VRAM. These clever scientist guys all have a million VRAM

And RAM. Oh keep 5000 images and combine them> Yeah lets just keep them in RAM.

Anyway thanks fo helps an GN, need sleep. 👋

[brazen sun]

nope, feels even slower tbh, using zluda if that matters

[ornate elk]

Hmm

[tribal anvil]

Sometimes loras don't get the characters right

I use BREAK

And I use the prompts I find on civitai

How do I fix this?

[austere pine]

Tried changing the resolution and nothing. I noticed my model wasn't being detected as Vpred, so I changed it to Vpred ZSNR manually, but now I'm just getting noise

Same thing for normal V-Prediction

[worthy zephyr]

Has anyone got Hidream working in comfyUI? I tried running it but it just wont work correctly. I left it on overnight only to see the generation was only 17% done in the morning which is insane. I thought maybe it had to finish downloading something but I get no errors when running it and task manager does show full utilization of my 4090.

[azure magnet]

Hi, where i can check the last version of comfyui manager?

[deft hinge]

Hi everybody! New to image generation, and have an issue you can maybe help me with.

I have tried reforged and comfyui (the desktop version), and both seem to only use my CPU and cause my pc to run hard and continuously (Unless I shutdown or restart my pc). Is there a fix for this, or a way to force it to use my GPU? I'd take a solution for either reforge or comfyui (Do people have a preference for one or the other for a beginner?).

I am running Two EVGA Gaming iCX GeForce GTX 1080 Ti 11 GB Video Cards, if that helps. Thanks in advance!

[wanton wagon]

I'm getting this error. I was told that this command shouldnt help but its telling me to do so.

OutOfMemoryError: CUDA out of memory. Tried to allocate 400.00 MiB. GPU 0 has a total capacty of 7.63 GiB of which 275.94 MiB is free. Including non-PyTorch memory, this process has 6.27 GiB memory in use. Of the allocated memory 5.74 GiB is allocated by PyTorch, and 414.57 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Time taken: 18.4 sec.

EDIT: I swapped to ComfyUI and I have no more issues with VRAM

[fair oxide]

i need a linux guide

for SD (a1111 or forge)

[brazen sun]

What is it supossed to do?

[calm moth]

I'm trying to Batch Hi-Res fix with ReForge. It works perfectly fine, importing all the Metadata from my images (Prompt, Negative, Step, CFG, etc). The only issue I'm having is that it isn't using the Upscaler I designated in Settings, instead it's just Hi-Res Fixing with "Latent". Does anybody know if there is something else I need to do or is this an issue in ReForge?

[calm moth]

Did you install CUDA Toolkit?

[calm moth]

What were you using before?

[deft hinge]

I have not, and I unfortunatly don't know what it is. To be fair, I was a Mac user who recently ascended to the pc master-race; So I am re-learning windows all over again.

[calm moth]

https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exe_local

Try installing this, then running your WebUI of choice again.

[deft hinge]

Is there any risk to installing this?

[calm moth]

It's nVidia Software

[deft hinge]

Gotcha, I'll give it a shot. 👍

[calm moth]

Try checking the CMD when you run the WebUI, see if it detects your Graphics Card.

[deft hinge]

Got it installed, but it's still making noise like it's running hard and is continually using the CPU.

[wanton wagon]

Automatic1111

[ornate elk]

Install the sd-forge-couple extension

[ornate elk]

Which webui?

[ornate elk]

Can you show the cmd log of ReForge?
It should work with a GTX easily

[ornate elk]

You have to set the upscaler in hires fix not in settings

[deft hinge]

Just tried it with comfyui, let me run a test with reforge. One sec...

[ornate elk]

Make sure your GPU drivers are up to date too

[fierce hollow]

Is it possible to run HiDream-ai/HiDream-I1-Full in Forge?

[ornate elk]

Thats an out of vram error.
Which settings did you used?

[ornate elk]

Nope

[fierce hollow]

I read about the recommended system requirements of 16 GB of video memory.
As I understand it, on the 3070ti (8gb) it won't even start, right?

[wanton wagon]

Medvram-sdxl -xformers.

I've also tried lowvram settings. Nothing seemed to work with the memory leaking. After one run it would crash without me resetting SD.

[ornate elk]

Yep, you would need to wait until Someone makes quantitised model version of it. They are smaller

[ornate elk]

I mean the image settings. Resolution steps etc

[ornate elk]

Looks normal

It used your GPU

[fierce hollow]

And which environment supports HiDream?

[ornate elk]

Probably only Comfyui

[fierce hollow]

Understood, thank you

[deft hinge]

Weird... It just seems odd that it would cause my pc to run hard and seemingly constantly (I haven't let it run like that for more than 30min, so I cant tell if it does stop on it's own).

Is this normal?

[ornate elk]

Have you other stuff running in the background like Wallpaper engine, Games or Rainmeter?

[deft hinge]

Not to my knowlege. I make sure it's the only thing running after I noticed the issue.

[ornate elk]

Okay, and what do you mean you can't tell if it stops on its own?

After the image is Generated it stops

The model is still loaded in the RAM and vram

[deft hinge]

Thats the thing, I make an image and all seems fine... and then: the pc fans start running and running, The CPU starts running around 100%, and wont stop until I shut down my pc; the longest I've let it run is for 30mins like this (I'm protective of my pc, and am worried about bricking it.)

[ornate elk]

The CPU goes to 100% after the image is Generated?

[deft hinge]

Yup!

[ornate elk]

What's your CPU and how much RAM do you have?

[deft hinge]

CPU: AMD Threadripper 1950X 3.4 GHz 16-Core Processor

Ram x 4: Corsair Vengeance LPX 32 GB (2 x 16 GB) DDR4-3200 CL16 Memory

[ornate elk]

Okay looks normal

Can you Screenshot what process uses the CPU the most when it goes to 100%?

[deft hinge]

👍 One sec...

[mental falcon]

is there a way to not automatically open up the web browser after running webui-user.bat? as I am using it as an api service, I have another software that open the webui-user.bat on launch.

@echo off

set PYTHON=
set GIT=
set VENV_DIR=

set COMMANDLINE_ARGS=--xformers --api --skip-torch-cuda-test --ckpt models\Stable-diffusion\dreamshaper_8.safetensors


call webui.bat

[ornate elk]

frequency is not usage

so its not doing anything but runs at a very high clock speed

[deft hinge]

Thats what I am lead to believe as well, as far as I can tell nothing is happening but for some reason SD makes an image and then causes my pc to run like crazy. No other program I've run does this.

[ornate elk]

you can disable it in the webui its the setting called autolaunch

[ornate elk]

have you tried an other browser? like firefox?

also open up a cmd as admin and run
sfc /scannow
That fixes system bugs

[deft hinge]

Just chrome, but when I run ComfyUI via the desktop app it does the same thing.

[ornate elk]

hmm are your energy options set to high performance or to balanced?

also do you have an Asrock mainboard? They are the worst to use with AMD CPUs, had problems with that combo in the past too

[deft hinge]

Balanced.
and I am running a: Gigabyte X399 AORUS XTREME EATX sTR4 Motherboard

[ornate elk]

okay

and the your fans dont stop spinning until the frequency is back to normal?

[deft hinge]

Fans wont stop until I shut down or restart the pc. So... Yes... I think? (Not completely pc literate)

[ornate elk]

okay very strange

make sure you runned sfc /scannow ones

[deft hinge]

Just did, no issues detected.

[ornate elk]

do you have a hdd in your system or only ssds?

[calm moth]

Sorry, I forgot to mention that I'm trying to Batch HiRes Fix in Img2Img.

[deft hinge]

SSD

[ornate elk]

so your using the sd upscale script?

[ornate elk]

okay, then im out of ideas. maybe a bios update can fix it or the latest chipset drivers for the mainboard.

[calm moth]

This is what I'm trying to use:

[deft hinge]

I'll do some work on that tomorrow (It's late/early here lol). Might be on tomorrow and let you know how that went.

Thanks for you assistance!

[calm moth]

Does HiRes Fix in Batch just fine, but only uses "Latent" as the Upscaler.

