ComfyUI for Intel Arc using IPEX

[marsh spindle]

for instance i have two versions of qwen3-14b, one uses 10 and works ok, another 12 - no go.. could be something related to b580 limit hard coded?

3090 vanilla, 9sec 1.08 it/sec
a770 PG, 17sec 1.72 it/sec
exact numbers for same prompt (z-img)

[earnest grotto]

gib sdxl

[reef ivy]

I always found lowvram slowed performance on the a750. Haven't tested since like last May though.

[earnest grotto]

Lowvram makes the text encoder run on the cpu.

[reef ivy]

Maybe its because I force that with a custom node.

[earnest grotto]

It's slower than running it on the GPU, decently slower for qwen, not much slower for the various CLIPs and not much relatively slower for the text encoders of the fatter slower models like wan, qwen/edit, etc.
But it's also way faster than running out of VRAM

[reef ivy]

Hopefully will be able to mess with stuff again soon

Reserve vram always helped me never oom.

And sped up gguf speeds a ton

[earnest grotto]

If you ran out of vram, then reserve vram will be faster yes

If you don't, it will always be massively slower

[marsh spindle]

seems BTW they fixed current pytorch, 2.10.0+xpu now runs ok with alchemist (just did what Li suggested) im now on python 3.13.9 Pytorch Version 2.10.0+xpu

[earnest grotto]

i'll have to check that out 🤔

gib sdxl @marsh spindle

[lunar thicket]

I should do a 2.10 install and see how it compares to the one im on

for B580

[earnest grotto]

might be mildly faster

don't remember if the speed boost came with 2.9 or 2.10

[lunar thicket]

I am on an old 2.10 nightly

[earnest grotto]

doubt much will be different then

[marsh spindle]

i feel now that it just one sec slower, a little bit, interesting to compare what you guys think

[earnest grotto]

why no sdxl

[marsh spindle]

dont have it atm, which one would you like me to download? sdxl turbo?

[earnest grotto]

are you interested in doing anime art

[marsh spindle]

not exactly ) but if for science

[earnest grotto]

if you don't want to then you can just get base sdxl https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors

i just wanna know what your performance is, with the a770 and the 3090

[marsh spindle]

ok i'll just choode that sdxl_simple_example

[marsh spindle]

all right i did the tests, so: a770 intel "vanilla" 2.55 it/set (pt 2.10), a770 intel "downgraded" 2.69 it/sec (pt 2.9) and 3090 3.69 it/sec (again on 2.9.1)

[earnest grotto]

how many steps, and with euler?

you know what they say, run twice, measure the second time

comfy tends to skew the it/s on the first run

usually skewing them higher i think

[marsh spindle]

yes i did a couple of runs

this exact preset no changes

[lunar thicket]

lower in my testing on the first run, and then subsequent runs all measure similar to eachother. At least for sdxl

[earnest grotto]

i see
odd that 2.10 is slower

that shouldn't be happening

and, i guess the b580 does indeed have 3090-tier speeds

hopefully int8 support comes to comfy kitchen sooner rather than later

with int8 and torch compile you can get z-image to be what, like 2x faster or more?

compile might take a bit of time to be fixed up for windows, or i wonder if things have improved now if 2.10 is also finally fixed for alchemist

[marsh spindle]

they also seems to have 2.11 (nightly)

[earnest grotto]

yeah

[marsh spindle]

any point to install that how d you think

[earnest grotto]

you mean any point in using sdxl?

the point to download and try it was to know your performance, so I can know everything is fine

if you're asking for using it longer term, for realistic stuff there is no point in using it over z-image

and well, the base model doesn't do anime. but the anime finetunes of it are basically the best models currently for doing anime

[marsh spindle]

no i mean try 2.11

[earnest grotto]

if you think it might be faster then try it out

otherwise there won't be much difference

[lunar thicket]

What I was responding to was the "measure twice" comment; in my experience the first run is slower than the second run.

I wasn't commenting on the 2.10 pytorch or anything else

[earnest grotto]

this

slower

sometimes the first run is slower for me, i might've been mistaking that with the it/s it shows initially

don't remember anymore

[marsh spindle]

hehe but 2.11 does not work at all

[earnest grotto]

could just be today's nightly

[marsh spindle]

somehow i noticed these strings Intel devs did in comfyui cfg speeds up things considerably: --disable-ipex-optimize --bf16-unet

[reef ivy]

Only 8gb VRAM so it was always faster lol. Sometimes had to adjust the number for different models etc. I think block swap fixes all that though, was only in kijais nodes last I tried though

Is 2.11 still broken on alchemist or did it finally get fixed

[somber trellis]

@earnest grotto I got Qwen3 TTS working for a little on my xpu using AI (Despite the script being slower than cpu for some reason)

The quality of that TTS literally matches or trumps vibevoice 7b

and it's a 1.5b parameter model

I'm gonna keep trying to get good performance out of it. Right now though, it's pretty slow.

[earnest grotto]

nice

[somber trellis]

Bet itt'l get better since qwen3-tts is an LLM. LLama-CPP should be able to run it

We already can build sycl llama-cpp from source

maybe ill try that route

[somber trellis]

This is with waves realtime-autotune added. Original audio above is Qwen3-TTS with transcribed cloning audio for Glados from the intro.

(random llm text, ignore that just listen to the voice quality)

[somber trellis]

Voice quality great, emotional quality? EH.

[somber trellis]

Actually, Qwen3-TTS has emotion control too. It might be the best all-round.

[earnest grotto]

@somber trellis What's the repo

[somber trellis]

https://github.com/QwenLM/Qwen3-TTS

get the 1.7b base model

that one is the one for voice cloning

voicedesign and customvoice are text based

[somber trellis]

I also dont believe I mentioned this, but a 100m parameter voice cloning model released before this one, and ive been using it since in games lmao

it runs on cpu so its no issue to run at sll

that 100m model is pocket tts

kyutais newest

not as good as qwen3 obv but its 17x smaller lmao

[idle siren]

Hey guys, I'm new to this and I wanna know the latest and greatest way to run comfyui on Intel arc b580 gpu.. should I just run it via playground or there's other good way?

[earnest grotto]

You can use AIPG, and you can install it with my script ^ if you want only comfy

@idle siren

[idle siren]

@earnest grotto can u give me more info please

[earnest grotto]

What do you not understand

[idle siren]

What's AIPH

[somber trellis]

uh vik the current qwen3 models do not support emotion instruct

[idle siren]

*AIPG

[somber trellis]

HOWEVER

[earnest grotto]

AI PlayGround

[somber trellis]

https://github.com/QwenLM/Qwen3-TTS/issues/14

^

This. They're releasing another model. Named Qwen3-TTS-25Hz-1.7B-VoiceEditing

[idle siren]

Oh.. thanks.. where will I get the script

[earnest grotto]

From the message I replied to that the arrow points at.

[idle siren]

@earnest grotto thanks I'll try it..

[earnest grotto]

cosyvoice was the last model i tried which supposedly does but i didn't test it that much. i just poked it a bit on modelscope

[somber trellis]

The upcoming voiceedit model will basically be a combine of the two other models qwen3 has, voicedesign (text input used to describe voice) and base (audio input used as reference)

[earnest grotto]

yes the issue says that

[somber trellis]

Which means we should be able to use a neutral expressive audio input, and then instruct it to any expression we want.

We could technically even take an angry voice, instruct it to be neutral

that would be interesting

[earnest grotto]

We should many things. We'll see how it pans out

I imagine it will mostly work with a neutral voice but am doubtful about making an angry voice neutral

[somber trellis]

https://github.com/kyutai-labs/pocket-tts btw is the small model im using

It's surprisingly good

[earnest grotto]

eeeeh, not quite there imo

[lament shale]

updated to the latest version and my old comfyui version breaks while vae decoding. Did you update the script later? and does the latest version of intel-comfy work good, because I may wanna just update it to the latest

[earnest grotto]

Be more specific

Show errors, what model, etc.

[lament shale]

it was working just fine before I installed the latest driver

https://www.intel.com/content/www/us/en/download/785597/intel-arc-graphics-windows.html

there's no error log, comfy just disconnects

after i gen an image with an illustrious model, i tried it with amanatsu v1.1

[earnest grotto]

So you updated the drivers, not comfy?

[lament shale]

yeah

i should update comfy

does your script still work?

[earnest grotto]

The script still works. It might not fix your issue

What GPU, show the workflow

[lament shale]

intel a750. it just disconnects

[earnest grotto]

What makes the "efficient" ksampler "efficient"

[lament shale]

i use that to use a hires script

and it's faster than normal ksampler

[earnest grotto]

Open task manager. Switch to the performance tab (2nd tab), and to the GPU, while keeping both the RAM and the VRAM visible, like this. Make it crash and show what task manager shows

[earnest grotto]

If you don't know what makes it "efficient", it's better to not use it and learn to use regular comfy.

[lament shale]

which is odd, since i have been using this flow for half a year

[earnest grotto]

You do not need any custom nodes to do hires fix

Nor to do some of the common not great hacks to speed up things like 1.0 cfg at lower sigmas

[earnest grotto]

Use tiled VAE decode instead.

[lament shale]

this ss is right after it crashed

okay

[lament shale]

okay, that worked now

i genned a pic

what is the normal vae decode method instead of the tiled that made it crash

[earnest grotto]

When shared memory starts being used, that's because you ran out of VRAM

[lament shale]

oho

[earnest grotto]

VAE decoding large images will use massive amounts of VRAM

Tiled VAE decode uses far less VRAM

[lament shale]

are there any drawbacks or so they give the same output

reagrdless, thanks for the help Vik

appreciate it

[earnest grotto]

The quality is the same. If you do not run out of VRAM, it will be slower.

[somber trellis]

or just cpu vae decode if you dont want checkerboarding from tiled decode

The checkerboarding and padding from tiled vae decode really isnt bad tho

[earnest grotto]

yeah 2.10 is still broken for alchemist, just in a different way

2.11 nightly more broken

let's see if newer drivers help either of them

[earnest grotto]

ok, works now

good

[amber fractal]

Can anyone tell me how to add a model from huggingface.co as on the website I cannot seem to find the exact link that the Playground 3.0 is asking me for?

