ComfyUI for Intel Arc using IPEX

[somber trellis]

i tried it non-lora 20-step

[earnest grotto]

yes, the styles it knows seem to have taken a hit, and it outright doesn't do style transfer from a reference image

[loud crescent]

are you using the pytorch nightly? if that is the case its probably this:
https://github.com/pytorch/pytorch/issues/162813
you can switch to torch==2.9.0.dev20250821+xpu or set this environment variable:
_ONEDNN_GRAPH_SDPA_FORCE_PRIMITIVE=1

[earnest grotto]

tried stable ^

wouldn't hurt to try stable again, i guess

[vocal swan]

the _ONEDNN_GRAPH_SDPA_FORCE_PRIMITIVE=1 env variable does fix it finally, but makes generations slow
I switched using
pip install torch==2.9.0.dev20250821 torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu

but I get the error in the image. Is the torchvision version wrong? Its torchvision-0.25.0.dev20250926+xpu. Perhaps I will try reinstall using stable again

[earnest grotto]

install stable, with --force-reinstall

[vocal swan]

seems to be working now after a bit of fiddling

edit: nvm, worked for a couple images then back to outputting black

Im on drivers 32.0.101.8132

[somber trellis]

suno v5 released

i hope in the future we get comfyui models to match these proprietary models

[zealous atlas]

Are you using new TextEncodeQwenImageEditPlus nodes? They work much better than old TextEncodeQwenImageEdit with this 2509 version.

[earnest grotto]

The model is worse at doing anything style-related even with the new text encode node, at q6, with no lightning lora, taking the time to do 20 steps with 2.5 cfg

It now has the tendency to not change anything at all sometimes

Other times it will work at least

Kontext had the same issue though it was worse

[earnest grotto]

For most everything else i've tried though it's better (usually?). Odd that they singled out styles like that

Well, I'm not sure about text yet since it had a worse result with the no book no life / read and find out above, but I'll chalk it up to the lora or prompting, didn't want to wait 10 minutes

[zealous atlas]

Photo to pixel art

[somber trellis]

I am.

[somber trellis]

Is that with 2509?

Seems you did. checked the embedded workflow that image had.

I don't know why it worked for you. The exact same prompt in that image with the same seed and schedular+sampler doesnt change the style like it does for you.

I wonder if it's the quantization causing this,since you used a non-gguf fp8 quantized model while i am using q8 gguf

Nope. Trying the fp8 changes nothing sadly. Looks like it's not that.

i literally do not know

NOPE I just did a picture of a dog and it worked

It's just deciding not to change certain images.

[earnest grotto]

Try changing seed. It's just refusing to do things like kontext

I wonder what is in kontext's, and now qwen 2509's datasets

I just can't imagine the model doing this without at least a few pairs of nothing changed whatsoever

[earnest grotto]

Hmm... I guess prompting does make a pretty big difference
Original
"Upscale this image" ❌
"Transform this image into high quality anime art" ✅

Well, I guess it did remove JPEG artifacts when asked to upscale but that's only half of it

[lunar thicket]

Qwen edit?

[earnest grotto]

yes

https://x.com/mokito_ou/status/1971114315676778689 found some sketches I like and am having fun trying to make something more finished out of them.

needs minor fixes, though I find often it's easier and works pretty well to edit the sketch itself. damn that's good

[lunar thicket]

Thats crazy how it cleared up those jpeg artifacts

(The image before last)

[earnest grotto]

Not sure if there's a limit to how low res and jpegged the input can be before it can't help it much or if I just need more detailed and clever prompting

Original, with that wacky prompt
Original, x1, with a more detailed prompt describing the white roses and copper metal on the side etc., and x3 (with some cherrypicking)

Better prompting and a better lora could probably get even better quality

Hmm, i think i'll spoiler tag this, though what is depicted in the images doesn't actually happen in silksong. dunno

[dark pasture]

Anybody having issue if using batch_size greater than 1 images are different comparing to generation of single image with seed+1?

[earnest grotto]

That's not an issue, that's how comfy works
the bigger batch is just a fatter latent with noise from that seed. the noise pytorch decides to give the extra dimensions is not the same as if the seed was +1'd
if you don't want that behaviour, use the add noise node (for every individual latent), batch the latents yourself, and use the advanced ksampler

[dark pasture]

that's interesting. thanks 🙂

cause i've stumbled upon this PR https://github.com/comfyanonymous/ComfyUI/pull/1454/, but applying it didn't give any result

[earnest grotto]

2023

it may be that whatever this pr changed is not applicable anymore

maybe the code it changes got deprecated. idk

[dark pasture]

yeah, patch applied well, but had practially no effect 🙂

[iron star]

Is this still relevant? I have tried getting A1111 running on my B580 to no luck. Using vino's fork etc but still nothing, this looks interesting but i see the post is rather old and was meant for the A series, Does anyone know the current best option with an API.

[iron star]

Ive just tried this comfy ui stuff aswell, nothing ever seems to work. Its all cuda this cuda that, pytorch not working here not working there ive tried getting ANY sort of image gen working for like days. id say 6hrs spent straight without any results, can someone help me out?

[dark pasture]

Windows?

[iron star]

windows 10

[dark pasture]

have you tried?
#1193952640225267802 message

[iron star]

No, ill give it a shot

[iron star]

Got it running, model didnt exist so i added the model, now it wont run, I know beggars cant be choosers but i know A1111 is possible and its my main goal

[dark pasture]

what is an actual error you got?

[civic charm]

A1111 stopped getting updates more than a year ago, it doesn't support new GPUs including new Nvidia CUDA ones

Use SD.Next or ComfyUI

[earnest grotto]

whoever told you to use A1111, stop listening to them
If you want a ui like a1111, there's sdnext

Generally, if you're listening to people on the stable diffusion subreddit... don't

[iron star]

Alright same issue again, seems sdnext is not using my gpu, i thought i read from the documentation it would auto detect but reading the documentation you can also point it toward ipex which i did but then when running i get this error
AssertionError: Torch not compiled with XPU enabled

I also get a bunch of python syntax errors

[earnest grotto]

--use-ipex --reinstall

[iron star]

Thank you, finally got it working, i reinstalled python and reinstalled as you said now its running on my gpu, however i have one final question, what should be put in the "negative prompt" section or is it self explanitory, like stuff you dont want in the image?

[earnest grotto]

What model are you using and what are you trying to do

Anime art? Realistic stuff?

[iron star]

i mean anything, i have a juggernautXL something something (more realistic stuff from what i have seen) i found on civit, and im testing out the prompting but im unsure of what the negative prompt is for

[earnest grotto]

The negative prompt is not literally negative. The intended usage is to leave it blank, models are trained to still work with a blank prompt. If the model understands, you can add "low quality, worst quality" or such to it, this is usually present for anime finetunes however, not for more realistic models like juggernaut. The model is ran twice, once for the positive and once for the "negative", and the difference is used to continue denoising. If there is something you don't want in the image, prompting for it in the negative prompt can help, however it's not magic. No model exists that knows what "bad hands" are for example, so adding bad hands to the negative is pointless.

[iron star]

Alright, thank you for the help

[earnest grotto]

@civic charm The windows autodetects seem to need the basekit, i wouldn't get autodetected if not for the oneapi folder in C:/Program Files (x86)/Intel/oneAPI i have which is practically empty (for some reason there's a few leftover files), I definitely don't have the others like sycl-ls or oneAPI directly in C, some possible alternatives for autodetecting intel on windows:

check if "C:\\Program Files\\Intel\\Intel Graphics Software\\IntelGraphicsSoftware.exe" exists (control panel app that comes packaged with drivers, you can choose to not install it but this seems like a good idea for a simple check)

subprocess.check_output(["C:\\Windows\\System32\\WindowsPowerShell\\v1.0\\powershell.exe", "(Get-WmiObject Win32_VideoController).DriverVersion"]) this command might give a few lines, Intel drivers are 32.0.101.number, I imagine number>5000 is gonna be a good filter for outdated iGPUs, latest is 8132 https://www.intel.com/content/www/us/en/download/785597/intel-arc-graphics-windows.html
Example output:

PS C:\Users\Vikto> (Get-WmiObject Win32_VideoController).DriverVersion
15.39.56.845
32.0.101.6989

(First one is virtual desktop)
Alternatively you can look for .Name and see if it starts with "Intel (R)" but doing that might hit old igpus? not sure if those are Intel HD (R) or such

listdir C:\\Windows\\System32\\DriverStore\\FileRepository, search for something starting with iigd_dch (iigd_dch.inf_... is integrated, iigd_dch_d.inf_... is discrete)

[civic charm]

added os.path.exists("C:/Program Files/Intel/Intel Graphics Software") to autodetect

[royal sparrow]

so after using my divine intellect google ai overview, the black output is caused by Not-a-number (NaN) values being introduced during generation (ksampler or whatever) and makes the output black in the end. this can be seen with --preview-method auto

to prevent this there is a custom node, Interrupt on NaN, that replaces these invalid NaN values with zero and prevents the output from becoming corrupted and giving a black output. clone the repository into your custom node directory, add the node to your workflow and it should prevent black outputs.
git clone https://github.com/Extraltodeus/Uncond-Zero-for-ComfyUI

there have been some hiccups here and there but no way near to how it was before. i am using the stable release of pytorch if that helps, it should work with nightly if you're using that. hopefully this issue is fixed in future releases.

[earnest grotto]

i have a strong feeling this is not gonna have the effect you think it will

like making patches of or the whole image undenoised or fried

[earnest grotto]

run furmark or timespy or something else
try older drivers, like something before 7000

[earnest grotto]

https://huggingface.co/lightx2v/Qwen-Image-Lightning/tree/main/Qwen-Image-Edit-2509

2509 lightning loras

[earnest grotto]

@somber trellis Highly relevant comparison. I was right

The old lora is so much worse

Transform this image into an anime style

Wonder how much better it is at upscaling too

[woeful sable]

anyone got a simple step by step guide, pdf, google doc, whatever, on how to install comfyui on B580?

[lunar thicket]

Check the pins for Vik's script, I think he has made it pretty seamless

if you are on Windows you can literally just install AIPlayground, run it, and it installs ComfyUI for you. Then there is a "launch ComfyUI" button in the app.

[woeful sable]

did that, but didnt seem to work for me.
Prestartup times for custom nodes:
0.0 seconds: C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\ComfyUI\custom_nodes\rgthree-comfy

Traceback (most recent call last):
File "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\ComfyUI\main.py", line 147, in <module>
import comfy.utils
File "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\ComfyUI\comfy\utils.py", line 20, in <module>
import torch
File "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\cenv\lib\site-packages\torch_init_.py", line 282, in <module>
load_dll_libraries()
File "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\cenv\lib\site-packages\torch_init.py", line 265, in _load_dll_libraries
raise err
OSError: [WinError 127] The specified procedure could not be found. Error loading "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\cenv\lib\site-packages\torch\lib\c10_xpu.dll" or one of its dependencies.

