ComfyUI for Intel Arc using IPEX

[earnest grotto]

also 20x faster than flux 2 i guess

[somber trellis]

I like z image's realistic fidelity, bet you could get some good images out of it

[earnest grotto]

you could get them even gooder given it's 6b

training loras with it is all but guaranteed locally doable without needing 64gb+ of RAM that's getting more and more expensive

just need training code

[somber trellis]

Yeah, the realism's already there.

[earnest grotto]

I really have to wonder now how much worse flux klein will be than this

[somber trellis]

z image kinda blows flux 2 away in realism standards

[earnest grotto]

total hunyuan image 3.0 situation

[somber trellis]

I feel how this guy looks with this model.

[earnest grotto]

EAT

[somber trellis]

EAT ART

[earnest grotto]

that one's pretty good

are you cooking him for thanksgiving

[somber trellis]

goin home to cook him rn

now we just need z image edit

Supposedly, this is a poop atronach.

The lizard has magic poop powers.

da bear wizard

[earnest grotto]

are you adding the black borders

[somber trellis]

nope

[somber trellis]

z image, 2048x2048 second pass

[reef ivy]

ngl, i need to see side by sides to know what models are actually better now. I haven't used anything since flux for images, if something is better and faster that'd be amazing.

[earnest grotto]

Z image turbo is better, faster and half the size of flux 1

No one wants to wait 5 minutes per image (or more). You're not using flux 2, and it's debatable if the results are better than z image anyways
When it comes to realism it seems to be... mildly worse than z image?

flux 2 just seems like the kind of model made so you can automatically generate a lot of images for your product to sell on amazon

not for a normal person to use

[earnest grotto]

it's probably more accurate to say z image is slightly slower than sdxl while being better quality than flux

and again, without torch compile. lumina benefitted a whole lot more from compiling when i poked on linux than I think sdxl does

supposedly it's based on lumina. how much, is that just the text encoder part, idk

just a typical case of dataset & training methodology (& architecture?) > moar parameters

it's so wild that flux2 is almost 5.3x the size of z-image

and like 10x slower?

despite also being a distill itself

[earnest grotto]

Supposedly one of the people behind Noobai

[earnest grotto]

https://www.reddit.com/r/StableDiffusion/comments/1p856z1/according_to_laxhar_labs_the_alibaba_zimage_team/

Illustrious Lumina and Animaestro aren't even out yet!

Damn

If this materializes before them...

So wild

[lunar thicket]

What kind of filesize are we talking with this zimage model? This usable for the low VRAM contingent?

[earnest grotto]

6B model

4B text encoder

1 parameter = 1 byte at 8 bits (1 byte) per parameter, fp8

IE, 6GB at fp8, 12GB at bf16

regardless of low vram, offloading is a thing too

[civic charm]

both of them are cancelled for a while already

[earnest grotto]

damn

[civic charm]

illustrious in general is pretty much dead

they lost their only trainer to pixai

[deft bloom]

I guess everyone is talking about Z image rn?

yeah, I just read the messages lol

this model is actually insane

for such low vram use and fast speeds to produce THAT quality?

[deft bloom]

the only issue I get is that my comfyui doesn't want to update thru comfy manager

[earnest grotto]

Also, this https://huggingface.co/NewBie-AI/NewBie-image-v0.1-exp-model-repo

[deft bloom]

thanks!

[earnest grotto]

well those aren't really relevant to you i think, you wanted more realistic stuff right

z image itself does that

i personally am interested in those because the best (local) anime models are still sdxl finetunes even after so long and so many other new models

disty might be interested if he doesn't know about those 2 already, and there's some other people here who liked doing anime stuff too

i guess novelai is good but it's not local

[hard sphinx]

How are you generating images under 30 seconds? I'm new to ComfyUI. I'm starting it with "python main.py --fp8_e4m3fn-unet" and have set weight_dtype - fp8_e4m3fn in the workflow. It still takes 216 seconds to generate 1024x1024 image with 20s/it. I'm using Arc 770M with 16GB Vram and 64GB ram. 9 steps.

[earnest grotto]

--fp8_e4m3fn-unet
don't do that

amount of time it takes depends on model and other things, what model
z image?

[hard sphinx]

yes

[earnest grotto]

if you're using z image and want it in fp8, do that from the selector

in the node to load the diffusion model

if you fp8 sdxl i'd expect it gets cooked

[earnest grotto]

run with --lowvram as well and say what happens

[hard sphinx]

Do I need to download different model? I see only bf16 here: https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/diffusion_models

I see fp8 is here: https://huggingface.co/drbaph/Z-Image-Turbo-FP8/tree/main I guess I will give it a try

[hard sphinx]

Still slow if I switch it here

[earnest grotto]

@hard sphinx

[earnest grotto]

You don't want to use the fp8 launch argument you do because of this

if you download and use the fp8 model, it will load a bit faster

your 20s/it times are not because of that

[hard sphinx]

--lowvram helped, now it is 2.3s/it

with fp8, bf16 is still slow, but I guess 16GB is not enough

[earnest grotto]

if you want block swapping, --reserve-vram 7

at bf16, it's 12gb model + latents + windows and your screen + browsers or whatever else, you are very close to running out of vram

there is no point in making the model get casted to bf16 entirely if it's the fp8 model though i'm not sure how you'd even do that

arc has no fp8 hardware iirc, or if it does it is not used in pytorch. comfy has its own special stuff that makes normally fp8-incompatible GPUs work

while stuff is stored in fp8, calculations are done in bf16

you can't squeeze out more precision from the fp8 model by upcasting it to bf16, that precision is gone

it's just for doing maths that the gpu can do

[hard sphinx]

ok, thanks for explaining, really appreciate

[earnest grotto]

just keep using the fp8 model and keep --lowvram

there is practically no reason to not have --lowvram

it makes text encoders run on the cpu. 99% of them are so fast that's something like a 2 second difference (total time). only exception is qwen image, the not plus (2509) one. which, with 2509 now, it's redundant so whatever

[lunar thicket]

I was wondering how much the offloading of clip/text to cpu mattered to speed

Good to know 👍

[deft bloom]

another example of the Z-image gen

it nailed the polaroid feel perfectly

[coarse whale]

I update comfyui by rerunning the script setup (to try z image), when i try to run my old workflow with tiled vae decode i get this now:
AttributeError: 'NoneType' object has no attribute 'cudaHostRegister'

[somber trellis]

@earnest grotto And since this is only the turbo model released, we still have the normal and edit models to wait for which is gonna be awesome.

I'm ready for it.

[earnest grotto]

I wonder how it will stack up against qwen edit

Same company different divisions(?)

If it can fill in the gaps that qwen edit can't do, without taking 5 minutes per image like flux2 that'll be pretty nice

or hell, even if it can't, i'll be able to train it at least

for loras, i really don't think making edit datasets is too hard

for the scale of a whole model yeah, but for loras, mmm

like, surely an edit model will be able to learn how to totally kill any sort of depth of field and blurring through blurry images -> impossibly sharp but still realistic 3d renders

the blender foundation has released and still releases a lot of free scenes

if you've seen this specific classroom before

[somber trellis]

2.82, yes

[earnest grotto]

yeah

[somber trellis]

I used blender for a period of time when modding bonelab

made a gun, weightpainted

etc

https://www.youtube.com/watch?v=9DNJ4GA8UGA

[earnest grotto]

nice

[somber trellis]

in return for that i essentially got my very own custom metrocop textures from someone as a trade for weightpainting halo models for them

[earnest grotto]

hl2's usp always has me expecting usps to be whiter and more matte

[somber trellis]

https://www.youtube.com/shorts/5GrgKl5fOXY

the real usp match isnt though, weirdly enough

it's a metallic slide

[earnest grotto]

yeah

[somber trellis]

https://www.youtube.com/watch?v=9nSkSPzClvg

oh yeah and uh i make music https://www.youtube.com/@dan9070

🤷‍♂️

[earnest grotto]

https://www.youtube.com/watch?v=ah3BFn0AvQI I asked you a while ago for a ranni voice

I'm happy with the result

[somber trellis]

I can't wait for RVC to improve even more

voice to voice is awesome

[earnest grotto]

neco arc singing videos are truly a miracle

kinda odd that no one has made dagoth sing but oh well

[somber trellis]

i did a skyfall 007 cover with TF2 spy but they took it down lmao

even though they took it off I still have it ofc

There are parts that can be fixed, like the fact that I couldn't seperated the top and bottom voices from the middle and near-end of this song

But I think the spy voice is pretty good

best parts are definitely the beginning and end

Z image just floors flux2, man.

[coarse whale]

tried to update from the new script and now i get black images generation and this error:
C:\Modding C\Stable Diffusion\Comfy_Intel\ComfyUI\nodes.py:1594: RuntimeWarning: invalid value encountered in cast
img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8))

[earnest grotto]

b580?

[coarse whale]

yes

i tested a bit im quite sure is the tiled vae

[earnest grotto]

update the script if you haven't (2.5 now), run the script again, install torch 2.8. don't use stable or nightly, those currently break with battlemage and produce black images. Hopefully soon™ this will be fixed in the nightly, which will eventually be the stable
the script should highlight 2.8, if it doesn't, tell me
you can also just press enter to pick the default option

[coarse whale]

ok cool i wasa on stable

the script was highlighting stable btw

0.2.5p

thanks that fixed it!