[earnest grotto]

This is a comfy thread, are you asking how to use it in comfy or in aipg

[amber fractal]

well I am trying to add a model into AIPG but it tells me to give a link but cannot find the right link on any huggingface page

[earnest grotto]

What model

[earnest grotto]

@vivid moat It works now

[amber fractal]

For example https://huggingface.co/sylvester-francis/typescript-slm-7b-reasoning-full

[earnest grotto]

This is completely unrelated to comfyui
Just use that link?

And,
https://discord.com/channels/554824368740630529/1245461432141873245

[amber fractal]

Sorry for the wrong thread

[earnest grotto]

Maybe this one actually https://discord.com/channels/554824368740630529/1265055106928476192

[reef ivy]

So 2.11 works on alchemist with latest drivers? Nice

[earnest grotto]

2.10 does, honestly didn't try 2.11 as well but i expect it does now too

[reef ivy]

Nice, interesting that drivers fixed it and not an update to pytorch

[earnest grotto]

I may have been low on space as well. In either case, it has been updated and was crashing differently

And less

[somber trellis]

huh wait does the latest nightly work

I can finally stop using 2.9.1?

I'll test the latest 2.11 nightly right now on the latest 8425 driver

[somber trellis]

Torch compile on windows still does not work on the latest

At least from what i can tel

[somber trellis]

@earnest grotto Traceback (most recent call last): File "C:\Users\dbs_5\Comfy_Intel\ComfyUI\execution.py", line 518, in execute output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) File "C:\Users\dbs_5\Comfy_Intel\ComfyUI\execution.py", line 329, in get_output_data return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) File "C:\Users\dbs_5\Comfy_Intel\ComfyUI\execution.py", line 303, in _async_map_node_over_list await process_inputs(input_dict, i) File "C:\Users\dbs_5\Comfy_Intel\ComfyUI\execution.py", line 291, in process_inputs result = f(**inputs) File "C:\Users\dbs_5\Comfy_Intel\ComfyUI\custom_nodes\ControlFlowUtils\ControlFlowUtils.py", line 2833, in garbage_collect torch.cuda.ipc_collect() File "C:\Users\dbs_5\Comfy_Intel\cenv\lib\site-packages\torch\cuda\__init__.py", line 1178, in ipc_collect return torch._C._cuda_ipc_collect() AttributeError: module 'torch._C' has no attribute '_cuda_ipc_collect'

on the latest nightly

found the line causing it, marked it out

My issues are only seemingly related to custom nodes. Now that they're fixed, it's operational.

[wicked fulcrum]

FYI AIPG 3.0 add models feature for chat has a bug. This will be fixed with beta in a couple of weeks.
Reason is, presets have filters the model list uses from a models.json file. We are fixing this so adding modelss gives you the option to tag the model duty the right data

[somber trellis]

I seem to be getting random issues with the latest nightly, where it just randomly closes itself.

I'm gonna go back to 2.9.1 for now.

Nevermind. The issue is not related to pytorch. I need to DDU.

[somber trellis]

Okay, guess this current driver isn't working right.

I just did a DDU, and both 2.9.1+xpu and 2.11.0+xpu just silent crash on the latest driver (8425).

Downgrading to latest WHQL driver and retrying.

[somber trellis]

Seems downgrading to the latest WHQL was the right choice. Three generations so far without crash.

[somber trellis]

Pytorch 2.11 nightly is operational on Alchemist, then.

Compile's still broken.

[earnest grotto]

Fascinating, so opening Blender seems to cause training with kohya to frequently segfault

I guess one of the incessant memory management issues finally presented itself in a more reproducible state, though this will probably still take me a while to better isolate

Could be down to my specific blend file

Also could be something newly introduced with the latest driver because I'm pretty sure this didn't happen before

[earnest grotto]

Could even be GPU subdiv 🤔

[reef ivy]

Is it torchcompile not working or the comfy node? Its calling cuda there.

There is a command you can use to test torch compile, i forget where it was maybe on the nightly pytorch page. Sorry it has been so long since I have been able to sit down and mess with AI stuff

[reef ivy]

Run this and see if you get the same error https://docs.pytorch.org/docs/stable/notes/get_start_xpu.html#train-with-torch-compile

[somber trellis]

It was tested via another script entirely outside of comfyui

[somber trellis]

I get the same exact type of error I get in 2.9.1.

[somber trellis]

Running that script gave this error on 2.11+xpu

This error would not occur on Linux.

[reef ivy]

Did you ever do this stuff? #1193952640225267802 message i feel like 2.9 should work unless thats when it broke, last I used was 2.8 though

If that does work still, maybe someone could pin it for people. Let me know if it works

[somber trellis]

I tried the latest release of that loader, and I also tried to use 1.20.2

Neither of those fixed the issue, despite the oneapi environment being activated.

Could be on my end though if you wanna try testing it.

[reef ivy]

Its likely just broken then, I won't be able to try myself

I take it going back to 2.8 it works? Could also be one of those driver updates

[round stag]

i'm a complete newbie in this

i'm facing this error while doing a fresh install of comfyui

Command: comfyui-backend startup
Exit Code: Unknown
Duration: 6.0s
Timestamp: 1/27/2026, 11:29:12 PM

Standard Error:
Adding extra search path checkpoints D:\Ai_Playground\AI Playground\resources\service\models\stable_diffusion\checkpoints
Adding extra search path clip D:\Ai_Playground\AI Playground\resources\service\models\stable_diffusion\checkpoints
Adding extra search path vae D:\Ai_Playground\AI Playground\resources\service\models\stable_diffusion\checkpoints
Adding extra search path unet D:\Ai_Playground\AI Playground\resources\service\models\stable_diffusion\checkpoints
Adding extra search path loras D:\Ai_Playground\AI Playground\resources\service\models\stable_diffusion\lora
Setting output directory to: C:\Users\Dhruvajyoti Dutta\Documents\AI-Playground\media
Checkpoint files will always be loaded safely.
Total VRAM 11940 MB, total RAM 16237 MB
pytorch version: 2.8.0+xpu
Set vram state to: LOW_VRAM
Device: xpu:0 Intel(R) Arc(TM) B580 Graphics
Using pytorch attention
Traceback (most recent call last):
  File "D:\Ai_Playground\AI Playground\resources\ComfyUI\main.py", line 148, in <module>
    import server
  File "D:\Ai_Playground\AI Playground\resources\ComfyUI\server.py", line 32, in <module>
    from app.frontend_management import FrontendManager
  File "D:\Ai_Playground\AI Playground\resources\ComfyUI\app\frontend_management.py", line 16, in <module>
    import requests
ModuleNotFoundError: No module named 'requests'```

[lunar thicket]

This looks like you were installing Intel's AI Playground, right? That uses comfyui but is its own front end. There is a section for it here: #1243956384052285560

Did you download it straight from their GitHub? It should simply install, from an exe.

[round stag]

i installed it from here

[lunar thicket]

Oh I see the line in the installer is what is failing. Probably still best to move to #1265055106928476192 so the right eyes are on it

[round stag]

ah kk didnt know where to post the report thanks

[lunar thicket]

*This channel you were originally posting in is for comfyui sans AIPlayground. Which is what I moved on to after using AIPlayground for months

[coarse whale]

ZIT base should work on the same workflow of the turbo right? For me images on same workflow look like this:

i tried with more steps CFG bug same result

any ideas?

[coarse whale]

fixed it was the turbo lora not working

[coarse whale]

i updated comfyui using the script, but the same workflow i used to use before that was taking li 10 gb vram now goes to shared gpu memory making it incredibly slow. What could be wrong? I updated also all custom nodes.

i tried both lowvram and lower vram. it seem the workflow now uses more ram than what it used to.

Tried a default workflow and same problem

[earnest grotto]

Show the workflow

And say how you're launching Comfy

[coarse whale]

here is one, i tried also illustrious one that were not giving me any problems beofre and same thing

i always lunch thorugh here. Tired both lower and default

also tried through here

when i updated through the script i chose all the default options, aside from the one of installing additional cusotm nodes

I'm gonna try to install with pytorch nightly instead of stable and see if it makes a difference

[coarse whale]

nightly same problem

tried also 2.9 and the memory management works (don't go into shared) but i get a bug i used to get a while ago where some images (not all) are black on illustrious (but turbo is fine)

for reference this is the same error i used to get on 2.9 here: #1193952640225267802 message

in summary:
on stable and nightly memory management doesn't work and slows a lot down generation when it hits shared memory
on 2.9 it works but i get problems of black images with illustrious

[earnest grotto]

hmm, zimage does seem to use much less vram with 2.9 for me than it does with 2.10

no idea why, definitely something broken

[earnest grotto]

sadly, the random black images bug is fixed in 2.10 and not 2.9, so not much you can do

[coarse whale]

Does illustrious work for you on 2.10? For me it also uses much more VRAM exactly as zimage

[earnest grotto]

i haven't paid attention to its vram usage

[coarse whale]

Okay, I think for what I notice is not about the model but in general something not working with comfyui from 2.10 on VRAM usage

[earnest grotto]

8.5gb with 2.9

10gb with 2.10

[coarse whale]

nice thanks for checking

it seems i need to revisit my workflow if i go with 2.10.

[earnest grotto]

hol up

one more version

ah, rip. also 10 with 2.11

[coarse whale]

did they add some variables that use more v-ram that could be disabled?

[earnest grotto]

no clue

[coarse whale]

I dont find any mention online so maybe is something that just affects intel

[earnest grotto]

there are already a few longstanding intel-specific bugs

you should assume any odd severe issues (of the crash/leak/corrupted outputs/etc. kind) on non-nightly versions are intel issues

[earnest grotto]

memory-related bugs

so another memory-related bug is even less surprising

[coarse whale]

is 2.11 a nightly or a stable?