[ornate elk]

where is the hires fix option there?

resize mode is not hires fix

you would need to enable sd upscale script or ultimate upscale extension

there you can select an upscaler

[calm moth]

It doesn't specify being hires fix, but the output is identical to standard txt2img hires fix with "Latent" upscaler

[mental falcon]

thanks, that's what I need.

[calm moth]

Is the ultimate upscale extension HiRes fix?

If so, I'll just install that and don't worry about this.

[ornate elk]

yep either install the extension or use the inbuild sd upscale script

[ripe tapir]

Why are the images I generate with Stable Diffusion so ugly and weird? The colors look strange, and the overall appearance is just bad. Did I mess up the settings? Where exactly is the problem?

[chrome rampart]

Hi guys hhow can I check if i have cuda(driver) installed ? And do I even need to. I have 3060 card

I do always install latest driver set for card via Nvidia site

[slow cedar]

let me see your ui settings

[ornate elk]

Your mostly generate at a lower resolution than the model got trained on

[ripe tapir]

Here is my settings

[slow cedar]

yeah i see the issue LOL

tis the whole thing

[ripe tapir]

No matter which models I use, the generated images always look like this — simplistic appearance and weird colors.

[slow cedar]

im not sure what dimensions cs1o would recommend but i use 832x1216 or 1216x832 usually

set your steps to like 25

then give it a try

i also have my sampling method set to 'euler a'

anyways try that, see how it works

should fix a lot of your issues?

you can also use hires fix

set the denoise to 0.4 tho

and only use 10 hires steps

[ornate elk]

Sdxl models are trained on 1024x1024 and your below that.
Even 768x1024 should look much better

[slow cedar]

it makes images better but takes longer

ah here, this is how i set up my ui, these are pretty good defaults imo

[ripe tapir]

The images have improved, but they still haven't reached the effect I have in mind — the lighting and shadows are still not sufficient

[slow cedar]

uhh if youre getting tired of resetting your defaults everytime you close and open your ui too, go to the settings tab, scroll down to default, hit apply, itll save everything you have set up as your default so you dont have to set up your ui every time

😅

[slow cedar]

you should get adetailer too ngl for the face

[ripe tapir]

How to get it please

[slow cedar]

https://github.com/Bing-su/adetailer

copy and paste that link

in your ui go to uhh

fuck

extensions?

then install from url

[slow cedar]

paste it in the first line, click install, restart ui. your txt2ig and img2img tabs should now both have adetailer

and just enable adetailer, you dont need to change settings for it either

itll fix faces ez

[calm moth]

If you're graphics card is strong enough, you can HiRes Fix. That really boosts the quality of outputs

[slow cedar]

ya i mentioned it earlier

[slow cedar]

right here

anyways, hope that helps

[ripe tapir]

Should I type https://github.com/Bing-su/adetailer.git in cmd to download it?

[calm moth]

Just do it in your WebUI

[slow cedar]

no

copy and paste the link hun

click the green button on the page

see the copy and paste icon?

the two squares?

click that

then go back to your ui and navigate to your extensions tab yeah? so you see a subtab called 'install from url'?

click that

and paste the link there

and just install and restart your ui

you got this

[ripe tapir]

I try this

[slow cedar]

good luck

[ripe tapir]

Thank you

[tranquil gyro]

Anyone here has HiDream running on blackwell? 🙂

[thorny iron]

Why is this happening? ':f

[azure magnet]

Should I add to system enviroment path, comfyui's path?

[woeful knot]

In reforge after a fresh install I keep getting the error "PIL.UnidentifiedImageError: cannot identify image file <_io.BytesIO object at 0x0000022A8AFFDC10>" when inpainting, anyone knows what that's about? It goes away if I change the mask slightly

Didn't do it before the reinstall

[azure magnet]

How many vram i need to set high vram inside comfyui?

[thorny iron]

Could not locate cudnn_ops_infer64_8.dll.

What does this mean? :<

@pure tusk

[ornate elk]

When does that appear?

[thorny iron]

Umm..

It happens when I click Generate, I get spammed with it.

I check the CMD and it says that my cudnn_ops_infer64_8.dll cannot be located.

I've never had this problem before.

I'm using Forge.

I was thinking maybe I need to install Cuda now

[ornate elk]

What's your GPU?

[thorny iron]

4080m

Everything worked fine before.

I don't know what changed this time when installing.

[ornate elk]

okay run the update.bat in forge, then delete the venv folder and relaunch

[thorny iron]

Where is Venv

[ornate elk]

inside the forge webui folder

[thorny iron]

I don't have one

[ornate elk]

thats strange

ah or maybe not

as the whole forge is a portable folder

[thorny iron]

Oooh..

[fierce hollow]

I tried to change the girl's hand in the photo using the Inpaint model: https://civitai.com/models/978482/agfluxfillnsfwfp8
That's the kind of mistake, what's the problem?RuntimeError: mat1 and mat2 shapes cannot be multiplied (4096x64 and 384x3072)

[visual pilot]

Hi, is there any way to automatically save each generation configuration from indexed db to history.txt? Every so often, my entire history is deleted, and I lose all those settings.

[tribal anvil]

How do I find sd-forge-couple im looking on GitHub but I can't find the link

[thorny iron]

So what should I do?

[thorny iron]

Tag Autocomplete: Could not locate model-keyword extension, Lora trigger word completion will be limited to those added through the extra networks menu.
2025-04-13 06:41:55,176 - ControlNet - INFO - ControlNet UI callback registered.
Model selected: {'checkpoint_info': {'filename': 'D:\\[Stable Diffusion 6]\\webui\\models\\Stable-diffusion\\ponyMegaMixXL_v20.safetensors', 'hash': 'af14d758'}, 'additional_modules': [], 'unet_storage_dtype': None}
Using online LoRAs in FP16: False
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
Startup time: 74.9s (prepare environment: 14.9s, launcher: 2.3s, import torch: 36.6s, initialize shared: 0.3s, other imports: 5.3s, list SD models: 0.3s, load scripts: 6.6s, initialize extra networks: 0.1s, create ui: 6.2s, gradio launch: 3.1s).
Environment vars changed: {'stream': False, 'inference_memory': 1024.0, 'pin_shared_memory': False}
[GPU Setting] You will use 91.66% GPU memory (11257.00 MB) to load weights, and use 8.34% GPU memory (1024.00 MB) to do matrix computation.
Loading Model: {'checkpoint_info': {'filename': 'D:\\[Stable Diffusion 6]\\webui\\models\\Stable-diffusion\\ponyMegaMixXL_v20.safetensors', 'hash': 'af14d758'}, '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}
Model loaded in 2.4s (unload existing model: 0.3s, forge model load: 2.0s).
[Unload] Trying to free 3051.58 MB for cuda:0 with 0 models keep loaded ... Done.

Moving model(s) has taken 1.53 seconds
[Unload] Trying to free 1024.00 MB for cuda:0 with 1 models keep loaded ... Current free memory is 9225.19 MB ... Done.
[Unload] Trying to free 7390.16 MB for cuda:0 with 0 models keep loaded ... Current free memory is 9224.50 MB ... Done.
[Memory Management] Target: KModel, Free GPU: 9224.50 MB, Model Require: 4897.05 MB, Previously Loaded: 0.00 MB, Inference Require: 1024.00 MB, Remaining: 3303.46 MB, All loaded to GPU.
Moving model(s) has taken 6.48 seconds
  0%|                                                                                                                                                      | 0/20 [00:00<?, ?it/s]Could not locate cudnn_ops_infer64_8.dll. Please make sure it is in your library path!```

[wild tendon]

CHECK your dms reaal quicketh

[azure magnet]

What is the different between nightly and latest inside comfyui manager?

