ComfyUI for Intel Arc using IPEX

[reef ivy]

1316 sec's. 512x512. Probably can optimize but just seeing if ti works. q4_ks gguf

LTXV lol

but that was probably done in a minute or 2

[earnest grotto]

install with my script or see the 4th pinned post on downloading disty's hijacks and applying them to comfyui

[reef ivy]

I bet if this could run in ipex 2.3 that it would take 10 minutes for that first one. No wavespeed support yet and only teacache available with kijai's nodes which likely won't work with arc.

Gonna try 2.7 again after this last generation and see if the memory error persists. Also adaptive guidance seems to have brought the speed from 60s/it to 54it/s so that is something.

[wicked fulcrum]

really intersting comparison

[reef ivy]

oh actually, after hitting 7 steps the speed shot up to 25s/it, this may be the issue with memory in windows

[wicked fulcrum]

I tried to get wan 1.2 1.3B going but got Ernno error 22 in KSampler. I think being inside AIPG structure I hit file path limit. Need more time to troubleshoot

[reef ivy]

lol yeah, ltxv is actually pretty good with closeups though, but actions like running and walking look funny (at least in the first model, didn't try this image with 0.9.1

[wicked fulcrum]

You do both on your A750?

[reef ivy]

might need 2.5 and above, i think aipg is on 2.3 still

[reef ivy]

yeah, I only have an a750.

[wicked fulcrum]

We move off IPEX 2.3 in the next release. Want to move to PT 2.6 if we can

If not then we go to IPEX 2.6

[reef ivy]

this is on 2.6, and it is slower still but things like wavespeed can make up for it but it's not available for everything yet

[wicked fulcrum]

Li has AIPG ComfyUI working on PT 2.6. I assume these vids include the workflow. I'll drop to him to test on 2.6

[reef ivy]

loaded the wrong image, but with this workflow it went from 20minuts to 12 minutes, adapative guidance seems to work.

[reef ivy]

they are workflows from reddit, i just changed the load nodes to gguf etc.

well the ltxv one is highly customized with llm's and stuff, not sure if that one will run right in aipg

that one is also old, and doesn't have wavespeed in it yet. I just posted to compare, oh and it's a complete mess lol

[reef ivy]

yeah, something is defintely wrong wtih memory allocation in 2.7. Gonna try with lowvram(usually just need reserve vram)

[reef ivy]

disabling the ipex hijacks gets it to run, but now it's extremely slow. I wonder if something needs to change in the hijacks for whatever they are changing in 2.7?

[vital hatch]

Hi there Bob.
I am encountering the same requirements-unknown.txt missing file issue.
The system I have is:

Processor Intel(R) Core(TM) Ultra 9 285K 3.70 GHz
Installed RAM 64.0 GB (63.4 GB usable)
Video AMD Radeon 7900 XTX

I have installed vc_redistbutables.
I uninstalled python 3.12 too in the trials.
Also installed as admin and as for local user.
None seem to work.

What other info you'd need for this?

Thanks!

[reef ivy]

If you are installing with the old method you need to download the requirements file from the post and place it in the comfy folder. Your best options though are Vik's script or ai playground, check the pinned comments for the script these are easier to install and more up to date.

[reef ivy]

If I up the length or resolution even a little it shoots to over 30minutes from 10. Tripled slowdown for some reason. 512x512 33 frames is the most it can do without going super slow. The slowdown doesn't seem relative to the vram increases.

[wicked fulcrum]

The issue from that discussion is specific to A730M. It was fixed with firmware driver update.

Your issue is you do not have an Intel Arc iGPU or dGPU on your PC. You do not meet min specifications

[earnest grotto]

Given your 7900 XTX, I suggest using SDNext with ZLUDA instead https://vladmandic.github.io/sdnext-docs/Installation/

I don't know how zluda is implemented in sdnext and don't know if there's any way to easily get it going in comfyui. Go to an amd discord and hope someone helps you there?

Alternatively, look into WSL2. It's extremely easy to set up

You should be able to use ROCm there now, that is, with latest drivers and such

[civic charm]

SDNext Discord has dedicated zluda thread in help channel: https://discord.com/channels/1101998836328697867/1207548125841989702

Zluda wiki: https://vladmandic.github.io/sdnext-docs/ZLUDA/

[civic charm]

invite link for the discord server is in the readme of sdnext

[reef ivy]

and... a new ltx video model released. video is coming in fast

[reef ivy]

new model definitely understands human motion better

messed up the prompt, had something left over from another gen about the earth rotating lol. might explain weirdness

[earnest grotto]

I think I like the first one better

[reef ivy]

me too, second (and the third one I just did) get wierd lol.

might need to change the prompting since they updated the model lol

[reef ivy]

Hunyuan img2vid is out now too, will probably do another comparison if i can get it to work

[earnest grotto]

I've updated the script with some setup for the 3dpack. Everything past the setup is completely untested on windows. 2.6 will not work (no torch-scatter for it). 2.5 might work.

#1193952640225267802 message

i also have no idea why pyopencl doesn't detect the gpu on windows. oh well, multithreaded cpu is probably good enough

[reef ivy]

hunyuan i2v.

I feel even ltx keeps the features better, but hunyuan still has better movement. Gonna try the q3 quant of Wan2.1 and see if I can still get decent results but be able to make longer videos without taking an hour for 4 frames.

Teacache may have still been used in that output, I feel comfy is having issues unloading anything that goes to ram atm.

[somber trellis]

It looks like wan 1.3b also can make single-frame images

[reef ivy]

https://github.com/bitsandbytes-foundation/bitsandbytes/issues/1545 was looking up bitsnbytes xpu support.

[reef ivy]

I wonder if you could make a single frame image, save the seed and then use that to generate a video. So basically doing img2video by getting the the image preview first?

[somber trellis]

It's just so interesting to me however that it is a 1.3b model

capable of decent imagery

cool ah mask helmet

[reef ivy]

How are the videos? I haven't tried the 1.3b yet since I'm more interested in img2vid

but, if I can control the output with an image like that quickly maybe it's worth it

[earnest grotto]

pixart sigma is 600m and I think beats it

for images

[somber trellis]

i mean the 1.3b model

doesnt even load the gif

lmao

thats the 1.3b model

imma try some weirder prompts.

[somber trellis]

monster man

more steps woulda did this some good

[reef ivy]

I am finding that frame length has more impact on speed then resolution. It may be you need to use what its trained for, went to 480p amd it was only 2 minutes slower. Went to 41 frames from 33 and it went to over 30minutes.

Also didn't seem ti use much more vram or ram.

[reef ivy]

starting to get things going, almost 14min, 480p 41frames. took me 30 minutes before at 512x512 and 41 frames for some reason.

I feel like you can do some really cool stuff with this model. Would love to see what people can do on the a770 with the img2vid

wish lowering the res would allow me to do longer videos quicker

[sly trench]

Hi @earnest grotto . I encountered this error.
My system is U7 258V. Is there any way to fix it? Thanks

[earnest grotto]

Scroll up and show a screenshot

[reef ivy]

New hunyuan model uploaded, fixed the glitch of not following the input picture

[earnest grotto]

@trim onyx Are you specifically looking for spar3d

[trim onyx]

Yeah, I prefer SPAR3D if possible.

[earnest grotto]

the comfyui 3dpack says it has stable fast 3d
I've added some extremely untested installation of the 3dpack to my installer script. if you want to gamble on potentially wasting your time, you can try that
but it'll be a gamble

[trim onyx]

Hmmm I may just try running it on the other GPU I have handy, though it'll take longer to set-up :I

[reef ivy]

anybody tried this? https://github.com/intel/intel-xpu-backend-for-triton/blob/main/.github/WINDOWS.md

[earnest grotto]

Hey, still here?

[reef ivy]

getting an error with the hunyuan clip vision model now in latest comfy update RuntimeError: Non-uniform work-groups are not supported by the target device I think it's either arc or torch 2.7 related? other clip vision models still work

[reef ivy]

yup, seems to def be an issue with 2.7 nightly.

[stiff kindle]

Hi. I'm trying to install this, but it's not working. When I get to the point where I run this command, python main.py --bf16-unet
It ends up giving me an error saying the frontend package is not installed. It tells me to install it using the update_comfyui.bat file, but I did that and after I still get the same error. Please help.

[somber trellis]

its a module in requirements.txt

look inside the comfyui folder

[stiff kindle]

For what?

[somber trellis]

pip install comfyui-frontend-package==1.11.8

[stiff kindle]

Thank you.

[somber trellis]

👍

I currently can't seem to use comfyui at all right now.

