🤝｜tech-support

[dusty sand]

We used the same images.

With the same seed.

[maiden coral]

I came back after a year, so idk what the stability matrix is, only heard about it yesterday, but ComfyUI just works.

[dusty sand]

Segmented Adetailer gets ignored

BBox is less detailed.

[maiden coral]

Alright, so if it updated or something then the self-contained environment of Forge might have some chan. Check python it used before and after. Also which version of Pytorch. Are you using Nvidia or AMD GPU? I forgot from the conversation.

[dusty sand]

It wasn't updated.

At least not manually.

My buddys Forge was untouched.

[maiden coral]

Could you give me the link on github repository of the forge you use?

[dusty sand]

https://github.com/lllyasviel/stable-diffusion-webui-forge?tab=readme-ov-file

or Stability Matrix

https://github.com/LykosAI/StabilityMatrix

[maiden coral]

I mean which one you use

[dusty sand]

2.3.1

The recommended one

on SM I run pytorch version: 2.9.0+cu128

[maiden coral]

And what's your exact GPU model? E.g. GTX 1060

[dusty sand]

5080

My friend has the same issue with a 2060

[maiden coral]

I can't view the archive, the github repository doesn't have the python embedded environment, so I guess it's in the archive of Ilyasviel version. Just check the compatibility with your CUDA and Pytorch version.

Also there is this, I'm not sure if it means auto updates every 90 days or whatever:
"Forge is currently based on SD-WebUI 1.10.1 at this commit. (Because original SD-WebUI is almost static now, Forge will sync with original WebUI every 90 days, or when important fixes.)"

[dusty sand]

Doesn't seem like Forge has gotten an update in months.

Unless I am missing something.

[maiden coral]

The only ideas I have for now (I'm on phone):

1. Check VENV_DIR parameter.
2. Check if you have Python 3.10 in the environment Forge uses python --version, never version won't work (since it's a fork and I pulled out from 1 year ago they won't change that) unless it changed, also check you cmd and see if there are any errors / fails

And recheck the pytorch version too:
python -c "import torch; print(torch.cuda.is_available())"

That's all from me for now.

Yeah, don't forget to actually go into the environment, id you just use python--version it will show what python u have on your PC but not in the environment forge usees, so in Forge folder there should be the environment if you have it, go there and open cmd from there and check the versions and if they fit

[ornate elk]

Do you have an example to test?

I will do some tests later

[dusty sand]

You mean which one I used?

I tried a few.

I can DM you a civit link if that's okay?

[ornate elk]

Yea sure

[dusty sand]

Sent you a friend request.

Can't DM you otherwise,

[primal root]

Does anyone know where I can download the sd3.5 model?

[vocal burrow]

https://huggingface.co/stabilityai/models

[cosmic furnace]

mb i shud run some torch specific install commands ? 🤔

Pretty sure i didnt had theese errors on same forge on lower hip sdk version & probab lil bit lower python

[maiden coral]

I think it needs Python 3.10

anything higher won't work since it's a fork of the old Automatic111

[ornate elk]

3.11.9 works fine but 3.12 not it seems

[maiden coral]

yeah, ComfyUI uses 3.13 in it's embedded environment

strange that Forge hasn't updated yet

[ornate elk]

yea comfyui is an exception

nothing else uses the latest python or recommends it

[maiden coral]

latest is 3.14

[ornate elk]

that got released 2 weeks ago xD

[maiden coral]

yup, also is there any good guide for the nodes that'd free up VRAM? I'd like to try a lightning model (like SDXL LCM models) but with WAN

but I haven't found a concrete video except one where is Purge VRAM node that I have, my VRAM is just very low for these models but still want to give it a try and at least make it load

[ornate elk]

hmm ,i dont know of any guides

whats your vram size?

[maiden coral]

GTX 1060 6 GB VRAM, I tried Wan 2.1 Camera Control Model, it works just fine
Wan 2.2 GGUF model 2-bit (the lowest), but it kinda pales since I only used 4 steps (the default settings of the comfy UI template workflow), at least 20 steps is necessary

also wondering if high and low noise model is necessary, everything is kinda new to me, so I'm trying to get sources for these things

[ornate elk]

you need boath models

high as model and low as refiner if i remember correctly

wouldnt even care about video gen or wan on a 1060 for real xD

even on high end gpus it takes very long for a simple lowres 5 second video

[prisma moss]

so one of my detailers decided not to detail anymore, wtf is with that
like it detects it, and does the 'steps' but makes no changes

[ornate elk]

yea that issue is since a few days, if the detail model is a segmentation one

[prisma moss]

yea it is

how dfq does that happen

is it forges fault?

since ive got a fork of adetailer

[ornate elk]

ive done a few tests and it seems like its a torch version issue

[prisma moss]

luckly i have the option to use bbox instead of segm

this bypasses the issue, good to know im not theo nly one

[ornate elk]

yea bbox works without issues

I did some tests on my old Nvidia rig (GTX1080) and there Segmentation models work with:
✅Auto1111 with Torch 2.1.2+cu121
✅ Forge on Torch 2.3.1+cu121
✅ReForge on Torch 2.7.1+cu128
Works not on:
❌ Forge Classic on Torch 2.8.0+cu128

On my AMD PC it doesn't work with Torch 2.7.0 and 2.9.0
Everything tested on Python 3.11.9

[prisma moss]

so why did it suddently happen

like, why did torch update or what

how did it update in the first place

btw im still on 2.3.1

[ornate elk]

idk as the old forge didnt got any update aswell as adetailer

[prisma moss]

and it doesnt work on segm

[ornate elk]

did you got any nvidia update in the last 5 days?

[prisma moss]

define nvidia update

driver, no

everything else, maybe, since i was installing forge neo, idk if it pulled some shit system wide

[ornate elk]

forge wouldnt use it anyway as it has its own files

[prisma moss]

figured

its probably a combination of gpus as well

since im on rtx 40xx

maybe it works with grx 10x

who knows

[ornate elk]

yea for now no one knows what it caused

Reddit post with more users that have the same issue:
https://www.reddit.com/r/StableDiffusion/comments/1oc3410/i_reinstalled_forge_and_reforge_and_now/

[tepid crypt]

Someone knows if there's any extension or tool for SD Forge, or similar, that allows to use an image of character as reference like in NovelAI? I tried inpaint or img2img but it always looks akward, no matter if try to change the pose or style or anything. Also tried kontext, but my pc can't handle it.
(ignore the part of the Style Aware, just want to use a single image of a character as reference)

[prisma moss]

i believe ipadapters are used for that

[storm knoll]

Hey I'm using Comfy and figured out that having WAS nodepack active breaks all the colors of my nodes and turns them blank
Is anyone aware of a fix for this?

[lucid yoke]

Sounds like the node pack is broken, id use alternatives

Or if you have multiple node packs it could be a incompatibility

[dusty sand]

Oh another one. Welcome to the club. 3 of my friends have it. So do some people on reddit. No fucking clue. I am more limited in my tests since I use a fucking 5080.

So I am pretty sure I need cu128+

[cosmic furnace]

#🤝｜tech-support message - btw theres too many files. which one i need for qwen clip ?)

[ornate elk]

Best would be the same as the gguf model.
I used Qwen Q3_K_M

[ornate elk]

I used the one from here:
https://huggingface.co/city96/Qwen-Image-gguf/tree/main

[cosmic furnace]

mb wrong torch or python version ? yet again this error. about same time + - running sd . not sure

on prev i have different size of this file c10.dll . but i also have lower hipsdk & python there 🤔

[ornate elk]

What's your hip SDK version?

[cosmic furnace]

6.4)

[maiden coral]

yeah, 43 min with video control, depends on the promtp and if I use PC, otherwise usually 1 hour I think, I also found out that people use something akin to LCM SDXL loras and use high or low denoise model (what I understood form my findings is that the high model is better at high denoise and low at low denoise levels) that's also why you se steps to half (half of the picture in each frame is processed with the high denoise model and low with lower)

not denoise model but Wan or whatever checkpoint you use*

[maiden coral]

What's your:
GPU
pytorch version including CUDA?

[cosmic furnace]

dont know cuda. its amd with zluda 7900. pretty sure shud be installed what fits . problem is that i might have run install rocm for this sd. or smthing like that. forgot . w/e >_>

[maiden coral]

I'm not familiar with Zluda, sadly, maybe CS1o or someone else would know.