[earnest grotto]

any error about mismatched tensor sizes or devices or whatever is of course someone else's problem

[earnest grotto]

nightly

[coarse whale]

In your opinion do you think something like this might get fixed in 2.11 or later? Or is it unlikely?

[earnest grotto]

i don't know the release schedules of pytorch and I don't know how far along 2.11 is but I'd assume,
unlikely in 2.11
possible in 2.12
too far into the future beyond that

if 2.11 is currently earlier than I'm assuming, it might get fixed in 2.11 too

[coarse whale]

Thanks, I don't follow them at all and I'm barely understanding how it works, so is good to get an external view. I will keep on ZIT on 2.9 for now

and test again the 2.11 in a month

[honest hull]

add --reserve-vram 6.0 to your launch arg python main.py --lowvram --reserve-vram 6.0

[earnest grotto]

huh, ok, comfy must've fixed performance with reserve-vram

the lowervram shortcut has reserve-vram as a launch argument if you don't want to edit things

[coarse whale]

that doesn't seem to do anything for me, this is the bat file i launch it from:

cd /D "%~dp0"
call conda activate ./cenv
:: bmg does not need forced slicing
:: No additional environment settings needed for Windows
cd ./ComfyUI
python ./main.py --bf16-unet --disable-ipex-optimize --lowvram --reserve-vram 6.0

i added --reserve-vram 6.0 to the end but it still takes much more vram on 2.10

[earnest grotto]

the lowervram batch file (and shortcut respectively) has reserve-vram

it helps when you're running out of vram

if you're not running out and stick to 2.9, it won't do much

it is basically mandatory for the bigger models like qwen or wan

[coarse whale]

yeah i thought it was a possible solution, didnt see you already included it in the lowervram shortcut

i mean i find it a bit annoying cause i cannot run my old illustrious workflow at the same resolution as before since in 2.9 is bugged and 2.10 consumes too much vram

but is not a big deal

[earnest grotto]

which is why li is telling you to use reserve-vram

[coarse whale]

yeah but the workflow still uses more vram

[earnest grotto]

and reserve-vram will make it use less

[coarse whale]

at the price of speed?

[earnest grotto]

don't know what the speed will be

reserve-vram used to make speed worse when not running out of vram. now it doesn't. i don't run out of vram with sdxl either way so i can't test and see

so you test and see

just run the script and pick 2.10 and run with the lowvram shortcut, test and see, repeat for 2.9 and regular

[coarse whale]

i still run out of vram with the lowervram (which is 7 reserved)

during the k sampler

´should i increase the 7 or decrease it?

[earnest grotto]

increase to 12

that's not good

[coarse whale]

like before this run without using shared memory, is like a lot more vram used

this is weird no?

[earnest grotto]

how's the performance

[coarse whale]

really slow

[earnest grotto]

yeah alright, was worth a try

[coarse whale]

thanks for the help anyways!

[honest hull]

--bf16-unet --disable-ipex-optimize you probably don't need those anymore

if you are using Vik's scripts.. @earnest grotto is your script also applying disty0's workarounds ?? https://github.com/Disty0/ipex_to_cuda this helps with memory management..

[fathom lichen]

when i try to run python main.py --bf16-unet it says " File "C:\ComfyUI\app\assets\scanner.py", line 5, in <module>
import sqlalchemy
ModuleNotFoundError: No module named 'sqlalchemy'" any sollution

[earnest grotto]

Yes

@fathom lichen Install comfy using my script ^

[fathom lichen]

okay bro

[fathom lichen]

Can u guide me if its possible

[earnest grotto]

What part of the instructions in that message do you not understand

[fathom lichen]

i download the file

now

[earnest grotto]

@fathom lichen well?

[fathom lichen]

It's done thanks bro

[earnest grotto]

Yes

[earnest grotto]

One of these is Flux 2 Klein 4B, the other is Z-Image
Man...
||Left is flux, right is z||

Granted I trained flux with just 0.59mp (768^2) images and z-image with both 0.25 (512^2) and 0.59

[somber trellis]

Left is clearly the better image.

Klein looks like a near-finished piece.

Z-image is giving that flash game drawing vibe

lmao

[somber trellis]

zimage base

Pretty good.

[earnest grotto]

shame that it seems klein is basically sd3 2 in terms of duplicating people's limbs

i guess it shows that it was made by former SAI people

decoupled diffusion transformer never

[earnest grotto]

i can't believe a good anime model just got shadowdropped

just like that

and it's not even done training and is already on par with/better than(?) the sdxl finetunes

without any of the issues of neta or others

and 2b

damn

[slender zodiac]

Which model ?

[earnest grotto]

https://huggingface.co/circlestone-labs/Anima

Native comfy support already (comfy likely already supported cosmos)

From my very limited testing, basically on par with any modern SDXL anime finetune but without the vpred fried colors / epspred incapability of bright/dark scenes, and with the sharpness the newer VAEs provide (qwen in this case), about the same speed but compile and other stuff will likely help that even more

And this already when anzhc is doing the flux2 vae adaptation of noobai

FINALLY some good new anime models

oh and of course, this has proper language understanding which the sdxl models of course lack, but unlike neta/yume it doesn't mysteriously ignore certain things or listen to them too little (like quality tags / artist styles, among others)

it is a bit undertrained and the anatomy is a bit iffy but hey, anatomy can already be iffy with sdxl anyways so whatever, it's only getting better

hopefully musubi tuner supports it soon, finally I might switch off of noobai vpred eqvae https://github.com/kohya-ss/musubi-tuner/issues/350

[earnest grotto]

hopefully the qwen instead of flux2 vae doesn't mean training loras is worse on it than sdxl 😐

[somber trellis]

Indeed. Been messing with this model.

It works with the wan 2.1 2xupscale vae, with comfyui-vaeutils node.

this is with the default sampler in their workflow, er_sde 100 steps cfg4 (easycache)

eh

[earnest grotto]

Noobai Epspred 1.0 | Noobai Vpred 1.0 | Noobai Vpred EQ VAE experimental ver.
Netayume 4.0 | Noobai RF Flux2 VAE 0.3 | Anima Preview

That last one could probably benefit from non-booru prompts for lumina and anima but oh well

[coarse whale]

I guess is cause anima doens't work on torch 2.9?

[earnest grotto]

Show the workflow

[coarse whale]

i got the model from here: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/diffusion_models
maybe i should try with the one on civitai instead

https://civitai.com/models/2356447/anima-fp8?modelVersionId=2650296

[earnest grotto]

that link is correct

you do not need an fp8 version

it's a 2B model.

[coarse whale]

ok

[earnest grotto]

is your comfy up to date @coarse whale

[coarse whale]

do i need to update it after running again the script to downgrade torch version?

[earnest grotto]

I don't know what is causing this issue

So, we're going through various things that could be

[coarse whale]

this is the way right?

[earnest grotto]

No.

[earnest grotto]

Your Comfy is old

[earnest grotto]

Run the script from the same location you previously did when installing comfy, and it will update it. Keep pytorch the same, 2.9 works and you shouldn't downgrade.

[coarse whale]

thanks now is better, i only get an error on vae.

Can you send me a workflow you tried so i can test with that?

when you get the chance to, no hurry

[earnest grotto]

Is this with the same workflow again

[coarse whale]

nevermind i tried again and it worked

it donest work with the tiled vae but i guess i dont need it

thanks for the help once again!

[earnest grotto]

You don't need most everything you've done in that workflow and the various parameters are off

e.g. 12 CFG is way too much

[coarse whale]

if you send me one i will use it!

[earnest grotto]

the model is high fidelity enough to not need upscaling

and if you were to upscale, you most likely shouldn't do it with usdu

[coarse whale]

what does "H,W (144, 113) should be divisible by spatial_patch_size 2" mean? I tried to switch width and height to have another aspect ratio

[earnest grotto]

it's supposedly not been trained too much on 1mp but it should work at higher resolutions better than sdxl models. for that matter, illustrious and by extension noobai have some training on higher resolutions so those too can be used more directly to upscale and you won't have to deal with ugly seams

[earnest grotto]

width and height should be divisible by 8, or probably even 16

since you were using some bad resolution with sdxl too,

⁨```
1024x1024 1.0 or 1.0
1152x896 1.2857142857142858 or 0.7777777777777778
1216x832 1.4615384615384615 or 0.6842105263157895
1344x768 1.75 or 0.5714285714285714
1536x640 2.4 or 0.4166666666666667

These are the 1mp resolutions people train SDXL models on, and I imagine what this model was also trained on

[coarse whale]

thanks!

[proper knoll]

Not sure this is the place but I've downloaded ai playground and when trying to install comfyui it keeps coming back saying failed 🤔

[lunar thicket]

What version of AIPlayground?

also there is an #1243956384052285560 area 😁

[proper knoll]

The latest version i believe

[lunar thicket]

what version

i am asking cause someone recently had this exact problem, said they had the latest, and in fact did not have the latest

installing the latest fixed the issue

[proper knoll]

Version 3.0.1-alpha

[proper knoll]

This is the error I keep getting

[lunar thicket]

huh interesting.
It seems to be telling you exactly what to do though 🙃

is D an external drive or formatted with a weird file system or something?

[proper knoll]

Its an external drive yes

[lunar thicket]

try installing on an internal drive

Or, run that git command it told you to run, and try the install again on D

[proper knoll]

Well I just clicked install on the ai playground so where would I run the command 🤔

[lunar thicket]

if you don't want to learn about git, then just install AIPlayground on your internal drive.

You would run the git command from the command line. But the fact you asked where, that tells me you probably shouldn't mess around in command line, unless you wanna do a little homework. 😁

[proper knoll]

What's the difference running it on the internal drive ?

[lunar thicket]

Because it is an external drive there are all sorts of possibilities for why it would throw this error (if could be, i assume, as simple as having it formatted FAT32 instead of NTFS. Just a shot in the dark here).

But it would be highly unusual for your C drive (or any other internal drive) to report the same error since it was installed and configured with your machine. Or at least, most typically is.