It blackscreens displayport after a few runs.

tested it on a new install with just pytorch 2.3.1 ipex and also 2.6.0, and the current 2.7.0 nightly.

this might be gpu driver related so i might just ddu it

[reef ivy]

Comfyui has separated itself from ui and backend I think, install instructions may need to change.

[somber trellis]

Is anyone else having blackscreening using split optimization

[reef ivy]

Haven't tried split optimization, do regular work flows still work?

[somber trellis]

it seems to work just fine, ipex to cuda auto swaps to it when you reach a certain vram limit

or maybe thats comfy idk

🤷‍♂️

[stoic forum]

Is there a way to change the install path for ComfyUI to C:\ and not to System32?

[earnest grotto]

Don't run the script in system32

The script installs where it is ran.

[stoic forum]

I don't run it from sys32 but from usb stick... so just put it on c:?

[earnest grotto]

Put it somewhere in your user folder and have it install there.

You can move the resulting thing around, or as it says it makes a shortcut for you that you can put anywhere.

[stoic forum]

From where ever I start the script, it says: A folder Comfy containing ComfyUI and Conda environement "cenv" will be created in C:\Windows\System32,

[earnest grotto]

You can shift+right click the background of a folder to open a command prompt in that folder.

You are probably opening an admin command prompt in system32 and not moving it elsewhere.

Open one it the script's folder. You don't need admin permissions.

[stoic forum]

I have the option to rightclick and open cmd with adminrights in every folder, do this since aged on my win pc via regedit

[earnest grotto]

You don't need admin and you don't need to edit the registry to have the command prompt show up with shift+right click.

[civic charm]

don't use admin rights, that will casue more issues on the ui side

aka ui can't access the venv etc

[stoic forum]

but why open the py script with cmd?

[earnest grotto]

You run the python script with the python interpreter

I contemplated making an executable. I coded some of it but utlimately decided distributing a black box exe is too sus

The script can be run just with double clicking, if you set up your python in a specific way

[stoic forum]

I downloaded the pyscript as suggested from the pin I replied to and run it with python

forget it... I'm already so annoyed, trying to get it running since 2 days with all suggested version of howto's from here, and it always ends with error. It's not that I have no clue from comfy, on my desktop I have everything running (ofc Nvidia, just saying)... but this is freaking me out now. I feel stupid not even I get a script runnin... 🤯

[earnest grotto]

If you want a very simple installation, you can use AIPG

[stoic forum]

no its to much stuff I don't need nor want

clean and simple comfyui, thats all I need 😉

[earnest grotto]

Just run the script from a regular, not admin command prompt, inside the script's folder, somewhere inside your user folder (e.g. Documents)

[stoic forum]

confused: wasn't it enough to install miniconda on default path?

[earnest grotto]

Shift+right click the background of the folder the script is in (e.g. Documents). Open command prompt. python ./Setup_ComfyUI_Intel.py

[stoic forum]

seems the python installation was broken/messed up... it's installing now

thanks!

to clarify... I tried another tut from the PIN before and had python 3.10.6 installed... after that how to doesn't worked, I uninstalled and deleted everything, but something messed up the uninstallation of python. Every try of a re-installation of python was very quick (I already wondered). After I checked the install path of python, I saw nothing was installed. I then did a repair install and that was it, after that I installed miniconda... so finger crossing that, when the script has finished, everything works

1. It runs now! Thanks for your script!
2. I get some warnings and some could not load.

Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended.
C:\ComfyUI\cenv\lib\site-packages\torchvision\io\image.py:13: UserWarning: Failed to load image Python extension: 'Could not find module 'C:\ComfyUI\cenv\Lib\site-packages\torchvision\image.pyd' (or one of its dependencies). Try using the full path with constructor syntax.'If you don't plan on using image functionality from torchvision.io, you can ignore this warning. Otherwise, there might be something wrong with your environment. Did you have libjpeg or libpng installed before building torchvision from source?
warn(

Could not load bitsandbytes native library: Could not find module 'C:\ComfyUI\cenv\lib\site-packages\bitsandbytes\libbitsandbytes_cuda121.dll' (or one of its dependencies). Try using the full path with constructor syntax.
Traceback (most recent call last):
File "C:\ComfyUI\cenv\lib\site-packages\bitsandbytes\cextension.py", line 85, in <module>
lib = get_native_library()
File "C:\ComfyUI\cenv\lib\site-packages\bitsandbytes\cextension.py", line 72, in get_native_library
dll = ct.cdll.LoadLibrary(str(binary_path))
File "C:\ComfyUI\cenv\lib\ctypes_init_.py", line 452, in LoadLibrary
return self.dlltype(name)
File "C:\ComfyUI\cenv\lib\ctypes_init.py", line 374, in init
self._handle = _dlopen(self._name, mode)
FileNotFoundError: Could not find module 'C:\ComfyUI\cenv\lib\site-packages\bitsandbytes\libbitsandbytes_cuda121.dll' (or one of its dependencies). Try using the full path with constructor syntax.

[earnest grotto]

@stoic forum Both of those can be completely ignored

They will always show up

Bitsandbytes currently doesn't support Intel GPUs, but work is ongoing

[stoic forum]

Alright! Thank you again! 👍

[earnest grotto]

Hence why the gguf custom node instead, for quantized flux

[stoic forum]

I can't thank enough after all I tried. This was the easiest solution!

[earnest grotto]

The image python thing was just a thing with IPEX IIRC. It doesn't matter

[stoic forum]

OK

[white steeple]

installing libjpeg-turbo and libpng before installing torch fixed on my side but nothing changed actually yeah can be ignored.

[stoic forum]

Ok

[reef ivy]

Found this https://www.reddit.com/r/comfyui/comments/1j72tn2/monitor_randomly_crashes_to_blackno_input_while/ no answer but similar issue

[earnest grotto]

As the person in the replies says, driver or psu, but i doubt psu

most likely driver issue

Could be the leaking i had seen on windows

Restart pc and check again

[reef ivy]

Is there a node to force the clipvison model to cpu?

[stoic forum]

One more Q... Does it matter which python version miniconda installs? Just saw it came with 3.12.x - But I installed as suggested python 3.10.6 - could there be conflicts? How to downgrade the py from miniconda, never used it before...

[earnest grotto]

You don't install anything, the installer handles everything for you

You need to have some modern-ish python to run the script and that's it

if you want to manage the conda environment yourself, press ctrl+c after launching with the shortcut, and you can pip list or do whatever else you want

[stoic forum]

quote: You will need conda, miniconda or some such already installed, preferably in the default directory (C:/Users/You/miniconda3)

[earnest grotto]

yes, you need some conda

whatever python that comes with is irrelevant

you don't install anything else

[stoic forum]

alright

[earnest grotto]

given that the script launched, you had conda, so

[somber trellis]

well we'll see if i crash constantly afterwards

I just did a safe mode ddu and reinstalled 6647

[reef ivy]

So they fixed the hunyuan model, but now teacache/wavespeed don't work and no gguf quant, so it's just as slow as wan and takes up more ram now lol.

[earnest grotto]

https://www.reddit.com/r/StableDiffusion/comments/1j7cr1y/new_clip_text_encoder_and_a_giant_mutated_vision/

[somber trellis]

testing it now

[somber trellis]

100% doubt on psu, mordhau brings my gpu to a very high power target at near 90 celsius

[earnest grotto]

driver

and probably not fixable by you

[somber trellis]

that's why i ddu'd

went into safe mode and did a full ddu uninstall

[earnest grotto]

i am on windows again and was poking comfyui. the leaks are still not fixed

well, "leaks"

[somber trellis]

Then my only alternative is linux

[earnest grotto]

not exactly a memory leak, at least not one that would show up in task manager

but an issue i've seen for a very long time

[somber trellis]

Wonder when SteamOS 3 is gonna be out of beta.

[earnest grotto]

you can restart your pc and try again

[earnest grotto]

you don't need some special steamos or whatever else, you can just use ubuntu

[somber trellis]

... I know that.

I want SteamOS for it being SteamOS.