[thorny iron]

I didn't receive any? :<

[lucid yoke]

dont click on any links

[lucid yoke]

anyone willing to help will do it here. in dm's its often malicious actors

[wild tendon]

hmmmmmm try again

[thorny iron]

I PM'ed you directly. :f

[wild tendon]

i am not malicious i am australian

[lucid yoke]

just a rule of thumb here, we get many scammers

[wild tendon]

i am a scammer

[azure magnet]

What i should choose

[wild tendon]

close

[lucid yoke]

what gpu do you have

hmm

[azure magnet]

4060 ti 16gb

[thorny iron]

Eurotypo do you know about the cudnn_ops_infer64_8.dll issue? :<

[lucid yoke]

the 5000 series needs the nightly but since your on a 4000 series the latest would be fine

[lucid yoke]

hmm probably not but ill read the error log first gimme a sec

[thorny iron]

Thank you so much, I have googled and I'm quite lost.. I can only think I need to install Cuda on its own at this point.

[lucid yoke]

ahh are you using A1111?

[thorny iron]

Forge

[lucid yoke]

hmm gotcha

[thorny iron]

Umm..

[azure magnet]

Sorry i don't understand, nightly is something releated only for 5000, other gpu should use latest, right?

[lucid yoke]

you could also use nightly but "nightly" is the unstable version

bugs ,newest features etc

[thorny iron]

webui_forge_cu121_torch231

This one here in the guide. :<

[lucid yoke]

https://github.com/SYSTRAN/faster-whisper/discussions/715#discussioncomment-10903540

its apperanly a Torch vision/torchaudio issue

hmmmm

[thorny iron]

Err..

[lucid yoke]

can you run the update.bat?

should be easier to repair i think

[wild tendon]

bats cant run they fly

[thorny iron]

Yeah I upd-.. can you leave please? :f

[lucid yoke]

the weird stranger? just ignore him

[thorny iron]

I updated and it just says it's latest, nothing happened.

[lucid yoke]

hmm

[thorny iron]

I'll try what the post suggests.. umm

https://github.com/SYSTRAN/faster-whisper/discussions/715

Over here.

[lucid yoke]

yeah but

[wild tendon]

rude

[thorny iron]

But?

[lucid yoke]

idk where you need to place the DDL in forge

[thorny iron]

Oh.

[lucid yoke]

ill try looking

[wild tendon]

says the guy that eats bats

[slow badge]

Is kohya or forge working with rtx 50 officially? or not yet

[wild tendon]

counterargument:

[thorny iron]

I'll see if this helps first, then try find where to put the files with the other post.

[wild tendon]

i am a mirror

[azure magnet]

Ok, last question, so nightly become latest after a while in testing?

[wild tendon]

ill point and laugh with you

[thorny iron]

Please this is tech support.. :<

[lucid yoke]

yes after some time

[wild tendon]

god damn i was just tryna be nice

what else you fool

[lucid yoke]

@wild tendon@magic hearth

@pure tusk sorry to bother but we got two trolls using up #🤝｜tech-support

[wild tendon]

typo point and laugh

i was doing it before you were born

more like 10/10 fact

[thorny iron]

Thank goodness blocking multiple people combines it all together to save the excess scrolling. :f

[wild tendon]

fucking capitalism agian

[thorny iron]

I've posted into the Issues at https://github.com/lllyasviel/stable-diffusion-webui-forge

Hopefully someone will be able to tell me what's going on, then I'll share it here.

@lucid yokeWhat the fuck

[lucid yoke]

whats up

[thorny iron]

... apparently it's because I had square brackets in Forges path..

[lucid yoke]

oh

[thorny iron]

[ ]

[lucid yoke]

lmao thats unique, i didnt see it

yeah windows doesnt like that

[thorny iron]

Well, it works now.. it USED to be fine though, I guess something this version made it unhappy.

Thank you for helping. You too @ornate elk

[thorny iron]

So I put an old image through and reproduced it. It's 99% correct but the face is slightly different.

Is that due to updates with the tool?

Or am I missing something.

Prompt hasn't changed at all.

[lucid yoke]

Are you using euler a?

[thorny iron]

Oh, no. DPM++ 2M

Hbbbt, frustrating.. faces are just all messed up now.

[thorny iron]

@lucid yoke

Apparently this was checked by default.

All fixed. ';d

What does this do though.

And is it helpful to use for hyper-realistic / real life pictures?

[ornate elk]

Its okayish for realism when used with codeformer.
But I wouldn't use it and instead use Adetailer

[thorny iron]

I see. Thank you so much.

[brazen needle]

guys is did anyone get the " Torch is not able to use GPU" error when installing sd?
zero coding expirience btw

[ornate elk]

What's your GPU?

[brazen needle]

rx6700 xt (i know amd aint optional 💔 )

[thorny iron]

Does your path have [ ] in it? ;d

[brazen needle]

whats my path? 😭

[thorny iron]

(Square Brackets)

As in c:\users\blahblah

[brazen needle]

oh

no

i use it on a lot of folder but not on this path

[thorny iron]

If you installed the program into a folder with a name that has weird things in it, it'll cause problems, which I've discovered today. >xp

[brazen needle]

dont think thats the problem

i guess its my gpu

[thorny iron]

https://www.reddit.com/r/StableDiffusion/comments/z6nkh0/torch_is_not_able_to_use_gpu/

[brazen needle]

yeah i tried that fix

atleast the first one but still nothing

[thorny iron]

Hmm..

What version of Stable Diffusion you using

[brazen needle]

automatic 1111

[thorny iron]

Did you go via this route

[brazen needle]

na i didnt

i'll try again with directml

[thorny iron]

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

Check this page out and pick what you want to use under AMD section.

[brazen needle]

perfect thanks

[thorny iron]

I am very new to this myself

I recommend you use Forge, because it's really easy to setup.

[brazen needle]

i think i read somwehere forge doesnt work with amd

might be wrong tho

[thorny iron]

Check these sections.

[brazen needle]

bet

wheres the difference between a1111 and forge?

is forge as user friendly as a1111?

[ornate elk]

Then you need to follow the AMD guides.
First link of the pinned messages

[thorny iron]

Forge is friendlier than A1111 in my opinion.

[ornate elk]

Dont use directml

Go for zluda

[brazen needle]

oh ok

[brazen needle]

nice i'll give it a try

[thorny iron]

It's the same UI almost, 95%

[brazen needle]

do yall think i will get decent generation times with my gpu?

[thorny iron]

It also has auto GPU detection apparently which might help your card.

[ornate elk]

When you use zluda yes

With directml nope

[thorny iron]

Forge apparently generates a lot quicker than A1111, I think I've noticed it too.

[brazen needle]

good to know

🙏 ❤️

[thorny iron]

At least, it does so with far less hickups.

Since I moved to Forge I'm much more happy.

I could only 90% recreate my old outputs from A1111 however, but that might be because I'm still new and some settings are active that I am unaware of..

So be warned if you have already established a style you like, you may need to tweak it.

Pro Tip: If you are doing anime / cartoon stuff, turn off Restore Faces, there is a search function in Settings. >xp

Is there a way to cancel a Generation quickly?

When I click cancel it takes like 2 mins sometimes for it to end.

I'm assuming that's because it's slowly trying to dump all the shit in ram?

[brazen needle]

this might sound stupid

but can i install the pro version and the adrenaline version drivers for amd at the same time?

[thorny iron]

.. humm..

.. I think you can, just not run them both at the same time.

[brazen needle]

guess i'll jus try

[thorny iron]

Pro's for Media, Adrenaline's for Gaming.

[brazen needle]

but what if you use both

Ichigo cooked

[ornate elk]

Double click on interrupt

Sometimes triple click

[thorny iron]

I do that, it still takes a while though.. perhaps my images are a little big. I notice Python goes up to like 40gb.

[ornate elk]

Hip SDK can be Installed and doesnt interfer with the basic Adrenalin Driver

Dont install the pro drivers

[thorny iron]

Listen to CS1o not me. >xp

I guess it makes sense that since Stable Diffusion isn't exactly a professional tool and isn't regulated as such..

[brazen needle]

so i did everything as explained so far

but where it should now work i get ("Torch not compiled with CUDA enabled")

any idea what the problem could be?

[ornate elk]

Can you show me.the full cmd log?