[earnest grotto]

what pytorch version did you pick

[woeful sable]

nightly.

[earnest grotto]

run the script again, select stable, try again and say what happens

[woeful sable]

Prestartup times for custom nodes:
0.0 seconds: C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\ComfyUI\custom_nodes\rgthree-comfy

Traceback (most recent call last):
File "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\ComfyUI\main.py", line 147, in <module>
import comfy.utils
File "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\ComfyUI\comfy\utils.py", line 20, in <module>
import torch
File "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\cenv\lib\site-packages\torch_init_.py", line 281, in <module>
load_dll_libraries()
File "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\cenv\lib\site-packages\torch_init.py", line 264, in _load_dll_libraries
raise err
OSError: [WinError 127] The specified procedure could not be found. Error loading "C:\Users\Porkey\Desktop\AI\Comfy\Comfy_Intel\cenv\lib\site-packages\torch\lib\c10_xpu.dll" or one of its dependencies.

with stable one

[earnest grotto]

Delete the folder the script creates, try again and say what happens

[woeful sable]

like the comfy_Intel one? would that do anything? as i deleted it before i did the stable one, and script hasnt changed so not sure what it would do. or what folder do i need to delete and run the script again or just try launching it again? im confused

[earnest grotto]

The script creates a folder, named Comfy_Intel. You delete that folder.
Yes, nothing is changed about the script, because this issue does not come (strictly) from the script and people have been having it with AIPG as well and I'd like to see what might fix it

Ah, how about this: After it finishes, if it still gives the same error, would you like to zip the Comfy_Intel folder, upload it somewhere and send it to me?

It's sizeable, ~2-3GB I think

Personally I use mega.nz

[woeful sable]

think about 5 but can zip compress it.

[earnest grotto]

I can't reproduce the issue with the same comfy you sent me. Just a lot of windows defender notifications. So probably a driver problem 🤔

@woeful sable What driver version are you using, and do you want to DDU and try an older one?

[somber trellis]

@earnest grotto

Saw that the lightning lora got updated, now I can't seem to run anything.

ipex_to_cuda being funny

[earnest grotto]

rerun script

you can also delete the ipex_to_cuda folder in comfy

[somber trellis]

dont use it at all

i have it installed via pip

[earnest grotto]

yes

[somber trellis]

I actually just realized I already deleted it.

It's already using the pip installed package

Also, I did rerun the script.

I'm on 2.9.0+xpu stable.

rerunning it again

woops

fed the ipex_to_cuda repo and model_management.py into ai studio and got a fix

prompt: remove the sword

works again on my end

[earnest grotto]

I can't replicate whatever issue you were having on a fresh install

[somber trellis]

I have no issue what the issue even was.

It just seems like a device mismatch

[earnest grotto]

It's possible an update broke the patching the script does to model_management and the script couldn't recover it afterwards, since I've had that kind of issue when manually screwing things up but no idea how that would happen with the script

[somber trellis]

🤷‍♂️ No clue either. I'm just happy Gemini fixed it.

No OOMs.

Outputs are working as expected.

Slower, because I'm on 2.9 but still

current 2.10 nightly has dynamic link mismatch with torchvision

because it doesn't have a matching version

Also with 2509 the pixel art lora from civitai works

[placid violet]

anyone got some ideas why comfy running my CPU to the max after update
I was using 0.3.42(or smth) then updated to the latest 0.3.64 using Vik's script

[earnest grotto]

Show workflow

[tough wharf]

it wasnt like this before update
but ive been using this workflow since and it doesnt really throttle my cpu
https://civitai.com/models/741620?modelVersionId=2283356

[earnest grotto]

Show the workflow that you're having high CPU usage with

[tough wharf]

yes, that was the workflow that I use

[earnest grotto]

Show everything the command prompt says when you launch comfy

up to when it finishes loading

[lunar thicket]

holy crap thats a lot of nodes lol

[tough wharf]

ive seen more

[tough wharf]

Here when its first time launch and when it run a single time

[earnest grotto]

This is the workflow from the screenshot you sent but you say it doesn't have higher CPU usage than before

[earnest grotto]

Needlessly complex and pretty pointless. Basically ~5 upscales and segments chained together with their controls scattered everywhere

Segmenting and upscaling isn't really much better than just upscaling the whole image, especially for a portrait. If it wasn't a portrait, different prompts for different parts of the image could help to e.g. steer the model towards a more coherent background and make sure it doesn't generate tiny verions of the character, but no one does that

[tough wharf]

yes

[earnest grotto]

Is this ^ your alt account?

[tough wharf]

yes, sorry 😅
I forgot to switch over before askijn

[earnest grotto]

Please show the workflow that you believe had unusually high CPU usage

Not the one that you say "fixed" it afterwards

[tough wharf]

wait, I didnt say it fixed
its it wasnt like this before

[tough wharf]

its this one

[earnest grotto]

Yes, you said "it doesnt really throttle my cpu"

Which is not very correct, but I take it to mean the issue doesn't happen anymore, it's "fixed"

[earnest grotto]

The image, and the workflow you linked, are the same

[earnest grotto]

[Thermal] Throttling means that your CPU overheated and started running slower. Which would be an issue of your cpu fan not working

Having 100% CPU usage is not throttling

It can be an issue, if something that should run on the GPU doesn't. It can also be normal

[lunar thicket]

I saw a back to back upscaling and detailers and was wondering how much all that is truly beneficial 🤔

I do think I need to explore a detailer or upscaling node of some kind for my workflow

[earnest grotto]

"detailer" is just sgemented upscale

[lunar thicket]

You can see some eye funkyness over here #1245461701667848352 message

[earnest grotto]

Is that qwen/flux or an anime sdxl finetune

[lunar thicket]

and thats hard to avoid with just straight up generation so i think i could benefit from something

[earnest grotto]

the style gives me qwen/flux vibes

[lunar thicket]

The latter, its illustrious

[earnest grotto]

Any upscaling (with the sdxl finetune) will fix the eyes

[lunar thicket]

Then i do need an upscale added to my process 👍 👍

The style is probably throwing you off due to being intentionally Disney styled, particularly in the second two images

[tough wharf]

oh yea, my apologies I shouldve rephrased it better

[tough wharf]

yea, this is what I wanted to say

[tough wharf]

Maybe, still pretty fun to use with all the options already available in a single workflow

[earnest grotto]

show the workflow where it's not working the way you think it should be

[earnest grotto]

you say using a different workflow made you not have the issue

[tough wharf]

im not really sure how to desribe it since I really only noticed it earlier today

let me try on a different workflow

[tough wharf]

Ill say maybe during this part of the image where it loads the Clip/Text it uses too much of the CPU process, I think?
like theres a lag between it (im not sure what im saying at this point 😅 )

[earnest grotto]

You noticed how it has always worked

This hasn't really changed

This is due to --lowvram. text encoding is done on the CPU. For every model up till now - except for qwen edit, the text encoders are fast enough that doing them on the CPU is a negligible performance difference, for the benefit of having a bit more free vram so you can do larger batch sizes or upscales or whatever

[tough wharf]

well ig this really explains it, thank you
I wasnt able to make a backup of comfyui when I updated it earlier so I couldnt make a comparison

[woeful sable]

@Vik, i got latest drivers, not sure how to DDU it. sry for late reply.

[earnest grotto]

https://www.wagnardsoft.com/display-driver-uninstaller-ddu

[earnest grotto]

Download older driver (let's say older than 7000), run this in safe mode, restart, install older driver, restart. Say if the issue persists after doing that... I got no idea

[somber trellis]

@earnest grotto the LLM model im using (gemini 2.5 pro) fixed it for me fixing the logic between ipex_to_cuda and is_intel_xpu arguments in the model_management.py

This piece of code is all it actually added. Two at around lines 202 and 1160.

something to do with how ipex_to_cuda passes itself off as cuda:0

and native torch.xpu library doesnt understand this

at least thats what it says

Oh, and this one too.

around line 1250

only mentioning this all over again because i reran the comfyui setup script to install the latest pytorch nightly

[earnest grotto]

So it happens on a clean install with the nightly?

Or when updating?

[reef ivy]

I feel so behind now, I think its been over 3 months since I have messed with anything ai related. Is the new pytorch fixed and still faster on intel?

[earnest grotto]

what is broken to be fixed

[reef ivy]

I remember hearing about black image outputs or something with the latest nightlies

But also speed increases

[earnest grotto]

That's more specific than just newer pytorch. It might also be battlemage only and specific drivers only. The person kinda left so who knows

I am not getting any black images

[somber trellis]

I haven't tested this. This is probably very likely something I did.

But at least I can make pixel art with 2509 now.

[earnest grotto]

Finally, implicit casting (to other comfy nodes)

[grim wyvern]

I apologize if this is something obvious. I'm somewhat new to all of this.
I've tried the installation procedure above and, when I type the command "python main.py --bf16-unet", I get an attempt to complete the installation by it starts going wrong with a message "Disabling PyTorch because PyTorch >= 2.1 is required but found 2.1.0a0+git70bcf7da", says "torchaudio is missing", and finally errors out with a message "No module named 'av'"
Any help on fixing this would be greatly appreciated. Thanks.

[earnest grotto]

@grim wyvern Install using my script ^

[lunar thicket]

What does this mean, what is implicit casting?

[earnest grotto]

In terms of nodes, this would be not needing to insert a conversion node every time you want to convert something to another type

it gets converted (cast) implicitly, rather than you explicitly needing nodes

can connect to an image, mask and float input simultaneously without having to spam convert nodes everywhere

though the ability to convert an image to a single element is usually not present in nodal editors i think. it's certainly nowhere to be seen in either base comfy or blender's compositor

blender has a lot of automatic conversion which is one of the many UI/UX things it has over comfy

[grim wyvern]

That did it. That script is an amazing piece of work. Thanks for it and the help.

[lunar thicket]

So like the Encode/Decode nodes, right?

[earnest grotto]

You mean VAE encode/decode? no

[lunar thicket]

I don't have my desktop operational to look at my workflows, so im kind of guessing at what nodes youre no longer needing lol

[earnest grotto]

Moreso the convert mask to image and image to mask nodes; the primitive node as well, since it's kinda... iffy sometimes and needs to be plugged into a chain of nodes sometimes to work

VAE encoding/decoding depends on a specific VAE. Usually people only use 1 VAE but there are cases where you might want to use multiple

e.g. comboing some sdxl finetune with more modern models, to get either the better understanding of composition of newer models or the better details, depends on situation a bit

The multiple ways to convert rgb into a mask don't matter, averaging it out is the best

and in cases where you want a specific channel, you can split the rgb

if you want to skip on plugging VAEs around, you probably want this https://github.com/chrisgoringe/cg-use-everywhere

Also I've messed around a bit with interpreting things as latents directly, not super tested yet but it could be useful. comfy has no node to tell it that an image should now be a latent, without conversion, or nodes for doing any sort of math with latents

[lunar thicket]

I had to RMA my RAM so I'm offline for a week or so

[grim wyvern]

There is one minor thing after using Vik's script. It drops the non-critical message that
"EnvironmentLocationNotFound: Not a conda environment: C:\Users\qohel\OneDrive\Desktop\Comfy_Intel\cenv"

I wasn't using conda before, so I think I may have missed setting up the cenv at that point.
I assume I'd just have to set that up, and might be able to figure out just how with the message content, but I'm wondering if anyone can help me "skip to the end."