I was an endeavourOS and Mint user

🤷‍♂️

Just sad that intel arc vr support on linux

is nonexistent

[somber trellis]

linux mint is probably what id do

with that clip

i mean it definitely seems to adhere better

i previously used a different custom vit-l for text enhancement

[earnest grotto]

this same person did some vit-l finetunes before this

[somber trellis]

that was probably the one i used before

[earnest grotto]

yes it's his again

[somber trellis]

ye

that one definitely made better images

so this newer vit-l is a nice sight

I'm still using base flux dev tho

[earnest grotto]

in the post he says he'll add some hf code, eventually that might hopefully translate into running the even better ones in comfy directly

[somber trellis]

are you running --use-split-optimization vik

my speeds have gotten much slower on 2.7 since I did a clean install of my gpu drivers

[earnest grotto]

no

[somber trellis]

im running python main.py --bf16-unet --disable-ipex-optimize --use-split-cross-attention --reserve-vram 4.0 --lowvram

with the latest torch+xpu 2.7.0 nightly

[earnest grotto]

and i've almost decided to not update my drivers anymore as they've been introducing too many issues since battlemage's launch

and i'm mostly on linux, just on windows for now

want to mod a windows game and it's less headache doing it natively than fiddling with protontricks

[somber trellis]

    import intel_extension_for_pytorch as ipex
    from ipex_to_cuda import ipex_init
    ipex_init()
    _ = torch.xpu.device_count()
    xpu_available = xpu_available or torch.xpu.is_available()
except:
    xpu_available = xpu_available or (hasattr(torch, "xpu") and torch.xpu.is_available())```

[somber trellis]

proton ge was good to use

[earnest grotto]

and for some reason blender on linux causes the intel drivers to start crashing after a while

[somber trellis]

ye for some reason im getting 30s/it with flux

imma just

go back to 2.3.1

[earnest grotto]

It's not a compatibility issue, not about a custom proton version. I want to be able to run a bunch of different things under the same wineprefix without much hassle, and there is simply hassle

[somber trellis]

unhassleable

wine moment

when you want wine to run all your windows app in one app

I mean, they are doing a great job at improving its emulation

I had to mess with wineprefixes for FL Studio when I was on mint

[earnest grotto]

it runs things mostly fine. i've still seen some issues but it's mostly fine-ish

[somber trellis]

I swapped back in the end tho for linux because proton wasn't giving me the performance in specific games that I was comfortable with

[earnest grotto]

and those issues might be more blameable on the mesa intel drivers than proton

[somber trellis]

mordhau being one of em

also gmod eye textures are broken in linux

specifically the layering and how the gloss texture for the eyes are used

linux no likely, causes the eyes to shine with a pink and black hue

lmao

[somber trellis]

ipex_init: (False, AttributeError("module 'torch.xpu' has no attribute 'Tuple'"))

Ipex-init is failing supposedly

[somber trellis]

unless im wrong and im supposed to ignore that

reinstalled your recent comfyui script, back to 3.3s/it on flux

But that error is there

@earnest grotto

[earnest grotto]

this error?

[somber trellis]

Yes.

Your current script gives me that, with both stable 2.6 and 2.7 nightly

[earnest grotto]

fresh or over an existing installation

[somber trellis]

fresh

[earnest grotto]

hmm

what version of the script are you running

[somber trellis]

0.1.5p

i installed 2.6.0 and then self updated to 2.7.0 by activating the cenv

pip3 install --pre --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu

and then fix any dependencies

[earnest grotto]

i

i'll just add a 2.7 option

[somber trellis]

did i do something wrong

[earnest grotto]

no

[somber trellis]

hmm

[earnest grotto]

wait like 10 minutes so i can reinstall

[somber trellis]

ipex_to_cuda is disty0's repo right

or are you using yours

[earnest grotto]

it's disty's
I forked it for some changes and very temporarily made the script download my fork (needed for the 3dpack)
0.1.5p of the script uses disty's

[somber trellis]

thats the one difference between yours and his

[earnest grotto]

i PRed, disty saw it and fixed stuff, i synced my fork afterwards which removed whatever changes

i haven't synced it since

[earnest grotto]

did you choose to install the 3dpack, the smaller set of working custom nodes, or nothing else

[somber trellis]

i hit y on that, didnt install the experimentals

let me just do a complete reinstall using your script

ipex_init: (True, None)

ok it was me somewhere that did the problem

[reef ivy]

Is this helping your speed?

[somber trellis]

it sacrifices a small bit of performance for better ram usage

[reef ivy]

Ah okay, I can't sacrifice anymore performance until they fix the performance lost since 2.1.

[somber trellis]

at least i think

[reef ivy]

2.7 seemed a tad bit faster than 2.6, but I got errors for in comfy with it, at least the nightly from yesterday

If on 2.7 try 2.6, they are definitely cooking something that is messing with stuff atm

[somber trellis]

yep 2.6 resolved it

true,none is back

latest 2.7 nightly broke something

[earnest grotto]

yeah that's kinda the point of nightlies

alphas and all that

[reef ivy]

I just wish they were on github, as some nightly 2.7's ran pretty well. Also wish I could see what was updated or when it was updated

Torch 2.7 is on there but no xpu

[somber trellis]

bruh

I forgot to disable UHD integrated after DDU

no wonder

[somber trellis]

ok it worky again

[earnest grotto]

there's another comfyui-compatible clip finetune as well, check the hf repo in the reddit thread

https://huggingface.co/zer0int/CLIP-Registers-Gated_MLP-ViT-L-14/tree/main

the xtreme one... hmm

[somber trellis]

that is the one i have

i have the balanced one

which is the recommended

[ripe pivot]

my decode takes like half a minute for image gen, is that normal?

[earnest grotto]

Show workflow

[ripe pivot]

will uploading image to discord retain the metadata? or do I have to upload it somewhere else?

[earnest grotto]

the metadata is retained yes

[ripe pivot]

is this fine?

[reef ivy]

Probably something with the facedetailer nodes makes it take that long. Never used them in comfy

[ripe pivot]

Not using it cuts it from 40 to 20 secs, makes sense since the detailer are just auto inpaint. The encode feels still quite long

[reef ivy]

what is your gpu?

you could also try setting vae decode to false and then adding a vae decode node? I haven't used that Ksampler before

[reef ivy]

I don't have the same model or a lora, but I got vae decode in 0.9 secs. So it could be the lora maybe? I am on an a750 btw so it could be longer on different gpu

[ripe pivot]

B580, haven't use the lora

[somber trellis]

ive been testing out different launch args on comfy and it seems --use-pytorch-cross-attention and --reserve-vram 6.0 is optimal

4.0 is slower

[reef ivy]

Yeah, i found 6.0 is now faster too with latest drivers and pytorch.

I am hoping we get flash attention soon(heard vik say its supposed to come in 2.7?)

[somber trellis]

flash attention will speed us up a nice bit

[reef ivy]

I wonder what changed from 2.1 to 2.3? 2.1 was significantly faster in flux than 2.3 and later.

[somber trellis]

I do miss the performance we had.

[somber trellis]

ok i just had a displayport disconnect again but on 2.3.1

this was while running juggernautxl with first block cache from wavespeed

using split-cross-attention --reserve-vram 16.0

lol does alchemist not like split cross attention

definitely doesnt seem to mind the same setup on pytorch cross attention

[reef ivy]

I've never gotten a monitor disconnect when generating, only sometiems when changing resolutions or going from windowed to fullscreen in games.

[somber trellis]

crashed again on 2.6.0 with pytorch cross attention

idunno

[earnest grotto]

generate more images, with more complex workflows, without shutting down your pc

[reef ivy]

So it's a common problem with intel? I don't typically do batches since swapping from a1111 years ago. it's too slow in comfyui

[earnest grotto]

no need for batching

i never do batching

[reef ivy]

honestly, with flux I typically get an image I like right away. But I mostly do samples to test out video models now though.