[ornate elk]

Have you tried running forge neo with rocm over an hour?

[cosmic furnace]

no 🤨

and btw what extension had soft inpaint in it ?)

[sick zephyr]

我想本地部署SD3模型，企业部署如何定价呢

[fair oxide]

that's not english, I think

[sly pike]

hello, can I use workflows in sd forge neo?

[vocal burrow]

workflows ? what do you mean by that ?

[prisma moss]

no such thing as a workflow in forge

[prisma moss]

just use normal inpaint, its better

[candid walrus]

Can anyone help me? I'm trying to generate x anime character but as if they were in an IRL photo and I can't seem to do it at all. I need help.

[peak agate]

Work communication

[candid walrus]

Please if anyone can help me ping me or something idk if its possible?

[prisma moss]

you need a realistic model and the character lora

[candid walrus]

The example images for all the loras of the characters I want to use are all anime art style though so that mean I can't make it how I want?

[prisma moss]

not necessarily

[candid walrus]

OK well I'd love to chat but I have class in the morning and should have gone to bed 2 hours ago if I'm to get 8 hours of sleep. Talk to you tomorrow and see what can be done about this?

[lucid yoke]

Try using a realistic illustrious finetune, then most likely the character you want is already in the dataset

[candid walrus]

What I'm confused? Sorry I've only done 3 image generations and a couple hours of troubleshooting install with GPT because I was told to ask it to help setup but its way easier to just watch a YT video.

I don't really know what is what or how to do what yet.

[alpine comet]

whats going on?

text form:

[lucid yoke]

Gpt installs will always fail, im home later and show you some examples

[candid walrus]

OK well I didn't know that, anyways its installed now so when we both get home later can you help me get image generation going?

[ornate elk]

What's your GPU?

[alpine comet]

2070

[ornate elk]

Okay then reinstall the webui using the guide of the pinned messages

[candid walrus]

?

[ornate elk]

anime character lora + realistic illustrious model

[lucid yoke]

and im not home yet, would help you when i was

[candid walrus]

I tried a lora of character and a model that does real life stuff the lora never showed up in the ora tab and I tried to write it in the prompt to force it and it didn't do that. I could be doing something wrong I don't trust GPT anymore (my help guide) but it said that's because most loras of anime girls are not compatible with real life models and I only saw the two loras and they both did work.

[ornate elk]

have you tried illustrious based loras and models?

[candid walrus]

Thanks so much for trying to help when do you get home?

Also I stupidly upgraded from my 4070 to 9070xt (it was less then $100 more then I sold my 4070 for after tax so I thought good upgrade deal) and 64gb system memory not 34.

Anyways its installed on my laptop and that what I should be using or should I use the desktop?

[candid walrus]

I tried what GPT told me when it didn't work and a third to have the time it tried to tell me install tips that didn't work till I went on YT I was stupid. Idk what I've been doing.

Starting a class test now have to go talk in 2 hours.

[ornate elk]

alright, and yea better dont listen to chatgpt... it doesnt have any clue about image generation webuis

[ornate elk]

what gpu does oyur laptop have?
the 9070xt can work with SD, you just need to follow my AMD setup guides from the pinnes messages

[lucid yoke]

just now lol

[candid walrus]

4060 did I not mention?

[candid walrus]

Extra vram helps but so does cuda does cuda help more then double the vram here or is the desktop better? It has 32gb of ram but I've been thinking of putting in 64 but only if it will help for AI. I think I upgraded wrong going from Nvidia to AMD here so my desktop might not get much use and downgrading 4gb of vram to an all around slower GPU on my laptop is not good. (Unless I have a 2070 and can somehow use that acceleration with the 9070xt Vram or something don't think I can but wouldn't that be nice).

[candid walrus]

Oh I'll be home in max an hour sorry in class but I can use my laptop a little I guess? Don't have the high performance charger just the 100w USB C one so it is slow plus people taking tests I can't just ram the fans. (It unfortunately does not except more or less from USB C)

[lucid yoke]

No worries i got time, no rush

[candid walrus]

Thanks first question (driving home now) is do I use my laptop still like GPT said or dies the extra vram make up for no cuda? I'll ping when home but after desktop setup?

[lucid yoke]

honestly i never worked with amd so why not give it a try?

at worst you lose like what? 10 min

setting it up

[candid walrus]

Plus the worse time to generate and the lower quality output according to gpt but yeah.

[lucid yoke]

not lower quality per se

refining and upscaling needs some tweaks but CS10 is king on amd genning

someone i knew genned with 6gb vram and made pretty good stuff, it just offloads

[lucid yoke]

dont make me tap the sign:

Responses from Large Language Models like ChatGPT, Claude, or Gemini are not facts.

They’re predicting what words are most likely to come next in a sequence.

They can produce convincing-sounding information, but that information may not be accurate or reliable.

[candid walrus]

Gpt said worse and slower output, I can try it I always assumed more vram more better but got said yes but cuda even more better. Idk I'll try when I get home and compare output.

[lucid yoke]

seeing you mention gpt often,
other then prompts for the newer models (like flux, sd3.5 etc) do not listen to chatgpt at all for anything image generation related stuff

every other day someone comes in with tons of errors because they listen to gpt

[paper dagger]

Use CS1o's pinned guides if you can. It should help.

[ornate elk]

The output isnt defined by the hardware you use. Its defined by the model and resolution etc

[candid walrus]

Yeah said something about I can only run worse models on amd idk? I don't trust GPT anymore after how much it butchered the setup process.

[ornate elk]

thats totally wrong lol xD i can run any model that is currently out

you can easily create images like this with either your nvidia or your amd gpu

[celest grotto]

Hello, I have encountered some problems that I need help with. My SD Automatic 1111 was doing fine then suddenly, it doesn't work anymore. Every time I tried to generate an image. It will always reach 95% then crashed my browser. The browser just shows "aw, shuck" or "aw, snap".

Everything was doing very well before

[peak agate]

Troubleshooting

[celest grotto]

How do you do that? I'm sorry if this is a stupid question

[ornate elk]

Hey, what's your GPU and what's in your webui-user.bat?

[cosmic furnace]

Dyu think the rock gonna be updated soon ? 🤔

[ornate elk]

it gets updated daily

you can see the last commit was 18 hours ago:
https://github.com/ROCm/TheRock/commits/main/

[cosmic furnace]

is it worth trying or very small fixes?

[ornate elk]

small fixes and stuff, i always switch to the latest to test

[cosmic furnace]

how hard it is to update it ? 🤔

[ornate elk]

not hard, its just using these 3 commands inside the cmd of the webuis folder:
venv\Scripts\activate.bat
pip uninstall torch torchvision torchaudio -y
pip install --index-url https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/ torch torchvision

it will always pull the latest working version

currently its from 21.10.25

Forge neo also gets updated daily

[cosmic furnace]

same problems , stutterings when optimizing, each res change results in new optimizing 🙁

have u tried ultimate sd upscale extension btw? seems like it can help to create higher res images faster on flux for example. because tiling image or smthing 🤔

mmkay, when trying to run qwen i get insta error 🙁

hmmm, on ssd it surely runs better 🤨

switching between txt2img & img2img doesnt seem like making delay & new optimization. Or its very fast 👀

trying to run qwen still makes insta error tho 😑

[candid walrus]

So which card do I use then? I know cuda does help like never say the same vram per card amd could be better but the AMD card does have more vram.

I'm just wondering if I keep setup to my laptop like GPT said or if I'll have better performance on my desktop?

[lucid yoke]

It just changes input, no difference there but seeing you running amd im unsure if it supports qwen edit

Even my nvidia card struggles sometimes

[ornate elk]

i would set up boath and compare

[cosmic furnace]

i cant run qwen even once. either setting up smthing or yet again "clip" doesnt match =\

[ornate elk]

load that image into png info to see what i used

[cosmic furnace]

are u sure that 9gb gguf file is the clip 🤨 it wasnt like that for flux

[ornate elk]

yes but you can also use a smaller one if you want

[cosmic furnace]

also text encoder + this file ends up having 19gb which is almost same as model of qwen itself. resulting in about 40 gb space? lol ?

[ornate elk]

dont use the large qwen model lol

try the ones i used and i will work

[cosmic furnace]

https://civitai.com/models/1864281?modelVersionId=2110043

19 gb one isnt large) 😊 theres 38gb one 😵‍💫

yet i see only text encoder & vae there. no clips

In comments someone had same issue , but i dont see how he got it solved -
SOLVED-answer inside Im struggling to find where to download CLIP that is in this workflow. huggingface has 3 part thing which is weird. how to load it in workflow? please any hints?