Technically possible you could grab a random drive from a different computer and install it in your machine without formatting but its less common lol

[proper knoll]

OK I'll uninstall and try again

[umbral nimbus]

Just want to send an apprectiation message to Vik and everyone else helping out the noobs. The script has some phenomenal ease of use and I'm glad I can get comfy as an Xe-core enjoyer

[earnest grotto]

https://cdn.discordapp.com/emojis/378059485173907457.webp?size=96

[somber trellis]

@earnest grotto

[earnest grotto]

the new ace-step?

[somber trellis]

Yeah.

it's actually starting to become good.

[earnest grotto]

bitrate still sounds lowish but definitely better

[somber trellis]

That is fixable

novasr (a stupid fast audio upscaling model)

could fix this

[earnest grotto]

hopefully it trains better than this

[somber trellis]

my ears

LMAO

[somber trellis]

the lsd monster

vik i hope you realize

this is like the lowest of the barrel outputs from this model

the prompts i used are like 5 words long

it describes nothing

like if a random user tried to use ace step

https://blog.comfy.org/p/ace-step-15-is-now-available-in-comfyui

Some of the examples shown here seem a lot better than these

Like, its nuts

like its almost as clear as suno 5 in this example

Got the workflow.

the aio turbo checkpoint's only 9.8gb

[somber trellis]

I had to use a custom node method of forcing it onto my xpu

lowvram kept shoving the clip encoder into cpu

now that its on my xpu, clip encoder takes very little time, a minute and a half or so

[earnest grotto]

that's what that does yes

[somber trellis]

i know

but thank you for reminding me

id like to mention

that we can generate songs faster that we can listen to them

I am fairly impressed.

[somber trellis]

People on the huggingface page stated that the stable maximum is around 60-65 steps

default for this model is 8 steps

its so fast that theres no reason not to do this

I'm having issues with vae decode, though. Gonna force that on cpu

[somber trellis]

Fixed it myself. Made AI code me a tiled audio decode.

[somber trellis]

This one was a failure, but interesting to listen to.
Mainly how I handled the lyrics.

[somber trellis]

I'm getting there, but I think I'm starting to understand a bit more on how to use this.

[somber harbor]

Hello everyone, I've made it so that the text encoder starts reading in seconds (previously, everything worked through the processor and took a long time), but as soon as the finish occurs, this error occurs and the program crashes. I have 32 GB of RAM and a 16 GB graphics card

[earnest grotto]

Open task manager and pay attention to your ram and vram. 2nd tab. You can see both at once when you're looking at the gpu.

[somber harbor]

the amount of video memory occupied 15 GB, and the RAM was half-used

The chronology is as follows: After starting the generation and reading of the text encoder, the RAM was immediately filled, but the graphics card was idle. Later, the RAM became less, and the graphics card was loaded with 15 GB of VRAM. When the text encoder was read, the program crashed. How to avoid this, I really liked the model), but the text encoder takes 23 minutes to load on the processor (((

[earnest grotto]

If you installed comfy with my script, run it with the lowervram shortcut.

[somber harbor]

I will try to do this now and write about the results.

I tried it.... No changes(((

[earnest grotto]

Alright, I'll see for myself later

[somber trellis]

I've gotten the hang of Ace Step, for the most part.

It seems for now, changing the keyscale at all causes adverse results to the output.

But setting temperature in the clip encoder to 2, helps.

[somber trellis]

I finally got a song with a good brass-sounding instrument.

I changed the wording of what I used.

[earnest grotto]

2.8 ipex lowers anima (cosmos predict 2b) lora training speed from > ~5s/it to ~3.3s/it, and vram usage from 13gb to 11gb (incl. random garbage in the background; 12 rank, 1024^2 only)

impressive

[somber trellis]

I'm very confused as to why these are happening.

Intel moment?

[earnest grotto]

I'v heard the comfy implementation of ace-step 1.5 is scuffed in multiple ways compared to their gradio version, not that that is issue-free

Good chance it's their problem

[somber trellis]

I'm also getting device lost crashes and bluescreens every few gens.

🤷‍♂️

[somber trellis]

I don't know if anyone has worked on making the gradio interface useable on xpu

[earnest grotto]

it has at least some explicit xpu support in the code

installation though doesn't seem good

[somber trellis]

It's hella slow on Comfy when using the 4b lm to clip encode

So I'd really like to find a different method, especially since the only working model on my end is the turbo model in comfyui

[somber trellis]

Yep, never noticed the requirements-xpu there

gonna run that and try it out

[earnest grotto]

Well, that's probably because
It was literally just added (3 days ago)

good on them, or whoever did the pr

it also supposedly has a lot more features than are exposed in comfy

besides lora training which in itself could be pretty big

interesting that they also apparently have compile enabled by default for intel

[somber trellis]

which ill have to disable since it just implodes

on windows

a

[earnest grotto]

one of the people who did the commit, did it on windows

[somber trellis]

...

[earnest grotto]

**Tested Device**: Windows laptop with Ultra 9 285H integrated graphics.

[somber trellis]

Okay, is there an up-to-date guide to getting torch compile on windows working on Arc?

[earnest grotto]

wonder if it's because of battlemage or what or why it worked

[somber trellis]

Because I'll read it. And I'll do it.

[earnest grotto]

i have no idea and that's what i'm interested in maybe finding out at some point

the person also specifically used 2.8+ipex to test

[somber trellis]

Oh.

I had torch 2.8 compile working at some point.

[earnest grotto]

yeah it might be that it works with 2.8+ipex

[somber trellis]

That was when aaron had it working.

🤷‍♂️

[earnest grotto]

and broken afterwards

[somber trellis]

I never bothered returning to it

compile didnt seem worth using due to the precompile times

Yeah, okay. It's quite a lot faster through the gradio interface.

Lol

I'm going to downgrade to 2.8, I guess. Doesn't seem like 2.10 has anything notable.

And it has it's own prompt enhancer.

Cool

[somber trellis]

I've actually been having worse results on the gradio UI.

🤷‍♂️

[reef ivy]

2.8 torch.compile definitely used to work. Not sure if it's actually faster than 2.10+ without it though.

[digital pond]

hello i have installed comfyui from the python script, how do i install comfy maneger?

[somber harbor]

Good evening, go to the custom nodes folder, run cmd from there, and copy - git clone https://github.com/Comfy-Org/ComfyUI-Manager.git

[earnest grotto]

ComfyUI has come with a manager built-in for a long time, what do you need the external one for?

[earnest grotto]

noobai vpred eqvae experimental | anima preview

Hmm... I really gotta ponder the pg-13 thing now

I have another grid I'd like to post

well, not quite a grid but still

Well, I guess I'll pass.
Anima seems to struggle with smaller details like fingers and lacks a bit of knowledge on more obscure stuff. The kind of stuff I'd expect from a slightly undertrained model, it is a preview after all, but it's already really good

[fiery current]

hey folks, does installing this makes any difference in comfyui?

[earnest grotto]

No. Save the space.

[somber trellis]

@earnest grotto Having an issue I've realized in both comfyui and the gradio ui implementations of ace-step

where the output gets more garbled the more generations you do

until you either "out of memory" or device lost

[earnest grotto]

I have no idea for now, sorry, but I'd expect this to be a bug with ace-step rather than intel. they may have done more fixes on their repo so update that and recheck

[fiery current]

I have this problem with this version of portable comfyui: https://github.com/YanWenKun/ComfyUI-WinPortable-XPU. For some reason I can't get the server to be connected on my other devices. Tried --listen 0.0.0.0 multiple time, but still not working. Any way to make this work so that i can access comfy on my phone? FYI, I tried normal way of installing comfy,but an update borked that installation. SO i went for that xpu portable solution

[earnest grotto]

I have a script that sets it up for you as well, and can update ^

The author of the repo you linked mentions flux.1 doesn't work and doesn't mention installing disty's hijacks, or any equivalent monkey patching, so, it's worse
dunno why someone would install 40+ custom nodes but omit the GGUF ones

ah, nevermind, so they install the gguf ones but still can't get flux 1 to work?

sad for them

[earnest grotto]

And what IP are you connecting to through your phone?

[fiery current]

my computers ip that show up on router, 192.168.0.200.

[earnest grotto]

Do ipconfig in any command prompt
Is that what's shown?

[fiery current]

same

[fiery current]

is there anything I do with this ui that pops up when initializing

[earnest grotto]

I have not made or used this and I can't tell you anything about it. Keep launching with --listen, you do not and probably should not specify an IP after that.

[fiery current]

ok

[wicked fulcrum]

Hey all, looking for the message that links to the latest ComfyUI install instructions that uses xpu. I think the pinned messages are out of date - still use IPEX

[earnest grotto]

My script (3rd pinned message) is up to date

Otherwise, comfy has instructions in its readme https://github.com/Comfy-Org/ComfyUI?tab=readme-ov-file#intel-gpus-windows-and-linux

[fiery current]

any possibility of having a portable version for intel?

[earnest grotto]

The thing you linked me is already portable

The comfy structure my script creates can be called portable (it does not install anything outside its folder, though you can't move it willy nilly without needint to edit things)

[fiery current]

I know but it kinda feels bugged and some of the scripts always report missing or backdated on it's console. Even comfyui crashes randomly

[earnest grotto]

And the comfy, on github, that you manually install, is also considered portable while not really being quite portable (you can't put it on a flash drive as the environment is in your PC's python install)

[fiery current]

you know what, scrap that. I am testing your setup script. Let's see how this one goes.

[somber trellis]

im assuming that could be resolved with a portable python install

then youd have your little on-the-go comfy instance

[warped tulip]

something was wrong with me

im have cpu intel 10400f and arc b580 (titan noc oc)

Instead of comfy I'll probably just download Anacode

[lunar thicket]

I saw you were using Gemini. It is probably assuming you have Git installed, but your command prompt looks like you don't.