[civic charm]

my attention hijacks already does splitting to save vram and also not trigger the 4gb block limit, why do you need to use split attention?

also if split atten is using something other than sdpa, my hijcaks won't make any effect on it

[somber trellis]

It's not that I need it, I just wanted to test out its relative performance and savings compared to base attention

eitherway this displayport disconnect thing is happening for me with pytorch attention as well

im kinda excited for flash attention tho

if that truly is going to be usable for us in 2.7

[civic charm]

OpenVINO has it and SDXL gives 2.3-2.4 it/s at 1024x1024

IPEX 2.3 gives 1.6 it/s for comparison

Non flash atten speed of OpenVINO was around 1.8-2.0 it/s

[somber trellis]

So by that comparison id expect a 0.4 it/s increase

25 percent

cool

Right now however, the current 2.7 nightly gives a tuple error

[somber trellis]

Eh sdnext isnt for me

lol

[reef ivy]

For me, sdnext was a bit too complicated to learn. A lot of settings to adjust

comfy seems complicated, but being able to look at workflows and reverse engineer them is better for me lol

[somber trellis]

i just prefer the ui in comfy over sdnext

🤷‍♂️

Something must be happening to cause a workflow like this to just abruptly crash

but even juggernautxl for me is crashing (but right now it's deciding to work so ok)

[reef ivy]

Might need to make a new environment

[covert cave]

hi, im doing this for the first time and try to start comfyui but always end up with this

(GenAI) E:\ComfyUI\ComfyUI>python main.py
Traceback (most recent call last):
File "E:\ComfyUI\ComfyUI\main.py", line 134, in <module>
import comfy.utils
File "E:\ComfyUI\ComfyUI\comfy\utils.py", line 20, in <module>
import torch
File "C:\ProgramData\anaconda3\envs\GenAI\lib\site-packages\torch_init_.py", line 262, in <module>
load_dll_libraries()
File "C:\ProgramData\anaconda3\envs\GenAI\lib\site-packages\torch_init.py", line 258, in _load_dll_libraries
raise err
OSError: [WinError 126] The specified module could not be found. Error loading "C:\ProgramData\anaconda3\envs\GenAI\lib\site-packages\torch\lib\aoti_custom_ops.dll" or one of its dependencies.

(GenAI) E:\ComfyUI\ComfyUI>

tried the first method without anaconda, doesn't work, jumped into anaconda and i got this. consulted chatgpt/deepseek for answers and tried various stuff but it somehow always leads me back to this

Running on a b580

reinstalling pytorch and running dependencies then running to start it up again with python main.py leads me to this

[earnest grotto]

@covert cave Use AI Playground, or install with my script ^

[civic charm]

fixed

[civic charm]

sdnext has multiple ui type options : )

mainly modern and old a1111 like type

rest are themes

default is the old a1111 like type

here is how the modernui looks like. panels are resizable and right panel is collapsable

[covert cave]

what a legend, thank you so much.

[earnest grotto]

for battlemage, AI playground will be faster but might not be compatible with newer nodes, it uses an older version of pytorch+ipex before some performance regressions, my script can't download that for you (for battlemage). the speed difference is less than 2x.

[somber trellis]

nightly 2.7 20250310 is operational again

👍 thanks @civic charm

[somber trellis]

it uses 2.3.110+xpu lol

i checked

[earnest grotto]

yes, but I doubt this person knows what that is or what each version means

[somber trellis]

🤷‍♂️

true

[earnest grotto]

They're using a custom battlemage build. The 2.3 available (last I checked) did not have any battlemage support, and it was released before battlemage after all

[somber trellis]

The latest version that supports battlemage

is 2.5.1

so yes

[civic charm]

ipex 2.6 is out today

[earnest grotto]

Yes, my script downloads that

[civic charm]

that also should have bmg support

[somber trellis]

That's... good.

Are there any benefits of IPEX 2.6+xpu vs base 2.6+XPU

[earnest grotto]

I could dig and see where they download their 2.3 wheel from but i don't wanna do that. seems rude

[somber trellis]

guess ill test for you then

[civic charm]

Couldn't test it because of this: https://github.com/intel/intel-extension-for-pytorch/issues/794

[earnest grotto]

performance differences are different on linux and windows 😛

[civic charm]

got yeeted by the bot lol

[somber trellis]

oh its 2.6.1

ok

[earnest grotto]

no idea what's up with that

[civic charm]

posted messages with links too quickly i guess

[earnest grotto]

It's happened to me a bunch of times
It is what it is i guess 🤷‍♂️

[somber trellis]

I'm getting 1-1.1IT/S on 1024x1024 juggernaut xl euler ancestral at 7 cfg

with ipex torch 2.6

swapping back to 2.7.0 nightly to compare

no contest

pytorch 2.7.0's recent nightly does 1.6-1.7it/s on the same settings

and openvino does 2-2.4

so

🤷‍♂️

another eh pytorch version

[earnest grotto]

I'm getting ~1.95it/s with some images I'm spamming right now to check what noobai's artist tags look like

With 2.3, on linux with kernel 6.12.3

[somber trellis]

NoobaiXL?

NAIXL

[earnest grotto]

896x1152 though

[earnest grotto]

https://huggingface.co/Laxhar/noobai-XL-Vpred-1.0/blob/main/NoobAI-XL-Vpred-v1.0.safetensors

[somber trellis]

100 cfg (7cfg mimic scale)

ye i have it

i have illustrious 1.0 as well

and pdxlv6

[earnest grotto]

seems pointless to me to use non-vpred anime models now that there is a vpred anime model

[civic charm]

i still use epsilon

v-pred models are too fried

[earnest grotto]

I'm happy with it

[somber trellis]

vprediction doesnt like overclocking cfg

lmao

[civic charm]

neither actually good models

cfg is a hack, good models shouldn't need it

[somber trellis]

ok guess every model ive tried that is enhanced with it are bad now

including flux

lmao

[civic charm]

flux doesn't use cfg tho

[somber trellis]

it can be forced to use cfg

by using dynamicthresholding at mimic scale 1.0

and also there is a dedistilled flux that can use cfg

https://civitai.com/model-versions/943787

[earnest grotto]

cfg improves prompt adherence yes
but subtracting one image from another isn't a good thing to do. and it is partially what results in multiple/fused limbs and so on

[reef ivy]

2.7 might have new optimizations, i wonder if there are differences between 2.6 xpu and 2.6 ipex?

[somber trellis]

it seems just as fast as base 2.6

just with added ipex support

[somber trellis]

This fused limbs issue you mention isn't something I've ever gotten from using CFG

[reef ivy]

i was just about to test xpu 2.6 in wsl 2*, finally figured out my errors were from bitsandbytes being installed

[somber trellis]

Unless it's from too much CFG causing burnin

ive gotten fused limbs more from using caching nodes than anything

[reef ivy]

One thing, there is experimental bitsandbytes for xpu, but it's only for ipex. might run on latest? have to compile from source so haven't tried

https://huggingface.co/docs/bitsandbytes/main/en/installation?backend=Intel+CPU+%2B+GPU#multi-backend-compile

might run in wsl

[somber trellis]

nvm it does windows

just idk the viability of this

[reef ivy]

oh yeah, it supports windows too

last time i compiled from source on intel it took like 3 hours lol. might give it a go though

[earnest grotto]

If the model predicts people in the same pose regardless of prompt, cfg won't make broken limbs
But if it predicts even slight differences, you can start to get broken limbs

[somber trellis]

ive noticed this happen eitherway at low denoises on second pass

ipadapters can resolve the issue due to bringing more context of the previous image in

and flux redux can deal with it entirely

[earnest grotto]

https://cdn.discordapp.com/attachments/1091563787749953647/1343168405422870590/denoiser_lq.mp4?ex=67d1628e&is=67d0110e&hm=76e801fff20ea15b38f751d2728b64224f5988bd62f489e954d7d1285befc3ed&

e.g., look at the result at the end of this video, then stare at the 2 predictions and the subtractions at steps 7, 8 and 9

they predict baggy clothes in different poses. subtracting that eventually finangles the model (animagine 4 zero) into thinking it's 2 hands

changing the negative prompt on just step 8 to have more of the stuff in the positive, makes a more similar prediction there near the hand and boom, no more double hand, baggy shirt restored

[civic charm]

i am also happy with epsilon

[reef ivy]

Going Wsl2 seems way slower so far, at least in Hunyuan i2v. Might need some different command args

[earnest grotto]

wsl2 doesn't completely fix the not-quite-leaks

or was dan having the blackscreening issues

[somber trellis]

was me

on windows

[reef ivy]

Gonna try wan in a minute, maybe flux first.

Then 2.7 nightly

Doesn't seem to want to unload models from ram or vram

[somber trellis]

sup

[reef ivy]

think i'd likely need 64gb of ram to make wsl usable, seemed super slow with the video models.

[reef ivy]

2.7 nightly seems about 40% faster atm than 2.6xpu in comfy. But this may be due to recent changes in comfy that made it slower

[reef ivy]

So now getting about the same speed in wsl2 as native with 2.7. trying out torch compile but not sure it helped or maybe it helped it catch up to native

[reef ivy]

You guys using linux do you constantly get errors and nodes not loading because of bitsandbytes?

[earnest grotto]

bitsandbytes does not currently support intel. you will get errors.

[reef ivy]

on windows i get warnings but i can still use nodes, on wsl 2 I have to completely uninstall it to use a lot of stuff

[ripe pivot]

how do I into i2v? does b580 adequate enough for that?

[reef ivy]

ltxv, wan2.1, and hunyuan i2v can all do it. I can do it on an a750 so should work on b580, you just need a workflow. you should be able to download the videos/images I posted and just drag them into comfy then download missing nodes etc.