3 Comments

20 days ago
Check the official ComfyUI documentation (https://docs.comfy.org/tutorials/image/qwen/qwen-image). There should be an example workflow that automatically downloads the text encoder, as well as a direct download link for the text encoder (https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors).

20 days ago
@VV24 im sorry how did you found the link of that fp8 text encoder? i was looking for an hour and found only 3split part 1 part 2 part3.. and i was confused. thanks!!!

20 days ago
@gemstonebro No worries, this stuff can be pretty confusing. The download link is right there in the documentation I linked, in the section "Model Download", along with all the other required model files: https://docs.comfy.org/tutorials/image/qwen/qwen-image#2-model-download

[ornate elk]

20gb is large, its to large for 24gb vram to handle if you load in a clip + vae

as the computation (torch) itself needs at least some free vram too

[cosmic furnace]

isnt it like standart size for qwen ? Also , what doesnt fit vram, goes into ram or pagefile or smthing ?🤔

Didnt thought that model file size shud be equal to vram size o.o

i mean theres no 38 gb vram GPUS... well except that very rare & expensive ones xD

[ornate elk]

yea and slows down everything or goes directly into oom error

[cosmic furnace]

I can see that all qwen models seems to be 19gb of size tho 🤨

[ornate elk]

gguf:
https://huggingface.co/city96/Qwen-Image-gguf/tree/main

text encoders:
https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/tree/main

[cosmic furnace]

btw how are theese different frome the one i gave link in civitai ?)

[ornate elk]

smaller

slightly reduced quality

[cosmic furnace]

are u sure that even 7900 with 24vram like yours can have issues ? Or just think speed loss & prob issues doesnt worth the extra quality ?)

[ornate elk]

yes im pretty sure xD and i also dont want to wait 10 minutes for one image while having a frozen PC

[cosmic furnace]

Indeed 🙃

[ornate elk]

also qwen, flux, chroma and wan are all bad in terms of anime images xD

so im not using qwen a lot, but with the files i used one qwen image takes 1:40

illustrious with much better anime quality gets me images in 16 seconds and even upscaled xD

[cosmic furnace]

Damn, I felt that illustr & pony still better 🙂

i guess that illustr lately even more popular than pony 🤔 . For an half a year mb or even year

[ornate elk]

yea i like it more than pony

[cosmic furnace]

so flux-qwen wont do good for anime-cartoon huh? & art images mb ? 🤨

[ornate elk]

realism and art yea

[cosmic furnace]

Have u tried anime oriented flux models tho ?))

[ornate elk]

nope

but i have looked at the new lumina type

[cosmic furnace]

and ? what was it like ?)

[ornate elk]

haven't tried yet, forge neo dev added its support 3 days ago

but i think illustrious is still the goto

[cosmic furnace]

yeah. i guess. SOmetimes it feels like its good enough understanding complex promts 🙂

also, all this new models also very bad at nsfw xD . y ?)

[ornate elk]

i havent tested

[woeful merlin]

Hello is there a way get res_2m sampler for forgeui/stable diffusion...? I hate comfy lol

[maiden coral]

desktop GPU is 50% more powerful

[candid walrus]

I mean I know cuda helps but I guess just the raw horsepower makes up for it?

[maiden coral]

yes, currently I'm bottlenecked with 6GB VRAM GPU on 14B model

Wan 2.2

[maiden coral]

try euler_ancestral, you can't go wrong with that

[candid walrus]

Wow how wrong is GPT, it said I'd be twice as powerful..then the LTT server said no du and openrouter server said things are Nvida for a reason. Guess they don't know what they are talking about? I trust the openrouter community the most they would know what works best for there software on what hardware.

My 4070 is this still better then that or should I have stuck with the 4070 and not have upgraded to the 9070xt when I did?

[peak agate]

Network configuration

[cosmic furnace]

after messing with forge neo & trying launch qwen. Forge zluda doesnt lauch for me. Even fresh-clean install =\ this error

mb by accident uninstalled & installed the one torch for the rock & forge neo. But why then freshclean install says same error ? o.0 . Dunno. Only thing i did besides it. Is installing ultimate sd upscale extension. 🤔

it says zluda version 3.95 . & all was working fine yesterday. dunno main problem is this i guess -
AssertionError: Torch not compiled with CUDA enabled

[maiden coral]

if you want it for AI gen. then big VRAM and power, there isn't much tot alk about, Nvidia currently has lead in AI, although AMD isn't far behind either

[maiden coral]

basically... check if the pytorch is built with CUDA support (it must be a version for GPU, not CPU)

if it isn't, then Zluda won't work

[ornate elk]

open up a cmd (not in any webui) and run this:
pip uninstall rocm rocm-sdk-core rocm-sdk-libraries-gfx110X-dgpu -y
then try again to launch forge with zluda

[odd blaze]

├── controlnet-openpose-sdxl-1.0
│   ├── diffusion_pytorch_model.safetensors
│   └── diffusion_pytorch_model_twins.safetensors
└── OpenPoseXL2.safetensors

I am using ComfyUI ControlNet for OpenPose. I downloaded 3 of the for SDXL, but only the twins work.
Why? I don't quite understand.
I'm using a specific SDXL-based model.

Shouldn't all of them work?

[ornate elk]

Ah okay, then can you show the settings you used when it crashes? My guess is that its an out of vram issue

[lucid yoke]

if your using illustrious, normal controlnet doesnt work

whered you get them from and what model?

[odd blaze]

I got the 3 SDXL OpenPose from the ComfyUI manager (there's only 3 of them).
I tried on various models, some works, some don't. I can only do trial and error to see which combination works.

Want to know if there is a better way of doing this?

Illustrious can do the right pose, but not sure if it's from the ControlNet.

I mostly used SDXL and its derivatives, including Illustrious.

How would you control the pose in Illustrious then?

[cosmic furnace]

launched , thx

[cosmic furnace]

i wonder what happened 🤨 , Is this because ive tried switching between zluda-the rock ? Or indeed messed up & used rocm install comand in wrong folder, zluda one ?)

[ornate elk]

it happens when you run the torch install command for therock in a cmd that hasnt opened a venv

[celest grotto]

Yeah, I think it is apparently that

But it was fine before?

[ornate elk]

maybe you have enabled upscaling and didnt set the hires steps to 10

but get yourself the tiled diffusion extension, its Tiled VAE option can solve the issue most of the time

[celest grotto]

How do you do that?

I'm sorry but I'm still quite new

[ornate elk]

ah no problem,
click on the extensions tab,
then on available, then on "load from"
then you get a list:
search for Tiled VAE and click install
also install adetailer and booru tag autocompletion
then restart the webui

[celest grotto]

Thank you. I already had adetailer and booru

Can you explain what does Tile VAE do?

[ornate elk]

at the last step of an image generation the VAE gets applied on the overall image as one large tile.
This step is the heaviest in form of vram usage in the image gen process.
The tiled VAE extensions tiles the vae into multiple smaller tiles and applies it that way, reducing the vram usage a lot

[celest grotto]

Oh thank you very much

[ornate elk]

no problem

[celest grotto]

Is it called TiledDiffusion with Tiled VAE

[ornate elk]

Yep

[celest grotto]

Ty, but when I ran the second image. The same problem rose again

[ornate elk]

Have you enabled tiled vae?

[celest grotto]

Yeh

[ornate elk]

Okay, which model, resolution, steps, hires fix etc have you used?

[celest grotto]

I’ll show the screenshot when I get home

Ty for helping me man

[ornate elk]

Alright np!

[ornate elk]

hmm looks okay but can you try with a 832x1216 resolution as test?

also how much RAM do you have?

[ornate elk]

i mean system RAM

not gpu vram

[celest grotto]

Oh, it's 16

[ornate elk]

ah okay, that explains the lag.
Then you need to increase the windows pagefile

Here is the Guide how to increase it:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Enable it only for C and make sure its not enabled for any other drive.
Set it to 16000 Min and 24000 Max for C
Then apply and reboot the PC.