[earnest grotto]

are you launching with the shortcuts or are you doing something else
did you move any of the folders around

[grim wyvern]

I am launching from the shortcuts and I didn't move any of the folders.

[earnest grotto]

Show what's inside the comfy_intel folder

@grim wyvern

[reef ivy]

Man, ddr4 sure did shoot up in price lol.

[grim wyvern]

Inside the ComfyUI folder on the desktop:
start_lowervram.bat
start_lowvram.bat
a ComfyUI folder (do you want to know what's in that?)

[earnest grotto]

Delete everything (keep your extra_model_paths if you changed it, i guess. I also suggest moving your models out of comfy in general and changing the extra_model_paths if you did download them in comfy. all the various webuis can use the same models). Right click the background of where the script is. python ./setup_comfy_intel.py > log.txt 2>&1
You might not see anything in the console, so, just press enter a few times. That picks the default options (set up comfy, regular pytorch, no custom nodes).
Keep the log.txt file around. After the script runs, show a screenshot of the resulting folder. You can take a cropped screenshot with winkey+shift+s. If there is a cenv folder in there, run comfy, say if it runs fine.

[grim wyvern]

I'm not sure I understand exactly what you mean by "delete everything". If I do that, how will I find the setup_comfy_intel.py script? Along those same lines, I do have a ComfyUI_Intel.py script. Was that what you meant?
Finally, good call on putting the models in an external path and using extra_model_paths. I'll definitely do that.

[earnest grotto]

The script is here ^

[grim wyvern]

Got it. And I think I may have discovered the problem right there:
It begins with a line
condapath = "replace this text with your conda directory"
I hadn't changed that to the actual conda directory [major embarrassment].
I'll do that and run the revised version.

[earnest grotto]

You don't need to edit anything in the script.

That line is in case conda can't be automatically detected; if the script can't detect it, it will tell you, and won't run

Run the script from the command line
python ./setup_comfy_intel.py > log.txt 2>&1
Show a screenshot of what's inside the folder it makes. Run comfy. And so on

[grim wyvern]

Will do. (In fact, am doing it right now.)

[grim wyvern]

Did it. Same results. Here's the log.txt file:

[somber trellis]

    conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r
    conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/msys2```

Enter each of these lines seperately, then re-run.

The reason this isn't working is because you haven't accepted these. It won't create a cenv until you do.

[earnest grotto]

Updated the script, hopefully that should not be necessary now, but you can also do the above too

[grim wyvern]

"conda" seems to be the problem.
I get a 'conda is not recognized as an internal or external command' message.
I have miniconda installed in ProgramData/miniconda3.
I see a _conda.exe in the main folder of miniconda3, but no conda command file.
Also, the path didn't include ProgramData/miniconda3, but I fixed that.
I was able to type in the commands by using the Anaconda Prompt app in Windows 11.

[grim wyvern]

I seem to have messed it up even further. Before it would still run, just not in the conda environment.
Now I get: EnvironmentLocationNotFound: Not a conda environment: C:\Comfy_Intel\cenv

Failed to run 'conda activate ./cenv'.
Traceback (most recent call last):
File "C:\Comfy_Intel\ComfyUI\main.py", line 14, in <module>
from comfy_execution.progress import get_progress_state
File "C:\Comfy_Intel\ComfyUI\comfy_execution\progress.py", line 5, in <module>
from PIL import Image
ModuleNotFoundError: No module named 'PIL'

(base) C:\Comfy_Intel\ComfyUI>

So it appears that it's getting into the conda prompt now and erroring out after that.

[grim wyvern]

So, in the time-honored trope of "Have you tried turning it off and then on again?", now that the conda prompt would load, I deleted everything and ran the Vik's updated script.
It worked.
So, also, in the time-honored trope of "I really don't know exactly what was wrong, but it appears to be working now.", I'm going to call it a win.
Thanks for all your help on this Vik and Dan. You're credits to the community.

[bold owl]

This has been such a cool tool to have lately because there is a new AGPL 3.0-licensed RAW image editor out, RapidRaw, that will integrate with locally-run ComfyUI.

[honest hull]

https://github.com/Comfy-Org/comfy-cli/pull/319

I made an update to comfy-cli for the Intel Arc installation... using comfy-cli to setup ComfyUI is straight forward too

conda create -n comfy python=3.12 
conda activate comfy
pip install comfy-cli
comfy install --fast-deps

then follow the instructions in the prompt

[bold owl]

For the script installation, I am getting a new error message when I start comfyui start_lowvram.bat:

"ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host
Exception in callback _ProactorBasePipeTransport._call_connection_lost(None)
handle: <Handle _ProactorBasePipeTransport._call_connection_lost(None)>
Traceback (most recent call last):
File "C:\Users\jmchi\Documents\Comfy_Intel\cenv\lib\asyncio\events.py", line 80, in _run
self._context.run(self._callback, *self._args)
File "C:\Users\jmchi\Documents\Comfy_Intel\cenv\lib\asyncio\proactor_events.py", line 165, in _call_connection_lost
self._sock.shutdown(socket.SHUT_RDWR)"

Any thoughts? I'm also running llama.cpp and OpenWebUI (the latter with Docker), and this is a new error but I haven't changed anything that I can think of.

I mean, it runs. Like, ComfyUI is running and I can still generate images, but this error message keeps repeating itself every like 30 seconds or so.

[earnest grotto]

you can ignore this error, any errors about bitsandbytes, and an error complaining about stuff along the lines of 127.0.0.1 != mhkmfkgkmkdfm

[earnest grotto]

this one. warning i guess, still ignorable

[dark pasture]

seems like some browser extension name ^_^

[bold owl]

Thanks!

[hollow cradle]

I'm going to cry 😭.. Thank you, you are an amazing person.. I was going crazy, I don't know English, no one in my country teaches how to do the installation.. I had already given up.. Thank you ❤️

[brave stirrup]

after the last command python main.py , i dont get any ip adress for comfyui

python main.py - after this i dont get any ip address

[brave stirrup]

someone pls help

[brave stirrup]

it says no module named 'av'

[earnest grotto]

@brave stirrup Install ComfyUI using my script ^, or using the script linked a few messages above you

[earnest grotto]

.

[brave stirrup]

i dont have arc, using igpu...your script says incorrect gpu

[earnest grotto]

Show a screenshot of what's in task manager's 2nd (performance) tab

[earnest grotto]

Show it without cropping and show what the CPU is.

[earnest grotto]

Show me exactly what the script said

[earnest grotto]

When was the last time you updated your graphics drivers?

[brave stirrup]

2 days back its upto date bro

[earnest grotto]

Download the script again, I've updated it, run it again

[brave stirrup]

ok let me try

same issue ..i think i need to have some gpu

[brave stirrup]

its same issue bro

thanks anyway

[earnest grotto]

Show a screenshot of what the script says

[brave stirrup]

same

[earnest grotto]

If it is the same then you did not redownload the script

[brave stirrup]

is it like updated link somewhere else, i again downloaded the script and right click save

[earnest grotto]

Show what it says.

[earnest grotto]

Yes, you're running the old version

Delete the old version. Refresh the webpage. Download it again.

[brave stirrup]

maybe ill wait for sometime ,,i closed the browser deleted old one again redownloaded ,,but still same

[earnest grotto]

If it says 0.2.4, it's old. It should say 0.2.4.1

[brave stirrup]

can you send me the link

[earnest grotto]

https://raw.githubusercontent.com/a-One-Fan/ComfyUI-Intel-Installer-Script/refs/heads/one/Setup_ComfyUI_Intel.py

[brave stirrup]

thank you

[brave stirrup]

its crazy your link says version = "0.2.4.1p" when i download and save ...i opened with notepad to crosscheck it says version = "0.2.4p"

[earnest grotto]

I have no idea why that would happen. At best, delete it and then download again

[earnest grotto]

If you can't figure out why you can't download it, you can try this ^ or use AI Playground https://game.intel.com/us/stories/introducing-ai-playground/ which is not comfy

[brave stirrup]

can i just put it in anaconda folder?

[earnest grotto]

Do what the script says.

[brave stirrup]

also do i need change the username lol

[earnest grotto]

Either install conda, to the default location in C, or find where you installed it and put that in the path at the top of the script

[brave stirrup]

it opened

can i intall all

gguf brushnet

[earnest grotto]

yes. don't use the experimental option

[brave stirrup]

yes or no

i did no

[brave stirrup]

this what is says now

i think its done

[brave stirrup]

lol ok thanks

[earnest grotto]

Sometimes Intel's servers also don't work, so you might need to try again

[brave stirrup]

i dont know that ,,,but thanks a lot man , you gave me hope ...will upate later

[earnest grotto]

And which version of pytorch did you pick

And, running the script again will be faster since most things are already downloaded

[brave stirrup]

not sure i dont think it asked me that , i just did no for EXPERIMENTAL thingy

[brave stirrup]

does having latest python matter or specific version ?

[earnest grotto]

doesn't matter

It still can't install it?

open the script in a text editor, search for COUNTRY = "us" #if chosen_ipex < 2 else "cn", remove that first #

[brave stirrup]

yes just now edited and running it ... selected STABLE pytorch

[brave stirrup]

is this done

there are 2 comfy ui which one to run

[brave stirrup]

how to change low vram to normal vram

[earnest grotto]

There is no point in not using lowvram with the majority of models

and the models where it did make sense (qwen edit), want the bigger vram usage reductions because they're too fat, and their text encode node while it was very slow if ran on the cpu, is not actually necessary to encode the images with so even that issue is a mostly non-issue

[brave stirrup]

ok thanks but it says set to : LOW VRAM - can this be changed to NORMAL VRAM?

[earnest grotto]

@brave stirrup ^

lowvram means the text encoders are ran on the CPU instead of the GPU.

That's it. They are very fast and running them on one or the other doesn't matter much.

If you don't want that to happen, you can open the lowvram batch file and delete that argument.