[earnest grotto]

https://www.reddit.com/r/StableDiffusion/comments/1j8h0qk/new_longclip_text_encoder_and_a_giant_mutated/

[ripe pivot]

I got the UR error trying the WAN demo

[reef ivy]

Post the error log, I don't know what that means

[somber trellis]

uh guys

the latest nightly is installing triton

pytorch-triton-xpu is included in the current nightly

[civic charm]

2.5 and 2.6 also has it

native pytorch will install triton with the correct backend on linux

tho windows build is a new thing

Welp, i guess intel is the first vendor to officially support triton on windows

you need to go through lots of hoops to get an unofficial triton build to run on windows with nvidia

[somber trellis]

also we got a tts that works ok on arc

spark-tts

example text:

output

(I regret this)

[civic charm]

nightlies has switched to torch 2.8

[earnest grotto]

That's a pretty good clone, damn

Tortoise tts would just turn anything british

[somber trellis]

0.5b parameters too

its only using 6gb of vram

with the ipex-to-cuda hijack

[earnest grotto]

link the custom node, i wanna try it

[somber trellis]

It's not a node

its a repository

its own thing

uses python 3.12

https://github.com/SparkAudio/Spark-TTS

There will most likely be comfy implements of it tho

ok this is gonna be really dumb

but im gonna try to do mario teaches typing

[somber trellis]

it tried

[reef ivy]

Triton on windows? So we get torch compile now? Sage attention is not compatible right?

Been on linux all day just trying to get pytorch installed, and then brand new stuff added to windows lol.

[somber trellis]

we'll see

RuntimeError: 2 active drivers ([<class 'intel.XPUDriver'>, <class 'nvidia.CudaDriver'>]). There should only be one.

[reef ivy]

Did you have to pip install sage attention? I tried it in wsl 2 and it wouldn't work for me

[somber trellis]

i did and it downloaded, but hte problem now is that i have two driver classes lmao

[reef ivy]

Are you getting the error with bitsandbytes? If so try and uninstall it

[somber trellis]

no the error is with triton

[reef ivy]

i had this issue in wsl2, ended up installing the experimental multi vendor version

Only with sage attention? It could also be it's trying to use the cuda fall back, there is alo a sage attention node that allows to use triton in the settings.

You should be able to use torch compile though if that doesn't work

[somber trellis]

ipex_to_cuda is confusing triton mainly

i think thats the issue

[reef ivy]

might need an update

[somber trellis]

Input tensors must be on cuda.

[reef ivy]

maybe disable and try a small res that fits in 4gb?

must be cuda only then

[somber trellis]

disabled ipextocuda

that was the error

[reef ivy]

We have to wait for flash attn (unless they added it?)

There are 2 sageattentions, sage attn 1 and sage attn 2 as well

I think there is a fork for amd on sage attn 1 anyway on rocm

torch compile should also add some speed though, if it works in windows

[somber trellis]

latest nightly gets 1.8it/s on SDXL now

without compile

AttributeError: 'torch._C._XpuDeviceProperties' object has no attribute 'L2_cache_size'

compile gives an l2 cache size error

[reef ivy]

with hijacks enabled?

[somber trellis]

ye

[civic charm]

Fixed this one

update

torch changed inductor to lazy import so my triton fix was failing

[somber trellis]

new error

[earnest grotto]

seems intel issue

[somber trellis]

and without sageattention the issue with compile is with a missing attribute, "AttributeError: 'torch._C._XpuDeviceProperties' object has no attribute 'L2_cache_size'"

[earnest grotto]

you can set that attribute yourself. but i imagine that wouldn't be a good idea and there would be more errors after it

[reef ivy]

i'm pretty sure this is an error on my part with intall but I can't get it to run with ipex hijacks ``` File "/mnt/A6765FA1765F70D3/Stable diffusion/ComfyUI/comfy/ipex_to_cuda/hijacks.py", line 163, in functional_linear
return original_functional_linear(input, weight, bias=bias)
RuntimeError: could not make an engine with allocator

could be I need the environment to run in home directory, who knows. Linux is just over my head lol.

[somber trellis]

could not create an engine for me usually directs me to looking at if my integrated graphics is enabled or not

that is what resolved that for me

a driver install re-enabled it

[earnest grotto]

my script should also work on linux

though i haven't tested it much recently 😛

and yeah, disable the igpu. there is a way to get ipex to ignore it i guess

[somber trellis]

gpu drivers keep trying to re-enable it

either that or ddu did it

🤷‍♂️

best settings for now when running SDXL is to use first block cache from wavespeed at 0.050

Allows a 50-step image to be done in 20 seconds

100 steps can be done in 26 seconds

once we get torch.compile working alongside this though

SDXL is gonna be hella fast

not even to mention sageattention

[earnest grotto]

"woman with pieces of jelly on her cheek"

[somber trellis]

nope

[reef ivy]

no intergrated graphics lol. I dunno what's wrong tbh

[earnest grotto]

i mean, i can assume it's supposed to be blood or pimples it's just that it looks like jelly 😂

[somber trellis]

r1 made it

the initial prompt was just liminal entities

[reef ivy]

could just be torch 2.8 or could be the driver install

linux is always like this for me, takes forever just to get it running and so many things could be wrong

[somber trellis]

2.8

dayum

[reef ivy]

latest nightly is 2.8

[somber trellis]

huh

for linux or windows

[reef ivy]

both I believe

[somber trellis]

no

im on 2.7.0 nightly that was released today for windows

[reef ivy]

you may need to update, looking at the site they have 2.8 wheels for windows too

[somber trellis]

Where.

Where are you getting these wheels?
I'm using the official pytorch nightly...

https://pytorch.org/docs/stable/notes/get_start_xpu.html

[reef ivy]

https://download.pytorch.org/whl/nightly/torch/

[somber trellis]

thats why

https://download.pytorch.org/whl/nightly/xpu/torch-2.8.0.dev20250313%2Bxpu-cp311-cp311-win_amd64.whl

or not

hmm

[reef ivy]

torch-2.8.0.dev20250313+xpu-cp310-cp310-win_amd64.whl

nightly is updated at random, they may have uploaded right after you downloaded today lol

[somber trellis]

i had to force it

[reef ivy]

although it may be broken so be weary

[somber trellis]

it wouldnt let me --upgrade normally

its there yeah it's just that pip package manager decided to think that 2.7.0 is 2.8.0

and that it was already there

[reef ivy]

i always have to --force-reinstall

[somber trellis]

if the version is higher numerically you usually dont

at least ive realized that when installing torch 2.3 to 2.6

🤷

[reef ivy]

looks like 2.7 is in stable preview as well

[somber trellis]

How does --upgrade in pip determine a newer package

[reef ivy]

gonna swap and see if it works in linux, or if I just don't know what i'm doing

[somber trellis]

Pip follows the Python packaging standards for version numbering. It compares the currently installed package's version number with the available versions in the package indexes (like PyPI).

ok

so im right

[reef ivy]

I dunno, I just always force it.

gonna try 2.7 in linux, then maybe swap back to windows as I am annoyed with this lol

[reef ivy]

what's the speed difference from before?

it's trying to download every version of torch again, I dunnow why it does this sometimes wtf

[somber trellis]

2.8's speed seems indentical to 2.7, hitting 1.7it/s on the same sdxl setup as the last

the reason why it didnt upgrade me to 2.8 was because torchaudio and torchvision aren't released for 2.8 nightly yet

but it doesnt matter atm

[reef ivy]

Having alot if trouble using the 2.7 preview, wants to give ne 2.5.1 lol. Linux!!!😡

[queen remnant]

was there an issue with speed on some specific recent torch versions using intel? i held off on updating because i had heard of it but i dont know how much worse it is

[reef ivy]

Everything after 2.1 is slower, 2.3 is the most balanced, 2.5 is the slowest, 2.6 and newer gets a little apeed back

If you are just doing like sdxl then no reason to update tbh, but if you want to try new models then you have too

[queen remnant]

ah alright, i also saw that triton for intel windows was being worked on, is that working yet?

idk how to update to test it myself

[reef ivy]

it's there but torch compile doesn't seem to work yet, but I haven't tried personally yet.

sage attention seems to be cuda only as well

[queen remnant]

oh ok

i'll wait until torch.compile is working then

[civic charm]

managed to get torch.compile working

1.3 it/s -> 1.6 it/s

Tho you are better off using OpenVINO imo

It will be faster and will take much less time to compile than torch.compile

also torch.compile has the same limitations as OpenVINO too

this is a torch.compile thing, not an intel one

[queen remnant]

ah, there arent any nodes for OpenVINO use in comfyui though right?

i could use SD.Next but i like my workflows ;-;

[reef ivy]

so 2.7 is in stable preview but they didn't add torch vision that's compatible lol. So get a dependency conflict and it installs 2.5.1

[reef ivy]

don't think we have an option for openvino in comfyui yet, torch compile is usually only used for video models by most people i think.