[maiden coral]

I think this depends on where your ComfyUI runs.

if you have ComfyUI portable on your second drive and pagefile is disabled, you run out of memory (a quick test with Wan 14B model)

[ornate elk]

nope

pagefile should be located on the fastest drive in the system

[maiden coral]

how come it does not, I have it automatically managed on my main SSD and HDD had pagefile turned off, I couldn't offload those 2 big 15GB models

but it isn't, as you can see

ComfyUI is on my 2nd drive (HDD), my system is on SATA SSD

[ornate elk]

you said pagefile is on your main ssd and not hdd

so its correct

it doesnt matter to much where the files are that get loaded and used, but it matters a lot where the pagefile is

because pagefile acts as virtual ram so you want the fastest read and write speeds possible.
Having comfyui on an hdd slows down comfyui loading times. Having a pagefile on a hdd can slow down the whole PC itself.

the best approach for every webui is:
Programm installed on an SSD.
Models stored on a HDD or SSD with enough space. Then linked to the webui.
Pagefile always SSD

[maiden coral]

oh, yeha I got you you meant it differently, though

it doesn't matter that much at all

unless you are training or something like that

[ornate elk]

you mean comfyui on hdd ?

yea its not too bad

[maiden coral]

once loaded, it doesn't use drive at all

[ornate elk]

yea but if possible i would still avoid that setup

linking models is easy

[maiden coral]

I'm severely limited, yeah, but hey, it worked in the end.

[lapis wedge]

trying to run stable diffusion, keeps giving No Python at '"C:\Users\USER\AppData\Local\Microsoft\WindowsApps\PythonSoftwareFoundation.Python.3.12_qbz5n2kfra8p0\python.exe'

help?

i have 3.10.6

[maiden coral]

What client are you using to run models? Automatic111 / Forge / ComfyUI

[lapis wedge]

A1111

[maiden coral]

it's too old, use Forge (at least according to other users) or ComfyUI

but Forge is just a fork of Automatic111, so idk about that one

[lapis wedge]

Comfy ai is genuinely VERY confusing for me, i used to use A1111 back some time ago and it ran perfectly and was simple

[maiden coral]

You can watch a tutorial on Yotuube for ComfyUI, its's very easy unless you want to do some more complex stuff.

[lapis wedge]

i just want to be able to post my base models, lora and vae

[maiden coral]

well, then you just do it

it's like Scratch io (simplified programming, but here we use nodes)

[lapis wedge]

since i can just drag and drop a template that preloads it for me, can you make me a quick one?, 3 lora slots 1 vae 1 base model slot?

[maiden coral]

This is what I found, you can search really quickly:
https://www.youtube.com/watch?v=NYVqcA9JvJQ&pp=ygUhaG93IHRvIGxvYWQgbG9yYSBtb2RlbCBpbiBjb21meXVp

essentially, you just double click into empty space and write "lora" or "load lora" and it'll show you the node and you just click it

and everything else is in the base rpeset that loads up in ComfyUI by default

[lapis wedge]

got it. thanks, how about adding a vae? same thing?

[maiden coral]

you can go into templates -> basic

and next time you either create a new one with your own things or use a workflow form templates, they are also separated by use case

you also need ComfyUI manager (can't do anything without it, or if you wnat to download everything manually, sure)

[lapis wedge]

comfyui manager just downloads stuff off civitai or no?

[lucid yoke]

absolutely not

[maiden coral]

and you put the git repository you donwload into ComfyUI/custom?nodes and restart ComfyUI, that's it, everything else you can find on youtube

[lucid yoke]

it downloads nodes from github from anywhere and anyone

so its not recommended to use random workflows from people you dont trust

[lapis wedge]

a vae kinda works like an image fixer no?

[maiden coral]

Nope, you download models etc. manually, but nodes can be downloaded via ComfyUI Manager. It just pulls git repository into custom_nodes folder and that's it.

[lucid yoke]

keep in mind using multiple custom node workflows will eventually break your comfy instance and you have to reinstall it sometimes

[maiden coral]

read these first, the very basics, I recently found it, wish I knew this a lot sooner:
https://comfyui-wiki.com/en/tutorial/basic

very easy to understand

[lucid yoke]

tbh if you want to use a UI and sometimes mess with comfy, SwarmUI is the way to go

especially if your not that experienced

[ornate elk]

uninstall python 3.12 from the microsoft store

it causes these issues

[lapis wedge]

not installed tho

i had it installed but removed it

[ornate elk]

then delete the venv folder of auto1111 and relaunch

[lapis wedge]

jes bueno

thank you man

[ornate elk]

no problem 🙂

[storm swallow]

Hello, can someone help me with the error, "RuntimeError: CUDA error: no kernel image is available for execution on the device CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with TORCH_USE_CUDA_DSA to enable device-side assertions."

[ornate elk]

hey, whats your gpu?

[storm swallow]

5070ti

[ornate elk]

and which webui do you try to install?

[storm swallow]

I did the automatic 1111 install from a tutorial video, Im not 100% sure. Is there a way I can check

[ornate elk]

ah thats probably the issue as auto1111 doesnt support rtx50 cards

[storm swallow]

ooooh that makes sense

[ornate elk]

best you can do is setup Forge Neo from my setup guides:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

its based on auto1111 but more updated

supports rtx50 by default

[storm swallow]

Awesome, ill give that a try, thank you so much!

[ornate elk]

np! feel free to ask if any step isnt clear

[storm swallow]

Thank you!

[lapis wedge]

All up and running on my end. is there an easy way to setup Anime style drawing WAN 2.2? img2video

[ornate elk]

nope, that only works in Forge Neo or comfyui

[lapis wedge]

oh dont worry im on comfy

[ornate elk]

ah okay

[maiden coral]

What is your GPU if I may ask?

[lapis wedge]

3060 rtx, prolly not the best for wan 2.2

[maiden coral]

it's doable

also you'd better have more RAM, I heard people do it like this

[lapis wedge]

i can load 1024x1024 with lora in 23s, is that good or not idk

[maiden coral]

for a video you'd better go lower

with gguf models it should be doable, you have twice the VRAM I have, any 6GB+ VRAM GPU can run Wan 2.2 14B with off-loading (I take it that 6GB VRAM is some decent gaming GPU, kinda older, but it's for gaming or more powerful things)

[lapis wedge]

what vae and other things that you would reccomend so that my lora and Base model would make finer images?

[maiden coral]

euler_ancestral is okay, also I recommend this workflow I tried (since generation takes extremely high time for me): https://civitai.com/models/2039923/not-so-simple-or-is-it-wan-2122-lightspeed-i2v-workflow?modelVersionId=2312835

it was made for a model in description, but they are quite huge, Wan 2.1 is easier to manage since you only need one model, but since version 2.2 there are 2 models, one for high noise and 2nd for low

[lapis wedge]

downloaded, json file, do i put it where in my workflow?

[maiden coral]

into ComfyUI/users/worfklows

D:\ComfyUI_windows_portable\ComfyUI\user\default\workflows

here

it'll look something like that

[lapis wedge]

my user directory doesnt have workflow folder

just folder that says default

[maiden coral]

well, you can create those folders

[lapis wedge]

done

[maiden coral]

It'll get created once you click on a template, load a template and then save a template, it's a bit confusing.

When you load a template, you can see that it's in a 2nd row next to the ComfyUI icon letter "C"

you click on it, save and the workflow folder gets created, so you did it manually now, it's alright

but you still need to understand the basics of what each component does in order not to get lost, so m advice is to check the wiki I posted on ComfyUI, it's a few basic tutorials

nodes can have whatever name, but in essence, you'll understand what each does, if you odn't, you open their github and read on what a node and its parameters do

[maiden coral]

get ComfyUI Manager through google, it should pop on the first page as first result

[lapis wedge]

i have a manager on the app already though

[maiden coral]

after you download it into custom_nodes folder (relative path: custom_nodes/ComfyUI-Manager) you open ComfyUI again (restart is needed for the new nodes) and go into Manager -> Custom Nodes Manager -> in filter check Missing nodes and download them one by one (click all of them first and it'll ask you to restart ComfyUI, you click on Restart in the Manager menu and that's it

[sweet grove]

I keep getting this error, RuntimeError: The size of tensor a (640) must match the size of tensor b (1280) at non-singleton dimension 0
The size of tensor a (640) must match the size of tensor b (1280) at non-singleton dimension 0

[maiden coral]

You must've chosen wrong latent / wrong models or nodes

show the workflow

[sweet grove]

I changed, the lora and it worked. mustve been something buggy with that one

[maiden coral]

you chose a lora that was meant for SD model and ocmbined with SDXL model or some other combination

[sweet grove]

I see, thank you very much

How do I check on civitai where that info is?