[earnest grotto]

updated script, now should highlight the correct option and also inform you if there's a newer version

[quartz kelp]

Just got Z_Image running on my A580, how much faster is the B580?

[lunar thicket]

Ill look into getting it set up with my B580 when I get home tomorrow. Then can compare, if no other B580 owner does first

[earnest grotto]

launch comfy with --lowvram --reserve-vram 7
if you're using my script, the lowervram shortcut does this

[quartz kelp]

I have a B580 coming in the mail today, I was just wondering if anyone knew off the top of their head. Thanks for offering though.

[earnest grotto]

You are highly likely running out of vram. You can make the a580 not run out of vram and then I'd expect about 4-8s/it?
And I'd expect the b580 to have... 2-4s/it?
Extremely rough numbers

[quartz kelp]

I got the generation time down to 30seconds using —novram and —reserve-vram 4.0

[earnest grotto]

Use the arguments I told you

[hard sphinx]

Is it possible to use QwenVL on Intel Arc in ComfyUI? KSampler node works fine and it runs on GPU, but QwenVL node from AILab only uses cpu. Is there other node that does not rely on cuda?

[earnest grotto]

Ah, so I take it Comfy's implementation is fried. I got ~9s/it

[civic charm]

their sampling is fried too

you have to hijack in diffusers and run diffusers sampling to get rid of the comfyui noise on z-image: https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler

[earnest grotto]

damn, that's pretty sad for comfy

[earnest grotto]

by the way, what are your lumina 2 speeds

[earnest grotto]

I was getting 1.6s/it without compile and 1.0/ with

[earnest grotto]

Ah, nevermind on the 9s. Comfy's new nodes are just crazy broken, that was at 2048^2 when the nodes were set to 1024^2. Actual speed at 1024 is 2.11s/it, still slower than diffusers though

[civic charm]

Exact same speeds as z image

Make it 2x slower for CFG

[hard kraken]

anyone still using ipex-llm? does it work with latest ipex 2.8 ?

[reef ivy]

Ipex llm doesn't actually use ipex, and ipex-llm is no longer being developed

[somber trellis]

i swapped to lm studio

since it's just a llama-cpp wrapper

[civic charm]

z-image on a770 with sdnq int8 matmul + torch.compile

[somber trellis]

takes me 30, so a 4x speed increase

nice

[civic charm]

1.75 s/it is the normal speed

[somber trellis]

im running the q8 gguf version and on an a770 it's 3s/it

[civic charm]

A770 has only 2x INT8, 4x is not possible

city96 gguf is slower than ram offloading btw

you should use ram offloading instead

[somber trellis]

So just straight safetensors?

🤷‍♂️

[civic charm]

yes

if you want to use real gguf, use sdcpp or llamacpp instead

city96 gguf isn't gguf

[lunar thicket]

anyone tried it out on B580 yet?

[lunar thicket]

get a chance to install that new GPU?

[somber trellis]

guess i gotta live with city96 in comfyui

all the models i have in comfy are gguf

[lunar thicket]

https://www.reddit.com/r/StableDiffusion/comments/1pchpjb/quick_psa_the_stablediffusioncpp_implementation/

I guess comfy fixed their stuff

This is outdated now. I originally wrote a somewhat wordy post comparing the performance of the two and speculating on why my card (2060 6GB) was faster on sdcpp. Since then comfy pushed a fix which remedied the performance discrepancy, and outperforms both by a wide margin.

[earnest grotto]

Why not try it out yourself? Nothing bad's gonna happen
It's basically SDXL speed (when Comfy's working properly), I doubt that'd be much different for the b580 and even if it were a bit slower than sdxl it won't be by a lot and it's still a good model

[somber trellis]

comfy stopped working the moment i updated it

lmao

I'm getting invalid argument errors for the clip text encode from multigpu

the latest update broke ggufmultigpu nodes

nvm im getting an invalid argument at ksampler now using normal safetensor models

[earnest grotto]

driver/pytorch/oneapi/... bug

intel issue

most likely

could be down to specific pytorch versions, have to est

[somber trellis]

Alright, let me swap to 2.9.1 stable then.

[earnest grotto]

steam decided to pop up the hardware survey as i was updating drivers

[somber trellis]

yep 2.9.1 is working, with gguf models

[earnest grotto]

yeah 2.10 bug then

torch.count_nonzero is broken

by the error I'm guessing someone is indeed like it says calling their kernel with the wrong number of arguments (or they made it take the wrong number), i'm sure they'll fix it

rather than it being completely unrelated like other errors

[earnest grotto]

https://docs.comfy.org/changelog

Fascinating

I had seen their training-related nodes floating about

[earnest grotto]

pytorch 2.9, a770, bf16 z-image with
--reserve-vram 11 (~6gb actually reserved on windows): ~2.6-2.7s/it
--reserve-vram 7 (~2gb): ~2.32s/it
--reserve-vram 6 (IIRC I've seen AIPG use this at one point, IMO it's too low at ~1GB): ~2.06s/it
no reserve-vram: 1.76s/it
no reserve-vram, comfy's own fp8_e4m3fn: ~2.07s/it
no reserve-vram, fp8_e4m3fn_fast: ~2.07s/it

I guess they did optimize it to be as fast as diffusers now that it's 1.76s/it

hopefully they fix or fixed the sampler too

will be interesting to check out comfy's own training nodes once the base model pops up

[lunar thicket]

Oh I intend to, I know nothing bad will happen. I am asking out of curiosity.

I don't get on my desktop every day, and I haven't had a free moment to go tinkering lately. So I just haven't gotten around to it myself.

[lunar thicket]

updated comfyui, ran the 'image_z_image_turbo' workflow it has by default and using all default settings. 2.31s/it rate for 1024x1024 9 steps

pretty darn fast and an impressive generation

[lunar thicket]

Crucially, with pytorch 2.8

Huh actually going back to it Im seeing variability in generation speed. Went as fast as 1.77s/it

Running 8 batches now to get a feel for the average.

as fast as 1.59s/it

[lunar thicket]

using the lower vram script (--reserve-vram 11 iirc) lead to great stability but slower generation. I believe it was around 2.4s/it. But using the normal script I was getting crashes occasionally and would have to reboot comfy.

[lunar thicket]

Good things are that I confirm that black image error is yet not hapenning with latest drivers, same as for yesterday version

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

(maybe) black image generation fixed for Battlemage in latest nightly pytorch

[earnest grotto]

If you want better performance you can lower the reserve vram amount, however i've upped it because intel still has some iffy memory management bugs
and even with it this high, i've had issues, and I've seen others having issues too. the driver(?) randomly decides to die sometimes, which will break comfy

going as low as 7 is generally ok. 6 gets risky. on windows, imagine you are reserving 5gb less than the number, so 7 would be reserving 2gb for the rest of the OS, and 6 would be 1gb

[lunar thicket]

I plan to experiment with it 👍 and find a good middle ground

[earnest grotto]

if you are on linux, try 6, 2 and 1 directly
hmm, I should make the script change the reserved amount when used on linux

[lunar thicket]

I am. I plan to try around 6 to begin and see how it goes

When running SDXL I just use the normal one with no reserve, but z image is a bit chonkier

[somber trellis]

z image works with this

flux 2 also works with the q8_k_xl quant of mistral 2506 instruct

all the gguf quants for llms in comfy are now that quantization unless it doesnt exist, then its q8

but for text gen in lm studio, q4_k_xl

i would do q8_k_xl ministral quants in lm studio but it's too big to fit in vram

[solid helm]

I finally got everything working well but using SDXL and a bunch of models based on it I can't get crowds or multiple people, 9/10 pictures doesn't get more than 1 subject when I enter such prompts. Any model recommendations?

Also the instructions I followed didn't download and use any clips just checkpoints and VAE is that fine?

[somber trellis]

Because SDXL checkpoints have them integrated. They are AIO models, all in ones.

If you're looking to do newer models, Z-image, Qwen Image and flux 2 are choices you could look at.

[solid helm]

alright, thanks

[solid helm]

do those also have everything integrated?

holy 64GB flux 2

[solid helm]

tried using Z-image. seems to need 17gb VRAM and crashes.
"torch.OutOfMemoryError: XPU out of memory. Tried to allocate 76.00 MiB. GPU 0 has a total capacity of 11.59 GiB. Of the allocated memory 17.78 GiB is allocated by PyTorch, and 99.50 MiB is reserved by PyTorch but unallocated. Please use empty_cache to release all unoccupied cached memory."

any solutions?
Same issue with 9 steps

[somber trellis]

Did you install comfyui through vik's setup script? It should already have --reserve-vram, ipex_to_cuda and such to prevent OOMs.

It should allow you to run most of these models with offload.

Also, I'm using gguf q8 models to half the vram requirements.