[brave stirrup]

thanks a lot .. i keep forgetting that i can edit or im scared ill screw it up..😆

[royal sparrow]

been getting this warning since installing and updating comfyui through the script, mostly curious about what it is, is it worth being concerned over, is there a fix?
I've been ignoring it and comfyui has been performing to my liking so i dont really have an issue with it. it only appears once when running a workflow for the first time then never again

Comfy_Intel\cenv\lib\site-packages\ipex_to_cuda\attention.py:123: UserWarning: SDPA Flash Attention backend is not supported on XPU, falling back to OVERRIDEABLE backend. (Triggered internally at C:\actions-runner_work\pytorch\pytorch\pytorch\aten\src\ATen\native\mkldnn\xpu\Attention.cpp:147.)
hidden_states = sdpa_pre_dyanmic_atten(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, scale=scale, **kwargs)

[cobalt radish]

Hello guys, I'm new to this idea of installing comfy UI on intel GPUs. I have ARC A770 and I plan to use QWEN Image editor and Wan 2.2. I'm thinking to follow this tutorial: https://www.youtube.com/watch?v=fQKOJVVi44E

Is it worth trying to set this up or it's just easier to use NVidia instead? Is this tutorial enough? What's your experience? Be honest please 🙂

[lunar thicket]

Thats Bob Duffy he is the resident AIPlayground guy.

I haven't watched this video but he is the one that posts updates to AIPlayground in #1243956384052285560

I mention this because thats a windows installation in the video. But AIPlayground will literally just install comfyui for you

You just download it and run and it installs auto-magically.

Latest release is here: #1245461432141873245 message

My experience is I have installed and used comfyui via this and it works great. I have a B580.

[earnest grotto]

You can ignore that. If anything, seeing this should remind you that Disty's hijacks are working, and you want them to be working so that the various custom nodes that do not use ComfyUI's built-in device getters and hardcode "cuda" will work anyways (e.g. ultimate SD upscale, or some of the ones my script installs).

[earnest grotto]

What do you mean easier to use Nvidia... You're gonna sell your GPU? I think that's a bit hasty
AIPG is very easy to set up. There's other options if you want just Comfy as well

In general, there is no Nvidia offering that has that 16gb the a770 has anywhere near the a770's price. Closest would probably be a 3060 12gb, if you can find one? Which is not much better, might even perform worse, certainly performs waaaaay worse than the b580. The 5060ti 16GB is actually better, but it also costs 2x as much usually

[royal sparrow]

thanks for the clarification!

[somber trellis]

There are some specific examples that I've seen with Nvidia that do just beat us performance wise.
I blame cudagraphs. Someone can run Chatterbox 3x faster than me on a 3060 because I can't get torch.compile to work with it.

Lmao

[reef ivy]

New LTXV is "out" not open source yet though I don't believe. Looks pretty good, hopefully it's not huge and slow like the last one released. Quality looks worse than wan but fine tunes could fix that, also has built in audio which is cool.

[somber trellis]

Mattvidpro discord has a few gens from ltxv2, and it's trained on some weird stuff.

Unprompted, it gave someone Garfield gens.

[earnest grotto]

who's gonna finetune a video model

finetunes pretty much stopped happening since flux

Quality does look better than wan 2.2 to me though, at least cherrypicked

[cobalt radish]

Thanks for helping Grok, that was clean. Appreciate it!

[lunar thicket]

My pleasure!

[reef ivy]

There have been plenty fine tunes of wan so far

[earnest grotto]

i'm probably no-true-scotsmaning a bit but, there aren't really any finetunes with significant amount of training, besides chroma, neta lumina and now pony v7
most of what isn't those 3 might as well just be a lora

and those 3 are... iffy IMO

chroma is the least iffy, looks to be fine if you want realistic stuff, but I still don't quite like it

neta is undertrained and might have some other issue understanding styles and other things, pony v7 looks to be fried

[young talon]

this is my first video i generated today. big shout out to Mr. Bob Duffy for creating the Video on MSI Claw 7. Using a distilled version of LTXV within my anaconda environment. Thanks. My results aren't much but here is my progress. Anyone else feel free to message me too. I'd like to learn more.

[young talon]

6 seconds long

[formal tusk]

What kind of generation rates are you all getting w/ Wan 2.2?

I've been having a lot of trouble with slow generations - i.e. around 30-45 min for a low resolution 5second video w/ a model that has lightning or a similar accelerator LORA baked in.

10700k, 64gb ram, A770 16gb vram

[earnest grotto]

Highly dependant on resolution and framecount

Did you install with my script?

[lunar thicket]

Vik- Your script is OS agnostic right (works for windows and linux)?

[earnest grotto]

yes

[lunar thicket]

gonna try it on kubuntu environment

got my ram back from RMA and installed today

[late glen]

Yes, wondering if I need to do a fresh install now.

[earnest grotto]

You mean yes for this?

[late glen]

Yes.

[earnest grotto]

For bigger models/more vram usage like this, launch with the lowervram shortcut/batch file

What framecount and resolution

Also, I take it, Bryce is your alt?

[late glen]

cd /D "%~dp0"
call conda activate ./cenv
:: xpu does not need forced slicing
:: No additional environment settings needed for Windows
cd ./ComfyUI
python ./main.py --bf16-unet --lowvram ```



560x720 @ 81 frames 

Yes, my professional account. Lot of unrelated stuff I try to keep separate. Hoping to get wan running at decent speeds, then use  wan animate for educational media esp. sign language for developmentally delayed kiddos

[earnest grotto]

650*720 is ~0.4mp. I recommend first starting off with 61 frames and 0.3mp (for your AR, about 480x620), to nail down a prompt or first/last frame or whatever, then increasing it back to 81 and 0.4mp or slightly more. Either way, the script you link is the lowvram one, use the lowervram script (which adds --reserve-vram 11).

--reserve-vram 11 means that in a perfect world, 11gb of vram would be free for your OS or other things to use. but this is not a perfect world (on windows), so, it's more like 6gb free. If you want even faster, you can lower it from 11 to 7. --reserve-vram 6 is kinda the cutoff point where if you have anything else running, you will start to run out of vram and have issues

[late glen]

Thank you. I'll give it a go.

[earnest grotto]

After the first run, this should be close to 1-2 minutes per video

[late glen]

FP8 or a quant like Q5?

[earnest grotto]

I use Q8 high Q4 low

Refer to the 4 videos here ^

You can also use fp8 for the high instead of q8

I have only 48gb of ram and even still, q8+q4 is kinda cutting it close on ram usage (i needed to crank up swap/pagefile a lot, it's 50gb now). You might be able to get away with both being fp8 with 64? not sure

[late glen]

Tried lowervram with the lower frame rates and it worked as you suggested.

[late glen]

both as 14gb models FP8

Was thinking before I asked that low and high loaded in vram independently and the goal was to load each fully, without swapping w/ ram. Finding a smaller model was tough!

[reef ivy]

If you are using gguf make sure you use the --reserve-vram argument. Block swap with kijai's nodes might help as well but never tried (also last I checked you had to edit a few lines of code to get his nodes to work with intel arc)

[earnest grotto]

No
Use reserve-vram if you're running out of vram (which you are with an fp8/q8 wan model)
It doesn't have much to do with gguf and will kill performance when you're not running out
Block swapping most likely performs much better, but I haven't tried that node either

[reef ivy]

It speeds up gguf in native, if block swapping exists in native though it is probably better.

[lunar thicket]

@earnest grotto got an interesting error, running your script in kubuntu. Seems to be failing on ID'ing my GPU, maybe.

[earnest grotto]

Run clinfo and show me what it says

ok, dm it i guess

[lunar thicket]

DM what?

[earnest grotto]

Some messages are automatically deleted, most likely to get rid of the scam bots

If you didn't notice, your message got deleted.

[lunar thicket]

OH weird it shows on my end still

ok

DM sent

[earnest grotto]

Odd that it didn't even detect the cpu as a cl device

Either way, the script checks for your GPU through clinfo. If you yourself can't see the GPU in clinfo then you have some issue with the drivers that needs sorting out

you can do clinfo | grep name if you don't wanna scroll through it though with 0 devices i guess there wasn't much to scroll

[civic charm]

clinfo requires manual installation

it isn't required or comes with anything by default

disty:~ $ cat /sys/class/drm/card0/device/device
0x56a0
disty:~ $ cat /sys/class/drm/card0/device/subsystem_vendor 
0x8086

8086 is intel

56a0 is the id for A770

https://dgpu-docs.intel.com/devices/hardware-table.html

[earnest grotto]

The script pokes the user to install it if it's not present
In this case it just flat out returned 0 CL devices which seems problematic to me

Though actual GPU detection might be pretty pointless at this point, and conda. I only used them for the older pytorches that were way more convenient to install with conda, and had specific versions for specific GPUs, now newer ones are simple to install and faster so there's no point in using the old ones

I might keep it around moreso to make sure the user's drivers all work?

[earnest grotto]

Since if not even the CPU is detected as a CL device, that seems very broken to me

[lunar thicket]

Vik helped me progress via DMs and I was able to run the script and install comfy. Passed the clinfo error.

but i ran into further trouble later, actually launching comfy. I forget the errors now, but I think I need to wipe out my drivers and do a fresh install. It was something about no xpu device, which doesn't make sense after all that.

But i read some folks having errors like that when they had a disjointed driver install like i apparently did

[civic charm]

Clinfo can't detect my CPU and AMD GPU either

Only A770 is detected

OpenCL is on its way out. Using /sys/class/drm or lspci will be more reliable

[earnest grotto]

Damn, alright then

And yeah, opencl is indeed on its way out

[earnest grotto]

Chrono edit is usable in comfy right off the bat. Needs prompt rewriting though

IMO, worse than qwen edit

That 2B model could be interesting though

If it's still good, would be amazing for something most people can actually train

[formal tusk]

Anyone else using Qwen3-VL for prompt generation? I have a modified version of the node that works with intel hardware if anyone is interested.

[earnest grotto]

Was wondering why comfy sometimes gets completely stuck when I'm doing large qwen edit images with long prompts(?). Well...
33223 is a prime and find_split_size looks for a number that the [in this case head size] is divisible by. so, 1
(ipex_to_cuda)

[solid helm]

hi, I'm just now installing openvino and comfyui and Stable Diffusion, and Grok's tellin' me to convert fp32 to fp16 since the fp16 installationdoesn't exist in stable diffusion's own dowloads page. is this something I should do? using a B580 btw

the command it game me:
from openvino.tools import mo
mo.convert_model('models/checkpoints/sd3.5_large.safetensors', --data_type FP16, --output_dir models/openvino/)

honestly I'm completely stuck with this installation

[earnest grotto]

Don't listen to LLMs

There's a custom node that does whatever conversion
Why do you want to use SD 3.5

[solid helm]

It's the newest (:

[earnest grotto]

No?

[solid helm]

oh

[earnest grotto]

Even if recency meant it's good, which it doesn't, qwen image is far more recent

[earnest grotto]

Did you ask grok for this too?