If you want to use Comfy use Vik's script here: https://github.com/a-One-Fan/ComfyUI-Intel-Installer-Script

as it says on that page, you will need to install git and miniconda

[warped tulip]

git and minicode downloaded

@lunar thicket sorry for ping

but please help

You can download the script from above, or from here, then right clicking and choosing "save as".

[warped tulip]

you'll explain what happens next

[lunar thicket]

run the script 😁

it is going to run in the console and prompt you with some options. When it is done you will have ComfyUI installed and Intel GPU compatible

[warped tulip]

only that it downloaded as a txt file, and I don't know what to do next

[lunar thicket]

It should be a .py file which is a Python script.

Do you have python installed?

[warped tulip]

yes the latest version

[lunar thicket]

If so, you run it by opening a command prompt and typing:

python SCRIPTNAME.py

replace SCRIPTNAME with the actual name

[warped tulip]

I'm not an advanced user,

I wanted to try comfy ui, but I see it's not that easy

because the problems are at the installation stage

[lunar thicket]

@warped tulip

it is true that setting up ComfyUI is not a casual experience

If you still want to use it but want an even easier installation, you can install Intel's AIPlayground.

[warped tulip]

I tried once but something didn't work out, I'll try it a second time at most

because this is the first time I'm trying to run AI on my computer

[reef ivy]

Replace "scriptname" with the actual name of the script

[warped tulip]

I installed it, just Intel AI Playground (less problems with it)

[reef ivy]

No doubt its good to use, but obviously won't be as up to date

[wicked fulcrum]

ComfyUI install setting in AI Playground updated.
TIP: For Intel Core Ultra you can remove or reduce reserve vRAM, but for dGPU this is important

[rough seal]

With dedication for mr @earnest grotto. For his great instalation script :), thank you dude! 🙂

[earnest grotto]

@rough seal No unsolicited friend requests, please.

If you have something to say say it here, if you have some reason to send me a friend request say it here or at least send a DM anyways

[neon tapir]

does anyone have a problem with efficiency nodes installing?
Edit: found the fix in the cenv i needed to run

pip install "numpy<2"

[reef ivy]

Hey yall, is anyone doing professional work using local ai? Advertising, freelance, film/animation, art/design? If so what models are the best and would 8gb vram and 32gb ram be sufficient with upscaling etc. I have been out of the loop for months because of personal matters but will be hopping back in soon to see what is actually possible with modest hardware, starting with intel but maybe moving to Nvidia if its too cumbersome. (Too bad about the b770, maybe b70 if I can actually monetize one day)

[reef ivy]

Also would updating to 48gb(ddr4) even be worth the 130$+ price for 16gb or just upgrading to a new gpu with 12gb or 16gb?

[lunar thicket]

I only use it for goofing around

[earnest grotto]

since you didn't get a response, i think it's safe to say no one does. if you want a model recommendation, flux 2 klein 4b and 9b for realistic image gen/editing, and consider qwen edit 2511 as well. lodestones (chroma author) is training a 4b not-chroma finetune, though I don't know what it's called. my interest is mostly anime models, and BFL are actively hostile towards that. they even made a tweet bragging about it and implicitly how they want businesses to be using their models (bragging about how flux 2 is less "vulnerable" than qwen image).
for video gen there's ltx2 but i haven't tried it.
IME 48gb does help decently over 32 but 16gb for $130 sounds too ridiculous to me and with an 8gb GPU you're likely better off getting a better GPU first, especially since the b580 at least is decently faster, and there's the 9060 xt 16gb / 5060 ti 16gb as well.

[queen osprey]

I upgraded comfyui to v0.15.1 I have a770 16gb but I have older dell server dual xeons amd 128gb ddr4, I have it all working apart from when I try upscale it doesn't work

[earnest grotto]

There are tons of ways to upscale, how are you upscaling, what workflow, what are you doing

[queen osprey]

No mate I have never been able to do it pc crashes hard if I try

Its cause the dell sever I use didn't have resizeable bar till I made it have it

[reef ivy]

9060 xt and 5060 ti are just so overpriced now. 5070 is about the same, much more performance but not enough vram SMH.

[earnest grotto]

new ltx

[reef ivy]

Looks like it is both better and worse than 2.0 lol.

[queen osprey]

Even the ai playground if i choose upscale it crashes

[reef ivy]

upscale what, and to what size?

[somber trellis]

He says, then I go in and try it.

🤷‍♂️

[somber trellis]

Better prompt + RIVE V49

[somber trellis]

@reef ivy I think the model is a clear improvement.

Using it alongside RIFE, makes it look very nice indeed at 640x480

[reef ivy]

I think slow motion is also one of the issues, but i also think that adding slow motion prompts to negative fixes it. I can onlu go by what I read from others though

[earnest grotto]

pretty bassy audio for a not so impressive scene

[earnest grotto]

this one's pretty good though

[reef ivy]

yeah, looking at some of the other generations, slow motion is a problem. Like the background will stay slowmotion even when the speech is at regular speed. But negative prompt seems like it can fix it easily

[somber trellis]

it was quite literally the second generation ive ever done with the model, and the first with an LLM enhanced prompt

It's interesting, for sure.

[somber trellis]

@earnest grotto I don't think you need ipex_to_cuda anymore. I've removed the import entirely from model_management a while back. Your setup script still installs it from my knowledge.

I haven't had any OOMs in the last two comfyUI runs I've had

[earnest grotto]

It has not been needed for basic comfy for a long time
It will be needed for custom nodes for a while

[somber trellis]

On the latest nightly

Oh. I didn't know some custom nodes needed it.

Apologies.

[earnest grotto]

A few of the ones my script installs might need it but I don't remember anymore. Most people do not use comfy's built-in device management and just resort to leaving things to be "cuda" or "cpu" or nothing. though some are not like that at least

the tts suite is good about it generally, though it's kinda per-model for some reason

Well, while I'm at it, I figure I might as well drop some pics in here

[proud lodge]

Can someone please help me I did everything i have Arc A770

[earnest grotto]

@proud lodge ^

use that

[proud lodge]

im new so how do i use it and where to put it?

[proud lodge]

@earnest grotto how do i use this?

[lunar thicket]

if you have conda/anaconda installed, and python installed, then you can just run the script from a commandline pointed at the directory you saved the script to. Where you save the script doesn't matter. You would type (without quotes):

"py Setup_ComfyUI_Intel.py"

Alternatively , if you have Py Launcher installed, you can run the script via a doubleclick on the script file.

The readme on here goes over it some more: https://github.com/a-One-Fan/ComfyUI-Intel-Installer-Script

[proud lodge]

im confused lol

mmm

[proud lodge]

Any ideas?

This my system

[lunar thicket]

is this second log from after running the script?

[proud lodge]

yes

[lunar thicket]

it should generate some shortcuts for you to use. It looks like it is installed. So then, wherever you downloaded and ran the script, there should be links called "ComfyUI" and "ComfyUI_Lowervram"

running those turns on Comfy. Then you just need to open a browser and navigate to your local IP (all this will be written in the commandline)

[proud lodge]

let me check

yes

so di i run it?

do

[lunar thicket]

doubleclick the "ComfyUI" one

if you run into VRAM problems, you can use the "ComfyUI_Lowervram" one

running models will use up a LOT of VRAM so its best you arent trying to do it while doing lots of other things

[proud lodge]

So how do i run the interface?

[lunar thicket]

that looks like it ended in error. Hmm never seen anyone with this one yet

AttributeError: module 'torch.xpu.memory' has no attribute 'CUDAPluggableAllocator'

what version of python do you have installed?

[proud lodge]

oh

10.6

[lunar thicket]

3.10.6?

[proud lodge]

yes

[lunar thicket]

that is pretty old. idk if that is relevant or not. but i might start there by updating python and rule that out

[proud lodge]

What version should i use

[lunar thicket]

i have 3.13 myself

3.13.7

[proud lodge]

ok so do i uninstall 3.10.6 and try 3.13?

[lunar thicket]

ill be frank: at this point, once the script is installed, my usefulness is mostly at an end 😅 (except a couple times i saw a familiar error and could help)
Vik is the guru who wrote it. I just guide people to its use occasionally. My suggestion at updating Python is a total shot in the dark

[proud lodge]

Ok thanks for helping I really appreciate it

[lunar thicket]

my pleasure. one more question, what version of PyTorch did you select during install?

[proud lodge]

it didnt get that option

[lunar thicket]

🤔 it has to present that option

you cant finish installing ComfyUI (using Vik's script linked in those hyperlinks above!) without that option

[proud lodge]

now i get this when i try to run the script

[lunar thicket]

hmm

when you installed Python did you have this highlighted line checked?
the yellow one about environment variables?

it looks like an environment variable problem

did you choose system32 for the Conda environment or was that default?
that is surprising to me (but i have never run the script in Windows)

generally speaking you shouldn't be doing anything in there

[proud lodge]

oh

i dont know if it was by default

still not working

[lunar thicket]

what folder did you save to, and run from, 'Setup_ComfyUI_Intel.py'?

[proud lodge]

Document folder

[lunar thicket]

ok

well im not sure. I'm gonna say wait for Vik to reply to your ping. Hopefully my questions along the way are informative for him

[proud lodge]

Ok

[earnest grotto]

@proud lodge I've updated the script. Download it again, run it again, you can run it from the same location as last time so it's faster (it won't need to redownload things as much)

[earnest grotto]

.

In the location you ran it here.

As the instructions say, please run the script in some folder you have permissions for.

[proud lodge]

@earnest grotto i don't understand

[earnest grotto]

Please run it from inside the documents folder or some other place that does not need admin permissions. Like you did the first time.

NOT in system32.

Shift-right click the background of the folder you made it in
Open powershell/command prompt here
python ./Setup and then press tab to autocomplete
run it

[proud lodge]

i ran it here @earnest grotto

[earnest grotto]

You ran powershell/command prompt/whatever from inside system32, pointing it to the script

If you decided to launch command prompt with admin permissions, that opens in system32 by default

[proud lodge]

@earnest grotto Can you remote into my pc

[earnest grotto]

no

[earnest grotto]

.