[ornate elk]

Open up a cmd and run
Pip cache purge
Then delete the venv folder and the .zluda folder in the stable-diffusion-webui-amdgpu-forge folder.
Then relaunch the webui-user.bat

[brazen needle]

rocBLAS error: Cannot read C:\Program Files\AMD\ROCm\6.2\bin/rocblas/library/TensileLibrary.dat: No such file or directory for GPU arch : gfx1031

rocBLAS error: Could not initialize Tensile host:
regex_error(error_backref): The expression contained an invalid back reference.
Drücken Sie eine beliebige Taste . . .

thats the error i get now

man i should have went with nvidia

lol so that was a scam

mf indians man

[brazen needle]

yeah so that didnt work out :c

[half island]

If it takes 2 min to end, you most likely attempt to generate something too large, and it uses ram as video memory most likely, which makes generations 40x slower

[half island]

You forgot to make the rest as backslash from this point and out

[brazen needle]

but i didnt type anything myself

thats what happens when i open the batch file

[half island]

Send the batch file in a zip here

[half island]

Open encironment variables.

[brazen needle]

done

[half island]

Click environment variables, and check in path in the second window to see if this part is there.
C:\Program Files\AMD\ROCm\6.2\bin/rocblas/library/TensileLibrary.dat

[brazen needle]

no

should i add that path?

[half island]

Yep. But should be more like this C:\Program Files\AMD\ROCm\6.2\bin\rocblas\library\

But first, make sure it's even present. Paste that into explorer adress bar and see if it hops into that folder

[brazen needle]

hm still same error

but the path is there

[half island]

Show a screencap

[half island]

I meant from path in environment variables

[brazen needle]

oh mb

[half island]

The whole window please. Like this.

[half island]

Scroll to where it says zluda

Or, is it vlad's sd you attempt to run?

[brazen needle]

no its this one

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-forge-webui-with-zluda

i tried installing forge webui with zluda

[half island]

And you followed every step, incluiding restarting your pc? Cause on vlad's git, he stated that zluda is still in very experimental stage. And may not work for everyone.

[brazen needle]

yeah i did everything

[brazen needle]

so i might just be fucked?

direcml is still an option ig but i heard that sucks

[half island]

I'd try to reinstall everything and try again. If it won't work the second time, driver on recommended, and everything, i'd just try sticking with regular rocm.
I might try zluda on my steam deck for kicks one of these days to see how that will go lol.

[brazen needle]

what is regular rocm? :c

[half island]

Zluda is an additional layer that uses rocm, or cuda translation layer to perform cuda tasks.

Rocm is like zluda, but the basic translation layer to what i just wrote. And Zluda from what i've gathered is another layer that appears to allow the amd card to perform near natively against cuda.

[brazen needle]

gotcha thanks

[thorny iron]

I am yes my card is 12gb and I go to 40gb

[ornate elk]

That bottom path needs to be removed

[ornate elk]

You need to replace the gfx files

[brazen needle]

i'll try reinstalling everything now

hope it works out

[brazen needle]

still this error, could it be that stable diffusion just wont work on my system?

[ornate elk]

It will work

But something Is still not correct

[brazen needle]

so chatgpt is helping me out rn

[ornate elk]

Have you downloaded the correct gfx files?

[brazen needle]

its telling me to reinstall the tensil libary

[brazen needle]

im trying to redownload them

but..

[ornate elk]

Please don't. Chatgpt doesn't hage a clue about this specific install

[brazen needle]

https://github.com/likelovewant/ROCmLibs-for-gfx1103-AMD780M-APU/releases/tag/v0.6.2.4

how to i download from here 😭

[ornate elk]

Scroll down

Click on assets

[brazen needle]

oh

[ornate elk]

Download the correct zip

[brazen needle]

i feel so dumb rn

lmao

[ornate elk]

Then do the steps like the guide says.
And then reboot the PC.
Then relaunch the webui-user.bat and it should work.

[brazen needle]

yes i'll reboot now

wish me luck

it worked out

wow finally

u my goat ❤️

[ornate elk]

No problem 🙂

Compiling can take a bit at first. Just let it run

[empty cliff]

HEEEEEEEEEEEEEEEEEEEEELP

I need to install flash-attention2 so I can use dreambooth via A1111 with my 5080

https://github.com/Dao-AILab/flash-attention/tree/2f9ef0879a0935c3ca852f7a6a7b7a9c24f41e96

Flash attention page says to install it using pytorch

I did some digging and it appears that the only pytorch compatible with this version of flash-attention is pytorch 2.2

The problem is that installing a specific pytorch version is a pain in the anus because it usually requires providing a URL

Can any braniac out there PLEASE help me install Pytorch2.2+cu128 on my A1111

[ornate elk]

hey dont try to install Dreambooth, its outdated and breaks A1111.
Better use Standalone Tools for Training like OneTrainer or Khoya_ss

[empty cliff]

Do you know if Khoya currently has support for Cuda 12.8?

I remember not using kohya for dreambooth because they don't let you pick the amount of regularization images to use per training subject. What can I say, Im old-school

[ornate elk]

idk, but the last update was 2 weeks ago

[empty cliff]

I'll go check

[ornate elk]

https://github.com/bmaltais/kohya_ss

[empty cliff]

Now I do see that DB for A1111 was updated about a month ago

[ornate elk]

what do you mean?

[empty cliff]

It means someone is still updating it, and it does seem to work fine

[ornate elk]

A1111 also doesnt got updated since a few months

[empty cliff]

Ya but the extension did

[ornate elk]

extensions are not maintained by auto1111 devs

[empty cliff]

However it won't run because flash-attention doesnt work on blackwell

I think A1111 doesnt get updated because its a finished product - it's literally perfection

Any changes would be detrimental at this point

It took an act of god to find an xformers that works on 50 series and the only remaining piece of the puzzle is a flash-attention version that works for it. If anyone has any idea on how to do this, a brutha would be in debt

[ornate elk]

yea the support for the 5000 series is pretty hacky atm

[empty cliff]

Oh btw I found a help article to run kohya on 50 series, I guess that will come in handy if this doesnt work: https://github.com/bmaltais/kohya_ss/issues/3096#issuecomment-2730229439

@ornate elk Brutha I know ur the OG you probably know this

You know how A1111 typically runs on its own venv

The blackwell one basically created a directory that it uses to run its own python, similar to a venv but more improvised

I followed instructions and used pip install pytorch etc etc but this is not installed on the venv, its installed in roaming/appdata/etc

Is there a way to force my A1111 to run from this new pytorch that I installed and not from the venv

Check out how they did it

Dont mind the name, I just nicknamed it A2222

Heres inside the system folder, essentially its a python container

do you think that going into my AppData and copying and replacing the python folder in here would work, or do you know how to reroute the installation to the right torch etc?

[ornate elk]

why not installed it into the venv of a1111?

or why not made a seperate venv and linked it in the bat?

[empty cliff]

Thats a great question honestly IDK and im scared to mess with it

because

everything breaks the blackwell integration, they did their own makeshift venv

but yeah this one doesnt have a venv

and trust me I tried over and over to get it to work on a venv and it doesnt

U know what let me do some testing first

Doesnt work, but I feel like Im close to solving it

So I did what this guy did:

So I just ran the pip install etc. from cmd

then I copied the newly installed torch into the A1111 python folder (from my appdata/roaming folder)

left xformers the same since they dont hurt nobody but now I get this error

So its essentially not recognizing my cuda and now I'm stuck in dependency hell bc I dont even know if this python version can run that torch

On top of that idk how to properly install that cu128 version of torch so my A1111 can read it

[ornate elk]

thats for the default auto1111 version not for your modified blackwell version.
If yo uwant to just get auto111 working with a 50 series do what the comment you posted shows.
But make sure you git clone the normal auto1111 and then run a cmd inside from the auto1111 folder and before inputting any command run this to get in its venv:
venv\Scripts\activate.bat

[empty cliff]

do I run cmd from the venv folder so it installs it there?

Oh I see

run that command first and then do the pip install

[ornate elk]

you have to click in the explorer adress bar of the auto1111 folder and there type cmd and hit enter, then the cmd is in the correct folder.
then run the venv\Scripts\activate.bat to get into the venv and there run the pip install torch.-..