[solid helm]

I came across the videos on intel graphics youtube with A1111 and stable diffusion and wanted to try it out, but I read here A1111 has lost support or something

[earnest grotto]

A1111 has not been updated in more than a year

[solid helm]

I just read that stable diffusion is the best at prompt interpretation somewhere

[earnest grotto]

If you want something that's like a1111, use sdnext

[solid helm]

alright. what about stable diffusion then

should I not use it

[earnest grotto]

https://github.com/vladmandic/sdnext

[earnest grotto]

Do you want to make realistic images or anime art

[solid helm]

realistic I guess

[earnest grotto]

qwen image then

[earnest grotto]

@solid helm You can download it from inside sdnext. You should be able to compile with openvino as well, go into the settings and see

[solid helm]

alright, thanks a lot

[earnest grotto]

Actually, how much ram do you have, what GPU

[somber trellis]

DyPE works in comfyui (Flux node that lets you generate much higher resolutions)

[civic charm]

torch 2.9 has the same speed as openvino, why do you want to use openvino instead of pytorch?

openvino is really good for onnx and captioners but dynamic things like stable diffusion doesn't really fit openvino

[solid helm]

B580 but only 16gb of RAM

[earnest grotto]

16gb is too low for... most things
even for gaming it's too low

[solid helm]

Been more than ok for gaming at 1440p so far lol

[earnest grotto]

I can game at 8k with my a770 with a specific choice of game

[solid helm]

Cyberpunk, RE8, Doom, CoD, etc. still fine for me. Also with current ram prices I'm not even gonna think about upgrading

[earnest grotto]

16gb definitely won't be enough for qwen image, flux, sd 3.5 large. probably won't be enough for sd 3.5 medium. what you can do with only 16 is sdxl

[solid helm]

I don't need like a batch of 50. I just want to do it as a hobby. batch of 1 works for me

[earnest grotto]

this has nothing to do with batching, these models are literally too big to fit in your ram even if we ignore that windows itself probably always uses up ~4gb

[solid helm]

I see

[earnest grotto]

juggernaut xl v9 is a realistic sdxl model you can check out

[earnest grotto]

and, this

I assumed openvino has some benefit but I guess not

[solid helm]

I thought it was the best option for arc specifically

[civic charm]

even sdxl is cutting it short with 16gb

you simply can't run openvino with 16gb

it needs 32-48gb ram to convert sdxl

[earnest grotto]

yeah

on the bright side, we know the b580 has pretty good sdxl performance (without openvino, compile, whatever), so you can have fun with that

3090 speeds?

I'd also suggest checking out the sdxl inpainting model. you can make up for what sdxl models can't understand super well in prompts, with inpainting

[earnest grotto]

Hmm, you might be able to barely fit sd 3.5 medium in 16gb of ram. you'll have to test and see, most likely will need to at least quantize the t5 which sdnext should be able to do

[reef ivy]

Man I wish I had upgraded my ram earlier this year, price is insane now. 40$ kit is now over 100$ smh

[lunar thicket]

My ram was $200 last December and now its $500. Insanity

[solid helm]

If you're curious, tried sd 3.5 medium (fp16) on comfyUI with openvino enabled, and it gave an error saying the VRAM ran out when the negative prompt was highlighted.

Then I ran this and it worked: "python main.py --disable-ipex-optimize --lowvram --force-fp16 --cpu-vae --reserve-vram 2"

I'm not even sure if I did the installation and everything correctly I think there's still a few missing files and stuff but I got it to generate finally lol

[earnest grotto]

don't use openvino

Show the workflow

[solid helm]

actually I lied I forgot to put it in the workload 😂

[earnest grotto]

Get the fp8 t5xxl

[somber trellis]

gguf q8 is better for T5

DyPE flux is pretty detailed

[earnest grotto]

How long did this take

Are the images you're sending the originals?

The images you've uploaded look to be compressed webps
I've had very few issues with cranking up Qwen Edit's resolution, worked well up to 8mp (~15 minutes), gave a sorta bad result at 16mp though it might've been a bad seed and that took an hour so I don't wanna try again

[somber trellis]

20s/it, 8 steps

These are both 2048x2048.

It can do 4096x4096, but it will take 93s/it.

Well, I did initially get 4096x4096 to work. Now my PC is imploding.

I will say, it's got issues.

I do also use qwen-image though

[somber trellis]

increasing to 16step seems to help

[somber trellis]

@earnest grotto What loras would you use with Qwen-image to achieve a realistic look? So far I've used flymy realism and a samsung lora.

Also looked at boreal portraits.

Right now I'm using flymy and samsung together with the 8step lora

[earnest grotto]

I got nothing other than wishing qwen edit had more loras sorry

[somber trellis]

thats kinda why i liked dype, because it's flux-based

qwen image got like 8 realism loras

🤷‍♂️

DyPE again

[earnest grotto]

qwen edit got like 8 useful loras total, and that's probably including the ones that don't work well like the upscale ones

ticks me off so much

damn 20b model

was hoping that maybe that pruned version of it people made that gets it down to 13b would be nice, nope, it fries most things and text is so, so broken

so sad since it's so good

[somber trellis]

for upscaling just use seedvr2

it's pretty good to be honest

I'm using the Q8 gguf versions of them

ignore the K_M

it's how they have the naming scheme done

^ upscaled

v original

original is a 250x326 image

I genuinely dont know of a better upscaler than that, it's pretty decent for details.

[earnest grotto]

Sure but I want something better than upscaling
I want a model that fixes fingers, a model that deblurs (especially when people already tried training loras for this for kontext, or defluxing), a model that delights or gives a UV pass or so much more

Even saw someone on the subreddit with ambitions to try training a delighting lora for kontext

RIP any of that

soon after kontext came out, someone trained a lora that made kontext colorize better. nsfw, but i'm pretty sure it helped in general
no such thing for qwen edit even after 2 months

now, qwen can colorize better than kontext with or without the lora but why not push it just that bit further

or why not train it so it can actually colorize with hints inside of the image itself instead of describing every single thing and hoping it gets it right, because it sometimes doesn't; or colorize with a reference

[earnest grotto]

Oh, I guess I can link you this one https://civitai.com/models/1927710/qwen-image-boreal-boring-reality-lora-for-qwen

Decided to browse the loras again just to get ticked off

Civitai still haven't added separation between qwen image and qwen image edit for some reason

[somber trellis]

DyPE again

[earnest grotto]

Qwen Edit with a more recent upscale lora.
4 steps, 1 step.

[earnest grotto]

holy speak of the devil https://www.reddit.com/r/StableDiffusion/comments/1os7ut5/lora_panelpainter_manga_panel_coloring_qwen_image/

literally 44 minutes ago

[somber trellis]

Seedvr2 definitely wins here imo, as the eye color and skintones are changed

Not bad tho

[earnest grotto]

I don't think they were?

[somber trellis]

It made the eyecolor more brown.

[earnest grotto]

Either way, Qwen can be prompted to also have a specific eye color, and can give different eye colors with different seeds so that's not much of an issue

[somber trellis]

ofc, but

i insta noticed the eye

lol

[earnest grotto]

I don't remember how vsauce m. here's eyes looked

[somber trellis]

guy has a doubled eyelid

https://tenor.com/view/vsauce-metal-chewing-gif-5994904747601623430

[earnest grotto]

nah, plausible with glasses

[somber trellis]

I can nitpick all day about these little things but they can all be fixed

[earnest grotto]

and the stuff you're pointing out is with the 4 step result

[earnest grotto]

which i... don't quite like

I guess just not denoising the rest of the way produces better looking skin

wild to think about

[somber trellis]

ye, 4step isnt ever enough UNLESS you do a second pass

what im doing rn is dype with two passes, one at 1024 and the second at 2048

alongside NAG too* so I can use negative prompts at CFG1.

[marsh bison]

hi, can someone please help me? I recently reinstalled torch to try and fix slower gens and now im getting this error

[earnest grotto]

ah, a comfy update literally a week or so ago broke that

i expected they'll fix it, i guess they didn't

I'll update my script to fix it

or i'll do a 1 line pr

[marsh bison]

is there a way for me to do a rollback then?

[earnest grotto]

Open ComfyUI/comfy/model_management.py
Navigate to line 1106

def pin_memory(tensor):
    global TOTAL_PINNED_MEMORY
    if MAX_PINNED_MEMORY <= 0:
        return False
    

add

    
    if is_intel_xpu():
        return False

after the block of code I pasted, while keeping the indentation intact

scroll down a few lines to def unpin_memory, do the same

[marsh bison]

ok that seems to have done it. thanks alot

while im here do you have any tips for getting faster gens?

[earnest grotto]

Depends on model

Only thing that will give you a definite speed boost in any scenario is getting the nightly pytorch, which will be ~10% faster than the stable one

[marsh bison]

i tried both the stable and nightly pytorchs they just give me black outputs

im on 2.5

[earnest grotto]

2.5 is massively slower

what gpu, what black outputs, what driver version, what model

[marsh bison]

arc b580; .8247,8136, and .6989 drivers; and it was an illustrious model

[somber trellis]

So Compass.1 Flux lora (a spatial enhancer lora for flux) doesn't work with dype. It messes up a lot.

[earnest grotto]

https://github.com/comfyanonymous/ComfyUI/pull/10691

hmm

[earnest grotto]

reinstall comfy with my script / redo your venv, try again

[marsh bison]

ok

[earnest grotto]

If you do it right now, you will also need to do this ^ again

[marsh bison]

this is probably a dumb question but all i have to do is run the Setup_ComfyUI_Intel.py again right?

[earnest grotto]

run it in a different location

if it doesn't work out, you can just delete the comfy it creates

everything it does is stored in the Comfy_Intel folder it makes

[earnest grotto]

Ok, if you redownload the script now you won't need to edit those 2 lines manually anymore.

It's updated

[marsh bison]

sorry i took so long to get back. But yea im still getting black outputs i tried 4 different models (3 illustrious and 1 noobai)

[neon tapir]

the black output thing was an issue for me too instead of using option 3 when using the script hit 2 for the ipex version

[somber trellis]

qwen-image-lightning, 16 steps (NAG)

[earnest grotto]

That's some sort of driver/pytorch bug and should not be happening
it doesn't happen for other b580 users

[lunar thicket]

Maybe foolish question but what is pytorch 2.XX+xpu ?

versus say pytorch 2.XX

[earnest grotto]

XPU are Intel's devices. GPUs, but not just

No xpu, no intel gpu support

[lunar thicket]

I recognize XPU as Intel devices, but I was under the impression that basic 'normal' pytorch now fully supported Intel GPUs

IE, pytorch 2.9 would 'just work' for Intel

so is it the case that pytorch 2.9+xpu is a separate package they made for Intel devices? Makes sense. I did preface this may be foolish

i see now on their site they even have multiple packages just for different CUDA versions.

😅

[reef ivy]

They also have Cuda pytorch versions as well, so not really sure but I always use the xpu

[lunar thicket]

@earnest grotto I could not for the life of me figure out why comfyui wasnt working in kubuntu. Except... i was on 25.04 plucky. Which may be the culprit.