[lapis wedge]

im all done, downloaded everything and loaded the wan with no problems

where do i set the lenght of video?

nvm

[maiden coral]

yeah, you have to download everything for WAN, it's not just downloading a model, it's something entirely new

but the author of that workflow has his own things there, you just click and download whichever text encoder you want to use, the base one, the lightspeed or ...

[lapis wedge]

ERROR: Could not detect model type of: C:\Users\user\Documents\ComfyUI\models\checkpoints\checkpoint03

?

[rose skiff]

I am trying to install comfyui on ubuntu. I followed the instructions here: https://comfyui-wiki.com/en/install/install-comfyui/install-comfyui-on-linux and when I try to launch it I get an error. I don't even have an nvidia gpu, I have a radeon 9070xt. How do I fix it?

[maiden coral]

Hi. Does anyone have experience with Sage Attention and knows if it can work on GTX 1060? (All I got from the internet is that it's untested or they have a wrong CUDA version without mentioning which version they even use; basically no info on that)

I have Pytorch 2.7.1+cu118 (CUDA 11.8) in my embedded ComfyUI environment (had to downgrade because of older GPU, ComfyUI comes with CUDA 12.1 or something like that).

[cosmic furnace]

Is there a way to resize to size be always same when i switch to it? lets say i want it 1400-1400 or smthing always when i switch to it in inpaint -_-

[ornate elk]

Sage attention is only available for rtx30 GPUs and newer

Rtx20 supports flash attention
And gtx10 only xformers

[cosmic furnace]

Dyu know any working automating mask extensions? Inpaint anything doesnt seem to work 😒 . Havent tried it on forge neo tho

[sour quarry]

just upgraded my GPU to a GeForce 5070 Ti. Now it seems pytorch isnt up to date for that card? its looking for cuda 120? Anyone have a solution to this problem?

[ornate elk]

which webui do you want to setup?

Auto1111 and forge doesnt support RTX50 gpus by default.
Best is to Setup Forge Neo.
You can find a guide in the first link of the pinned messages

[cosmic furnace]

i cant see forge couple anywhere o.0 wtf ?

[celest grotto]

I tried this and it worked for a couple of three or four pics before the same problem came back. Do you think it is because I am using laptop for this?

[ornate elk]

Its probably caused by to less ram or Vram. Try lower the resolution a bit. Also make sure you have 24GB free space on the C drive

[cosmic furnace]

i mean rly, wheres forge couple ?

i remember it shud be here 🤨

[ornate elk]

It always has to be installed manually with install from url

It was never included by default

[cosmic furnace]

why its not in aviable list ? o.0

indeed manually ? lol ?

what u use btw? couple or reg promter? Or both ?)

[ornate elk]

Couple

In the past I used reg prompter but couple is easier

[maiden coral]

yeah, found out the hard way

and xFormers only version 0.0.20

no other version works

because the newer versions need pytorch 2.8.0 or aren't compatible at all

[granite bough]

I got this error when I opened webui when trying to install forgeui

[ornate elk]

Hey, this error is relativly new.
Go inside the stable-diffusion-webui-amdgpu-forge folder.
Then click in the File Explorer bar (not searchbar) and type cmd then press enter.
Then you copy and paste these commands one by one:
venv\Scripts\activate.bat
pip install onnxruntime
pip install optimum
Then close the cmd and relaunch the webui-user.bat
Let me know if it worked

[granite bough]

same thing

[ornate elk]

which python version do you have ?

[granite bough]

3.11

[ornate elk]

and whats inside your webui-user.bat?

[granite bough]

--use-zluda --cuda-stream --attention-quad --skip-ort --skip-python-version-check

[ornate elk]

can you try remove --skip-ort and then delete the venv folder and rerun the webui-user.bat again?

[granite bough]

same thing

[ornate elk]

hmm okay, ill test it on my install real quick. maybe it needs a different pip package

[sweet grove]

what does this error mean? RuntimeError: The size of tensor a (640) must match the size of tensor b (1280) at non-singleton dimension 0
The size of tensor a (640) must match the size of tensor b (1280) at non-singleton dimension 0

[ornate elk]

incompatible files, like an sdxl model + 1.5 controlnet or lora

[sweet grove]

I keep getting that error, is there a way to filter out the loras I use so I dont keep getting this?

Like how do I tell if its compatible or not?

[ornate elk]

on civitai in the info box of the lora or model you see "base" that has to be the same for it to be compatible

[sweet grove]

the same with what?

OH, the other loras?

[ornate elk]

same with every file

models, loras, embeddings, controlnet

[sweet grove]

Sorry, Im new to this. I dont really understand

[ornate elk]

so i found the fix. do this:
Go inside the stable-diffusion-webui-amdgpu-forge folder.
Then click in the File Explorer bar (not searchbar) and type cmd then press enter.
Then you copy and paste these commands one by one:
venv\Scripts\activate.bat
pip install optimum[onnxruntime]
Then relaunch the webui-user.bat

[ornate elk]

models checkpoints are have a base version:
on the right box you see Base Model: Illustrious

So if you now search a lora, it has to be also Base: Illustrious
then boath are compatible

[sweet grove]

AHHHH THANK YOU

[ornate elk]

best is to make subfolders in the lora folder to not get them mixed up. same goes for checkpoints

[sweet grove]

thats a smart idea, thank you

[ornate elk]

no problem 🙂

[sweet grove]

What are all the types? are there LOTS or just a few?

[ornate elk]

1_5, SDXL, Pony, Illustrious, are the main ones

[sweet grove]

ty so much

[lucid yoke]

if you use a lot of different checkpoints, i recommend putting those in subfolders aswell

[sweet grove]

So the checkpoint and Lora have to have the same base model? or just the loras?

[ornate elk]

loras only work together with checkpoints

[sweet grove]

I see

Thank you

[granite bough]

ty

[ornate elk]

np

[celest grotto]

Thx

I’ll do that. Thanks a lot man

[storm swallow]

Quick question, for some reason my images are not coming out nearly as sharp or detailed as they were yesterday. I ran a test with all the exact settings and I get the same image just a way worse quality version of it

[robust spear]

Isn't it supposed to be at 80% or 100% when generating? I installed yestarday and been trying to make it use gpu

[peak agate]

What's available

[alpine comet]

what going on? i tried opening 'waiillustrious' model and i got thi error in cmd when loading it.

[ornate elk]

Hey, do other models work?

[alpine comet]

I reenabled it and it worked. This is on a old laptop. Very low frame so I can do generation's anyway

[ornate elk]

Ah so a restart fixed it?

[alpine comet]

Nah same session, just enabled the same model again

[zenith seal]

haiii, i got this issueUsing pytorch attention for VAE
Traceback (most recent call last): File "C:\Stable Diffusion\SD-Zluda\stable-diffusion-webui-amdgpu-forge\launch.py", line 54, in <module> main() File "C:\Stable Diffusion\SD-Zluda\stable-diffusion-webui-amdgpu-forge\launch.py", line 50, in main start() File "C:\Stable Diffusion\SD-Zluda\stable-diffusion-webui-amdgpu-forge\modules\launch_utils.py", line 675, in start import webui File "C:\Stable Diffusion\SD-Zluda\stable-diffusion-webui-amdgpu-forge\webui.py", line 23, in <module> initialize.imports() File "C:\Stable Diffusion\SD-Zluda\stable-diffusion-webui-amdgpu-forge\modules\initialize.py", line 29, in imports shared_init.initialize() File "C:\Stable Diffusion\SD-Zluda\stable-diffusion-webui-amdgpu-forge\modules\shared_init.py", line 19, in initialize from modules import options, shared_options File "C:\Stable Diffusion\SD-Zluda\stable-diffusion-webui-amdgpu-forge\modules\shared_options.py", line 9, in <module> from modules.onnx_impl.execution_providers import get_default_execution_provider, available_execution_providers File "C:\Stable Diffusion\SD-Zluda\stable-diffusion-webui-amdgpu-forge\modules\onnx_impl\__init__.py", line 6, in <module> import optimum.onnxruntime ModuleNotFoundError: No module named 'optimum.onnxruntime' Press any key to continue . . .