[empty cliff]

Nicee hoping for the best and thanks

Typical domino effect BS:

That happens when trying to render a generated image, willing to bet some dependency issue. Jesus what an absolute mess switching to a new generation causes

Im gonna have to stick to the makeshift venv one and find a way to inject that pytorch version in there, even then there is no guarantee that its flash-attention will work on training. Appreciate the help anyways and if anyone has any insight it is highly appreciated

Do yall know if xformers and flash-attention need each other to work?

Back in the day they used to work independently of one another, but now flash-attention comes inside xformers or am I smoking crack?

[marble heron]

help me guys...

Launching Web UI with arguments: --lowvram --precision full --no-half --skip-torch-cuda-test
Traceback (most recent call last):
File "Q:\stable-diffusion-webui\launch.py", line 48, in <module>
main()
File "Q:\stable-diffusion-webui\launch.py", line 44, in main
start()
File "Q:\stable-diffusion-webui\modules\launch_utils.py", line 465, in start
import webui
File "Q:\stable-diffusion-webui\webui.py", line 13, in <module>
initialize.imports()
File "Q:\stable-diffusion-webui\modules\initialize.py", line 15, in imports
import torch # noqa: F401
File "Q:\stable-diffusion-webui\venv\lib\site-packages\torch_init_.py", line 122, in <module>
raise err
OSError: [WinError 193] %1 не является приложением Win32. Error loading "Q:\stable-diffusion-webui\venv\lib\site-packages\torch\lib\cublas64_11.dll" or one of its dependencies.
Для продолжения нажмите любую клавишу . . .

i`m stuck...

[brazen needle]

is there a fix forlora and checkpoint previews not showing?

[verbal burrow]

Do I place Nova Anime XL under the Stable Diffusion folder?

[open raven]

Can i use a 6600xt using the directml webui? im seeing some things where i cant but also can so im just confused.

[silver briar]

can someone help me fix this?

my reforge wasn't working and i did a clean re install and now im getting this all of a sudden

[tidal steppe]

Start by ensuring PyTorch 2.3.1 and CUDA 12.1 are correctly installed, then clean your environment and reinstall the package. If the issue persists, check for JSON schema errors or try a different package variant. Let me know if you need further assistance!

[ornate elk]

Yes but you should better use zluda instead of directml.
Its faster and needs less vram

[open raven]

Do you have any videos u could link?

[ornate elk]

Nope there is no good video, thats why I made the AMD Guides.
You can find them in the first link of the pinned messages

Text based but if you have any question feel free to ask

[open raven]

Will do thanks.

one more thing tho, do u recommend normal webui or comfyui, or others?

[ornate elk]

With 8gb vram your better with Forge Webui or Auto1111

Its also easier to get into than Comfyui

[open raven]

yeah comfyui is quite confusing but im getting used to it

i think ill try forge webui

[open raven]

@ornate elk im doing the normal auto1111 but im a bit confused could u help?

wait i might have figured it out actually

[hybrid magnet]

I have already installed Python latest version. So what i do

[ornate elk]

Which version?

You need 3.10.11 64bit for the setup.
3.13 wont work with ai stuff

[marble heron]

@ornate elk

can u help me pls

[marble heron]

with this

@vocal burrow @echo hill >< im gonna leave in soon...

[ornate elk]

What's your GPU?

And please change your profile pic

[sand marlin]

Hi all. Can i ask question about model training in sdxl whit my dataset here?

[marble heron]

GTX 1070

[brazen needle]

CUDA out of memory. Tried to allocate 30.66 GiB. GPU 0 has a total capacity of 11.98 GiB of which 6.96 GiB is free. Of the allocated memory 2.70 GiB is allocated by PyTorch, and 145.79 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

what does this mean

as soon as i change the default width and height this problem occurs

i also cant upscale anything

[ornate elk]

Then follow the install guide from the pinned messages

Try Forge Webui or Auto1111
Also don't add --skip-torch-cuda-test --no-half --lowvram to the launch args.

[ornate elk]

That means your running out of vram by about 30gb xD

What settings did you tried?

[brazen needle]

what can i do about "OutOfMemoryError: CUDA out of memory. Tried to allocate 15.62 GiB. GPU 0 has a total capacity of 11.98 GiB of which 4.14 GiB is free"

bro what

[ornate elk]

If you use hires fix you need to Always set the hires steps to 10.
And then also it depends on which resolution you used

[brazen needle]

uhm dpm++2m karras

no hires fix

default resolution

[ornate elk]

Share your txt2img settings

[brazen needle]

like the whole window?

i can generate regular images but as soon as i start inpainting or upscaling i get the out of vram error

[ornate elk]

Yes because you didnt used the right settings.
Try enable hires fix.
Set the hires steps to 10,
set the denois to 0.5
Set the upscaler to Resrgan4x+ Anime6b
Set upscale by 2
At the bottom enable Never OOM for Tiled VAE only

For img2img Inpainting:
Set the Inpaint area to "only masked"
Set the denois to 0.5
Set the resolution to 768x768

[brazen needle]

bet i‘ll try that 🙏🏽

[vast crest]

hi, are there any problems on google colab for creating Lora?

[sand marlin]

Hi all, need help please.
Trying to train a Stable Diffusion XL (SDXL) DreamBooth model on Windows 11 with an RTX 4090 (24GB VRAM) using kohya_ss GUI and sdxl_train.py.

Training at 1024x1024 resolution with:

train_text_encoder = True

mixed_precision = bf16

no_half_vae = True

gradient_checkpointing = True

mem_eff_attn = True

use_8bit_adam = True

Dataset: 51 images × 15 repeats (765 steps per epoch), targeting 300 epochs.

Problem: always getting CUDA out of memory on the first training step, even with batch_size = 1.

Tried:
PYTORCH_CUDA_ALLOC_CONF values like max_split_size_mb=64, 128, 256, 512
cache_latents_to_disk
disabling text encoder training
switching attention methods
8-bit optimizer

Goal: make SDXL training run on a single RTX 4090, even if it's slow. Once it managed to run fully in system RAM (very slow but completed), looking for a way to replicate this reliably.

[brazen needle]

any idea why it keeps skipping nsfw preview image downloads from civitai?

i checked the box but they still get blocked

[marble heron]

not work, and i`m use webui auto1111

U can, give theory how fix that. And in next day i`m try this. Now, i gonna sleep

[ornate elk]

Delete everything and try using the install guide from the pinned messages. First link.
There follow the Auto1111 with Nvidia Guide.
It should work.

[brazen needle]

can i use a pony based checkpoint with a illustrious lora?

[ornate elk]

You can try.
It can work or not.
Same goes for the other way round

Best is to use the same versions

[brazen needle]

do you have any tips on how to consistently get better generations

9/10 i just get weird slop and i dont know how to fix it with inpaint

[wild tendon]

imposible

[brazen needle]

so just keep trying until something works

damn

[open raven]

edit the webui python part and add the path to your python

thats what fixed mine

[ornate elk]

Do you have an example?

[wild tendon]

sorry but its impossible to not get slop

[brazen needle]

guys im stuck on the cuda out of memory again

most people here probably dont have 25gb vram so there must be another solution right? 😭

[ornate elk]

Show your txt2img settings

[brazen needle]

one sec

thats nothing crazy right?

[ornate elk]

Do you run anything in the background like Wallpaper engine, games or multiple Browser tabs?

[brazen needle]

ye

all of the one mentioned lol

[ornate elk]

Really? xD

[brazen needle]

well atleast theres no games open

but yeah a lot of tabs

and wallpaper engine, discord etc

[ornate elk]

Wallpaper engine is the biggest problem.
It can cause freezes and PC crashes when its running while stable-diffusion runs

Even if its set to paused

[brazen needle]

turned it off

but still not working

can i change the vram usage?

[ornate elk]

You have to relaunch the webui and then enable the Never OOM for tiled VAE at the bottom

[brazen needle]

maybe because its trying to use all 12gb

[ornate elk]

No, it takes what it needs to perform the task

But you usually dont use resize to in img2img, its to vram heavy

[brazen needle]

i was following a workflow tutorial xD

[ornate elk]

Normaly you would Upscale using hires fix in txt2img. Or SD upscale script in img2img

[brazen needle]

but doesnt hires fix change the whole composition

[ornate elk]

You can define that with the denois strength

[brazen needle]

so lets say i have a fairly decent image but want to tweak some things and upscale it

should i send the image to img2img or inpaint?