[lunar thicket]

yeah im no expert but the --lowvram parameter i believe is supposed to solve this problem by segmenting the workload and offloading part to CPU

[solid helm]

A bunch of stuff kept giving errors so I updated to the latest versions. I'll try adding reservevram and lowvram on launch and try it out again later

[solid helm]

ok so lowvram alone didn't give an error but the whole thing crashed and disconnected

Same when adding reserve vram 2

[somber trellis]

I personally use --reserve-vram 8 but i dont kind the slight slowdown over ooming

[solid helm]

Welp paging file is too small error. 16gb RAM (:

I'll try the gguf q8

[earnest grotto]

How much ram do you have

Oh, nevermind

16 is too little

do get the q8 or an fp8 version

[solid helm]

I'm getting edged over here. Fp8 was going well but crashed at vae decode

[earnest grotto]

get the text encoder in fp8/q8 as well. if you're still running out of ram, get lower quants or run comfy with --cache-none. This makes it reload the respective model every time then unload it immediately when done, this will make things much slower but you won't run out of ram

[solid helm]

should this say bf16?

[earnest grotto]

the vae is like 200mb at bf16. and if you quantized it any lower it'd probably break

[earnest grotto]

Do what I said

Get lower quants, and if you're still crashing then --cache-none

You are not running out of ram because the vae is too big, it's most likely because the latents are too much for your ram and there's not much you can do about that, other than use a lower resolution
or maybe tiled decode but i kinda doubt tiled would change much

[solid helm]

https://tenor.com/view/victory-winning-yes-gif-14139223485814601631

I just used cached none before moving on and it worked. 2 s/it btw.
I'll try to see how much less vram i can reserve before it crashes again

Thanks a bunch

[lunar thicket]

Congrats! Oof 16gb of ram yeah that was crucial info. GJ getting it running on that

[frank stump]

Hi! I’d like to ask for your help. I’m getting the error message shown in the picture. Could you tell me step by step what I need to do? ComfyUI was installed before and it worked, but now I’m stuck. I tried to fix it based on the error message, but it didn’t work.

[solid helm]

pip install av

Worked for me

I spent wayyyy too long doing absolutely unecessary stuff for like 3 days trying to fix it. All it took was one line. 🥲

[frank stump]

Thanks a lot for the reply. I tried it, but I’m still getting an error message. It looks like the comfyui-frontend-package isn’t installed. Can I install this separately, or something else?

[solid helm]

Try downloading the latest versions of everything. That solved some of my problems. I don't exactly remember exactly I downloaded the latest versions of but I remember doing that for intel's pytorch extension and the requirements-ipex.text

This was one thing I did

python -m pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/xpu
python -m pip install intel-extension-for-pytorch==2.8.10+xpu --index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/

[frank stump]

Thanks! It’s looking much better now, but there’s still some kind of issue. I’m leaving the screenshot here.

[solid helm]

I remember this error specifically. read the instructions above the error

Go to the file location it tells you and run the python.exe as admin

Then copy paste the install command

I think I might have ran through the exact sequence of problems I just don't remember very well 😂

[frank stump]

Thanks a lot! It is working!! 😁 😎 Have a good day, Bro!

[solid helm]

[solid helm]

so I removed lowvram and reserve vram altogether and it still works

it's gone down from 130 seconds to 78

batch size of 4 that is

or is that because I was using the same prompt over and over without reloading the UI

Nvm those were all tests on the first run

Forgor

Still though itbworks just with python main.py now

[earnest grotto]

you shouldn't need reserve-vram if you're using the model at fp8 or the q8 quant, or lower, since it will be small enough to fit in vram
lowvram is redundant if you have cache-none

[reef ivy]

If kijai has nodes I would recommend them for block swap, just have to usually edit line of code for a series to work, but battlemage shouldn't have issues at all

[hard sphinx]

Few days ago I've asked if there is anything similar to Qwen3 VL custom node that can run on Intel GPU. Turns out it is super simple to modify ComfyUI-QwenVL node to use XPU instead of cuda. Now it runs super fast compared to cpu 🙂

[coarse whale]

anyone have problems running z image with lora? it turns only black images for me if i use it with a lora

tried these loras: https://civitai.com/models/599757/velvets-mythic-fantasy-styles-or-flux-pony-illustrious
https://civitai.com/models/938811/neurocore-anime-shadow-circuit-by-chronokight-flux-il-zit

[earnest grotto]

try with q8 ZIT

[coarse whale]

whats that?

z_image_turbo-Q8_0.gguf ?

[somber trellis]

zit = z image turbo so i assume he means fp8 or q8 gguf

[coarse whale]

fp8 is what I'm using, i can try that gguf version i never used it

the images i take the workflow from use the fp8 i think

[earnest grotto]

not fp8, q8

[earnest grotto]

it's out https://huggingface.co/NewBie-AI/NewBie-image-Exp0.1

[coarse whale]

is the only q8 the gguf one, or there is a "normal" one? I tried it but i get this error: "size mismatch for x_pad_token: copying a param with shape torch.Size([3840]) from checkpoint, the shape in current model is torch.Size([1, 3840])." when loading the gguf model

[earnest grotto]

Update comfy and the gguf custom nodes

[coarse whale]

the new script 0.2.6p doesn't open for me, just the cmd window flashes for a brief time and closes. I think i have all requirements already (python, miniconda) as I used the old version of the script in the past

[earnest grotto]

Right click the background of the folder the script is in, open a command prompt.
python ./Setup-Comfy-Intel.py

And show what it says

[earnest grotto]

you forgot ./

[coarse whale]

i had it in the first

[earnest grotto]

Type in ./S and then press tab to autocomplete it

underscores instead of dashes

[coarse whale]

or if I type it out all:

[earnest grotto]

odd that you don't have that installed

ok then

python -m pip install requests

[coarse whale]

thanks, is running now!
Is it a new dependency or did I break something?

[earnest grotto]

it's something i expected to exist by default in python

[coarse whale]

also should I leave stable or 2.8?

[earnest grotto]

apparently it's not in by default

[earnest grotto]

use stable

[coarse whale]

Thank you again, now the lora work with the normal fp8 model!

[earnest grotto]

apparently their diffuers PR isn't accepted yet and they rushed it a bit. also said PR has a hard requirement for flash attention even though they state it and triton are optional ("recommended")

and autodownloading doesn't work

damn

[hard kraken]

I tried to run z-image template from comfy but got this error..

[lunar thicket]

How did you install Comfy?

CUDA in those error messages is a reference to Nvidia cards

[hard kraken]

it's ok, i just code it over now it works

[reef ivy]

what pytorch are you using? xpu? have the ipex to cuda hijacks?

[earnest grotto]

disty fixed it

nice

certainly getting undercooked vibes from that image

[earnest grotto]

huh, so it fries at 2.5 cfg

[earnest grotto]

1.0 CFG, seed 42
1.5 CFG
1.75 CFG
2.0 CFG
2.5 CFG
1.5 CFG, seed 43
+<copyright>...
<copyright>..., 2.0 CFG
<>, 2.0 CFG, 50 steps

rip. worse than neta

[earnest grotto]

will local anime models ever move on from sdxl finetunes. jesus

getting an urge to try out HDM and compare

[civic charm]

Lumina2 is a dead end imo

It is somehow slower than the cursed hunyuan Image 3

[earnest grotto]

lumina 2 itself? i can definitely see that
zit seems to make the lumina 2 architecture work though

maybe i'm just holding onto my hopes that onoma's cancelled lumina 2 finetune would've worked out well. even when very undertrained i feel it was promising

unlike neta

[civic charm]

zit is fixed and improved lumina2

it is not lumina2

different architectures

zit also runs 2x faster than lumina2 even tough it is almost 3x larger

[earnest grotto]

i skimmed the 2 papers and the 2 diagrams look near identical, except different rope which they say is for image editing, makes sense, and of course different text encoder, dimensions of stuff and training etc.
hidden dim of 3840 is pretty cool for intel with int8 i guess
dunno. i'm kinda out of my depth

[frank stump]

Hi! I’m experimenting with Z-Image. I can run one generation at 1.94 it/s, but the next one only runs at 70–80 s/it. Can you tell me what I need to change so I can consistently get the same performance?

[earnest grotto]

how much ram, what gpu

are either or both z-image and qwen fp8 or bf16 or

[frank stump]

B580, 32Gb 3200Mhz Ram

[earnest grotto]

don't use the new nodes

they're broken

most of those nodes could have different values than what they show

and a lot of things are not visible (e.g. the setting to not change seed every time is missing with nodes 2.0)

[frank stump]

Should I try another workflow?

[earnest grotto]

turn them off

they're purely visual. the visual is broken and worse

[frank stump]

It is off

[earnest grotto]

it was off, or you changed it to off now

[frank stump]

It was off

[earnest grotto]

open task manager. go to the 2nd tab. select your GPU. move task manager somewhere onscreen
run the workflow to when it slows down to 70s as you say, show pics of both task manager and the whole workflow
don't crop out task manager. the ram is normally visible even when you have the gpu selected

[earnest grotto]

what launch arguments

launch with --lowvram

[frank stump]

python main.py --bf16-unet

you mean python main.py --lowvram --bf16-unet?

[earnest grotto]

yes

always use --lowvram. there is pretty much no point in not using it

[frank stump]

With --lowvram i have this feature 😅 I already tried it, but because of this I’m not using it.

[earnest grotto]

in the future - if your vram spills out into shared memory, you're running out of vram

[frank stump]

but i think this can come when i have not too much space on disc

[earnest grotto]

you're using it at fp8 anyways. just download some pre-existing fp8 quant of the model

https://huggingface.co/drbaph/Z-Image-Turbo-FP8/tree/main

you can also make a workflow with purely a model loader plugged into the save model node and just set it to fp8

besides not running out of ram, it will also load much faster since ukno, model will be half the size

[frank stump]

Thank you for your help! The downloads are in progress. I can’t test it today anymore, but I’ll definitely follow the advice and continue this way.