If you're struggling this much running a command prompt, you should probably just use AI Playground instead. It's simpler.

https://game.intel.com/us/stories/introducing-ai-playground/

[proud lodge]

@earnest grotto I will keep trying thanks for the help

[earnest grotto]

Do this then ^

[proud lodge]

ok

@earnest grotto do i just wait?

[earnest grotto]

You did not press tab to autocomplete.
Press ctrl+z to exit

Pressing tab will autocomplete ./Setup to ./Setup_ComfyUI_Intel and save you a few seconds of typing and me the possible error of getting capitalization wrong or something

[proud lodge]

got it

@earnest grotto

[earnest grotto]

Just wait

[proud lodge]

ok

@earnest grotto

@earnest grotto thanks so much you're awesome

[reef ivy]

How is the new 'dynamic vram" stuff in comfyui? Does it even work on intel?

[earnest grotto]

it doesn't

same as comfy kitchen that added fp8 and nvfp4, without adding int8

(though comfy already did fp8)

[somber trellis]

Have any of you been able to get any higher than 640x480 with the distilled LTX 2.3?

[long karma]

With the conoda method for installing, when i try to launch with "python main.py --bf16-unet", it gives me a bunch of errors. Does anyone know how to fix this

[lunar thicket]

what is conoda method?

your error mentions comfy-aimdo. This is explicitly Nvidia GPU only. The error states this, but also the comfy-aimdo project page states this too.

It also makes references to CUDA; this is Nvidia only.

Assuming you have an Arc GPU since you are in this server, you should not try to run Nvidia-only tools. You should probably uninstall this conoda method whatever it is.

[reef ivy]

I believe he meant conda

[reef ivy]

Use viks script from one of the pinned comments. It will install everything you need to get going. You can also just install the ai playground which has comfy under the hood.

[reef ivy]

If i am not mistaken, you need that dynamic VRAM stuff in comfy, so if its not working you probably just oom.

https://github.com/Comfy-Org/ComfyUI/discussions/12699#discussioncomment-16036578 "we hope to support xpu as soon as possible." So there is hope

[somber trellis]

I pray

I pray for the comfy gods to give us 1920x1080 intel arc ltx 3 videos

[earnest grotto]

anima lora trained on screenshots of a remaster of a ps2 game (ty the tasmanian tiger 2)

[earnest grotto]

https://x.com/ZeevFarbman/status/2033928611632206219

Zeev Farbman (@ZeevFarbman)

Why Big Tech Is Abandoning Open Source (And Why We Are Doubling Down)

X · Mar 17, 2026, 3:29 PM

[reef ivy]

My conspiracy had been they bought up all the ram not because they needed it but to make consumer ai more expensive and closed source seem more lucrative with monthly fees

[wicked fulcrum]

Got LTX2.3 working from AI Playground install of ComfyUI. Works nicely on B60

[coarse whale]

Tested again today with nightly pythorch, still really high vram use compared to 2.9, as 2 months ago
#1193952640225267802 message

[somber trellis]

first pass 1280x720 2nd pass 640x480

not accurate ofc but still

🤷‍♂️

[earnest grotto]

those artifacts at the end that everyone gets with the distilled model
man

[somber trellis]

im bluescreening every few gens

ive taken an alternate method of first pass at 1280x720 and downscale to 640x480 for second pass

tried 1280x720 second pass with latent upscale but its hella slow at like 5-10 minutes per it for 20 seconds of video.

[earnest grotto]

sana video 2b

linear attention

[somber trellis]

huh

is it any good though

i see the hf page

[earnest grotto]

normally, attention is quadratic. that means 2x the data processed is 4x slower.
linear attention is linear, 2x the data is 2x slower.
linear attention models are dumber, however it should be way faster, and in either case they've done other stuff to make sure it's still a good video model and their results seem to surpass other video models so
in that vein, the constant memory kv cache should be helping performance a bunch too. i doubt comfy did kv caching up to this point for any of the other video models, if it didn't then this means even more speed, if it did then this means even less memory usage

[somber trellis]

if it competes with ltx 2.3 while being essentially 10x smaller

then yeah people will be hyped if its decent enough

[earnest grotto]

it competes with the other 5b, 2b etc. video models 😔

but it's better

[somber trellis]

now we just need a san 22b

I'm still just waiting for dynamic vram

I genuinely hope that's the thing we need to do higher resolutions with higher framerates

still interesting to mess with

[reef ivy]

They have a distilled lora you can use with the full model and see if it doesn't have those artififacts. Have you checked out Banadoco discord and workflows and stuff people there have posted?

[somber trellis]

it does

i already know because i first used dev + lora before i went to distilled

the same ending artifacts are there

[reef ivy]

You played with the lora weights?

[somber trellis]

Nope, just used the recommended.

[somber trellis]

What has to be changed to get 2.9 working again in ComfyUI?

figured it out

[reef ivy]

Try and lower the weights and see if you get better results.

[somber trellis]

I don't even have it anymore. I just swapped entirely over to distilled only, because it's faster.

I initially thought the latest driver whql was causing my problems, but actually the latest pytorch nightly seems to not work.

Latest driver + current stable or 2.9 works fine

[earnest grotto]

https://x.com/modelscope2022/status/2035652120729563290?s%3D12

ModelScope (@ModelScope2022)

The answer eveyone is waiting for is here: there will be more open Qwen models!🚀

In Today's ModelScope DevCon @Nanjing, Jingren made a public appearance and confirmed that Alibaba is committed to continuously open-sourcing new Qwen and Wan models.

🌟Stay tuned!👀

X · Mar 22, 2026, 9:37 AM

Say the line, Bart!

[lunar thicket]

Lmao stay tuned!

[lunar thicket]

https://x.com/soraofficialapp/status/2036532795984715896?s=20

Sora (@soraofficialapp)

We’re saying goodbye to Sora. To everyone who created with Sora, shared it, and built community around it: thank you. What you made with Sora mattered, and we know this news is disappointing.

We’ll share more soon, including timelines for the app and API and details on

X · Mar 24, 2026, 7:57 PM

adios

[earnest grotto]

putting the good in goodbye

[reef ivy]

gonna dedicate the resources to military propaganda I guess

[earnest grotto]

@somber trellis New TTS, this time by mistral, Voxtral tts
4b

https://huggingface.co/mistralai/Voxtral-4B-TTS-2603

[somber trellis]

saw it never tried it

it's non-commercial

🤷‍♂️

[somber trellis]

guess who this is

[earnest grotto]

petah don't clone the dog petah

nyeh nyeh nyeh nyeh

[earnest grotto]

false alarm, apparently voice cloning is api only

[somber trellis]

rip

still awaiting dynamic vram support

[rancid root]

Need help how to solve this issue

[reef ivy]

Delete all that and just use this #1193952640225267802 message

[rancid furnace]

Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned

[earnest grotto]

https://huggingface.co/CabalResearch/Mugen

Anzhc and Bluevoll's (not-so-)noobai(-anymore) Flux 2 VAE RF adaptation of SDXL is getting even better

[lunar thicket]

What is an RF adaptation?

[earnest grotto]

The Flux 2 VAE is both higher quality than and more learnable for models than the SDXL, Flux 1, or (presumably) Qwen VAE (in particular, the SDXL VAE is much more learnable than Flux 1)
You can assume Rectified Flow is like a far better version of v-prediction - without the overcooked colors, while still not being limited to grey images like epsilon. They're reformulations of what exactly the model predicts (e.g. predicting the added noise vs predicting the clean image vs ...) and it turns out that makes quite a lot of difference
Training the model to learn a new VAE / new TE or any other bigger architectural change, like going from vpred/epspred to rectified flow, is adapting the model to that. takes a while

Here's BFL's graph of VAEs. I've heard Flux 1 is only marginally better than doing pixel-space diffusion

They just took Noobai and changed the VAE and trained it on rf instead of vpred/epspred (don't remember which noobai version they chose)

It's almost done learning. At this point it's getting pretty coherent

Anima is still more coherent despite it itself being decently undertrained on even 1024x1024 but it's still good stuff

due to rf, it can make nice good dark images without dying like noobai vpred often does

or various other specific compositions. I think eps models also struggle with image borders, which is sad because I like outside border (AKA the 3d border effect)

[earnest grotto]

Here's an example with an earlier version of mugen
IIRC, these are respectively: Noobai epspred, vpred, vpred with EQ VAE, Neta Lumina, earlier version of mugen (noobai rf flux2 vae 0.3), anima preview 1 (we are on preview 2)
likely due to natural language, neta and anima expect a bit more prompting to make the whole image dark but they very much can do it

you can see how epspred just flops like crazy

and vpred did do fine, this time. it's just that sometimes it doesn't do fine

second image is the first preview of anima, with more prompting

[lunar thicket]

"flops like crazy" lol I actually can not. Not sure what that means.

I mean I see different styles in all these. Some look a little wonky, like image #3 in particular, and #5 has some crazy going on

do you mean the difference in art style between the background and subject in #1? cause that is pretty clear

[earnest grotto]

It is not dark

It is supposed to be dark, in particular like #2, 3 and 5

lemme fetch the prompt

masterpiece, best quality, 1girl, solo, emilia \(re:zero\), white hair, long hair, purple eyes, white dress, detached sleeves, white thigh high boots, sitting against tree, outdoors, grass, tree, dark, black sky, chromatic aberration

[lunar thicket]

ohh the darkness of the image is all you are trying to show with these examples? Lol ok

[earnest grotto]

it is not a stylistic difference. there is nothing you can do prompt-wise to get noobai epspred or other ones like it to spit out a truly dark image

you can vae encode a pitch black image and denoise at 1.0 and will get a dark image

[lunar thicket]

now that i know chromatic aberration was part of your prompt i retract my comment on #5 lol

[earnest grotto]

top have VAEs that are too lossy to do visible chromatic abberration. bottom have VAEs that can do it

#5 has a very sketchy appearance because the previous version of mugen was very undertrained. undertrained models look like that, they're noisy

newer one is still slightly noisy, still has artifacts but is almost acceptable now

I'm training some loras to see how well it takes to them

[earnest grotto]

lr may have been a bit too high. In either case the model is still a bit too artifacty for my taste
noobai vpred eq vae
mugen
anima preview 2
I like that noobai does the eyes correctly consistently. that being said, it's exacerbating what most of the dataset contains which I'm not quite sure is a positive (same can be said for the facial expression that mugen does). Though anima does do the eyebrows correctly quite a bit. Anima might be too undertrained on 1024 to know that eyes can have detailed coloration, as it notably struggles with other small details like mikoto's tokiwadai emblem or beatrice's butterfly eyes or other stuff

otoh anima has the capacity to properly learn when fan artists screw up the long lock of hair

[earnest grotto]

nevermind, looks like anima's eyes issue was just a dataset problem from me keeping stuff around from pv1 to make the model not colossally forget everything

[reef ivy]

Hey guys, what is the recommended pytorch version now? Still issues with latest?