My solution: i updates to quokka lol 25.10. Deleted the comfy intel folder and ran script again, now it works

I am getting some intermittent black images though. Someone else was complaining about that in here. I wonder if that is a pytorch version problem. May try to downgrade to 2.8

[earnest grotto]

I have absolutely no idea why b580s are getting black images and I wish I did

If you downgrade to 2.5 you're leaving huge performance gains on the table

If you downgrade to 2.8... I don't know if that will fix it. Performance will be only very slightly slower than 2.10

[lunar thicket]

All I know is I had zero issues with 2.8... in Windows. So I know it is kind of apples and oranges. But still worth a try.

[neon tapir]

oh nvm i didnt see your message sorry

i would love to have the performance upgrades but almost always either with the first ksampler or when it upscales and goes into the second ksampler it blacks out, im getting decent speeds but it sucks that i cant have the best version, also i cant for the life of me get qwen to run without crashing my comfy

[lunar thicket]

what version of pytorch you running?
B580 user?

2.5+ipex?

i am gonna try installing 2.8 which worked fine for me in windows

[neon tapir]

b580 and yes the 2.5+ipex

[cursive sable]

im getting this error

[earnest grotto]

@cursive sable Install comfy with my script ^

[lunar thicket]

https://github.com/pytorch/pytorch/issues/162813

fyi for folks

Versions
2.10.0.dev20250910+xpu

By the way, this also occurs with 2.9.0.dev20250909+xpu

seems like the 2.9 issue all 4 B580 owners are having is acknowledged

https://github.com/comfyanonymous/ComfyUI/issues/10756
related

[somber trellis]

I got qwen image nailed now. Two loras and a different vae encode/decode (wan 2.1 vae upscale 2x imgeonly real)

[cursive sable]

can flux work with intel arc?

[earnest grotto]

yes

[earnest grotto]

.

[somber trellis]

Any unet model compatible with native ComfyUI should work. If it's from a custom node it would depend.

So flux, sdxl, sd3.5, qwen, lumina, chroma, ltx, wan

aka most models work

[Nano Banana 2 released. Got someone to gen this.]

[earnest grotto]

an actually interesting gen rather than 'here is an image of the original spyro. regurgitate your dataset and generate an image from toys for bob's remake'

[cursive sable]

tysm ill try that right now

[cursive sable]

i installed my comfyui with cmd only and its working perfectly fine with other models but whenever i load flux 1 schnell fp8 model it crashes without any error

[earnest grotto]

You probably ran out of ram.

Use the Q4 model.

[cursive sable]

okayy

is 16gb not enough for the flux 1 fp8 model?

[earnest grotto]

it's not, yes

16gb is low in general

https://huggingface.co/city96/FLUX.1-schnell-gguf/blob/main/flux1-schnell-Q4_K_S.gguf

my script can also download the Q4 model for you

[cursive sable]

i see

tysm for the link ill try downloading with the link since scripts and all that seems very complicated to me 😭 im new to these stuffs

[earnest grotto]

windows usually takes up something like 4-7gb
fp8 flux is 11gb
you are on the brink of running out already
fp8 t5 to encode the prompt is ~5gb IIRC
then there's the VAE and the image latents which combined are probably ~0.5-1gb. and if you have lots of browser tabs open, that's using up RAM

[cursive sable]

i see

should i download the encoders aswell of gguf?

or will it work with fp8 t5 which is already installed?

[earnest grotto]

q4 flux is roughly half the size (a bit more) of fp8 flux

you should have barely enough

[cursive sable]

i see. ill upgrade my ram in future then

ty for telling me all these

[cursive sable]

does anyone know what im doing wrong here?

[earnest grotto]

iirc it needs clip to also be gguf

[neon tapir]

when i use the installer script for option 2 to get the 2.5+ipex (since its the one that works without blacking out my images) it errors here and times out, is the package not available anymore?

[earnest grotto]

Stuff like this happens sometimes. No idea why. Part of the reason why I made the script keep trying to download.
Open the script in a text editor, search for COUNTRY = "us", replace us with cn, and run it again

say if it works

[neon tapir]

ill try that right now

[neon tapir]

unfortunately it errored out again

[earnest grotto]

hmm

ok, i can fix it this time i guess

@neon tapir New script version, 0.2.4.3p

Download, delete old one

[neon tapir]

thank you very much for the quick work, your the only reason i can keep generating on my new card reliably and it means alot

[earnest grotto]

You can keep an eye on this issue https://github.com/pytorch/pytorch/issues/162813 for progress on the black images bug

[lunar thicket]

When is the last time you tried the Nightly?

re: the issue above, it is possibly fixed. I installed the nightly in a new location two days ago based on this but i haven't been on my desktop again to actually test.

[lunar thicket]

well decided to go check: can confirm it is not fixed in 2.10.0.dev20251116+xpu

[lunar thicket]

I see actually they added a comment since i was on there that indicates theyre holding off on the update

we found several issues in OneDNN v3.10-rc branch acceptance test. Issues have been reported to OneDNN and we are waiting for their fix so that we can upgrade to v3.10 formally

[lunar thicket]

i will say the nightly build seems to produce less black images. I ran a test, 8 runs of 10 images, using 2.10.0.dev and 2.9.1

2.10 produced 16/80 black images (0.2)
2.9 produced 23/80 black images (0.2875)

same workflow, prompt, settings.
i closed the browser between runs. reopened for the fresh second run.

[deft bloom]

hello everyone!
I've been trying to install and use ComfyUI for the past several days and now I think I've got closer than ever to punching a hole in my monitor. When I'm trying to launch Comfy, that's what pops up:
`## ComfyUI-Manager: installing dependencies done.
** ComfyUI startup time: 2025-11-19 23:41:59.316
** Platform: Windows
** Python version: 3.10.11 (tags/v3.10.11:7d4cc5a, Apr 5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]
** Python executable: D:\Comfyui\ComfyUi\venv\Scripts\python.exe
** ComfyUI Path: D:\Comfyui\ComfyUI
** ComfyUI Base Folder Path: D:\Comfyui\ComfyUI
** User directory: D:\Comfyui\ComfyUI\user
** ComfyUI-Manager config path: D:\Comfyui\ComfyUI\user\default\ComfyUI-Manager\config.ini
** Log path: D:\Comfyui\ComfyUI\user\comfyui.log

[notice] A new release of pip is available: 23.0.1 -> 25.3
[notice] To update, run: python.exe -m pip install --upgrade pip

[notice] A new release of pip is available: 23.0.1 -> 25.3
[notice] To update, run: python.exe -m pip install --upgrade pip

Prestartup times for custom nodes:
4.7 seconds: D:\Comfyui\ComfyUI\custom_nodes\ComfyUI-Manager

Traceback (most recent call last):
File "D:\Comfyui\ComfyUI\main.py", line 147, in <module>
import comfy.utils
File "D:\Comfyui\ComfyUI\comfy\utils.py", line 20, in <module>
import torch
File "D:\Comfyui\ComfyUi\venv\lib\site-packages\torch_init_.py", line 262, in <module>
load_dll_libraries()
File "D:\Comfyui\ComfyUi\venv\lib\site-packages\torch_init.py", line 258, in _load_dll_libraries
raise err
OSError: [WinError 1920] Доступ к этому файлу из системы отсутствует. Error loading "D:\Comfyui\ComfyUi\venv\lib\site-packages\torch\lib\aoti_custom_ops.dll" or one of its dependencies.`

What I've tried:
Installing, repairing and reinstalling Microsoft VC Redistributable (newest and older versions)
installing OneAPI toolkit (however no guides told me to use it at the end of the day)
installing different versions of python
doing all of the above, while remaking the venv for 20 times.

what could be the other issue that stops me from using Comfy?

[deft bloom]

also, is the process of installing comfyUI for Arc Battlemage series any different to installation for Arc Alchemist?

nevermind; I've remade the whole installation process using conda and now it works 😉

[earnest grotto]

@deft bloom you should install comfy using my script
for the time being you should also use pytorch 2.5 not 2.9 (current stable) or 2.10 (current nightly)

you do not need the basekit

^

[earnest grotto]

Anyone else getting device lost very commonly? Even had llama cpp crashing, probably due to the same issue. I'll try older drivers

[lunar thicket]

Why not include a pytorch 2.8 option in your script @earnest grotto ?

[earnest grotto]

why yes include a 2.8 option
does it not have the blackscreening?

[deft bloom]

where can I find your script?

[earnest grotto]

^

Here ^ #1193952640225267802 message

[deft bloom]

ty

[lunar thicket]

Nope it does not. Seems to have surfaced in 2.9.

I know keeping a menu option for each version isn't really sustainable, or reasonable. And having user input complicates + opens an avenue for user error.

But iirc we were doing testing in here months ago and 2.8 was leagues above 2.5 in performance.

[earnest grotto]

yes i'll add it, it's not a big deal

[deft bloom]

@earnest grotto when I'm trying to run the script, I get this error on line 259:

line 259, in get_gpu videocontroller = subprocess.check_output([POWERSHELL, "(Get-WmiObject Win32_VideoController).Name"]).decode(TEXT_ENCODING, errors='replace') LookupError: unknown encoding: 65001

I tried to google and asked GPT, couldn't find a proper solution...

[earnest grotto]

In general, if you have spaces in your username or are using a non-english language for windows (not just as an input method), some things will break. I'm pretty sure conda breaks (or used to break recently enough) if you have spaces in your username on at least 1 OS
The error you're getting is due to the non-english language. I'll ponder what caused that specific error

[deft bloom]

Okay

[earnest grotto]

@lunar thicket @deft bloom New script version, 2.5
redownload, delete old one
has pytorch 2.8
hopefully fixes the text encoding bug above

[lunar thicket]

Im gonna compare its speed using the same workflow i did with 2.9 and 2.10dev

if it is close, but generates 0 black images, it'll be a win for sure

[deft bloom]

hey, the encoding bug is fixed. I still have one issue left (it says TorchCompileDiffusionOpenVINO and TorchCompileVAEOpenVINO nodes are missing), but I'm pretty sure I can just find these nodes separately

[earnest grotto]

there is no point in using openvino to compile

and the normal torch compiling is not included in the windows pytorch, it needs the oneapi base toolkit and even then i think they've misconfigured something.
it works on linux for non-sdxl models.

[deft bloom]

okay, if there's no practical assistance in using openvino here, I'll just ignore it

[lunar thicket]

Yeah so I ran 2.8 for 80 images using the same prompt and etc. 0/80 were black images.

[earnest grotto]

current 2.10 nightly is still slower than 2.8+ipex at training a lumina 2 lora

hopefully this time the base model netayume 3.0 is actually decent enough for my lora to be usable...

[earnest grotto]

12.17s/it vs 9.60s/it at step 50

[lunar thicket]

significant slowdown

in my testing of simple SDXL runs I found 2.8, 2.9, and 2.10(from a few days ago) all within a very small margin

[earnest grotto]

the only case so far where using ipex actually speeds things up, ipex drastically slows down everything else i've tried

2.8 ipex makes both sdxl lora training and inference decently slower

[lunar thicket]

What makes you curious about IPEX testing still?