[ornate elk]

That depends on what you want to do

[brazen needle]

upscale it first

[ornate elk]

Better regenerate it with hires fix

[brazen needle]

hm okay

[ornate elk]

Or send it to img2img and use the SD upscale script.
But the best is to first use hires fix aa you want a larger resolution to begin with

[brazen needle]

do i need to download that script first?

[ornate elk]

Nope its already integrated

Its at the bottom of img2img.
But if you use it you have to set the resolution to a square. Like 512x512 if you use 1.5 bases models and 1024x1024 for sdxl based models.
Then set the denois to 0.2
And select an upscaler at the bottom

[brazen needle]

bet

ill try that

do you think getting a cloud gpu is worth it?

i mean my generation speeds arent terrible but def not optimal either

[ornate elk]

If you want to do image or video generation professionaly yes, but IG you just want to generate some nice images or Wallpaper then nope

[brazen needle]

yeah makes sense

but its not really that expensive isnt it?

basically the same price as if you went with a website sub

[ornate elk]

Yea but you have to set it up yourself. And after a reboot its all resets.

Not every provider does that, but some do

[brazen needle]

that sucks

so i tried your settings

now the image renders up to 50% and then i get the error again

[ornate elk]

What have you tried? The SD upscale script?

Or resize by 2?

[brazen needle]

your hires fix settings

resize by 2

[ornate elk]

In txt2img?

[brazen needle]

yea

[ornate elk]

Have you enabled Never OOM for Tiled VAE at the bottom?

Thats needed

[brazen needle]

yeah

still the same

i dont get it

[ornate elk]

Can you Screenshot everything?

[brazen needle]

sure

[ornate elk]

I used a 6700xt 12gb a year ago and it worked with these settings

[ornate elk]

You didnt set hires steps to 10

[brazen needle]

oh true

[ornate elk]

And denois 0.5

And where did you get the fidelity FX upscaler lol xD did forge added it?

[brazen needle]

hm

still stuck at 48%

[ornate elk]

Check the cmd

Upscale steps take longer

[brazen needle]

it failed

[ornate elk]

Is Wallpaper engine and everything thats not discord or SD closed?

[brazen needle]

ye

[ornate elk]

Try Resrgan4x as Upscaler

[brazen needle]

same thing

its so weird

always stops at 48%

[ornate elk]

Can you show me the error messages or the cmd log?

[brazen needle]

and why is it trying to use 25gb

OutOfMemoryError: CUDA out of memory. Tried to allocate 25.00 GiB. GPU 0 has a total capacity of 11.98 GiB of which 6.08 GiB is free. Of the allocated memory 3.60 GiB is allocated by PyTorch, and 161.85 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

Time taken: 38.0 sec.

[ornate elk]

And the cmd?

[brazen needle]

noob question can you make yourself vulnerable with a cmd log?

i feel like im doxxing my self lol

[ornate elk]

Depends on the information it contains.
But with the cmd log from these webuis you can't.

[brazen needle]

got it

[ornate elk]

Can you relaunch the webui.
Then try these settings:
Sampler euler a
Res: 768x1024
Steps: 25
Cfg 7
Hires fix:
Hires steps:10
Denois 0.5
Upscale by 2
Upscaler REsrgan4x+
Never OOM Tiled VAE

[brazen needle]

on it

[ornate elk]

Oh you used an 1.5 based model then change the resolution to 512x768
And try again

[brazen needle]

generation is working so far

but damn thats slow

[ornate elk]

Yea that resolution was for sdxl based models

[brazen needle]

damn

it crashed again

[ornate elk]

What was your GPU again?

[brazen needle]

ill try the other res

[brazen needle]

ye

still at 48%

should i try using --medvram

[ornate elk]

That doesnt work in Forge

[brazen needle]

oh

[ornate elk]

It should use it when needed

[ornate elk]

Crashed again?

[brazen needle]

yup

[ornate elk]

Damn, you should show me that with screenshare tomorrow

I think we can get it fixed

[brazen needle]

bet thanks 🙏

[ornate elk]

How much RAM do you have?

[brazen needle]

32gb

[ornate elk]

Ok good

[brazen needle]

its really weird

[ornate elk]

Yea it is

[brazen needle]

must be something with hires fix cuz without i dont get the error

[ornate elk]

Upscaling needs more vram but it should work on a 12gb GPU

[brazen needle]

yeah easy

i upscaled 5min video footage already

so a image should be possible

[ornate elk]

I'm off now, tomorrow we can check again if you have time

[brazen needle]

ye thx

🙏

[ornate elk]

Np

[pulsar vault]

is there any guide anywhere or something i can google to be able to use SD with a 5000 series card? installing a1111 now but im hearing i shouldnt bother at this point

[gritty hornet]

I am having the same issue, let me know if you find a solution

[pulsar vault]

i would 100% not click on that link

[gritty hornet]

I clicked and joined another discord, lol

[pulsar vault]

change your password

[gritty hornet]

discord?

just clicked and left

so I wasted a few days of sleep just get may hands on 5000 series and built a new pc to support it and I still cannot generate naughty anime pics?

[gritty hornet]

fixed install 12.8 worked for me

pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

[pulsar vault]

ill install and see what happens

[gritty hornet]

i was getting this error before. (venv) C:\stable-diffusion-webui>python -c "import torch; print(torch.file); print(torch.version.cuda); print(torch.cuda.get_device_name(0)); print(torch.cuda.is_available())"
C:\stable-diffusion-webui\venv\lib\site-packages\torch_init_.py
12.8
NVIDIA GeForce RTX 5070 Ti
True

alternate was to download complie and do bunch but that was to much work

@pulsar vault let me know how it goes

[pulsar vault]

oooooooooooo it works and its speeeeeeeeeeeeeedy

[balmy plover]

hi

[placid lynx]

is openpose compatible with forgeUI, or do you need a different ui to use it

[ornate elk]

It can work with forge ui

[placid lynx]

okay, because when I've tied using it nothing happens and the controlnet unit is completely ignored

[ornate elk]

try use the controlnet union++ promax model

[placid lynx]

oh thats the one I was using I think

[placid lynx]

maybe I didn't enable it properly? This is what the controlnet module looks like, and here was a sample generation

@ornate elk ?

[ornate elk]

You have to set Preprocessor to None because your input image is already processed into a pose

Then it should work

[placid lynx]

so like this?

it looks like it works, I'm gonna test it a few times to make sure

thanks

[ornate elk]

no problem 🙂

[placid lynx]

hmmm, it has stopped working

does the prompt need to describe the pose too?

[ornate elk]

It helps

[placid lynx]

does the openpose consider the orientation of the image or just the specific pose the character is in?

[ornate elk]

To size itself i guess, so better enable crop and resize

[cold bronze]

I'm using the automatic1111 ui. I get identical (and terrible) results no matter what VAE I use. I don't think it's actually using the VAEs.

[cold bronze]

This is "woman with blue hair" with the model v1-5-pruned-emaonly.safetensors and a size of 512x512. It's not supposed to look like that, is it?

[orchid forge]

Hi hi, I'm completely new to this stuff, having trouble with startup. I followed the WebUI install guide at github for AMD Comfyui with ZLUDA as well as AMD Automatic1111 with ZLUDA

I've got a 7600 GPU and 7600 processor, 32GB ram, Win 10

Can anyone help me troubleshoot?

I did ComfyUI first, did not open browser. I rechecked the steps, restarted again, still nothing, so I downloaded Automatic1111 and experienced the same symptoms. Everything seems to be downloaded, path is updated, I was a bit confused about which drive to save to since it just says "your drive" at first and then specifies "C Drive" for the last bit where you copy the 3 .dlls from your unzipped ZLUDA folder? So I've got a copy of each on both C: and E: I tried E: first.