[earnest grotto]

I think 2.11 nightly still increased my VRAM usage when I tested it out training the middle lora here, compared to 2.9

[lunar thicket]

I think im on some 2.11 nightly still and it's been fine. I have not done any compare/contrast to see how it has affected my vram usage

[elfin current]

Hi, I'm having problem installing ComfyUI due to ModuleNotFoundError: No module named 'sqlalchemy'. Can't seems to solve it on my own. Anyone knows how to fix this?

[reef ivy]

Check the pins and scroll to viks script, easiest way to install outside of using ai playground

[elfin current]

Did that but now I got "comfy-aimdo failed to load: Could not find module" when I tried to run it. Am I doing it wrong?

[earnest grotto]

There is nothing you can do about that warning, other than use comfy normally and forget about it because it doesn't break anything and comfy will work fine.

[elfin current]

So am I supposed to wait for the rest of it to run? Cause currently it's just stuck at that message and nothing seems to be happening.

[earnest grotto]

If you are using an HDD, you should consider an SSD.

[elfin current]

It's actually in my M.2.

I left it for more than 30mins and still nothing lol. Maybe I should go back to SD.Next, but I couldn't get Zimage to work there

[earnest grotto]

Are you on windows? I might test it later then

Run the script again, it could be some broken commit comfy had that got reverted/fixed

You don't need to change the location or delete anything

Running it from the same location will update faster

[river crater]

Its weird, I can physically hear something (I guess the fans turning?) whenever comfy runs its queue

[river crater]

Actually it can't be the fans running

it sounds like a dvd being read from, or something moving

[lunar thicket]

coil whine? DVD being read gives that impression vaguely

[river crater]

is that what that sound is? I think it feels right

Mine ASrock challenger dual fans config. and from internet searhces, yeah seems like it is a coil whine issue. This is prob the first time ive ever heard of this

[lunar thicket]

Is your case on your desk?

I have that card (asrock challenger b580) and when I was testing it before closing it up, I could hear some faint coil whine. But once I closed it, and especially now that its under my desk, I don't hear it

anyways it is normal and fairly common

[river crater]

Yeah, on my desk

but its noticable even 3 meters away from it

[lunar thicket]

hmm sounds pretty loud

That's gonna be a "is this too annoying for me to stand?" judgement call

[river crater]

its not to the point of where Im annoyed, but it gets very noticable during comfyui runs

its more of "Oh damn, is my GPU actually dying from this?" kind of reaction

[earnest grotto]

Pretty late but, I tested again and I can't reproduce whatever issue

@river crater You're sure it's not windows scanning it or whatever?

[honest hull]

Portable ComfyUI zip for Intel Arc Graphics and Core Ultra series 1/2/3 processors with integrated Intel Arc graphics is now avaiable https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.19.0

As well as integrated in ComfyUI Desktop 2.0 https://github.com/Comfy-Org/ComfyUI-Desktop-2.0-Beta

[earnest grotto]

You should probably pin this

[honest hull]

@earnest grotto you might want to update your script so that it can pull the standalone build directly, then apply your optimizations 😉 the setup should be much easier now

https://github.com/Comfy-Org/ComfyUI-Standalone-Environments/releases standalone ComfyUI can be found here, or you could use the portable one

[earnest grotto]

I'm not sure, I'll check it out. I generally don't use Comfy Desktop and there aren't really much optimizations that my script does, it just installs disty's hijacks, patches some issues that arise in comfy when using them, and patches a few custom nodes which I recently decided are redundant or too outdated now

Base comfy doesn't need disty's hijacks, they're mostly for custom nodes, and I'm not sure how much people install those with comfy desktop, though with how easy it is to install some maybe they do

[honest hull]

I actually really like ComfyUI Desktop, it's an electron app. If I choose one of the workflows from the templates it prompts me that models are missing, but with one click it downloads all the required models and put in the proper directory for you.

[lunar thicket]

Interesting news

[earnest grotto]

The prompt to download them exists on the web version though I can't convince it to pop up right now to check if it can put them in the right folders or just downloads them into the browser's downloads folder

It's more likely it just puts them into the downloads folder

[lunar thicket]

shame the portable is Windows only

still good to see though

[earnest grotto]

Comfy wants to drum up hype https://comfy.org/countdown

Another 16-32B video model, realism only, that takes like 20 minutes per gen and barely runs on my hardware?

[reef ivy]

Since it's comfy, probably some API stuff's. Makes me think I should wait to try it out again as everything will likely break after the update lol.

Maybe the vram optimizations for intel arc? 🥸 lol

[slender zodiac]

Cloud based computation to offload your data from you.

I mean, to offload your processing

[earnest grotto]

https://blog.comfy.org/p/comfyui-raises-30m-to-scale-open

So apparently the countdown was for this...?

Not too bad, that would depend on how they spend it, but the countdown was a bit unwarranted IMO

[reef ivy]

well, that was definitely an announcement of all time. It's good they are doing good, but not sure why they had a countdown to tell us they are doing good lol.

[slender nymph]

Hi people!

Returning ComfyUI user having some issues, hoping to ask if i missed something 🙂

I tired to install comfy again recently and getting this error, even with Comfy Desktop.
Had it working.... one or two months ago? But had to formatt the M.2 it was on.

If any other kind soul want to give tip to what i might have missed is greatly appreciated 🙂

• Redownload script and try again
• Deleted folder and tried again
• Uninstalled minicoda from my USER/minicoda3 folder and treid again
• Tried Comy Desktop, both 0.2 and 0.18, same thing.
• Tried on another drive, same thing.

The install says it is sucessful, could it be something with either latest build of comfy or intel driver something messed with?

[earnest grotto]

Which pytorch version

Which GPU

[slender nymph]

PS C:\Users\marcu> pip list | findstr "torch"
pytorch-triton-xpu 3.5.0
torch 2.9.1+xpu
torchaudio 2.9.1+xpu
torchvision 0.24.1+xpu

B580

have i forgotten to update pytorch? -,-

[earnest grotto]

run my script again from the same location, no need to delete or move anything around, pick stable or nightly instead of 2.9, it will install faster, try again, say what happens

[slender nymph]

Usually run Stable when i do these install but ill test Nightly, one sec

i got that it completed but still same issue unfortunatly

[earnest grotto]

what driver version

hmm

[slender nymph]

Should be the latest i think?

[earnest grotto]

You say the issue also happens with the Comfy desktop package?
If yes, would you like to try an older driver like 8250? https://www.intel.com/content/www/us/en/download/785597/869290/intel-arc-graphics-windows.html

[slender nymph]

Looks to be pretty similar on Comfy desktop yeah, alot of fluff about extra search path but the error looks the same i think.

yeah could probably try on a older driver, long time since i downgraded graphics driver, know if i need to do any specific other than just run the installer?

[earnest grotto]

Nothing special is necessary, you can just run the installer as-is

[slender nymph]

Well same issue unfortunatly, running the script again to see what happens after that

Nah same issue still, might look into doing a complete reinstall of Conda and Python later but that is gonna be a later thing to try.

Thank you for the assist so far! 🙂

[earnest grotto]

Don't do those things
in the conda environment, a specific python version (3.10) is installed. your OS python doesn't matter

are you sure there is enough space on Q:

[slender nymph]

Yeah should be 😄

[earnest grotto]

hmm

wanna try it in documents in c

[slender nymph]

I did try it on my C:/USER but i can toss in the documents folder, sure

[earnest grotto]

nevermind then

stick with the old driver. i'm thinking

[slender nymph]

Installing the oldest version i can of Comfy desktop to test see if that runs.

[earnest grotto]

run comfy from the shortcut like this again, wait for it to crash
type in (or paste, but not with ctrl+v, use ctrl+shift+v)
python -c "import torch. print(torch.__version_\_)"
and press enter
show what it says

[slender nymph]

seems to be in the cenv but get syntax error

(Q:\AI\Comfy_Intel\cenv) Q:\AI\Comfy_Intel\ComfyUI>python -c "import torch. print(torch._version_)"
File "<string>", line 1 import torch. print(torch._version_)
^
SyntaxError: invalid syntax

[earnest grotto]

oops

python -c "import torch; print(torch.__version__)"

replace . with ;

[slender nymph]

Diffrent message:

(Q:\AI\Comfy_Intel\cenv) Q:\AI\Comfy_Intel\ComfyUI>python -c "import torch; print(torch.version)"
Traceback (most recent call last):
File "<string>", line 1, in <module>
File "Q:\AI\Comfy_Intel\cenv\lib\site-packages\torch_init_.py", line 287, in <module>
load_dll_libraries()
File "Q:\AI\Comfy_Intel\cenv\lib\site-packages\torch_init.py", line 270, in _load_dll_libraries
raise err
OSError: [WinError 127] The specified procedure could not be found. Error loading "Q:\AI\Comfy_Intel\cenv\lib\site-packages\torch\lib\c10_xpu.dll" or one of its dependencies.