[earnest grotto]

i just want to try and squeeze something out of lumina 2

neta lumina, and netayume, a finetune of it, have been the only notable new anime models since illustrious/noobai

onoma (illustrious) were training a lumina finetune and cagliostro (animagine) were going to release an sd3.5 anime finetune in "Q1 to Q2" of this year IIRC, aka april to september, and... september passed...

i really want netayume to work but the results I'm getting are just worse than with the random noobai vpred finetune i'm using

worse at styles, worse character knowledge, anatomy as bad or worse. understanding more complex prompts like "left is white right is black" is cool but i feel it's kinda pointless when the model is worse at more fundamental things

i know from my past training attempts ipex was faster at lumina 2 and kontext lora training

hopefully cagliostro didn't decide to give up training their model or something. lumina 2 lacking CLIP might be a pretty massive architectural flaw

tried to mess around with the conditioning from the llm, or even the features lumina 2 makes out of them that it then inserts in its fat vector of text+image features, and i couldn't get anything close to how you can just make sdxl/1.5 listen to a token more

only less, or frying the model

and iirc clip still giving useful conditioning even if not as complex was part of the reason why sai and bfl still used it in sd3 or flux

[slender zodiac]

Happy Sailor Moon is the goal 🤖

[earnest grotto]

netayume should be able to make good looking happy sailor moon

well, I won't test now cuz lora training and I'll be off my pc

[earnest grotto]

damn, 2 new local video gen models in 1 day

and apparently after the meme 80b model, now hunyuan are releasing a distilled model as well right off the bat

[earnest grotto]

noo, discord killed it

i guess i might need lower lr for lumina

Half resolution, a bit harder to make out how the neta results are soft

0 570 1140...4560 steps
Netayume 3.0
Noobai Vpred 1.0 with the EQ VAE

[slender zodiac]

It looks like someone has mixed Sailor Moon and the Powerpuff Girls 🤖

[earnest grotto]

The style I'm trying to replicate can be similar to powerpuff girls i guess, but a bit more pointy rather than rounded

Also this image isn't in my dataset or if it is i upscaled it, don't recall

I'll leave it to cook again but this time I won't bother with overtraining it past 2000 steps i guess

[slender zodiac]

Nice style 👌

Are you LORA training with ARC?

[earnest grotto]

yeah

a770 16gb

[earnest grotto]

Here's 3 more styles + that style again

I must've missed something with the top left, issue with removing alpha

but, fixable

retraining bottom left to see if different captions make better results, or if I can squeeze something out of netayume

I can whip out even more. Been training quite a few styles

[slender zodiac]

Looks like Mikoto cameo in Witchcraft Works 🤖

I'm looking at training one for a character I play. Any recommendations or things I should know?

[earnest grotto]

Besides installing comfy, my script can also set up kohya_ss for you for easy training (though i haven't tested this much recently, I rarely update my kohya, either way it should work)
You will also find this https://github.com/jhc13/taggui pretty useful for managing your images and their captions
training sdxl's text encoder is kinda pointless
Pick a model suited for what you want to do (anime model if you want to do anime art, ...)
Caption the way the model is prompted. anime models are prompted with booru tags, taggui has autotaggers. they work ok but can sometimes produce garbage. captioning very incorrectly (e.g. image with black background captioned "white background") will break the lora, though autotaggers generally don't caption THAT bad.
Loss as displayed in the trainer is meaningless. I bothered making some much better loss visualization myself, you will not see anything too informative or surprising if you had the best truest loss shown to you
Usually training for 2000 steps is enough per subject/style for the model to learn it, if it's nothing too alien. I trained Kumo a while ago but don't remember now if i needed to train that for longer
if your dataset is not diverse, the model will pick up on that. E.g. all in-game screenshots -> strong in-game screenshot style, dataset has mostly the same poses -> model will make those poses.
<10 images, you will fry the model to where it's almost unusable. ~20 images, low but lora can be usable. ~50 is a good amount.
I don't think flipping your images will save you from frying the model but it's something good you should do after reaching the above count just to get a bit more images for free

If you have some specific issue with training feel free to poke me or post in the kohya thread i guess, though it's basically just me and disty there (or here)

[earnest grotto]

With Qwen Edit now a thing, you can use it to buff up your dataset like removing text or HUD in case your game doesn't let you do that (deliberate HUD removal might be a bit harder though)
I don't know if anime-ifying your dataset with Qwen/Kontext would be good for an anime model 🤔

Training on video game screenshots works, but I would not recommend it for anime if you can avoid it, like in the above ^ case (there was and is lots of katress art). In your case though, it might be good enough. Not sure how visible it is in the 2nd image but the lack of unique katress poses is very strong in the lora

Tagging every single screenshot as "video game screenshot" and "3d" likely helped. This is what happens when prompting with those 2 captions instead of putting them in the negative...

might be a tad overtrained, don't membe anymore

[earnest grotto]

This being said, I've also had some success in doing this to a real-life monument with an anime model in order to anime-ify it before kontext/qwen edit. It... sorta worked

[lunar thicket]

50 images is a good dataset for lora training, did I understand that right? I would have guessed more

Are there diminishing returns that you're aware of after a certain quantity?

[earnest grotto]

More is better, I have just not bothered to specifically test 20, 30, 50, 100, 300, 1000 etc. images all with the same dataset for a specific concept the model doesn't already know. Gathering 50 screenshots is pretty feasible, getting more screenshots that are meaningfully unique gets a bit tedious. If you're scraping boorus though, or right clicking on twitter or whatever, sure
With less images you will start to overfit sooner and more disastrously, and more images are the opposite.
Anecdotally, loras with more images combine with other loras better, that's likely down to more diverse dataset
Since usually 2000-ish steps are enough, though you can train more for a stronger effect, going over 2000 images for a single concept the model already understands is kinda pointless

If you're training truly outlandish stuff then you will need more. However you probably will need far more than even 10000 steps or images, so

E.g. rewriting the model's understanding of language, adapting it to some vae, doing architectural changes, teaching an anime model about real stuff proper

best comparison i can give you is, 30 images vs.... ~200 but it's not exactly 1:1 since a good amount of the images i added also include another character and i don't think these 2 match the step counts 1:1
you can see how with more it's not breaking with further training

[earnest grotto]

comfy has no support for reserve-vram with hunyuan video models

so 16gb are just barely not enough with the default sittings with the 720p model

at 0.4mp, fp8 scaled, 81 frames and the cfg-distilled model, i'm getting a blazing fast 60s/it

make that 59s/it

(not running out of vram, ~14.8gb used)

[slender zodiac]

That 16GB on the LE is working out well. No way I would have got the 8GB version

16GB or . . . lool

[earnest grotto]

hunyuan 1.5 vs wan 2.2

maybe with a 4step lora and working reserve vram it might be good. some of the reddit results i've seen looked alright

but damn...

[reef ivy]

Wan is king atm

[earnest grotto]

I should probably also mention, doing masks (if what you're training is maskable) is generally good

[earnest grotto]

When training kumo (a spider), no masks often introduced too many limbs (like in the above monstrosity, and that's humanoid!), doing masks helped massively, it still got them wrong often but much less so

[earnest grotto]

And Qwen was already pretty damn big...

Jesus

Flux 2 is 32B with a 24B text encoder

And once again dev has the bad license and the distilled model they'll be releasing later that I have a strong suspicion will be fairly bad, is the one with an apache license

[indigo urchin]

Are there any downsides to accessing Comfy via the workflow button on ai playground as opposed to just installing comfy by itself via your script @earnest grotto ... I built a new pc and currently my comfy is having some issues like no manager button and custom nodes not showing up when everything was working on my previous build(with same GPU).
Just wondering if it wouldn't be easier to uninstall ai playground/comfy and just install with your script. I don't use playground except to access comfy

[earnest grotto]

What this script does is give you the ability to pick pytorch version, install a bunch of good custom nodes, and installs disty's hijacks so that custom nodes and a few other things work
newer pytorches are a bit faster. they also break with battlemage
you can also more easily manage your comfy this way (everything is contained within the Comfy_Intel folder this script creates, managing the conda environment should be easier, etc.)
aipg has its own monkey patching but I don't think it's as extensive as what disty did. i don't know for certain

[indigo urchin]

by "break with battlemage" what do you mean. I'm using a b580.

Sounds like I'll delete comfy and playground this evening and just install your script and hopefully be back up and running

[earnest grotto]

pytorch 2.9 and the current 2.10 nightlies will sometimes produce black images

aipg might be sticking with 2.8 due to this

2.10 has a ~10% performance boost over 2.8

[indigo urchin]

So just so I'm understanding right(sorry I'm new), I should be fine following your script? Or it will more than likely cause issues?

with my b580

[earnest grotto]

You should not have any issues with aipg
if you follow the defaults with my script you should also not have any issues

[indigo urchin]

Thanks for your help!

[lunar thicket]

Can confirm that 2.9 and 2.10 on battlemage produce black images roughly 20% of the time (i too have B580)

Vik in his awesomeness produced a 2.8 pytorch option in his script for this reason

it is worth it to go 2.8 because yes the others are a couple seconds faster per generation but losing 20% of your outputs easily eats up that speed difference

[earnest grotto]

Someone made GGUFs for the flux 2 base model. now to wait for the text encoder

actually, now to wait for the smaller model and/or ltx loras. 1 minute on a 4090 is not good

[reef ivy]

is the black output b series only or does it also happen on alchemist?

[earnest grotto]

only battlemage

you can use the 2.10 nightly

or 2.9 stable

[civic charm]

I get random black images on A770 with loras too

[earnest grotto]

thonk
i don't

hmm

[civic charm]

only happens when the model is quantized

doesn't happen on amd

[earnest grotto]

any model? sdxl? i have not had any black images with the gguf quantization so something in sdnq must be triggering it

i wanted to try sdnq with lumina dimoo but it ended up using massive vram anyways and I think that might've been dimoo's fault. needing a 40gb gpu for an 8b model certainly makes me raise an eyebrow 🤨

[civic charm]

sdxl

[lunar thicket]

does 2.8 alleviate the issue for you too?

[civic charm]

2.8 straight up segfaults with triton / torch.compile

so it doesn't start / run at all

[deft bloom]

these generation speeds seem to be very slow for fp8 quantized models. I assume there's insane cpu offloading, but does anyone know how to avoid such huge offloads and load models fully on arc's b580 VRAM so that it flies and not drags me into buying subscriptions?

371.9 sec per step is insanely slow

[earnest grotto]

yes, you ran out of vram

flux (and flux krea) are a bit more than 11B, which is a bit more than 11GB at fp8
and windows uses a little extra

launch comfy with --reserve-vram 7

if you're using my script, that's what the lowvram shortcut is for

[deft bloom]

yeah, I'm using your script and by default it's set as low vram mode; you want me to use lower vram script?