[orchid forge]

Nope

No responses yet

[lucid yoke]

Spam

@orchid forge dont join them

[orchid forge]

That's so rude :/ Ugh

Thanks for the heads up, Euro

[lucid yoke]

Cs1o is the amd expert, he pops in a lot

No worries

[orchid forge]

I figured I'd give it a night haha

I understand nothing about this. Usually a much better troubleshooter. Sorry in advance for being a n00b

[lucid yoke]

No worries, ai is a pretty new and difficult field

[ornate elk]

Don't use the 1.5 EMA pruned model its just bad and old.
Go on civitai.com and check for newer models. Checkout Juggernaut or Juggernaut XL as start

[ornate elk]

Hey, can you show me the environment path entries?
Also remove C:\zluda from E.

Then show me the full cmd log of Auto1111 when launched with the webui-user.bat

[orchid forge]

I have SD-Zluda in both C: and E:, the other folder is ZLUDA and only in C:

Both SD-Zluda folders contain both ComfyUI and Automatic1111 installs

And I'd love to show you the full cmd log but I don't know how to make it persist, it automatically closes after run

Googled it. Wait one, rerunning

Still running, but that's the start

Of course it worked this time... Sweet.

💀

Thank you for the effort.

[cold bronze]

I've tried other models, and I don't generally get good results. Also, shouldn't the result still change with the VAE?
Is there some specific model and seed and prompt I can use to test my system?

[cold bronze]

That one does seem to work. I think it's just ones with baked VAEs that work on my machine.

[fair oxide]

How do I stop Auto1111 from auto Updating? Everything works fine when I fresh install SD+Extensions+Xformers, but breaks after closing and reopening SD.
A1111 keeps updating and destroying my Nightly/Dev build for 5080.
Im using the Blackwell Installer on A1111 Github. https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/16818

[ornate elk]

The VAE doesn't affect the image its only for colour correction

[ornate elk]

Which extensions do you have installed?

[ornate elk]

You need different VAE files for different model types too

[ornate elk]

So auto1111 works now?
And what about Comfyui?

[fair oxide]

Even without Extensions, it does the same thing. Automatically updates and removes the 12.8 Nightly update that allows rtx 5000 gpus to use it.
So we can say, No extensions now.

[ornate elk]

What's in your webui-user.bat?

[fair oxide]

Nothing now. Clean.
I tried the ---skip-install arg but that doesnt work either.
Every time I run Stable, It has to be a fresh install, or it updates, and removes the Nightly 12.8 for rtx 5000

[ornate elk]

Hm okay, then either add the commands to install the nightly into the webui-user.bat or change the requirements_version.txt
You can add there the exact version of torch

Or remove the entrys there of the old version

[fair oxide]

A bit confused here ngl. How do I tell A111 to always use pytorch 12.8 in requirements version? (ive never edited that before)

[ornate elk]

In not at the PC to check the file

[cold bronze]

It looks exactly the same.

[fair oxide]

Just got done trying several things but none worked. How would I add the commands to install the nightly into the webui-user.bat ?

No matching distribution found for torch==12.8

[ornate elk]

Torch isnt 12.8 thats the cuda Version it is for

You See the torch version at the bottom of the webui after you installed nightly

[ornate elk]

Yea sometimes a VAE is the same thats already build into the model

[ornate elk]

What's your GPU btw?

[ashen loom]

Hey everyone! Thank you so much for helping me over the past year! I wanted to ask how difficult it would be to set up my desktop to use virtually? Like from my phone.

[half island]

Hoi, I just reinstalled windows, installed python, conda, pip, python, cuds toolkit and git, but when I try to boot up comfyui's existing Ben on another drive, it suddenly complains none of the modules existed? Is there a reason for this? As I thought that was the entire point. Contained virtual env to where it wouldn't need dependencies installed elsewhere.

[fair oxide]

@ornate elk does higher cuda version cause issues with lower cuda supported python libs. say 12.8 cuda and py libs of cuda 12.6

[ornate elk]

yes if some librarys doesnt support them they will not work

[ornate elk]

Hey, take a lok at RustDesk. Its a Remote Desktop App where you can easily control the PC with the Phone or an other PC.

[fair oxide]

yk i think tensorflow require 12.3 and there is no 12.3 in py torch. what to do in such a situation? how do i manage so many cudas?

[ornate elk]

What are you trying to setup?

[balmy plover]

Hey everyone, I’ve been trying to set up Python and CUDA for my AI setup, but I’m running into some issues. Here’s what I’ve done so far:

I installed Python 3.10 (made sure to check "Add Python to PATH" during installation).

I added necessary environment paths for CUDA and Python, including Scripts and bin folders.

I confirmed CUDA is installed and working by running nvcc --version (showed version 11.3).

I upgraded pip with the command python -m pip install --upgrade pip.

I tried installing ControlNet using pip install controlnet, but ran into an issue where Python isn’t being recognized in PowerShell, even though it's installed.

I’ve checked the paths and reinstalled Python, but PowerShell still doesn’t recognize it. Any thoughts or similar issues? Appreciate any help!

i can provide screenshots if needed

[ornate elk]

Hey, why are you trying to setup everything manually?
Just Follow the install guides of one of the many webuis.
Check the first link of the pinned messages

Or did you tried to setup Automatic1111 ?

[balmy plover]

the first link of the pinned messages?

automatic1111

[ornate elk]

In this channel click on the Pin needle icon

[balmy plover]

sorry to be a dope, cant see your pin needle icon

[ornate elk]

Are you on phone or desktop?

[balmy plover]

desktop

[ornate elk]

Top right then

[hazy pumice]

I don't really know what I am doing wrong but for some reason I can't find the 'models' folder for SD SwarmUI. I just redownloaded and installed the software and after that I also downloaded a fresh model and lora but I don't know which directory to place them in. Help? Thank you!!!!

[ornate elk]

here

and here the link:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

You dont need to setup cuda manually

just follow the install guide for Auto1111

[ornate elk]

Hey, its SwarmUI\Models\Stable-Diffusion and for Lora its SwarmUI\Models\Lora

[hazy pumice]

mmm... thank you for the reply but I'm just not seeing these folders. Do I need to create them?

[ornate elk]

nope normally they should be already there

[hazy pumice]

weird.....

This is all I've got

no models folder 😦

[ornate elk]

you have to launch it once

[hazy pumice]

oh

😛

[balmy plover]

@ornate elk ty very much, i will try this later tonight

[ornate elk]

no problem 🙂

[fair oxide]

Got everything Extensions+X formers working on rtx 5080 by freshly cloning the Dev branch of Stable Diffusion. Then I installed a Beta version of XFormers. (lemme know if I should write out the instructions for other ppl, I'm a bit shy)

[ornate elk]

oh great!
If you dont mind write them down, it can help a lot of 50 series users

i then can also add them to the guide list

[ornate elk]

@brazen needle hey if you have time we can check your install

[brazen needle]

could we maybe do it in an hour?

[ornate elk]

sure, dm me then

[empty cliff]

Does any kind soul out there know where the A1111 settings are saved to?

Im talking all settings in general

Im currently experimenting with many A1111 builds and would love to be able to just back up the settings

Also seeing plenty of Blackwell users here, nice

Have any blackwellers been able to train using those versions?? ^^^^

[ornate elk]

config.json and ui_config.json

[empty cliff]

CS1o may I ask u a question bro

Are you a human? Cuz Ive been here since 1.5 dropped and youve been here ever since

Do you just like being a helper? kekekekeke

[ornate elk]

yes im human xD

[empty cliff]

Oh and btw thanks. A life saver as always

[ornate elk]

yea, i like to help if i can

SD is not that easy to get into

[empty cliff]

Yeah I remember also hearing that you turned down a mod position

Rare to find folks like you tbh, this channel is lucky to have you. Most people out there in other rooms are hella toxic and unwilling to help out the noobs

[ornate elk]

thx, and true

[balmy plover]

hey cs10 im ib the webui bat file i think, but i cant find the set COMMANDLINE_ARGS= --xformers --no-half-vae

correctin, can somoene let me know if ive done this correctly

:accelerate
echo Checking for accelerate
set ACCELERATE="%VENV_DIR%\Scripts\accelerate.exe"
if EXIST %ACCELERATE% goto :accelerate_launch

:launch
set COMMANDLINE_ARGS= --xformers --no-half-vae
%PYTHON% launch.py %*
if EXIST tmp/restart goto :skip_venv
pause
exit /b

:accelerate_launch
echo Accelerating
%ACCELERATE% launch --num_cpu_threads_per_process=6 launch.py
if EXIST tmp/restart goto :skip_venv
pause
exit /b

[ornate elk]

your mostly in the webui.bat and not in the webi-user.bat

[balmy plover]

which one is the user bat?

i'm guessing this one

@echo off

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

call webui.bat

is that right?