[somber trellis]

im confused

how are you getting 1.3it/s on sdxl and im getting 1.7it/s

without compile btw

sdxl dpmpp 2m sde karras

[reef ivy]

probably sampler and cfg

i am getting about 1.33s/it on a750 now with sdx without torch compile

!!! Exception during processing !!! RuntimeError: Failed to find C compiler. Please specify via CC environment variable. what are you using to do torch compile in comfy? this is the default node with sdx

[somber trellis]

on 2.7.0 nightly?

How do I get it operational?

[reef ivy]

I believe he is only on linux (unless things have changed)

I found you have to add the c++ compiler to path https://github.com/comfyanonymous/ComfyUI/issues/5216 but I am now getting another error

updating visual studio to see if it helps

[reef ivy]

hmm, i don't know what i need for torch.compile to work on windows... get this error when using visual studio c++ compiler as path returned non-zero exit status 2.

maybe it needs dpc+++ ?

[reef ivy]

looks like it may need the basekit and level zero?

[earnest grotto]

Pretty sure one of the required compute packages (if the wsl ubuntu doesn't have it) is level zero

Or, maybe it has an older version idk

[somber trellis]

i have oneapi 2025, 2024.2 and 2024.1 installed

ipex-init is initialized at the beginning of comfy

[reef ivy]

did you set the c++ compiler to path like in that comfy thread I linked? I think we need both, I am downloading 2025 now

[somber trellis]

the cpp compiler is automatically initialized on initializing oneapi variables i thought

C:\Users\dbs_5>ipex-init :: initializing oneAPI environment... Initializing Visual Studio command-line environment... Visual Studio version 17.10.5 environment configured. "C:\Program Files\Microsoft Visual Studio\2022\Community\" Visual Studio command-line environment initialized for: 'x64' : advisor -- latest : compiler -- latest : dal -- latest : debugger -- latest : dev-utilities -- latest : dnnl -- latest : dpcpp-ct -- latest : dpl -- latest : ipp -- latest : ippcp -- latest : mkl -- latest : ocloc -- latest : pti -- latest : tbb -- latest : umf -- latest : vtune -- latest :: oneAPI environment initialized ::

[earnest grotto]

IMO there has not been enough improvement outside of LLMs

[reef ivy]

It's not a oneapi compiler, it's this C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.43.34808\bin\Hostx64\x64\cl.exe

I may be wrong, but about to see in a minute, still installing

[earnest grotto]

that's MSVC

[somber trellis]

imma do funny

if i had a better sample lmao

spark-tts gets much better outputs if you align the audio input with the prompt text input matching the audio

that way the AI knows what words match to what more easily

[earnest grotto]

rip

just dm me

[somber trellis]

20 seconds of low quality ninjago intro

momjago

[rustic sonnet]

So Compile on Windows supported on 2.8 nightly?

[somber trellis]

That we don't know.

🤷‍♂️

Disty is an intel user

on linux

[rustic sonnet]

Hmmm

[reef ivy]

maybe, but 2.8 won't install as torchvision has 2.7 as a dependency

[reef ivy]

bout to send you the error

[earnest grotto]

For spark tts specifically, it's close, way better than tortoise's zero-shot stuff, but I wouldn't say it's too big of an advancement in practice. Functionally it still lacks the features I'd want from good tts/still has the old issues... And those would probably need better datasets to resolve.
Voice sounds slightly crackly. No emotion control. No control over non-speech vocalizations, like grunts, screaming, sobbing, etc., and no alternative voice from the same speaker (whispering, fake impersonation voice, shouting, etc.)
With stuff like RVC I guess you can at least do the emotion yourself so that's one thing I prefer with it

[somber trellis]

It still has that RVC issue

The artifical breathing and high-end

[reef ivy]

are you prompting for the background music?

also seems to have some slight sound effects

[somber trellis]

no

i used uvr to seperate the music and voice

i did the voice then readded it to the base music

[reef ivy]

ah okay, I need to start learning voice stuff.

[reef ivy]

So just did a wan2.1 video that was taking about 16-20 minutes in about 10 minutes, not sure if it's just windows memory issue not happening or the new torch nightly. probably just the memory issue didn't kick in on that one, back to over 16minutes second run.

[civic charm]

Linux still didn't recover the speed with 2.7 / 2.8

[rustic sonnet]

https://github.com/codeplaysoftware/cutlass-fork

This fork seems to have flash attention for Xe

[earnest grotto]

Li already confirmed flash attention is coming for 2.7

So, should be soonish

[reef ivy]

hmm, does anything work? It looks like you hav 2.5.1 cuda and not xpu or ipex?

[ripe pivot]

t2i works fine, I think I already installed the xpu 2.5.1. I don't know any other preparation on i2v so I just slap the workflow and prompt it

[reef ivy]

What speed do you get? it may be running on cpu?

**PyTorch Version:** 2.5.1+cxx11.abi i believe this is cuda

sometimes nodes can overwrite your torch version if it's older than the latest, latest is 2.6

[ripe pivot]

how do I install the xpu/ipex then? I always thought the pip that I got since it was xpu, the link kind of similar

[civic charm]

pure pytorch doesn't have the +xpu tag

it is listed as this

[reef ivy]

Should be able to install it the same way you did before, it probably got overwritten when installing a node.

just load into your enviornment and then install this https://pytorch-extension.intel.com/installation?platform=gpu&version=v2.6.10%2Bxpu&os=windows&package=pip

I believe 2.6 is still latest, so it shouldn't override anything atm.

[reef ivy]

is it still able to run on xpu?

[civic charm]

yes

[reef ivy]

hmm, everytime I've had it install itself it wouldn't work

if that's the case then maybe he doesn't have your hijacks and it's calling for ipex?

[ripe pivot]

1.68 it/s on 1024*1536, seems like it's working on my gpu

[civic charm]

Tested 2.1.40 without any optimizations:

Changed TAESD previews to approx previews:

[reef ivy]

Do you have the ipex to cuda hijacks installed? It could be you just need to update comfy and your nodes to the latest maybe.

[ripe pivot]

I don't know what's that so probably not. how do I put it

[reef ivy]

#1193952640225267802 message only change is you need a different edit to the management file when not on ipex

    from ipex_to_cuda import ipex_init
    ipex_init()
    torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)
    _ = torch.xpu.device_count()
    xpu_available = xpu_available or torch.xpu.is_available()
except:
    xpu_available = xpu_available or (hasattr(torch, "xpu") and torch.xpu.is_available())

Using Vik's script is probably the easiest way to install

it does all this for you i belive

[ripe pivot]

do I need different comfy install or what? I think I'm on phyton 3.11

[reef ivy]

no, just git clone the repository to the comfy folder and then add that line to the file it mentions

you don't need to change anything with your install

just skip to step 4 and use the edit I posted instead of the one with ipex, or take out the import intel extension for pytorch line

not sure if the sdp line is still needed but I still have it just in case(it was for speed).

[ripe pivot]

do I add or do I modify?

damn I got the error and forgot to back-up the script

[reef ivy]

are you using vik's script? If so just let it do everything for you, just chose what pytorch/ipex you want etc.

[ripe pivot]

I've to redownload everything on that then? or can it "repair" my installation

[reef ivy]

Did you install with the script? I believe it doesn't delete what is already installed. You can also backup your nodes/models and outputs and just drag them back in as well.

Vik will be more help with that

[ripe pivot]

2.6 doesn't have the IPEX, can I proceed with 2.6 or no?

[reef ivy]

you can use 2.6, he probably just needs to update his script to add it.

[ripe pivot]

how do I change my drive? I don't think my C: has enough space

[reef ivy]

That I am not sure, probably need to run the script on that drive

You can also move your models to that drive and add them to your model_paths file in comfy folder

extra_model_paths.yaml file in main comfy folder, there is an example as well that shows how to add them.

[reef ivy]

so far it seems setting reserv-vram to 7 from 6 fixes this.

[ripe pivot]

the progress stuck without any download info or new lines, is it safe to restart the script?