[hard kraken]

no longer need to use --disable-ipex-optimize flag?

[earnest grotto]

it's not needed yes

[earnest grotto]

lol zimage lora training speed with musubi when i run out of vram is comparable to lumina 2 with kohya when i don't (though with ~0.6mp images vs 1mp)

[earnest grotto]

165 steps later and the broken memory management kills me. dammit

[lunar thicket]

So I installed using Vik's script in a new location. Downloaded the nightly (2.10.0.dev20251205+xpu) and ran 52 images of the default SDXL workflow with default settings and prompt. The one using the base model and refiner.

No black images, with B580

Looking like they truly fixed that problem 👍

[civic charm]

Same thing on RX 7900 XTX runs at 1.38 it/s

A770 with INT8 is very close to RX 7900 XTX

[lunar thicket]

wow 2.10dev is like 7-9% faster at SDXL than 2.8 in my brief testing

thats dope

[earnest grotto]

Damn that's fast

[earnest grotto]

speedup seems to be so big, z-image with compile and int8 with cfg will likely match lumina 2 without

that base model is getting ever so exciting

[earnest grotto]

Time to try with the dedistill 🤔

[earnest grotto]

Oh yeah

Now we're talking

[somber trellis]

makes me think i should set up wsl2 and use linux compile

[earnest grotto]

compile speedup depends on model

[somber trellis]

I still should though. Shouldn't take long either, since it just means I'll move my comfyui setup to wsl

[earnest grotto]

lumina, z-image and probably newbie but you don't wanna use that model all get a very big boost from compile alone

z-image is also structured in a way that makes intel's int8 perform pretty well with it

z-image should be 3800-something idk scroll up or check their paper
qwen and flux are 3072 or so

[somber trellis]

@earnest grotto Well I have comfyui set back up through linux, with torch 2.9.1+xpu

Are there any specific things needed to be done?

I'm currently getting slower speeds on the Q8 Z-image model I'm using, from 3 to 3.6 seconds.

[earnest grotto]

what launch arguments

show workflow

[somber trellis]

it's the comfy examples z-image workflow with ggufmultigpu loading nodes

--bf16-unet --disable-ipex-optimize --lowvram --reserve-vram 8.0

[earnest grotto]

why are you using reserve-vram with zimage

[somber trellis]

cuz i dont just use z-image

[earnest grotto]

also, reserve-vram works correctly on linux, it's off by 5gb on windows. if you were using 8 on windows, that'd be 3 on linux

[somber trellis]

Oh.

What compile node are you using?

2.6s/it with no reserve vram

on the gguf q8

[earnest grotto]

i haven't used any recently since i haven't used linux in a bit. use the default compile node.

[somber trellis]

I swapped back to windows due to compile errors on 2.9.1. Probably something I caused, but it's all good.

Speeds were basically the same non-compile, don't know what I expected.

[earnest grotto]

Left a zimage lora to train overnight for ~3600 steps.
it crashed again

this time at 460 steps

amazing

so much for assuming I didn't need to have a script to constantly restart training running anymore

[earnest grotto]

mm... even with the dedistill, the turbo model is way too unstable

rip (well, maybe better captioning will save it)

[sinful dome]

Hi, I have the B580, i was wondering if i can run Z-image model withing the AI Playground or its better to just follow the instruction above which gave me connection error, probably because of low VRAM ?

[earnest grotto]

AI playground has comfyui

The instructions at the very top of this channel are outdated and not necessary, things have been simplified

[sinful dome]

okay! so which model do i need to download for my b580

and all the dependencies for comfy ui are already installed right?

[earnest grotto]

https://huggingface.co/drbaph/Z-Image-Turbo-FP8/blob/main/z_image_turbo_fp8_e4m3fn.safetensors

https://huggingface.co/Comfy-Org/z_image_turbo/blob/main/split_files/text_encoders/qwen_3_4b.safetensors

https://huggingface.co/Comfy-Org/z_image_turbo/blob/main/split_files/vae/ae.safetensors

these 3

[lunar thicket]

Yes when you install AIPlayground it installs ComfyUI completely and simply.

AIPlayground is actually just a simple GUI stacked on top of ComfyUI

[sinful dome]

do i need any workflow file? I put all those 3 files in comfyUI folder then how do i work with it

[earnest grotto]

Comfy has a z-image workflow built in, by default
Here's one with its group node ungrouped. replace the bf16 model with the fp8 one.

[sinful dome]

disconnects after i press run

[earnest grotto]

How much ram do you have

[sinful dome]

16GB DDR5 6000MHz
and B580

[earnest grotto]

You ran out of ram.

Is this with aipg or a standalone comfy

[sinful dome]

aipg

[earnest grotto]

https://huggingface.co/vantagewithai/Z-Image-Turbo-GGUF/tree/main
https://huggingface.co/worstplayer/Z-Image_Qwen_3_4b_text_encoder_GGUF/tree/main

Get versions of these that will fit in your ram. close everything else open to free up ram. Consider getting more ram... In 2 years when the insane ram prices go down?, if they go down... Given micron's exiting, things are looking pretty grim. 16 was already kinda low for gaming nevermind AI, wonder if newer games will get more reasonable on RAM usage or what

For Qwen, don't get IQ, only Q3, Q6 or Q8. not sure if the IQ ones will work

Lower number after Q = less ram, less quality. Q8 is fastest, Q4 is second, all the other ones like Q3 or Q5 are slower.

Also, consider upping your pagefile

[sinful dome]

cant seem to use the GGUF files, it says only safetesensors are needed

[earnest grotto]

Use them with the gguf loader nodes

[sinful dome]

from here?

[earnest grotto]

Install the gguf custom node https://github.com/city96/ComfyUI-GGUF

[sinful dome]

got this then tried to install comfyui manager then got another error
[Installation Errors] 'ComfyUI-Manager': This action is not allowed with this security level configuration.

[earnest grotto]

I dunno how to fix that within AIPG's Comfy, sorry
I have a script that installs comfy for you, standalone. 3rd pin in the channel

[somber trellis]

@earnest grotto it seems pytorch 2.10+xpu is released under stable now

and it's also got invalid argument errors like the nightly

[earnest grotto]

rip

i guess i should be more proactive in reporting issues like this

ah, well, good
https://github.com/pytorch/pytorch/issues/170166

[somber trellis]

until then i'll just use pip3 install torch==2.9.1+xpu torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu

[primal hatch]

please anyone give me the instalation script for comfyui...

latest version

[somber trellis]

its pinned

literally the third message

[primal hatch]

Not working 🥹

[somber trellis]

Oh.

Are you getting invalid_argument errors in comfyUI?

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

^

The current version of the script installs the latest stable (2.10.0+xpu) if you choose stable. Both nightly and stable right now are broken on ComfyUI.

In order to remedy this, you need to downgrade the version of pytorch installed by the installation script.

pip3 install torch==2.9.1+xpu torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu

You need to go into the comfy_intel directory, enable the cenv and install torch2.9.1+xpu

[primal hatch]

This is very confusing. I don't understand much of it, but thank you.

[somber trellis]

Where are you stuck in terms of installing via the script?

[terse osprey]

so can I just use z image directly in ai playground or how does that work?

[lunar thicket]

I don't believe so, because AIPlayground is curated so unless an update has gone out to add z image in the fields, it won't be there.

But AI Playground has a 'launch comfyui' button that sends you to the standard interface. It will work from there

I havent used AI Playground in several months. I noped out of Windows officially

[primal hatch]

when i run that bat file it shows error

[earnest grotto]

That's not a bat file, it's a python file, and you're copypasting it into the command prompt
Do what the instructions say, download it, don't copypaste it into the command prompt, and run it from the command prompt

Or set your python up such that it can run the script by double clicking

[primal hatch]

which one should i select?

[earnest grotto]

3

[primal hatch]

can you give me z iamge turbo workflow?

i have downloaded comyui 🥹

[earnest grotto]

^

Double click somewhere on the workflow's background, type in gguf, use the diffusion model and text encoder gguf loader nodes

[primal hatch]

i have downloaded bf16 already, Do I have to download that gguf fp8?

[earnest grotto]

Sorry, I mistook you for the other person
Just set the model to load at fp8

[primal hatch]

crashing

[earnest grotto]

What CPU, what GPU

[primal hatch]

Ryzen 5700x ARC A770

[earnest grotto]

How much RAM

[primal hatch]

32GB

[earnest grotto]

Close everything that's eating up ram and start over
Especially if your browser is hogging 20gb

[primal hatch]

ok

😭

i have restarted PC but same problem

[earnest grotto]

Open task manager, go to the 2nd tab, select the GPU, show a screenshot without cropping such that the RAM usage is also visible just before and after the crash

[primal hatch]

before

after

when it comes to load the model it crash

[primal hatch]

.

[earnest grotto]

https://huggingface.co/drbaph/Z-Image-Turbo-FP8/tree/main
Download and use this already-fp8 model instead of the bf16 one. say what happens

[primal hatch]

both are same?

[earnest grotto]

use e4m3fn

[primal hatch]

now working fine with e5m2

[primal hatch]

but what was the problem with bf16?