[ornate elk]

this one

[ornate elk]

yes

[balmy plover]

appreciate your help, ty agian

[ornate elk]

you should enable "show file extensions" in windows file explorer

then you can see the .bat

also helps in not downloading and running a secret.pdf.exe

[balmy plover]

done and ty

last question for today

I'm setting up Stable Diffusion with the webUI, and I'm just making sure I'm installing the required Python modules correctly. I'm following the steps, but can anyone confirm if I’m doing it right? The packages should be installed using the requirements.txt files, but I want to be sure that I’m placing them in the correct directory and running the proper command.

Should I be running the following command inside the stable-diffusion-webui folder?

bash
Copy
pip install -r requirements.txt
And is there anything else I should check to make sure everything's being installed properly? Any advice would be appreciated!

[cold bronze]

NVIDIA GeForce RTX 3070

[ornate elk]

no pls dont

you only need to run the webui-user.bat
Everything will then be installed correctly and automaticly

[ornate elk]

ah 8gb vram, then make sure you have --xformers --medvram-sdxl in the webui-user.bat

[tame sun]

Did the AMD 9070 GFX files drop yet.

https://tenor.com/view/starving-alien-gif-21681479

[fair oxide]

To get Stable Diffusion A1111 working on RTX 5000 GPUs just create a Fresh clone of the Dev Branch. For XFormers, use a Beta version for now.

1. Navigate to the Folder Location where you want it to be and Open a Terminal/CMD there (or cd command there)
2. Type git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git webuirc then Enter
3. cd webirc then Enter
4. git switch dev then Enter
5. Locate webui-user.bat in the Location you chose to install A1111 and open it with a Notepad.
6. Overwrite and Add the following Lines starting at Commandline Arguments, then save and Exit.
   set COMMANDLINE_ARGS=--force-enable-xformers --xformers
set XFORMERS_PACKAGE=xformers==0.0.30.dev1005
7. If you have Extensions you want to Copy over, a little risky but do it at this time. A couple of my Extensions failed to install but it worked for me, A1111+XFormers still ran smoothly. (Most likely will work). Copy your Models too.
8. Run Webui-user.bat
9. When SD A111 Webui opens in the Browser, Go to "Optimizations" in settings and for Cross Attention Optimization select XFormers.
10. Generate an Image. (Will take a while for the first Generation. After that, should be very fast generations)

You probably want to make a Backup or/and try to stop stuff from UPDATING and destroying the installation of A1111 too. (ReForge normal install and 1click install works with rtx 5000 series until exiting, which is why some of you need this solution.)

in Webui-user.bat, Change your Commandline Arguments line to set COMMANDLINE_ARGS=--force-enable-xformers --xformers --skip-version-check --skip-python-version-check --skip-install

More help here https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/16818?sort=new

[ornate elk]

yes

you can follow the AMD Guides to get it working with Zluda

[ornate elk]

yea, i helped 2-3 guys already to get it work

[balmy plover]

@ornate elk runing now ty i really hope this works

[ornate elk]

if you just did what the guide lists (and nothing more), then it should work

[balmy plover]

i think i did, i may have accidently ran the script early

[ornate elk]

Thank you! will add it to the Nvidia Guides in the next days

[balmy plover]

at cs may i send you a private dm?

[ornate elk]

Sure

[tame sun]

@ornate elk I had meant to say optimized GFX files.

[ornate elk]

Ohh, then nope

[tame sun]

Pain

[mystic basin]

idk if this goes here or in the comfy channel but just a quick question, does comfy support HiDream model yet or it doesn't?

i guess not, either i get a clip error and if i change the clip to the same one i use for flux, comfy just crash

or i get this error if i change the other clip to t5xxl fp8

[mystic basin]

that's the issue i've beenn getting before i decided to to try other things

oh wait quadruple clip loader

[safe inlet]

yup

[mystic basin]

can't find the quadruple clip loader

[safe inlet]

update comfyui probably

[mystic basin]

the manager fail and the gitpull fail

i know that i modified one file so it use my gpu instead of my igpu but even if removed the change i did, it still doesn't work

wait it worked now

but i still can't find it

[rapid birch]

could someone please explain the 0 90 180 270 settings are for and when to use them. it doesn't seem to do anything.

sorry. i thought i was in facefusion. please disreguard

[mystic basin]

i guess the quadruple clip loader only work with CUDA for now

[brittle lantern]

@ornate elk I was told to come to u for help I installed wildcard extension on forge on Mac and the tab at the top doesn’t show but on extension it shows it’s checked and on hidden settings UI when I checked it doesn’t show wildcards what should I do

[brittle lantern]

Thanks, the wildcards tab showed up after idk how many restarts just decided to show up

[gleaming gale]

Hi, I've been looking in for controlnet models for illustrious but I only seem to find independent models for different methods. Is there an all in one model for ILL like the one for sdxl or pony?

[ornate elk]

Hey, yes there is the controlnet union++ promax model

That works with sdxl/pony and ill

[gleaming gale]

that's the one I've been using. It works well with pony but not so much with ILL

[ornate elk]

Hmm okay

Sometimes it doesnt appyl the poses

[robust siren]

Hi, does anybody have any suggestions for achieving the same workflow as " first frame" on runway? Ie. To stylize the first frame of a video and apply it to the rest of the clip? How can this be achieved in SD?

[faint minnow]

hey could anyone help me out with installing python for stable diffusion, ive been getting error 9009s and Incompatable versions or its not even detected

[ornate elk]

Hey uninstall everything from python.
Then install python 3.10.11 64bit

Then relaunch

[faint minnow]

Still got the error 9009 😦

been trying to fix this issue for the last 2 hours

[ornate elk]

Rerun the Installer.
Then click modify
Click next
Select "add python to environment variable's"
Then relaunch the webui-user.bat

[faint minnow]

it was already on but i will try

[faint minnow]

Didnt work

same error

[ornate elk]

Can you open up a cmd and run
Python --version
And then
where python

And screenshot pls

[spring slate]

When working with image to video (WAN 2.1) in ComfyUI is there a way to export the video to something other than a webp file?

[formal trout]

Hi, I was trying to generate with Wan 2.1 img2vid, I have a 12 GB VRAM card, RX 6700, I was wondering which model I can try that don't go OOM

[ornate elk]

You can try wan 2.1 1.3b

These are much smaller

But Idk if they work with 12gb

[formal trout]

great, ty

[calm loom]

it works with 12gb I run it off of my 3060 12gb

[ornate elk]

you can try to add teacache to the process to speed it up, but yea waiting 30 mins up to 1hour isnt worth it

[calm loom]

I moved it to #💬｜general-chat

[wet cedar]

I'm really confused since the release of SD3? I have the old one installed before SD3 ....just downloaded SD3.5...where does this go in the original folder? Is this a completely new thing to set up from scratch again?

[ornate elk]

nope its just a model like the others. which webui do you use?

[wet cedar]

Ummm

webui-user.bat?

@ornate elk

So I put the entire folder in the model section?

[ornate elk]

automatic1111 webui doesnt support sd3.5 i think

[wet cedar]

Does 3.0?

[ornate elk]

yes but its not great

[wet cedar]

Hrm. So I should just stick with the one before 3.0

[ornate elk]

your better of using sdxl based models

[wet cedar]

I was wanting to try Image to Video stuff

[ornate elk]

or Illustrious models

[ornate elk]

then you would need to use ComfyUI