[reef ivy]

show a screen shot

[ripe pivot]

this one, it has been 10 minutes

[reef ivy]

it's downloading torch 2.7, if it still hasn't moved maybe you need to restart it. Or if your internet is extremely slow it could be it.

[ripe pivot]

there's no tell, like isn't it usually there's a download bar below it. I don't know the estimation without any information

[reef ivy]

try and restart it

[civic charm]

you can also click on the link and install with the file

you can do pip install path_to_the_file

[ripe pivot]

do I have to activate the environment for it?

[civic charm]

yes

[ripe pivot]

uh, how do I make the setup.py?

[civic charm]

path_to_file is the .whl file

you are trying to install a folder

you can just copy the file and paste it into the terminal

it will paste the path

[civic charm]

is the python version 3.10 on that venv too?

that file is for python 3.10

[ripe pivot]

idk actually, the script should using the 3.10 right? Why won't it uses 3.10 files if it's on 3.11

[civic charm]

python --version will tell the python version

compiled .whl packages are specific to the python version

cpxxx tells the python version of the package

cp310 is python 3.10

[ripe pivot]

it's 3.11...so it stucks because of different python version

uh, how do I change within the script though

[reef ivy]

The script should detect your python version I think? But honestly IMO 3.10 is the most widely used with webui's still, so I stick with that.

Also, if python is in conda it should be able to install python 3.10 in the environment anyway, just remake your enviornment with 3.10 would be the easiest solution for you i think. *edit Actually the script should do that for you when it creates the environment.

Did the file not download when your ran the script again?

Could be issues with the preview servers? It is getting updated to 2.7 atm.

[ripe pivot]

it's stuck on this part now

[reef ivy]

hmm, I didn't even know it had openvino in comfy. It doesn't seem to be stuck downloading anything, not sure what's going on. Probably need to ask Vik. Does make me curious to try openvino though

13.5 minutes with enhance a video added to workflow.

same seed about 10 minutes without enhance a video lol. Think teacache was a little higher though

I think I may be getting close to the expected speed with this level gpu compared to nvidia counterpart.

prompt A beautiful blond warrior in a blue dress raises her magical sword high into the air and screams out a battle cry. Lightning crashes behind the mountains in the background.

probably will add an llm to make better prompts lol

gonna try a longer video and see what happens

[reef ivy]

nightly updated to 2.8 in windows, seems about 15-20% slower, Torch.compile still not working yet also

[earnest grotto]

Outside of the 5 missing dependencies, this was a very easy install. I'm happy

[earnest grotto]

soxr soundfile einx librosa transformers>=4.45 if anyone wants them

[reef ivy]

this is called spark tts?

[earnest grotto]

https://github.com/billwuhao/ComfyUI_SparkTTS

Zero shot, i.e. no training needed to clone a voice
Clones well and does TTS, speed looks good and I think it should work with fairly low vram usage. But that's about it. No 3 minute long neco arc singing invisible by duran duran with this

Honestly I'm not big on music with singing but I love the neco arc stuff, and similar

oh, and, since i got confused
bottom text of the clone node is your transcription of the cloned speech, top is the new text. you don't need to transcribe it but the results are way better if you do.

[somber trellis]

ye the webui is the same

transcription matching audio

i like that they used a noise reduction module tho

I like the comfynode far more.

Thank you for providing a link vik

[earnest grotto]

Where's dagoth ur, neco arc and miku

And maybe ranni

[somber trellis]

I really

really wanted to try to get the imperial watchguard working

I can't seem to nail him with spark

I need to make a decent sized audiofile of him

[earnest grotto]

Hmm, it works with very long reference audio?

[somber trellis]

I've done up to four minutes of audio without transcription

but i think you can do more

but as you can see

transcribed audio does indeed help like you said

that middle part

is spot on imperial

[ripe pivot]

does this pre-startup failed means that I can't make the i2v workflow works?

[earnest grotto]

You installed the experimental 3dpack stuff
It's experimental
Something broke. You probably picked a pytorch version that's incompatible with torch_scatter. You should have gotten a warning about this. I dunno if the setup would then end.

If the setup ended at that point, delete and reinstall without the 3dpack.

[ripe pivot]

I closed 3dpack midway because it's stuck on openvino for so long and restarted without the 3dpack, does delete and reinstall works?

[earnest grotto]

Yes, 2.7 is very much incompatible, there are no CPU torch_scatter builds for that

[earnest grotto]

Yes, delete everything and reinstall. The script does not have any infinite loops or anything like that. If it looks stuck, it's not, it's downloading or your internet cut out.

[earnest grotto]

Nice

[somber trellis]

second one is spot-on

The fact it's 0.5b

Astoundingly good

[earnest grotto]

angled chairs 🤢

[earnest grotto]

🤔 seems very sensitive to audio quality

[reef ivy]

Do you guys know if any of these programs can CREATE voices? Say combine multiple voices and createa unique one? I am really against using anyones voice or likeness tbh, outside of just learning the process of course.

I've seen a new paid model that could do it, so it must be possible, just not sure if open source can

[reef ivy]

Can't get the comfy node to work, and don't really understand how to add voices anyway. FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`. WeightNorm.apply(module, name, dim) Missing tensor: mel_transformer.spectrogram.window Missing tensor: mel_transformer.mel_scale.fb

[ripe pivot]

Requested to load WanVAE loaded completely 8344.23828125 242.02829551696777 True !!! Exception during processing !!! The paging file is too small for this operation to complete. (os error 1455) Traceback (most recent call last): File "E:\Comfy_Intel\ComfyUI\execution.py", line 327, in execute output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) File "E:\Comfy_Intel\ComfyUI\execution.py", line 202, in get_output_data return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) File "E:\Comfy_Intel\ComfyUI\execution.py", line 174, in _map_node_over_list process_inputs(input_dict, i) File "E:\Comfy_Intel\ComfyUI\execution.py", line 163, in process_inputs results.append(getattr(obj, func)(**inputs)) File "E:\Comfy_Intel\ComfyUI\nodes.py", line 910, in load_unet model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options) File "E:\Comfy_Intel\ComfyUI\comfy\sd.py", line 1027, in load_diffusion_model sd = comfy.utils.load_torch_file(unet_path) File "E:\Comfy_Intel\ComfyUI\comfy\utils.py", line 68, in load_torch_file raise e File "E:\Comfy_Intel\ComfyUI\comfy\utils.py", line 55, in load_torch_file with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f: OSError: The paging file is too small for this operation to complete. (os error 1455)

does this mean that my vram isn't capable?

[reef ivy]

Paging file is system ram, it's when you run out of system ram and use your ssd/hdd for ram

if you set it manually let windows handle it instead, or you could be low on disk space. How much ram do you have, i recommend 32gb minimum for ai stuff

also, try tiled vae

[ripe pivot]

my ram is 32gb

[reef ivy]

Try tiled vae, a smaller model, lower frames, and check and make sure you page file is enabled and handled by windows and you have enough disk space on whatever drive it's using(you can change the drive but make sure it's a ssd)

[reef ivy]

where do you get input voices from?

[somber trellis]

I just take input audio from youtube, and cut it down to samples.

If the audio has noise, I use UVR and mdxnet instvoc to seperate vocal from instrumental sounds

[ripe pivot]

is this normal?

[reef ivy]

For wan you will need teacache and adaptive guider to make it decently fast. You can take the workflow from this image, change the size and step count if you want. It's messy but it will work, you can also look online for some that are better, this is one I adapted from one I found.

Hopefully we get torch.compile and flash_attn soon as well.

[ripe pivot]

how do I grab the i2v workflow?

copy paste gets me that

[reef ivy]

download the image then just drag it into your comfyui window

if missing nodes, just go to manager install missing nodes

you can do this with most images/videos made in comfy as the meta data is saved with the files by default

[ripe pivot]

where do I grab the gguf? don't think the examples mention that

[somber trellis]

trying to figure out why im being errored out

spark-tts moment

[ripe pivot]

being unable to install nodes because of random reason is annoying

[somber trellis]

Lisha

click on the name of that node

grab the github link

go to custom_nodes in comfyui

and git clone it

Also, run comfyui as admin.

Lotta dumb things happen for me if I don't. Permission problems mainly.

[reef ivy]

that looks like some sorta oom

[somber trellis]

thats what im trying to figure out tho

[reef ivy]

there are 2 comfyui nodes for spark, only got one to sorta work

[somber trellis]

the one i have does work

it's just that certain situations I get a kernel size error

and some times i just get a device lost error

[ripe pivot]

comfyUI doesn't detect my git clone. you do it like this right?