[ornate elk]

hey, to fix this issue:
Go inside the stable-diffusion-webui-amdgpu-forge folder.
Then click in the File Explorer bar (not searchbar) and type cmd then press enter.
Then you copy and paste these commands one by one:
venv\Scripts\activate.bat
pip install optimum[onnxruntime]
Then relaunch the webui-user.bat

[zenith seal]

Thank you so much!! you are saving me ahah

i just don't get any preview :/

[ornate elk]

make sure the webui is whitelisted in any browser adblocker, and no opera vpn is enabled

[zenith seal]

Thankks!! but restarting the ui fixed it

[ornate elk]

Nice no problem!

[zenith seal]

OutOfMemoryError: CUDA out of memory. Tried to allocate 1.21 GiB. GPU 0 has a total capacity of 15.98 GiB of which 5.54 GiB is free. Of the allocated memory 8.36 GiB is allocated by PyTorch, and 412.10 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

okey i have this issue now.. @ornate elk

[ornate elk]

whats your gpu?

and which settings have you tried like model, resolution, upscaling?

[zenith seal]

6800xt

and i just loaded a lora

usualy work

[zenith seal]

just a base gen with illu, a base xl res and a lora

[ornate elk]

can you share a screenshot ?

Also make sure you dont have Wallpaper Engine running in the background or other heavy programms, games or streams

[zenith seal]

sure

[ornate elk]

i dont see the lora mentioned in the prompt. And is red lilly a model (6gb) or is that the lora? If lora, then its in the wrong folder

[zenith seal]

nono red lily is a checkpoint, and i just restarted the ui

and didn't put any lora for the sreenshot

[ornate elk]

ah okay

try use this sdxl vae:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main
it doesnt use as much vram as the other

[zenith seal]

it's the one i use already :/

[ornate elk]

oh okay

how much normal system ram do you have?

[zenith seal]

i have 16 of ram

but it only started today

i can gen with no issue normaly

[ornate elk]

ah okay yea 16gb ram is to less for sdxl or illustrious based stuff

you will need to increase the windows pagefile

Here is the Guide how to increase it:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Enable it only for C and make sure its not enabled for any other drive.
Set it to 16000 Min and 24000 Max for C
Then apply and reboot the PC.

Then try again, and make sure nothing else is running in the background, other than SD and discord

[zenith seal]

Okey thanks a lot!! i will restart my computer now and let you know

[zenith seal]

@ornate elk I think it's working no, thank you so much !

[ornate elk]

Perfect, no problem 🙂

[storm swallow]

For some reason my images are not coming out nearly as sharp or detailed as they were yesterday. I ran a test with all the exact settings and I get the same image just a way worse quality version of it

[ornate elk]

hey, load up the images in the PNG info tab and compare the meta data, then you see if they are really the same, sometimes there is a small difference that could change it

[storm swallow]

Here are the two images with the meta data

[ornate elk]

wow thats a strange issue

its the same meta data

ohhh wait

nope

its not

the upscaler changed!

Latent causes the blur, on the left you used Esrgan

[storm swallow]

Oh?? Let me try to fix that! and see

[ornate elk]

also on the left hires upscale is set to 1 and on the right its on 2

[storm swallow]

You are actually a life savor, thank you so much! That was the issue!

[ornate elk]

no problem 🙂

[lapis wedge]

someone help me run wan 2.2 on comfy

does anyone have workflow for it?

[zenith seal]

OutOfMemoryError: CUDA out of memory. Tried to allocate 1.21 GiB. GPU 0 has a total capacity of 15.98 GiB of which 5.88 GiB is free. Of the allocated memory 8.01 GiB is allocated by PyTorch, and 416.67 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
Time taken: 32.5 sec.

ohh well i still have the same problem :c

[ivory sierra]

The memory is very large

[rose skiff]

What is the best guide for installing this on ubuntu 25.10? I have tried a few guides so far and none of them worked.

My GPU is a Radeon 9070XT

[vocal burrow]

depends what you mean by "this"

[vocal burrow]

you re most likely mix and matching sd1.5 stuff with SDXL

[ornate elk]

Motion modules are only compatible with animated diff model 1.2 and not with 1.3 or xl

[rose skiff]

any form of stable diffusion

I don't care much about which frontend it uses

I have a 9070xt and I want to use it

[vocal burrow]

I have no problem using ComfyUI or Stable Diffusion Web Neo at home with a 9070 XT and CachyOS (arch linux).
The important part is to make sure that you upgrade the pytorch used by either comfyui, sdweb, etc after installing. As pytorch 2.10 brings Rocm7 support. Which brings massive performance improvement compared to Rocm6.

really inserting 0 instead of O

be more creative bots

if anyone clicked the link that the bot just spotted before I clicked, leave their server. They are scammers.

[vocal burrow]

If you need help setting things up I should be able to help you in 2 hours ish when I get back home.

[rose skiff]

Ok, thanks

[vocal burrow]

did you mangage to make it work ?

[rose skiff]

not yet

[vocal burrow]

what's the problem ?

[rose skiff]

I thought I did everything right for AMD GPUs but it doesn't look like it sees it

[vocal burrow]

what's the full log ?

what did you do ?

[rose skiff]

I read this and tried to do it: https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#manual-install-windows-linux

then pip said it needed a venv so I made one and installed the stuff there

maybe I should just rm the whole folder and restart

[vocal burrow]

ok this is probably one of those rare times where I wouldn't recommend doing what the readme says

(as it's outdated)

[rose skiff]

I will delete the directory and restart, then

What should I do instead?

[vocal burrow]

let me write everything down, gimme a couple minutes

[rose skiff]

thanks

[vocal burrow]

(I'm gonna assume you know your linux basics)

[rose skiff]

I know most of the basics, yes

I know how to do most things in bash

[vocal burrow]

#getting comfyUI
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI

#setting the venv for comfyui to use
python -m venv venv

#create a .sh script to launch it with some env variables and parameters if needed
nano launch.sh
#####################################
#Copy paste the following in it
#!/usr/bin/env bash
export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
venv/bin/python main.py 
#####################################
chmod u+x launch.sh

#let's activate the venv it's gonna use 
source venv/bin/activate #might source venv/bin/activate.fish or whatever depending of your shell

#install the pytorch 2.10 preview package
pip install --pre -U torch torchvision torchaudio pytorch-triton-rocm --index-url https://download.pytorch.org/whl/nightly/rocm7.0
pip install --upgrade pip wheel setuptools

#then let's install the rest of the requirements for ComfyUI
pip install -r requirements.txt

#exit the venv
deactivate

# That should be all, run the launch script
./launch.sh

Something around those lines should work

let me know if you see some errors along the way

looked like you had python 3.13 on your system so it should work with it.

It might be a good idea to use pyenv to assign a specific version of python for your ComfyUI folder. But for now let's forget about it.

cause who knows maybe one day your system will update to python 3.whatever and it won't be compatible with ComfyUI

(there's basically no performance to using "old" python)

[rose skiff]

if the point of sudo usermod +x launch.sh was to make it executable, I used chmod u+x launch.sh instead, I don't think usermod works like that on ubuntu

[vocal burrow]

yeaaaah my bad

righting everything from the top of my head

just got home

that's why I asked if you knew your linux basics XD

usermod is totally not the right tool here, fixed it

and also you're not forced to use nano of course to edit the launch.sh

[rose skiff]

yeah, but nano is my preferred text editor anyway

looks like the same error

[vocal burrow]

ok can you send the whole log please

[rose skiff]

ll

wrong window

I'll try to find it

[vocal burrow]

there's no log file.

the log is basically what's printed in your terminal

but I assume there's more than what you screenshot

[rose skiff]

just the line where I said to run it

[vocal burrow]

ok then....

what amd drivers did you install for your gpu ?

Like. What does glxinfo | grep "OpenGL renderer" gives you ?

[rose skiff]

OpenGL renderer string: AMD Radeon RX 9070 XT (radeonsi, gfx1201, LLVM 20.1.8, DRM 3.64, 6.17.0-6-generic)

[vocal burrow]

k looks good (not the latest but that's to be expected for ubuntu)

what about rocm-smi ?