[earnest grotto]

go into Q:\AI\Comfy_Intel\cenv\lib\site-packages\torch\lib\, does c10_xpu.dll actually exist

[slender nymph]

'Looks like it

[earnest grotto]

conda install libuv
then try the above command again. we'll be using it to see if it's fixed

[slender nymph]

sorry for possible stupied question but i am tired now xD

Presuming while still in cenv?

[earnest grotto]

yes

just keep that one open

without the \, discord is being annoying

[slender nymph]

yeah finnished, run the lnk from the command line?

[earnest grotto]

well, that tells me nothing will change so I'll keep thinking of what else to add

Wanna try installing the latest MSVC redistributable as well? https://learn.microsoft.com/en-us/cpp/windows/latest-supported-vc-redist?view=msvc-170

You are on x64.

[slender nymph]

sure, seems i have it installed but ill try repair 😄

Same unfortunatly :/

[earnest grotto]

do
pip list | findstr intel
and
pip list | findstr oneapi
and drop the results here

[slender nymph]

the oneapi seem empty

(Q:\AI\Comfy_Intel\cenv) Q:\AI\Comfy_Intel\ComfyUI>pip list | findstr intel
intel-cmplr-lib-rt 2025.3.2
intel-cmplr-lib-ur 2025.3.2
intel-cmplr-lic-rt 2025.3.2
intel-opencl-rt 2025.3.2
intel-openmp 2025.3.2
intel-pti 0.16.0
intel-sycl-rt 2025.3.2

(Q:\AI\Comfy_Intel\cenv) Q:\AI\Comfy_Intel\ComfyUI>pip list | findstr oneapi

(Q:\AI\Comfy_Intel\cenv) Q:\AI\Comfy_Intel\ComfyUI>pip list | findstr oneapi

(Q:\AI\Comfy_Intel\cenv) Q:\AI\Comfy_Intel\ComfyUI>

[earnest grotto]

yeah ok it's normal for there to be nothing oneapi

[earnest grotto]

just do pip list and paste everything here

[slender nymph]

Here you go

[earnest grotto]

thanks

actually, better idea
do you wanna put the cenv in an archive and upload it somewhere and send it to me

Q:\AI\Comfy_Intel\cenv

This folder

[slender nymph]

Sure ill zip it and send you a message, about to pass out so you'll have a message tomorrow though xD

Thanks for the help so far!

Edit: assuming Mega should be fine?

[earnest grotto]

yeah

alright, tomorrow

[slender nymph]

Hey @earnest grotto , sent you a direkt message with the env zipped on mega 🙂

[earnest grotto]

I'll check it out in a bit

[slender nymph]

its working?

[earnest grotto]

I'll check it out in 2 hours.

My expectation is, it will work for me and I'll have to figure out what is wrong on your end outside the conda environment

At that point, it might be a good idea to also try AI Playground, which will likely also fail, and so you can ask other peeps for help as well as the issue is likely broader than just some specific comfy install. But that's just my guess right now

Do you have an iGPU
I doubt that could be causing issues but still

In an absolute worst case scenario, installing the oneapi base toolkit (~40gb IIRC) and shoving activating it into the script to launch comfy might solve the issue, but the base toolkit should not be needed. it's just an issue that was happening when it was not activated and I suspect it could be similar

[slender nymph]

i would not be suprised if it works for you 😂

yeah ill go ahead and try that.

Should not have iGPU, doublechecked performance tab and does not indicate it.

Think there could be anyting that needs to be cleaned from the appdata folders if there is something old there causing issues?

[earnest grotto]

you should not need to be touching anything else

it's supposed to just work

[earnest grotto]

yeah ill go ahead and try that.
You mean AI Playground? Do try that

If you mean the basekit, don't try that, not before you see that AIPG also doesn't work, you ask in its support thread at least https://discord.com/channels/554824368740630529/1265055106928476192 and you are desperate to find a way to make things work anyways

[slender nymph]

Yeah meant AI playground 😛

But hey, error keeps coming flip table 😂

[earnest grotto]

Post this in https://discord.com/channels/554824368740630529/1265055106928476192

[slender nymph]

I kinda went rouge but hey it worked xD

ill write what i did longer message

[earnest grotto]

anima is better than whatever sdxl anime finetune

though torch compile is iffy on windows and it helps anima's speed a lot

at least there is an official turbo lora now

[slender nymph]

haha you just made me Jameson laugh that you first point that out directly xD

But that was the SDXL model i used last when i used this (its a old one that i liked the effect of when i started with AI... 2 years ago?)

What i did to try and fix it:

Uninstalled these both from system settings:
Python
Conda

Went manually to my user folder, deleted miniconda3 and .conda

went into Appdata.

local:
deleted folders realted to python + conda that was left behind, if i saw any Comfy folders they also got trashed.

Did not have anything related to any of these in locallow

Went into Roaming and deleted what i could find with python, conda, anaconda or miniconda and ComfyUI.

after that was deleted, did a reboot on my system, made a temp folder in q:/ai and put the script in there, ran it and worked perfect

[earnest grotto]

hmm, alright then, strange

oh well. get anima

i am, after all, training an anima lora right now

[earnest grotto]

.

[slender nymph]

haha on my way, any recommendation or should i stick with the offical first and then check out trained checkpoints?

mostly gonna do DnD European horror stuff 🤔

[earnest grotto]

I don't use random aesthetic finetunes. If you can't come up with an artist that's close to the style you want, search for a lora

https://huggingface.co/circlestone-labs/Anima/tree/main/split_files

[slender nymph]

on its way 😉

[earnest grotto]

don't forget the text encoder and vae

anima uses gelbooru tag names, in case an artist is renamed on dan. which is great as dan is a pain to browse anyways

artist names are prefixed with @

you can't mix artists as well as with sdxl (clip-based) models

prompt weighting needs much higher numbers than you'd expect. 2.0 for anima is about equivalent to 1.35 or so

it understands natural language (you can say "x on left y on right"), it has a 16ch VAE so it's much sharper, and it has way better knowledge of tags with really low amounts of images

[slender nymph]

oooooh that is nice, my previous experiance is with Forge with mostyl SDXL, did get good results but the results got funky when trying to explain positions

[earnest grotto]

you will likely want to open up danbooru's wiki on poses instead of thinking of a way to explain them in natural language

[slender nymph]

Prompt executed in 28.83 seconds

Second try with the "test" prompt that is in Comfy, that is not a bad result

[earnest grotto]

Anima is decently slower than SDXL. However at the same time, torch compile speeds it up almost 2x. That being said, even without torch compile, if you didn't choose to install a bunch of custom nodes with my script, get this https://github.com/ruwwww/comfyui-spectrum-sdxl
If you did choose to install custom nodes, it should be included

Use these settings. Stop_caching_step is your number of steps - 3 or so, e.g. for 35 it's 32, 33 is likely fine too though.

It says SDXL, it works for anima.

And this is the official turbo lora. 1.0 cfg, 8-16 steps. I don't like it very much outside of chibis https://civitai.com/models/2560840/anima-turbo-lora

Don't use both spectrum and the turbo lora, doesn't work well

[slender nymph]

oh that looks intressting, i did skip all the extra nodes this time to have a slightly faster install, if i want to install them i should just rerun the script and select i want them right?

[earnest grotto]

Yes

You can also change your pytorch version this way. Newer pytorches increased VRAM usage for some reason, no idea why. I don't know if the nightly fixed this yet but it should still be an issue with stable, and not with 2.9

specific pytorch versions also might work better or worse with compile, however I couldn't get compile working on windows either way. I am on linux right now

[slender nymph]

Just a quick image and i do like the result straight out the gate.

probably gonna take some getting used to prompting a new model since i was used to SDXL back in the day but yay it works 🙂

[earnest grotto]

tag-based prompting still works. anima is specifically trained on both

you can also mix them up

I guess, you were likely using epspred models. anima can make properly dark images so you don't need to fight it in prompting to have that

there's also a few extra custom tags, like most models do, and that's about it

all the custom tags there are, are the usual masterpiece, best quality, ... human score tags, pony aesthetic tags score_9, score_8, ... which are the only ones with underscores, year 2025 and so on, and newest, recent, mid, early, old

also devinatart/ye-pop images are prefixed with deviantart and a new line / ye-pop and a new line

[slender nymph]

That i honestly could not tell you if i used epspred models, not that experienced and had a long hiatus with this when i got my kid 2 years ago 😂

i did miss when pony got off (NO PUN INTENDED!) but yeah i do remember the masterpiece and things.

do you know if anima is good with "ugly" characters? i know there was trouble getting them to appear disfgured earlier, might still be an issue?

[earnest grotto]

noobai vpred came out about a year ago. before that all the anime models were epspred

[earnest grotto]

do you have any reference images for how you expect that to look

ugly, "ugly", and disfigured, are pretty different

[slender nymph]

fair 😂

no not currently any references, might come up when im prepping for my DND game but then i guess ill just have to try it out 😂

[earnest grotto]

if you're not sure how to prompt danbooru-tagged models, especially given you were talking about coming up with descriptions for poses,
Browse danbooru's tag wiki https://danbooru.donmai.us/wiki_pages/tag_group%3Aposture
While browsing gelbooru https://gelbooru.com for the actual images. danbooru forbids searching for >2 tags at the same time and searching for a few select tags

[slender nymph]

will bookmark these, cheers!

Oh yeah these sites, i member!

[earnest grotto]

gel/dan need you to search with underscores, but do not use underscores when prompting (outside of score_)

[lunar thicket]

Doesn't prompting register an underscore equivalent to a space?

[earnest grotto]

No

They are different tokens

If you prompt an SDXL (CLIP) model with underscores, it too will perform worse

Just only slightly worse, rather than much worse like anima

similarly, anima is also more sensitive to out-of-order character/series/artist names and to typos than sdxl/clip

[lunar thicket]

What are the differences between tokens?

[earnest grotto]

Different tokens mean different things to the model

[lunar thicket]

Like if I did green_eyes instead of green eyes, SDXL doesn't equate those things?