[earnest grotto]

you ran out of ram and have no pagefile

there's a spike in ram usage while loading a model. that's when you ran out.

[earnest grotto]

specifically, 32gb is fine, you should just go set some proper pagefile, unless you're extremely worried about your SSDs/HDDs

[terse osprey]

where is this button? I can't find it, I'm on version v2.6.0 tho

[lunar thicket]

Click these red areas in order from top to bottom

Gear, Image, Workflow

Then that blue Open ComfyUI link will be there (right above the workflow button)

[terse osprey]

thank you

[terse osprey]

z image works flawlessly on my A750 and 16GB RAM

[terse osprey]

do you think it's worth trying q8.gguf version over fp8.safetensors?

[somber trellis]

q8 gguf has less perplexity (meaning it's more true to fp16) than fp8, but it's slower

Both are very usable

[terse osprey]

but is it visibly better than fp8?

or just different

could you tell which one is which if you didn't know which is which?

i know there's a difference between let's say q8 and q6

but I don't know how fp8 compares to q8

I'm just curious

I don't exactly know how quantusations work so I don't know what technically I lose

sorry if my questions are annoying

[somber trellis]

@terse osprey ^ This is a comparison with flux as an example

[earnest grotto]

anime is probably not great for a comparison like this

[lunar thicket]

Actually a very illuminating example tbh

[reef ivy]

always found it crazy that the smaller gguf's are slower than a q8 quant with most models.

[earnest grotto]

You are not going to be using the model for anime, and these non-anime art models are very weak at anime/art. Flux dev needed cfg to make a darker skinned anime girl as another example. Yet, I expect that it can generate a realistic image fine without forgetting the skin color at lower quants

[earnest grotto]

also i think longer text is more affected by lower quants, this speech bubble isn't super long

since it's weaker at longer text too, naturally

[earnest grotto]

.

To even get it to generate that skin color in the first place you have to go out of your way and actually use cfg for a model that was not intended to be cfg'd

nevermind very much not intended for art

z-image is at least a bit better in that regard

but it's still ultimately not an anime model

And, different models respond differently to quantization, you can't quite compare them the same way. E.g. Wan 2.2 I2V's low noise model you can quantize down to q4 with basically no loss in quality

the high noise model otoh had more of a difference, with anime, but hey, there's no anime video model

[earnest grotto]

Another tts voice cloning model, Cosy Voice 3
https://www.modelscope.cn/studios/FunAudioLLM/Fun-CosyVoice3-0.5B
Huggingface space but chinese ^
Topmost text box: text to generate
2 radio buttons: 3 second fast inference | instruction-guided
Box to drop the text file is in english. Don't use audio >10s for the 3 second inference?
Textbox below it is automatic transcription, and last textbox is instruction (for shouting, etc.?) (only applicable when instruction-guided)

[somber trellis]

GLM-tts as well btw

https://huggingface.co/zai-org/GLM-TTS

cant get it running tho

havent found any good alternatives to Vibevoice yet

and that model is slow

[terse osprey]

I tested both fp8 and q8 version and q8 is like 3 seconds slower than fp8 on average on my PC

44 vs 49 seconds

but q8 does a visibly better job at generating text

there are two Q8 and two FP8 images

just by looking at the small text you can tell which one is which

[reef ivy]

I think i found that with most models q8 and fp8 where about the same, but stuff like q4 would be much much slower than both. i guess because of decompression

[terse osprey]

guys i tried to use seedvr2 but I can't download custom nodes it says "User provided device_type of 'cuda', but CUDA is not available." in comfyui manager. I don't know how to go around it

[earnest grotto]

Install comfy using my script

3rd pin in this thread

[terse osprey]

aw man

how many ComfyUIs do I have to install

the one from the AI playground won't work?

[lunar thicket]

How did you install it?

if you just grabbed something and ran it, it probably defaulted to CUDA. Because that is 98% of the userbase

So (guessing here) you installed something that installed expecting CUDA and now CUDA is not present

AIPlayground is not intended for you to go installing custom stuff its just supposed to be click and play. You can expand the available ComfyUI options but you're on your own with it.

Vik has created the most curated Intel GPU option for running ComfyUI that exists anywhere. It just requires a little more work to setup vs AIPG which is just an exe

[terse osprey]

idk if the workflow is good

i just copied it from reddit

@lunar thicket

I saw someone here using SeedVR2

[lunar thicket]

They probably used Vik's script lol

[lunar thicket]

Just grabbing random stuff off reddit is likely gonna give errors 👍

[terse osprey]

@somber trellis how did you set up SeedVR2 upscaler?

[reef ivy]

Probably best to post in the ai playground section for that version

[remote estuary]

Unsure what I am doing, trying to generate anything and it keeps kicking back with out of memory errors.
Is it using any of my memory?

loaded partially; 0.00 MB usable, 0.00 MB loaded, 1639.41 MB offloaded, 168.76 MB buffer reserved, lowvram patches: 0
Exception during processing !!! level_zero backend failed with error: 39 (UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY)

[earnest grotto]

Show the workflow, explain how you installed comfy and say what arguments you're launching it with

[remote estuary]

Attaching workflow...
Followed the pins and when installing

pip install -r requirements-ipex.txt (if using Intel Arc dGPU ie A750, A770)
It gave a error, saw your script, ran and installed ComfyUI correctly and its currently "running" with the simple image generation template " v1-5-pruned-emaonly-fp16 "

[earnest grotto]

what driver version

restart your pc, try again, say if it's still broken

[remote estuary]

Driver version of what? and ill restart

[earnest grotto]

your gpu

[remote estuary]

Restarted and still kicks back with the same error

loaded partially; 0.00 MB usable, 0.00 MB loaded, 1639.41 MB offloaded, 168.76 MB buffer reserved, lowvram patches: 0
level_zero backend failed with error: 39 (UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY)

Intel® Arc™ B580 Graphics
Version
32.0.101.6739

[earnest grotto]

and what pytorch version did you pick when installing
stable?

[remote estuary]

There is a update available for the gpu and yes, i did all the default the script had.

[earnest grotto]

When you launch comfy, in its command prompt it says, among other things:

ipex_init: (True, 'Skipping IPEX hijack')
Total VRAM 15931 MB, total RAM 49075 MB
pytorch version: 2.9.1+xpu
Set vram state to: LOW_VRAM
Device: xpu:0 Intel(R) Arc(TM) A770 Graphics

what pytorch version does it say

[remote estuary]

ipex_init: (True, None)
Total VRAM 11944 MB, total RAM 32592 MB
pytorch version: 2.10.0+xpu
Set vram state to: LOW_VRAM
Device: xpu:0 Intel(R) Arc(TM) B580 Graphics

[earnest grotto]

run the script again from the same location, but choose 2.8
no need to delete anything, it will just change the pytorch version, it will be faster than installing from scratch

say if it works then

[remote estuary]

ok, thanks

[lunar thicket]

oo boy did they break a new nightly for B580 😅

[remote estuary]

Not nightly, running stable

[lunar thicket]

Oh did they launch 2.10 officially? Neat

i am running on one of the dev releases of 2.10 still and it's working well for me (B580)

[remote estuary]

Looks like it works, thanks

[earnest grotto]

do you intend to do anime art or realistic stuff

[remote estuary]

Unsure yet might be a little of both

[earnest grotto]

models that can do one well struggle hard at the other

[remote estuary]

I was mostly just trying to find out why it wasn't producing anything, so now I guess ill poke around a bit

If you have some suggestions for me to grab ill do so and try them out

[earnest grotto]

for realistic stuff i recommend z-image. has workflow in comfy.
for anime i recommend NoobAI v-pred 1.0 with EQ-VAE. you can look that one up, I won't link it
each is prompted differently. z-image takes natural language, the anime model takes booru tags + a few extra tags not in boorus (masterpiece, best quality, great quality, low quality, worst quality / year 2024/2023/etc., see noobai's readme)

[remote estuary]

Awesome, ill try those two out thanks for the help and the script vik!

haha I see

[lunar thicket]

Maybe I try noobai over illustrious?

[earnest grotto]

vpred means the model can generate a dark/bright image. regular epspred tends towards making a greyish image though there are ways to remedy this, like noise offsetting. the downside is noobai vpred can also go crazy a bit and make blown out images, however that's the default noobai vpred.
the eq vae is less noisy than the normal sdxl vae, which makes the model train better. (doubt this is meaningful for lora training)
the noobai vpred with the eq vae model someone finetuned, doesn't have the broken coloration issue and is pretty decent
if you're using any illustrious loras, 99% chance they will not be working for this model. i train pretty much all the loras i use myself

(this one is with some edits afterwards)

And one without any loras. This might've been with regular vpred, don't remember

[somber trellis]

Is there a guide to getting torch compile working through linux/WSL?

I am rather frustrated that a 3060 can get near 5x my speeds on certain text to speeches, because they have cudagraphs.

[earnest grotto]

It works by default, what are you doing, what error are you getting

ah, hm, it works for image gen

sdxl was formerly broken but apparently works now

[somber trellis]

On Linux only though, right? Not Windows.

[earnest grotto]

Aaron had it working on windows. Things are partially set up on windows however I gave up trying to get their compiler to compile with the right flags and to figure out which of the various C++ compilers the basekit includes I should be using. Let's assume it doesn't work on windows

[somber trellis]

Well, I'm going to set up a WSL 2.9.1+xpu environment

And try to get it working.