[earnest grotto]

The script creates 2 shortcuts. One titled "ComfyUI", and one titled "ComfyUI_Lowervram". Use the lowervram shortcut.

[deft bloom]

okay, I'll try

isn't it going to offload even more to my RAM and make it slower?

or proper offloading actually benefits to performance

[earnest grotto]

.

[lunar thicket]

suboptimal

[deft bloom]

yeah, I'll have to still learn a lot about running ai models locally...

[lunar thicket]

this is a good channel for learning

im still new as well

[deft bloom]

I dived a bit deeper, turns out that Comfyui installation on HDD is a huge no-no. I think that the bottleneck doesn't even happen because of my gpu, but because of the installation location of Comfy. Potentially, I can speed up the loading process for up to 10x

[earnest grotto]

hdd is an upside only for small models, like if you have tons of loras

otherwise you are going to be waiting quite a while for models to load

[deft bloom]

that (and the fact that the gpu load rallies like a saw waveform) explains 500s diffusion model load

[deft bloom]

I'll get to loras a bit later, firstly I need to master generic workflows 🙂

[earnest grotto]

do you want to make realistic stuff or anime art

[deft bloom]

primarily realistic stuff, I'm thinking of trying to do ai promotional videos

I know that 12gb of vram is a hell of a choice for it, but I couldn't afford more atm

[somber trellis]

Uh-huh.

[lunar thicket]

[somber trellis]

nobody has a gguf quant of the mistral 3 flux2 encoder yet

[earnest grotto]

the person who did the base model quant said he will quant the text encoder later

[somber trellis]

Yeah, I read that.

Honestly not sure why I'm downloading base flux.

With a model this big I should just wait for the gguf quant of the schnell version of flux2

[earnest grotto]

no schnell

[somber trellis]

Nevermind then

just gotta wait for a 4 step lora

[earnest grotto]

klein means small. they're distilling the model into a smaller one of a currently unknown size

and i don't think they've said when, either

could be a few days. could be a month.

[somber trellis]

Big-ass model with big-ass requirements.

[earnest grotto]

holidays are nearing. kinda doubt they're rushing to release a smaller model now

[somber trellis]

Also, in your testing is Hunyuanvideo 1.5 worse than Wan A14B 2.2?

[earnest grotto]

^

See for yourself

[somber trellis]

LOL

What even

[earnest grotto]

It's possible it fries hard outside of its intended resolution. even if that's the case, at the time i tested it, reserve-vram was broken for it and it was still crazy slow even at the smaller resolution when not running out of vram

20 minutes for that video

4-step loras and fixed reserve-vram might save it. the videos i've seen on reddit seem... comparable to wan. sometimes better. however those aren't here now

[somber trellis]

Wan 2.2 already being very flexible on resolutions, I think I'll keep that downloaded for now.

0k, what.

[earnest grotto]

yeah, i like to be able to make a video in under a minute to preview how the model understands my prompt

[somber trellis]

Flux2 is multiple times bigger.

And I'm getting 20s/it?

15 seconds?

[earnest grotto]

possibly normal. I've seen someone claim ~6-7s/it on 3080ti

[earnest grotto]

wdym?

[somber trellis]

Still going down.

I mean compared to flux1.

[earnest grotto]

well, yes
11b vs 32b

[somber trellis]

I would've expected the time to be much longer given it's FP8 size is 30 gigabytes.

[earnest grotto]

slow offloading and slow gguf is less impactful for a model that is naturally fat and slow

[somber trellis]

fat and slow + fat and slow = somewhat less fat and slow

Yeah, it's down to 12s/it on 1024x1024.

[earnest grotto]

not bad

[somber trellis]

Alright, it's dipping at 11.6 to 11.7s/it

Generation successful.

5 minutes per gen

I'm fine with that, given the size.

[earnest grotto]

god i hate that sloppy anime style

[civic charm]

I get 8 s/it on A770 and 5.5 s/it RX 7900 XTX

[earnest grotto]

damn, that's sad for the 3080 user

[civic charm]

RX 7900 XTX drops to 4.2 s/it with INT8 matmul

[earnest grotto]

pretty good numbers though

[somber trellis]

I'm getting 11s with reserve vram set to 8

Windows 11 🤷‍♂️

[civic charm]

I was using diffusers with SDNQ tho

[earnest grotto]

comfy and the gguf custom node are less optimized

[somber trellis]

SDNQ

bigbrain

[earnest grotto]

and sdnq is better than gguf

[somber trellis]

Your SDNQ implementation is the only reason I got vibevoice working.

[earnest grotto]

linux vs windows performance now should be near identical generally

[somber trellis]

SDNQ Comfyui Node?

[earnest grotto]

i'll try it again, however that'll be only for 8 bit int, no 7-bit/6-bit/4-bit etc., since those are for diffusers only

[somber trellis]

int8 would be fine

[earnest grotto]

and no fp8 either

[somber trellis]

int8 is what i used for vibevoice, couldn't do int4

it really didnt like it

[civic charm]

uint4 is better than int4

Also svd quantization has much better quality

But slower speed

[somber trellis]

does flux2 fit in 16gb of ram at uint4?

[civic charm]

No

[somber trellis]

But I assume it's much faster.

Much less memory having to be moved.

[civic charm]

Haven't tried running flux2 in int8

RAM usage is too much

[earnest grotto]

it's 32b (rounded down i'd assume)
exactly 4 bit quantization would mean 16gb

[somber trellis]

multiplicative quantization be like

[civic charm]

https://huggingface.co/Disty0/FLUX.2-dev-SDNQ-uint4-svd-r32

18.7 GB

[earnest grotto]

oh yeah, sdnq does the text encoder too

[somber trellis]

I kinda do wish we just had a node that we could pass any model through for sdnq quant

🤷‍♂️

But I got a gguf quantized version for now
ittl do

i clearly need to read more into how quantization works

linear and zeropoint quantization

Cool first image.

[earnest grotto]

that doesn't matter, what's stopping sdnq in comfy is that comfy has its own custom separate from diffusers inference, and that the inplace quantization (8 bit only) need(ed?) a bit more stuff added to work
or you can just use sdnext directly

i'm gonna curse the eu hard if that klein model turns out to be bad, bigger than 12B, or worse yet both at the same time

need to check if the eu is giving money to bfl like the us is to closedai

[lunar thicket]

Lol closedAI

[earnest grotto]

for a company named "open" AI they don't release open models very often

pretty sure meta has released more than them

they used to, but not now

[somber trellis]

sam cultman

Testing 2048x2048 now.

40s/it atm. Expected slower, but not too bad considering it's 4x the pixels.

[earnest grotto]

will be interesting to see the comparisons with qwen edit 2511

[somber trellis]

2511 got me excited

[earnest grotto]

also, dunno if you've seen this but it's probably worth noting, they might have explicitly trained it with json prompts and color codes
which is good https://docs.bfl.ai/guides/prompting_guide_flux2

[somber trellis]

I'll take that, throw it into gemini with my current prompt and make it output it in this format

[reef ivy]

Torch.compile and teacache used to help me with flux. I used the q4 models though on a750. Also reserve-vram * edit also turbo loras

[earnest grotto]

torch compile doesn't work on windows out of the box and intel may have broken something in newer pytorches, pointing to the various C/++ compilers in the basekit doesn't work and each one breaks in unique ways

[earnest grotto]

works at q4

wonder if i should make some image to taunt the subreddit that it works on intel

looks to have the same discoloration and shifting present in qwen edit

[lunar thicket]

Holy shit json prompting sounds dope

[earnest grotto]

hmm
getting a strong feeling now that flux2 might have been trained on genshin

or some anime-style game

(other models do not retain the video game odd shadow burn mark on that wall, my bad, flux turning into a bad shadow was its own creation)

also i guess worth mentioning if you're seeing me use that image for the first time, those dudes are wearing caps, not weird cap-shaped haircuts

JESUS

nope this is cooked with text qwen beats it 1000%

literally as bad as the 13.7b pruned qwen edit model with the lightning lora and that lora wasn't even trained for the pruned model nor did the pruned model itself get any further finetuning

1 more try just in case

it's cooked

deep fried butter, deep fried ice cream
that's how cooked it is

2 original images (I don't remember which I used when, imagine this is just 1 image)
flux 2
13b pruned qwen edit 2509, without lightning lora
random qwen edit 2509 and original qwen edit gens, probably with lora

the chatgpt looking one is from the initial, not 2509 model
oops actually it might be a different prompt

the 13b pruned model was fast and all (at least, faster than regular qwen edit) but to me it degraded too much so i chose to not mention it here. well, evidently flux is worse

flux looks to be a bit more responsive to "enhance/upscale the image" stuff though

at a glance, not noticeably better at colorization than qwen, just on par

[deft bloom]

jic if someone wondering: after changing the directory to my SATA SSD from HDD, the diffusion load, DualCLIP load and CLIP text encode went down enormously (i.e. diffusion loader went from 560s to 30 seconds!)

[civic charm]

HDDs are as slow as a Gigabit internet connection

You are better off downloading the models from huggingface every time than using a HDD if you have a gigabit internet

[earnest grotto]

Meanwhile alibaba will be releasing ANOTHER model besides qwen; "z-image"

6B

[reef ivy]

Really? I heard that newer pytorch was faster is that compared to with or without torch.compile?

[earnest grotto]

it's faster compared to old pytorch

[reef ivy]

I mean compared to old pytorch using torch.compile or without?

If not then might not be worth it to upgrade tbh.

I think i last I used was nightly 2.7

[earnest grotto]

I don't use torch compile on windows

[somber trellis]

Flux.2 is capable of doing style transfers.

Which is very nice.

[earnest grotto]

...every model can do pixel art

i wouldn't be surprised if cosxl or instruct pix2pix could turn an image into pixel art

and for a 32b model this is not nice at all, it has the typical not-quite-grid-aligned errors

with the json prompts though, i wonder if you can specify an exact limited color palette

seems kinda out of scope. the vibes i'm getting is they intended this to be a model for autogenerating massive amounts of product images with some text around

[somber trellis]

that image above has a lut and betterfilmgrain applied, this is the original

but yeah, im not entirely agreeing that every model can do it. Image edit 2507 really hated it.

As for grid-aligned issues, I could just use a pix2pix downscaler

[earnest grotto]

https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files

it out

[somber trellis]

supposedly it aint that bad

and its pretty small

[earnest grotto]

yes, 6b

[somber trellis]

fp8 is very fast, 9 steps for the image.

22 seconds

[earnest grotto]

close to sdxl speeds

comfy says it's based on lumina. lumina had a pretty massive boost from torch compile, from ~1.6s/it to 1s/it. dammit. really don't wanna reinstall the basekit again and test out every compiler included just to see how each has different broken include directories passed to it...