[somber trellis]

you clone it into custom_nodes

once you do that and launch comfy it might be import_failed

go into comfy manager and try fix it under the import failed tab in custom nodes

[reef ivy]

go to custom nodes folder, then type "cmd" directly in the folder address to bring up command prompt, then just git clone whatever custom nodes repo

[somber trellis]

i found the issue

it doesnt like mp3

[reef ivy]

what do you use to get videos off youtube? The program I used isn't working, and I need node.js to fix it and build the source but it's a pain lol

[somber trellis]

jdownloader to get the m4a off of youtube

then i convert it to wav/ogg

[reef ivy]

thanks, i will check that out

[somber trellis]

also i swapped back to 2.7

the preview build

because i was crashing on 2.8 nightly

[reef ivy]

It was slower for me yesterday so I swapped back, i will try again today and see if anything is changed

[ripe pivot]

nope, doesn't change anything

[somber trellis]

if the custom node is in your custom_nodes folder and comfy isnt detecting it then idk

[ripe pivot]

Traceback (most recent call last): File "E:\Comfy_Intel\ComfyUI\nodes.py", line 2140, in load_custom_node module_spec.loader.exec_module(module) File "<frozen importlib._bootstrap_external>", line 883, in exec_module File "<frozen importlib._bootstrap>", line 241, in _call_with_frames_removed File "E:\Comfy_Intel\ComfyUI\custom_nodes\sigmas_tools_and_the_golden_scheduler\__init__.py", line 1, in <module> from .sigmas_merge import NODE_CLASS_MAPPINGS File "E:\Comfy_Intel\ComfyUI\custom_nodes\sigmas_tools_and_the_golden_scheduler\sigmas_merge.py", line 29, in <module> class GraphScale(enum.StrEnum): AttributeError: module 'enum' has no attribute 'StrEnum'

uh I hate the mismatch phython

seems like the nodes are asking for 3.11 and I'm on 3.10 probably

[reef ivy]

nah, i am on 3.10, so that shouldn't be a problem

[somber trellis]

Yep judging by what gemini responded with the issue is a python version mismatch

[ripe pivot]

got this weird alert too

[reef ivy]

yeah, i get that all the time but it doesn't hurt anything

[somber trellis]

The "invalid workflow against Zod schema" error in ComfyUI means that your workflow file (likely a .json file) doesn't conform to the expected structure defined by a Zod schema. Let's break this down:

[ripe pivot]

you probably got the older version of the nodes so the python version already adheres idk

[reef ivy]

comfyui prob updated the format recently

I have 1.0.0

if that's the case, git checkout an older commit

[reef ivy]

herer you go https://github.com/Extraltodeus/sigmas_tools_and_the_golden_scheduler/issues/7

kinda strange to require 3.11 when 3.10 is still the most widely used.

there is a fork, tbh, I may have done that a long time ago i dunno.

[reef ivy]

also, don't even bother, you don't need this node

just disable it

it's to make a graph of the sigma for whatever reason, this is from the workflow I downloaded but don't think it's needed to make a video

[ripe pivot]

now for the unet loader, where to download it. i don't see any on https://comfyanonymous.github.io/ComfyUI_examples/wan/

[somber trellis]

wan is supported natively, lisha

comfy core has the nodes

[reef ivy]

it doesn't show in missing nodes?

[ripe pivot]

I mean the gguf model, it takes me a year with comfyUI workflow

[reef ivy]

oh, you need to download the quants from city96. https://huggingface.co/collections/city96/gguf-image-model-quants-67199ef97bf1d9ca49033234

i am using i2v 480p....q4_k-s you may be able to use a larger model with 12gb vram.

Make sure it's i2v for this workflow, as text 2 vid won't work.

you should be able to use the vae here https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main

probably keep him bookmarked or something, he typically has quants for all the models that come out

[somber trellis]

comfy-org, kijai and city96 are the three main ones

[somber trellis]

@earnest grotto

[reef ivy]

converted mp4 to a wav and still get unrecognized format error.

guess I can try the stand alone version later

[earnest grotto]

Convert better
What did you convert with and why was it not ffmpeg, audacity, blender or handbrake

Or maybe davinci resolve or some Adobe™️ product

[somber trellis]

ye ffmpeg ftw

also audacity (both are good)

also handbrake (also good)

but blender for audio conversion is an interesting one

[earnest grotto]

In a worst case scenario, if you're already familiar with it, it has ui to make it doable. It uses ffmpeg behind the scenes for a lot of things anyways

[somber trellis]

Right now it seems I'm getting a ton of device lost errors using spark.

🤷‍♂️

I don't think it's offloading at all.

[reef ivy]

I used audacity with the ffmpeg addon

y2mp3 isn't working anymore, that would convert it straight to mp3 or wav. Need to update it's depencies but getting node.js working was more of a pain that I wanted to try.

Thinking, maybe they have a node in comfy that can convert with ffmpeg.

[somber trellis]

Even the current 2.7.0+xpu preview branch causes me to crash during vae decode to upscale and then second pass

it always seems to be during vae encode/decode or upcsale using model node

It might be due to the very specific upscale model im using

one that doesnt work in 2.3

nomosuni otf compact

I confirmed my suspicions, using the upscale (using model) node breaks comfy

[reef ivy]

maybe the download got corrupted

[somber trellis]

Nope i tested it with a diff upscale model and rechecked my comfy install

crashed

[reef ivy]

is it the native node?

[somber trellis]

ye

upscale image (using model)

for now ill just force the node on cpu

line 51 in nodes_upscale_model.py in comfy_extras

[reef ivy]

testing, never used this one doesn't seem like you can tell it what size to upscale by

[somber trellis]

you cant

thats why i used a version called "upscale image by (using model)

https://github.com/TheBill2001/comfyui-upscale-by-model

[reef ivy]

The native one worked for me (didn't upscale anything but no crash)

looks like that node hasn't been updated in almost a year

[somber trellis]

i intend on not using it anymore

lmao

what node do you use for upscale models?

[reef ivy]

I usually use sd ultimate

[somber trellis]

hmm, i could just force a single tile run

[reef ivy]

you can use the one from impact pack as well, and load in a model

Upscaler(SEGS) is the name in comfy

nvm, that one also uses latent. I forget but I had one that worked with just upscale

okay, it's Upsclaer(SEGS/Pipe) i think edit (nope). I have to go through some old workflows lol

[somber trellis]

no longer crashing as i used gemini pro to rewrite the nodes_upscale_model for cpu

https://pastebin.com/epPLr26X

[reef ivy]

nice, also I think this is what I used to do https://www.reddit.com/r/comfyui/comments/192w0ei/upscale_image_using_model_to_a_certain_size/

[signal patio]

someone installed https://github.com/Gourieff/ComfyUI-ReActor on comfy with arc?
im having this error when run install.bat from C:\Comfy_Intel\ComfyUI\custom_nodes\comfyui-reactor

install.bat

plz help

[earnest grotto]

Activate your environment and do python ./install.py, in that custom node's folder.

[civic charm]

python 3.13 is not supported by that node

[signal patio]

how to check which version of python is my comfyui using? to make sure

[civic charm]

Activate its venv

and run python --version

[signal patio]

sorry im noob, how to actiavete it?

[earnest grotto]

are you using my script

[signal patio]

yea

[earnest grotto]

launch from the shortcut. then press ctrl+c.

[earnest grotto]

then go into that custom node's folder and run the python install file.

ctrl+c kills comfyui, it doesn't copy, if that's why you're sending me the screenshot

cd ./custom_nodes/ComfyUI-ReActor

python ./install.py

there's also a ton of other face detect nodes that don't need extra arbitrary setup and MSVC and whatever else

[signal patio]

thank u man

u just figured out that if i stopped the script i can keep enter commands in this env

[earnest grotto]

I specifically made it so the shortcut keeps the command prompt open when comfyui dies
this is by design

[earnest grotto]

Read the node's install instructions. It needs a specific version of visual studio installed.

Visual studio is not small and takes some time to install.

[signal patio]

thank u, i used prebuilt Insightface package

but something is wrong when i trying to use pulid

[earnest grotto]

Show the whole stack trace

[signal patio]

this? or full error log?

[earnest grotto]

Move the onnx files into the parent folder: antelopev2/*.onnx```

Do that.

You should probably not be using a 2-3 year old custom node, I'm sure there's newer face detectors, since there's a ton of nodes for that

[signal patio]

it worked!

[signal patio]

i thought that im using the one and only pulid implementation for comfyui

new error

[earnest grotto]

You've chosen the wrong model in some node.