[earnest grotto]

If compile works for z-image/lumina for you, with the newest torch, on linux, then whatever doesn't work is probably a bug you won't be fixing

[somber trellis]

I didn't ever get compile working on linux.

[earnest grotto]

well, i guess 2.10 is broken for alchemist so eyah use 2.9

[earnest grotto]

Install comfy with my script and tell me what is not working when you try to compile, with the default compile node

[somber trellis]

what's the linux pip package for intel basekit?

just intel-basekit?

[earnest grotto]

you don't need that

[somber trellis]

Oh.

[earnest grotto]

only windows theoretically needs it for compile to work

[somber trellis]

i got it working

thats with the q8 gguf of z image

[earnest grotto]

use the same environment for whatever you were doing with tts

you should be able to just ctrl+c or ctrl+z and be in the environment

[somber trellis]

i needed a lot of new variables to be set for it to work

[earnest grotto]

qwen-image-layered is out

with GGUFs

https://www.reddit.com/r/StableDiffusion/comments/1pqnghp/qwenimagelayered_released_on_huggingface/

VAE: https://huggingface.co/Comfy-Org/Qwen-Image-Layered_ComfyUI/tree/main/split_files/vae
GGUFs: https://huggingface.co/QuantStack/Qwen-Image-Layered-GGUF/tree/main

Their HF space

hmm

If you want to get the same error, https://huggingface.co/spaces/Qwen/Qwen-Image-Layered

[earnest grotto]

no comfy just yet 🤔

god i hate the sd subreddir

it's like 80% of posts about amd are condensed ragebait misinformation, and I even saw some of that spilling over into intel too

[reef ivy]

Had to edit some files, don't actually remember off the top. Have to do a search as it never got pinned. No clue if latest pytorch/oneapi still work though

[earnest grotto]

not a oneapi issue

they most likely made it worse with newer triton/torch

it gives wrong arguments to each compiler and i didn't wanna bother

[slender zodiac]

Do you have any experience with comfy and AMD cards? I see posts about DirectML and such. Asking for a friend who has an older AMD card.

[earnest grotto]

How older

I don't have personal experience, neither do people in the SD sub. I'm just not colossally misinformed

If it's an rx 580 or something then yeah sure, give up on it

[slender zodiac]

Radeon Vii It was purchased for some desktop work but now the AI things are here they want to try out things. They've put in some considerable time and effort into getting things working. However, it's really hard going.

They have the DirectML working to a point but everything is so focused towards Cuda it's a PITA

I was hoping for some solutions that would be more in line with the IPEX Intel ARC we have here.

Searching the internet is just like searching through $#!+ for some things

[earnest grotto]

Is the person willing to dual boot Linux

[slender zodiac]

They are very competent (PhD in engineering)

[earnest grotto]

I don't know if the vii is supported but it probably is, for windows in particular tell the person to try ZLUDA
SDNext has a zluda guide

If not, ROCm on native Linux

VII is also a bit oldish so i'm not super sure on what it supports

[slender zodiac]

Thanks for the input, much appreciated.

[earnest grotto]

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

damn, zluda does go as far back as including the rx 580

the guide doesn't name the vii and i don't know gfxwhat it is but, yeah

[slender zodiac]

I recommended the card to him a while ago because it has some nice memory bandwidth for the tasks he has. It was also very reasonably priced for a 16GB card with 1TB/s memory.

[earnest grotto]

yeah it was a decent choice back then

if only amd had been more proactive about compute support

[slender zodiac]

I also have a couple Radeon Vii semi-retired in workstations. I should be able to test things out for him.

[earnest grotto]

If you do, no wsl for rocm.

Zluda on native windows, or rocm on native linux

90% chance rocm on linux works

it will probably perform better than zluda

[earnest grotto]

man

https://www.reddit.com/r/StableDiffusion/comments/1pqy27v/noobai_flux2vae_prototype/

bleach post for that amd post i saw

[wicked fulcrum]

New sneak peek at AI Playground 3.0 with release of 3.0.0 alpha. See in AI Playground General thread
https://discordapp.com/channels/554824368740630529/1245461432141873245/1451701978643038221

[civic charm]

directml is dead and is total pain when it works

directml only supports fp32

and is very slow

use native rocm on linux

radeon vii should be one of the oldest cards still supported by rocm

it has fp16 and int8 but no bf16 support

llms should work very well on it via vulkan

1 tb/s is still very competitive

[earnest grotto]

Preliminary result
Most likely the leftover noise is something that will be fixed

strongly considering just going to sdnext for this one though, or doing it in a standalone script with sdnq. all the extra performance will help. yeesh.

[earnest grotto]

A little better. Still has leftover noise though, even at 30 steps. Yeesh. Getting the feeling that this is undertrained

[honest hull]

https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installrad/windows/install-pytorch.html

ROCm torch on windows

[earnest grotto]

Isn't that only for rx 9000s and some 7000s?

[honest hull]

yeah only for newer Gen GPUs, and some integrated GPUs

[earnest grotto]

https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2511

[earnest grotto]

yoooooo ltx has lightning loras already trained

https://huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning/tree/main

(also fp8 scaled version of the model is there too)

[earnest grotto]

Doesn't work in comfy. I have a feeling that the block swapping/reserve-vram is once again broken

Gonna try with GGUFs https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF

works

looks better than 2509, as expected

[earnest grotto]

ah, not issue-free though. that ghosting that was present in Q4 and below of 2509 seems to be happening with Q6 of 2511 now (only sometimes??? when asked to upscale? not sure)

[earnest grotto]

Ah, the fluxkontextmultireferencelatentmethod node set to index_timestep_zero helps

[earnest grotto]

I guess this node just got renamed to "edit model reference method"

[somber trellis]

How does 2511 compare in terms of style transfer @earnest grotto

[earnest grotto]

Let's start off with the negative. There is still no zero shot style transfer. First row - the 2 reference images, "Transform the image of the woman in white clothes with red eyes into the style of the image with the woman with long curly hair and blue eyes"
The 2 other rows - 4 different seeds

And the positive. You probably already know these images. Whatever they did, they've slightly veered away from that disgusting sloppy look most post-SDXL (even base SDXL) models have when asked for "anime", besides just better quality in general.

Still gave venom an eye when his eye was closed but hey, it didn't remove the eyepatch

[earnest grotto]

rip stylegan

I really hope this sort of stuff doesn't just get forgotten or whatever

Anyways, in general the model is just better

It didn't get worse or something

You can just go download the ggufs, the new lightning lora, add the new reference latent method node after the text encode, and done

[somber trellis]

RuntimeError: Given normalized_shape=[3584], expected input with shape [*, 3584], but got input of size[2, 77, 2560]

Im using the GGUF q8 model, probably missing something in either the workflow or maybe an outdated node

[earnest grotto]

You might need to update the node and comfy

[somber trellis]

I actually already did a git reset --hard, git pulled the current repo and then reinstalled via your setup script, overwriting the pytorch to 2.9.1+xpu

[somber trellis]

Fixed the issue. Improperly named mmproj for qwen 2.5 vl.

[hot crescent]

Friends with the b580 using comfyui and qwen image edit, which quant are you using and how much ram do you have? I'm maxing out vram with the Q3_K_L

[earnest grotto]

How much ram do you have

[terse osprey]

I updated ai playground to the latest version and now when I try to generate an image using z image turbo in ComfyUI I'm getting this error. Can anyone help?

[lunar thicket]

Post in #1243956384052285560 sections

https://github.com/intel/AI-Playground/issues/366

Looks like you aren't alone

[terse osprey]

okay thanks

[hot crescent]

16gb system DDR5 🤯. But another 16gb stick coming tomorrow, I’ll be interested to see if that makes any diffeeence

[reef ivy]

Man, thats like 300+ dollars now lol. I should have just upgraded my ram early this year smh

[earnest grotto]

Only 16 is low yes

When you get more, up your pagefile as well. you should be able to at least use q4

also, --reserve-vram 7

[earnest grotto]

Show the whole error

[umbral nimbus]

Hope it's ok i ask here since it's a niche forum. My CD 'directory' path seems to do nothing and conda can't find my requirements file. How do I proceed?

[earnest grotto]

to cd across a drive, either use powershell, do cd /D letter:wherever/, or first cd letter: then cd to the folder.
In either case, the instructions you're following from the top of this thread are very outdated. I have a script that will install comfy for you, 3rd pin in this thread.
#1193952640225267802 message
@umbral nimbus

[umbral nimbus]

Thanks so much! swift answer!

[left glacier]

Hi. I've been out of touch with open-source developments for a while. Currently, what's the best generative video model that works flawlessly with Intel Arc A770? I tried AI playground but it doesn't work. it just give me "level_zero backend failed with error: 45 (UR_RESULT_ERROR_INVALID_ARGUMENT)"

[earnest grotto]

AI Playground 3.0 installs a broken version of pytorch (2.10) that doesn't work with alchemist GPUs.

Eventually, that version will be fixed. Here's an issue you can look at to see when it might get fixed. https://github.com/pytorch/pytorch/issues/170166

If you want to use comfy with a working pytorch (2.9), I have a script that sets it up for you #1193952640225267802 message

You will need some modern version of python, doesn't matter which, git, and conda installed (ideally miniconda in your user directory)

@left glacier

[left glacier]

Thank you. I'll try that.