[rose skiff]

============================================ ROCm System Management Interface ============================================
====================================================== Concise Info ======================================================
Device  Node  IDs              Temp    Power   Partitions          SCLK     MCLK     Fan  Perf  PwrCap       VRAM%  GPU%  
              (DID,     GUID)  (Edge)  (Avg)   (Mem, Compute, ID)                                                         
==========================================================================================================================
0       1     0x7550,   22577  47.0°C  32.0W   N/A, N/A, 0         1679Mhz  456Mhz   0%   auto  304.0W       11%    100%  
1       2     0x13c0,   16360  40.0°C  0.023W  N/A, N/A, 0         None     1800Mhz  0%   auto  Unsupported  6%     0%    
==========================================================================================================================
================================================== End of ROCm SMI Log ===================================================

I have an integrated gpu also, but it's the 9070xt that I want to use for this

[vocal burrow]

yeah that's what I was checking.

That could be the problem. It might be trying to use your integrated graphic instead of your dedicated gpu.

[rose skiff]

I'll have a second (smaller) dedicated gpu in a few days too, for my last three monitors

[vocal burrow]

There's an env for that to tell it which device to use, let me find it

in launch.sh, try to add export HIP_VISIBLE_DEVICES=0

before venv/bin/python main.py

[rose skiff]

same error

[vocal burrow]

what the... ok let me try to run all that again in a different install folder

Yeah that just works for me. Strange

After installing rocm did you run sudo usermod -a -G video,render $LOGNAME to set everything up regarding user privileges ?

(and I do mean usermod this time)

not sure if rocm-smi would even be able to see the gpu if you didn't but it doesn't hurt to run it again.

[rose skiff]

I'mm sure I ran that at some point yesterday, it didn't appear to change anything when I ran it again today

[vocal burrow]

it's not a verbose command, it will just add yourself to the group video and render

[rose skiff]

I got the same error after running it

although maybe I need to reboot

[vocal burrow]

I'd doubt it but you might as well try that while I search for answers

[rose skiff]

nope, that was it

I am pretty sure that changing groups only applies on the next login, and ctrl-alt-t doesn't count as a login

thanks for the help

[vocal burrow]

welp that's easy to check, reboot and try it out.

[rose skiff]

I did reboot and it launched

[vocal burrow]

everything is working ?

[rose skiff]

well, I need to doownload a model again now but I think it's working

the gui is on my browser

[vocal burrow]

Nice, let me know how it goes

For references here's some quick tests I did with an sd1.5 and sdxl model

your numbers should be about the same

[rose skiff]

Do I want normal SDXL or lightning or hyper?

[vocal burrow]

As you wish. lighting and hyper models are designed to output in 1, 2 or 4 iterations. So they're faster but their results tend to be ... different

[rose skiff]

ok, I'll do normal

[vocal burrow]

use it with the SDXL VAE fp16 variant while you're at it.

[rose skiff]

It took 9 seconds to make an image after I found all the things

well, 9 seconds to execute the whole prompt

[vocal burrow]

Speed will depends of your workflow. As you can see I got mine in 5 seconds with the 16bit VAE and without using any refiner.

The simpliest SDXL graph, bare minimum

[rose skiff]

yeah, it was faster after the first one

probably had to copy something to ram

Anyway, it seems to be working now

Thanks for the help

[tawny ore]

Guys help how can i fix it

[ornate elk]

Hey, what's your GPU and can you show the full cmd log?

[tawny ore]

3070ti

[tawny ore]

I've never had problems with Stable Diffusion, but recently I noticed that generations have become very slow. The reason was that xformers wasn't working. In an attempt to fix it, I reinstalled Stable Diffusion, after which I couldn't launch it. I tried ChatGPT and googled the problems, but nothing helped. It's stupid that I have to fix something after a clean installation.

I dont no why i cant just fucking install and use it

[lucid yoke]

A recent update or installation of something else using python etc could break it

Please do not use chatgpt for stuff like this, its 9/10 times wrong

[tawny ore]

i install clean windows and clean Stable Diffusion from git, i dont install python or anything

[tawny ore]

idk why i think that chatgpt can help me

[lucid yoke]

Personally i dont recommend stable diffusion webui since its been a year without updates, you could try looking into forge neo instead
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#nvidia-forge-neo

[tawny ore]

Oh, Ok

[lucid yoke]

The ui is similar but it gets updates

[tawny ore]

i just noob in AI generation and i find SD is simple and cool

thx, i'll try it

[lucid yoke]

I hope it works out for you, but if you follow cs1o's guide it cant go wrong

Especially since its a fresh install of win 11

[tawny ore]

make sense, my problem with xformer that SD cant work with new version of python

[lucid yoke]

R
The ui you used uses python 3.10 iirc this one uses. 11

[tawny ore]

@lucid yoke thx bro ❤️

[tawny ore]

@lucid yoke Guys help i have errors everywhere ( cant use lora's)

[tiny panther]

Guys, I think I'm dumb, because I can't generate or configure the SDXL in comfy-UI. I think I configured it correctly, but it simply won't display any image.

[prisma moss]

how do i choose wehich python the webui will run?

since i have both 3.10.9 and 3.11.9 for different purposes

im trying to install forge classic, but its trying to pull with 3.10

even with this is still doing 3.10

ok i got it

[lucid yoke]

Any cmd errors?

[tawny ore]

when i'm trying refresh "Lora" - AttributeError: 'FakeInitialModel' object has no attribute 'is_sdxl'

[lucid yoke]

And your sure the model you downloaded is a sdxl model?

[tawny ore]

i have bunch of them

@lucid yoke Nevermind bro, i just delete Lycoris, now is all working, sry)

[lucid yoke]

No worries!, soz for late response i was on my way to work

[ornate elk]

You have to delete the venv after this change.

Only then it will use the new python from the Path

[prisma moss]

It didnt but ok

As i said i got it

[chilly trail]

If you wanna make short videos locally based on photorealistic photos , would a 4080 16GB be enough? Thinking 5 too 10 second long.

[lucid yoke]

Yeah, might take a few min though

[granite bough]

Why does my flux take 2 minutes to generate a 512x512 photo

[ornate elk]

whats your gpu again?

[granite bough]

Rx6950

[ornate elk]

and which size has your flux model ?

[granite bough]

I think 16gb

[ornate elk]

then you should use a smaller one

like the Q8 or Q6, or Q5 GGUF variants

https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main

[granite bough]

Ok

[green plank]

Hello friends, I'm trying to get SD-Next working on my desktop with ZLUDA. I've downloaded sd-next forge ui and followed the directions on github but its not working for me. I keep getting the error "the procedure entry point could not be located in the dynamic link library C:\ProgramFiles\AMD\ROCm\6.4\bin\rocsolver.dll"

Could someone with more knowledge give a helping hand please??

[green plank]

I got it sorry, i had the wrong version of hip installed facepalm

[celest grotto]

@ornate elk so, I managed to fix the last problem you helped me with. Thank you very much. It worked fine for a while and now new problem arose. It was fine generating normal pic without Lora. But whenever I tried to input a Lora in, it shows the same screen “aw, shucks” as before. Everything worked very fine before btw.

[ornate elk]

Would recommend using forge with Zluda over SD next if you want to an easier usability

[ornate elk]

Hmm thats strange. And error in the cmd ?

[celest grotto]

It’s pretty late where I am right now. But I’ll reply soon. Thank you very much for helping me man. You’re the most helpful guy around.

[analog finch]

I managed to run stable diffusion finally but no clue how to get my AMD GPU to run. New stable diffusion or wont work? I followed guide, it said Python 3.10.6

No torch for amd found

[maiden coral]

Get comfyUI, it's ready for AMD and should work if you use a card mentioned in readme for amd users (on github)

[analog finch]

I did get comfyui but when it loads up it reads cpu, no gpu, and I tried a random text to image like wan2 and it "rented" a gtx 4080 24 vram or something

Settings I have to tweak? I am running a Radeon RX 9070 XT

[ornate elk]

Hey, please follow the setup guides from the first link of the pinned messages.
Go for Forge with Zluda.
I will add the new Comfyui guide there later too

[analog finch]

Alright thank you both a lot! Can I run any models with a correct Webui setup or does some models have to be converted to ONNX?

[ornate elk]

You can run any model. No need for onnx

[analog finch]

@ornate elk Much love, got webui to run now, its compiling. Should I ignore below or download in venv?
Legacy Preprocessor init warning: Unable to install insightface automatically. Please try run pip install insightface manually.

[ornate elk]

Can be ignored. Until you want to do something with faceswap

[analog finch]

Okay, I download it in venv right?