[earnest grotto]

@somber harbor ^ Link and some instructions there

[somber harbor]

2.8, or stable?

[earnest grotto]

Use what the script tells you is default. If you have an alchemist gpu like an a770, that's 2.8.

You can also just press enter without typing in anything to get the default

[misty rain]

Has anyone installed SeedVR2 in comfyui? I'm using torch 2.11.0.dev20260103+xpu and everything I tried with it has worked so far except this.

It doesn't import. The Upscaler node seems to work but the LoadDitModel and LoadVAEModel nodes are not.

[tough wharf]

check this out
https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler/issues/429

I wouldnt want to download a random zip file, but if youre willing to test it on a sandbox

[earnest grotto]

ltx2 out

and new comfy update breaks fp8. amazing.

[earnest grotto]

comfy-kitchen/backends/eager/__init__.py
line 30,
all_devices = frozenset({"cuda", "cpu"})
->
all_devices = frozenset({"cuda", "cpu", "xpu"})

I'll see about reporting it

[earnest grotto]

I imagine it'll be fixed in less than a day. There's already a PR for it and other bugs, if anyone wants to keep track https://github.com/Comfy-Org/comfy-kitchen/pull/8

[earnest grotto]

Merged

[somber trellis]

some examples of it seem decent

[queen remnant]

i've been away from comfyui for a long while and am having trouble updating comfyui to a newer version via comfyui manager after running a new version of vik's installation script on my ancient install (0.2.1p -> 0.2.6.1p). hitting "Update ComfyUI" just dumps the following into the console and does nothing:

[ComfyUI-Manager] Failed to checkout 'master' branch.
repo_path=F:\AI-NVMe\Comfy_Intel\ComfyUI
Available branches:
        master
ComfyUI update failed

[ComfyUI-Manager] Queued works are completed.
{'update-comfyui': 1}

After restarting ComfyUI, please refresh the browser.

i can't add, update, or remove custom node packages through comfyui manager until i am updated as stated by this message:

[Security Alert] ComfyUI outdated. Installations blocked (update allowed).
Update ComfyUI for normal operation.

and if i recall correctly there are patches that are done during the installation that i would have to reapply manually if i were to force an update, so im a little stuck on what i should do

any help would be appreciated

[earnest grotto]

The code block you posted is not output of the script

What does the script say
Did you run it from the same location you did the first time

The script has no "update comfyui"

[queen remnant]

poor wording mb

"Update ComfyUI" is in ComfyUI Manager

i can run the script again to get the output

[earnest grotto]

The script will update comfy for you, don't use comfyui manager for that

[queen remnant]

does the script install a specific version of comfyui or is it meant to be the latest release/dev build

[earnest grotto]

it installs/updates to the latest

[queen remnant]

oh im way more outdated than i thought

[earnest grotto]

?

[queen remnant]

the latest release for comfyui is listed as 0.7.0 on github, it shows 0.3.43 here

rerunning your script and posting the logs here in a moment

apparently im on a detached head

would running git switch main inside the comfyui folder and running the script again be a bad idea? im not too familiar with git

[earnest grotto]

open a command prompt inside the ComfyUI folder (e.g. right click the background when you have the folder open)
do git branch and tell me what it says

[queen remnant]

PS F:\AI-NVMe\Comfy_Intel\ComfyUI> git branch
* (HEAD detached from v0.3.42)
  master

[earnest grotto]

do git reset --hard and tell me what happens

[queen remnant]

PS F:\AI-NVMe\Comfy_Intel\ComfyUI> git reset --hard
HEAD is now at 170c7bb9 Fix contiguous issue with pytorch nightly. (#8729)

[earnest grotto]

ok, run script again now

[queen remnant]

said the same on trying to pull again in the script and is still the same version after running it

[earnest grotto]

what does git branch say now

[queen remnant]

PS F:\AI-NVMe\Comfy_Intel\ComfyUI> git branch
* (HEAD detached from v0.3.42)
  master

same as before

[earnest grotto]

git reset --hard
git pull
git status

[queen remnant]

PS F:\AI-NVMe\Comfy_Intel\ComfyUI> git reset --hard
HEAD is now at 170c7bb9 Fix contiguous issue with pytorch nightly. (#8729)
PS F:\AI-NVMe\Comfy_Intel\ComfyUI> git pull
remote: Enumerating objects: 5, done.
remote: Counting objects: 100% (5/5), done.
remote: Compressing objects: 100% (2/2), done.
remote: Total 5 (delta 3), reused 5 (delta 3), pack-reused 0 (from 0)
Unpacking objects: 100% (5/5), 2.33 KiB | 44.00 KiB/s, done.
From https://github.com/comfyanonymous/ComfyUI
   fc0cb10b..c0c9720d  master     -> origin/master
You are not currently on a branch.
Please specify which branch you want to merge with.
See git-pull(1) for details.

    git pull <remote> <branch>

PS F:\AI-NVMe\Comfy_Intel\ComfyUI> git status
HEAD detached from v0.3.42
Untracked files:
  (use "git add <file>..." to include in what will be committed)
        .gitBACKUP/
        SDXL_trainer_output/
        comfy/ipex_to_cuda/
        huggingface/
        loras_tags.json

nothing added to commit but untracked files present (use "git add" to track)

[earnest grotto]

git checkout master
git pull
git status

[queen remnant]

looks like that might've done it

[earnest grotto]

ok, run the script again

[queen remnant]

woah

ui is way different

works thanks for the help

[earnest grotto]

Are nodes 2.0 enabled?

If they are, consider disabling them

at least for now

they're buggy

[queen remnant]

they werent but i popped them on to look at them

thanks for the warning

[somber trellis]

I wonder when torch compile on windows will ever be fixed.

[reef ivy]

Is there still a speed increase compared to older pytorch with torch.compile? Or is it about relatively the same?

[somber trellis]

For me, Z-image was about 30-40% faster on compile when i tried it out via wsl2

[somber trellis]

@earnest grotto I got ltx-2 working

10 minutes per video gen tho (for 720p 24fps)

[somber trellis]

lol and i just realized i had it still set for the normal model (20 steps) when i should be using 8

not even just that, I'm on cfg 4 and not 1 for the distill

[reef ivy]

How much ram/vram needed

[earnest grotto]

should be doable with pretty low vram, but ram is a concern now

[reef ivy]

Wish I had upgraded early last year smh, oh well. Probably be cheaper to get a 16gb or used 20/24gb gpu now lol

[lunar thicket]

Interested to know what the render time is once you correct the steps

This is still on A770?

[somber trellis]

20 second 640x480 video took almost 9 minutes

4-5minutes for 5 seconds

It'd be faster if I had an fp8 gemma 12b cuz a nice percent of that time is the clip text encode

[earnest grotto]

gonna be honest, all the voices i've seen from ltx are atrocious

or i guess heard

the video matching the audio is pretty good and there being audio is good but man, how is it such a step back from even TTSes from 5 years ago

good thing it has audio input i guess

[somber trellis]

agreed

im not very sure if im excited over it

or more excited over the wan 2.2 lora svi pro's infinite video capability

[earnest grotto]

what's the ram usage

[somber trellis]

reserve vram 8 set

[earnest grotto]

https://www.reddit.com/r/StableDiffusion/comments/1q7dzq2/im_the_cofounder_ceo_of_lightricks_we_just/

They plan to release a 2.1 that will have better audio, and a bigger increment like a 2.5 that will better preserve details in motion, better latent space

[somber trellis]

I'll wait for that, then. Until then wan is better overall imo

[somber trellis]

until then

imma jury rig a workflow I found for SVI pro

so I can use T2V alongside SVI pro's I2V infinite video

[somber trellis]

@earnest grotto

I think there are examples where LTX-2 can provide some interesting shots

The screen is black, then a slow push-in reveals a desolate corridor where dark crimson blood crusts the concrete floor, glistening with wet sheen beneath cracked walls littered with rust flakes. Flickering fluorescent tubes cast long, trembling shadows over dried blood residue, painting the scene in cold, uneven light. From 30 centimeters above the ground, a medium-low angle camera tracks a skeleton—smooth, pale white bone like porcelain, devoid of joints or tissue—moving at a deliberate 0.5 meters per second. Its spine is arched at 45 degrees; its femurs elongate and shift in unnatural sequence, the right hip rotating ten degrees clockwise before the left hip follows. As it steps forward, each movement ripples through the dried blood, followed by deep bone clicks and groans that echo every three seconds—low-frequency vibrations beneath a steady electrical hum. Faint drops of blood fall at five per minute, silent but inevitable, like punctuation in a slow, unresolved sentence. No voice, no sound beyond the ambient bleed of wetness and mechanical resonance—the skeleton moves with quiet inevitability through the silence.

[reef ivy]

i'm seeing some very cool ltxv audio examples on the banodoco discord. Not able to keep up with it, but think I saw someone talking about voice cloning

[earnest grotto]

apparently video continuation is also possible

and jesus what is this lora

Surprising that it can do seemingly infinite (at least 2 minutes) video without any burn-in or other such issues

good stuff

[somber trellis]

no clue how to do that yet

[somber trellis]

Some of the outputs are quite good, to be honest.

[somber trellis]

Got GGUFs working, as a recent merge to comfyui-gguf has it functional.

[somber trellis]

I kept ooming on gguf

[earnest grotto]

This time it might be a comfy issue

[somber trellis]

fp8 works fine, so i assume its to do with the gguf nodes themselves

[reef ivy]

Try third party nodes like kijais, I remember a while back native gguf nodes had issues im comfy from what I heard

[earnest grotto]

"Hmm yes surely comfy implemented Blender's group nodes correctly this time :^)"

^ Clueless

Changing a subgraph and finding out that the other 5 are unchanged is unreal

[somber trellis]

kinda stinks that it takes me 12 minutes per 20 second 640x480 gen 🙁

but it works

[lunar thicket]

Im too impatient to try that stuff

those times are crazy

[reef ivy]

12 minutes for 20 sec's is amazing tbh. it's usually 4 or 5 sec's for wan

[earnest grotto]

I can generate 5 second 3mp (640x480) video with wan 2.2 in around... 3-4 minutes?

[lunar thicket]

well that is on track with Dan if you just scale up to 20 seconds

[reef ivy]

yeah, 20secs would be way more than 12minutes though.

[somber trellis]

so it seems i can do 1280x720x484 frames

on distilled, with res_2m sampler, two pass (8 step and 3 step) with second pass using ic detailer lora

Just barely too, but this is with --reserve-vram 16

Not ridiculously slow, either. Might actually take 12-15 minutes as well.

Okay, well second pass is taking 8 minutes per step. LOL

[somber trellis]

oh that's a lot better
swapped to ggufmultigpu nodes (which work for ltx-2 now it seems)

[somber trellis]

(generation took 30 minutes at 1280x720 20 seconds)

[somber trellis]

With the distilled gguf and multigpu nodes, 640x480 20 seconds takes almost 10 minutes

[marsh spindle]

has just compared z-image turbo in ai playground 3 with z-image comfyui nvidia same prompt, a770 vs 3090 -> 17 and 10 sec, not that bad!

[lunar thicket]

ok creepy

[somber trellis]

weird models are always great with horror

Seems like I need to edit my system prompt, though. The LLM I use to enhance prompts is adding the speech in the output prompt.

[reef ivy]

How much does the video change with resolution and the same seed? IMO, that wouldn't be a bad time for a final cut if you can do lower res versions.

[somber trellis]

No clue, but it's not like the quality of 640x480 outputs are immediately bad

It's an odd, mixed bag.

Like it's actually

quite movie-like in some shots

[reef ivy]

Once/if some loras come out the quality will get better. Same with wan 2.1 when it first released.

[earnest grotto]

flux 2 klein 9b base is gated. off to a great start.

The 4B one looks to not be gated

They really couldn't resist slapping that license on both 9Bs huh

[earnest grotto]

4b distilled | 4b base

Colorize the image. The girl has blonde hair and blue eyes. Her uniform is dark green. She has gloves. She's sitting on a brown couch.

The uh, skin-colored pants seem to be an issue with quite a few models

(For that matter the gloves are too but I decided I will specify them this time. From my experience with qwen, prompting "she has gloves" instead of something like "the gloves are black" is likely more prone to producing additional glovey artifacts)

Mm, a bit like qwen it's a bit better at higher resolution

Yeah, at 2mp the pants are mostly green

[earnest grotto]

Welp, I'd say it's overall much worse than qwen for editing.

Very touchy when it comes to CFG amounts

Not that I like CFG as it's mostly done now anyways

Upscaling, it has the tendency to fry the image at 5.0, 2.5 works. Edits where it needs to generate more new stuff from scratch though, needs 5.0 otherwise blurs

I wonder if z-image-omni-base will once again drop really soon now and shame bfl again

[fiery current]

can anyone help on this error?

[slender zodiac]

Create a virtual environment named comfy-env

python3 -m venv comfy-env

Activate the virtual environment

source comfy-env/bin/activate

[earnest grotto]

@fiery current @slender zodiac Use my script that automates it for you and does some extra patching to Comfy ^

[slender zodiac]

Did you activate the env?

[earnest grotto]

Also, https://github.com/Comfy-Org/ComfyUI/commit/2108167f9f70cfd4874945b31a916680f959a6d7
3 hours ago, comfy support for z-image-omni-base

[slender zodiac]

@earnest grotto

Thanks for the script

I'm just old and stuck in the past before script times

[earnest grotto]

sooooonnnn™©™

[earnest grotto]

Comfy can be installed without my script now, however my script adds disty's hijacks which help for custom nodes

And also certain pytorch versions break with certain GPUs

or right now in particular pytorch 2.10 breaks with alchemist, so my script handles that

used to be that 2.9 and 2.10 would break with battlemage iirc

[slender zodiac]

Is there a standard workflow for benchmark?

Just wanted to check everything is ok 😊

[earnest grotto]

any sdxl model at 1024x1024 with euler ancestral and more than 20 steps, 1 batch size

for a770, you should expect ~2.45it/s
For b580 it was 3.4it/s I think

run twice (two prompts/seeds, don't restart comfy). measure second run

[slender zodiac]

Thanks 😊

[lunar thicket]

b580 should be able to hit around 4it/s with pytorch 2.10

im still running one of the nightlies from December

[marsh spindle]

is that possible to make a script that will modify comfyui portable installation to run in intel arc? in theory

[earnest grotto]

The script makes a portable installation

I don't intend to make it use python venvs instead of conda, for now

[lunar thicket]

Hey Vik maybe a question you have answered before but: if you run the script on an existing installation location, say you wanted to change pytorch versions or update something, does it wipe your models?

[earnest grotto]

No

[lunar thicket]

for safety ive just been moving my location whenever running the script again, but I figured that was probably the case. Thanks 👍

[earnest grotto]

You can plop an existing comfy inside the file structure the script makes, you won't lose anything other than possibly any manual edits to comfy (and only comfy)'s code

Don't recall if it will install requirements for all present custom nodes or only the ones i've set it to install (if you choose so)

That being said, I keep my models in a separate external folder, for other web UIs or such

[lunar thicket]

I haven't used any other tools that would benefit from such an arrangement

[fiery current]

well I managed to fix it by manually adding 2 packages I traced from the error log and then installing the default requirements.txt that comes with comfy. But thanks for the script though.

[reef ivy]

Had latest pytorch been fixed yet?

[earnest grotto]

https://github.com/pytorch/pytorch/issues/170166

[reef ivy]

Seems its fixed but not added as a commit for some reason?

[honest hull]

if you really need the portable version... for now you can just download the nvidia portable.7z and modify one of the batch file in ComfyUI_Windows_Portable/update/update_comfyui_and_python_dependencies.bat

change 'cu130' to 'xpu', save the file and run the batch file to update the pytorch dependencies.

ideally we would work with comfy to get it integrated into github workflows for releases

[earnest grotto]

If you have an alchemist GPU, like an a770, you will also need to specifically use pytorch 2.9 because 2.10 is currently broken

And, this won't cover certain custom nodes (e.g. ultimate SD upscale will be very slow)

[quasi cypress]

I've installed the torch but it says this

[earnest grotto]

@quasi cypress You have not installed torch. You can just install comfy using my script ^

[quasi cypress]

what to do

[earnest grotto]

Show the whole command prompt

[earnest grotto]

You don't have permission to create your Comfy install in C:/Windows/system32.

And you don't want to do that anyways.

IDK how you're running the script to convince it to try to set up inside system32, but, don't
Right click the background in the documents folder, "Open command prompt/powershell/terminal here", "python ./Setup_Comfy_Intel.py"

[earnest grotto]

Check to make sure that this is where you saved the file and that's how it's named.

[quasi cypress]

ah yes it was missing the UI in name

thanks

[quasi cypress]

I've installed with the script then clicked the icon that appeared after the installation

[earnest grotto]

Update your GPU drivers

What driver version are you using

[quasi cypress]

updating rn

[earnest grotto]

hmm

[quasi cypress]

same error

[marsh spindle]

tnx! i also found portable release one chinese guy did (on github) YanWenKun. He did some tweaks to vanilla portabe

im getting like 60-70% perfomance with it of my rtx 3090

[earnest grotto]

Run the script again from the same location, this time choose stable pytorch instead of 2.9. It will install faster as it will only change the version. Say if it's still broken

[earnest grotto]

How many it/s with the 3090 with sdxl at 1024x1024

[marsh spindle]

i'll give exact numbers tmrw, but i compared with z-image turbo 1024x1024

[marsh spindle]

right now i can say however, that comfyui version bundled with AIP 3.0 still performs better, 17 vs 23 sec for 1024x1024 but could be they used somehow modified z-image workflow

models seems the same, output quality same, same steps but somehow AIP3 is faster

[earnest grotto]

What pytorch version are you using

[marsh spindle]

portable using 2.9.1+xpu and AIP i downgraded to 2.9.0

[earnest grotto]

@quasi cypress Well, did it work?

[honest hull]

in AIPG we added --lowvram as launch arg, so that could improve performance as well

[quasi cypress]

timezone diff sorry couldnt reply, yes bro it did thank you so much

[marsh spindle]

i noticed that, and re-configured to "--highvram" back again, thought you added it to be on a safe side with b570 and a750?

[marsh spindle]

may i as you another question then ) testing "advanced chat" i noticed that it always crash when trying to load large models - which theoretically should fit in a770 vram, but when it goes even slightly above 12gb vram, it crashes