[ornate elk]

You have to install it to the venv like this:
Go inside the stable-diffusion-webui-amdgpu-forge folder.
Then click in the File Explorer bar (not searchbar) and type cmd then press enter.
Then you copy and paste these commands one by one:
venv\Scripts\activate.bat
pip install insightface
Then relaunch the webui-user.bat

[maiden coral]

Is it as easy as that? I had to download a wheel file and compile it

if needed: https://github.com/Gourieff/Assets/tree/main/Insightface

[ornate elk]

yea now it should be easier than in the past

[maiden coral]

strange, I did try it but it always failed and told me that pip couldn't find insightface

it was really annoying, I found it by a chance on reddit that I need to download it from repository and compile it

[ornate elk]

when was that?

[maiden coral]

a week or two ago

I may find it a sec.

https://www.reddit.com/r/comfyui/comments/1k2fgvw/no_module_named_insightface_neewbie_looking_for/

[ornate elk]

ah okay

yea there are two ways for it to work, either with the precompiled wheel that always works. Or if you have visual studio build tools installed, it will work with just the pip install method.

if he still gets the error then he would need the files, or install vs build tools.

[analog finch]

Ill check it out after the 1st gen compiling is done, its taking a bit long. Its saying "comp in progress. please wait" x 14 times down

normal I hope? I read it would take a while

It defaulted to Euler A sampler. Isnt DPM++ 2M the best now?

[vocal burrow]

there s no best sampler, just like there s no best brush in photoshop

Euler A is usually the fastest one.

[analog finch]

Thanks

[mystic basin]

I'm using comfy with zluda, and since i had to change the repo I git pull from because the repo changed, well i have ram issue
in my Start-ComfyUI.bat I have this : python main.py --auto-launch --disable-xformers --use-quad-cross-attention --reserve-vram 0.8

if I modify it so it become this : python main.py --auto-launch --disable-xformers --use-quad-cross-attention --reserve-vram 0.8 --lowvram --no-half-vae
it just doesn't launch so i'm wondering if there is anyways to make comfy use less ram or make it so it doesn't crash (my gpu has 12gb of vram and my pc has 16gb of ram and I already use custom node to reduce vram and ram usage)
windows pagefile could be too small which could be the one causing the issue but the thing is that I'm scared that if I change the pagefile that other program running on my pc will have issue because of it

(os error 1455) is the error i'm receiving, it says pagefile but yeah i'm trying to avoid needing to modify it if I can avoid it (sorry if my explanations are not very good)

nevermind, I'm not sure what I did but now it crash when i tried generating an image

[ornate elk]

With 16GB of ram the best you can do is increase the pagefile size. The whole PC will benefit from that. No downside other than a bit more usage of SSD write cycles.
Make sure the pagefile is only enabled for the C drive and no other drive. And set it to 16000 min and 24000 max.
Then reboot the PC

[mystic basin]

alright, if there's no downside then I'll do it

[foggy niche]

hello i am constantly getting Cannot import 'setuptools.build_meta' error

any ideas?

i tried a lot of thing with chatgpt but no fix

[mystic basin]

it stopped crashing but I still have the issue i had before

UNETLoader
The paging file is insufficient to complete this operation. (OS error 1455)

[mystic basin]

nevermind it crash again

[green plank]

Hey friends i thought i had sdnext working with zluda but everytime i generate a image it disspaears at the end, it appears to be working, i watch the preview evolve but then at the end theres some error and i get no photo

something about float

[green plank]

RuntimeError: expected scalar type Half but found Float

[ornate elk]

How much free space do you have on C ?

[ornate elk]

Hey what's your GPU?
I would recommend using forge with Zluda over sdnext

The float error comes from the vae

[mystic basin]

105gb but i managed to fix the crashing

but

i have a new issue

[ornate elk]

can you post the full cmd log?

[mystic basin]

yeah

[ornate elk]

make sure your comfyui is updated. and when does that error appears?

[mystic basin]

it seem to happen when loading the diffusion model

[ornate elk]

how large is this flux model?

[mystic basin]

11.08gb (fp8)

[ornate elk]

okay, thats the issue then ,as you only have 12gb vram, filling 95% of the vram doesnt help, because it still needs vram for the lora and the image generation process itself. so better go with an smaller variant like gguf Q4

[mystic basin]

I didn't seem to have that issue before when I cloned LeagueRaINi repo like 10 month ago, well maybe the new version that I had to clone from patientx repo requires more ram and vram

but yeah I'll test with a gguf one to see if it work

thanks

[proper grove]

Cannot import 'setuptools.build_meta'. Can anyone help with this problem?

When installing a clip during the first launch of stable diffusion a1111, Installation fails with the following error.

Environment

• OS: Windows 10

• Python: 3.10.13

• pip: 25.3

• setuptools: 67.8.0

• wheel: 0.45.1

• torch: 2.2.0+cu121

• torchvision: 0.17.0+cu121

• CUDA: 12.1

• numpy: 1.25.2

I have already tried:

• Upgrading pip, setuptools, and wheel
• Reinstalling torch and torchvision
• Using different numpy versions

Torch and torchvision are working correctly and torch.cuda.is_available() returns True.

[ornate elk]

hey, whats your gpu?
The issue comes fro mthe Python installed from the Microsoft Store.
Uninstall it and install Python 3.11.9 64bit with a normal python setup exe.
Then delete the venv folder and relaunch.

[cosmic furnace]

Holy shit theres some good stuff lately & videos are improving)

how do i share nsfw link to civitai ? xD . Idlike to know yur opinion Cs1o

He says its a wan video but damn)

[ornate elk]

you can dm me the link

[cosmic furnace]

I kinda not sure this can be easily repeated 🤔 . Or mb theres some tricks he uses. Or Wan video models become so good now ?o.0

[analog finch]

@ornate elk

I downloaded multiple to q4 q6 q8

[ornate elk]

looks correct

set the frames to 17 and lower the resolution to 720x1080 maybe

[analog finch]

I did 720x720 and 512x512, CFG was 6 standard, did 1 now

Did flower as test

any wrong setting?

[ornate elk]

can you show the other settings?

[analog finch]

RuntimeError: Configured ROCm binary not found at C:\TheRock\SD-TheRock\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_rocm71.dll

Relaunched

Worrysome?

It locates GPU

[ornate elk]

nope, you can add --disable-bnb to the webui-user.bat to disable bitsandbytes because it doesnt work for AMD

[analog finch]

ROCm GPU Architecture?

[ornate elk]

ROCm is AMD yes, but bitsandbytes is only supported on linux for AMD, but is not needed for image gen

[ornate elk]

you have to select WAN at the top left. Always select the model version your using

[analog finch]

Right, any specific samp/sched? cfg 1-1.8 or higher for video?

No clue on Shift

[ornate elk]

I didnt used wan before so no clue about shift or how many frames xD

[analog finch]

I got the same AttributeError: 'NoneType' object has no attribute 'is_webui_legacy_model'
'NoneType' object has no attribute 'is_webui_legacy_model'

[ornate elk]

automatic

[analog finch]

oh my days 🤦‍♂️

[ornate elk]

oh you didnt selected the refiiner

[analog finch]

Switching ui applied fp8 and fp16 while removing vae

might work now

[ornate elk]

yea every preset has its own values

[analog finch]

im on 6 hours sleep 😛 lets pray

[analog finch]

Ill play around with highest doable fps/reso on q6 then try q8

see what 16gb ram can do

chatgpt said with some offloading it could work, it might OOM. q8 might hit 15.2 or smth vram

q8 could potentially run 24gb, i didnt read it fully

@ornate elk I am lost
FileNotFoundError: [WinError 2] The system cannot find the file specified

it generated to finish 100% and nothing generated + that

[ornate elk]

ohh you need to have ffmpeg setup in your system

it needs that to generate videos out of frames

[analog finch]

Do I have to open path for that / venv it to TheRock?

[ornate elk]

here is a short guide how to:
https://phoenixnap.com/kb/ffmpeg-windows

[analog finch]

Gyan dev, windows logo?

[ornate elk]

?

[analog finch]

Website give this

[ornate elk]

wrong file

[analog finch]

but guide say something weird windows related

gyan.dev and opens latest git

fuck me im tired, i think i found it

[ornate elk]

scroll upwards

it jumps down to the git links but the builds from guyan dev are at the top after clicking on the logo

[analog finch]

apparantly i downloaded ffmpeg 2 days ago for img

never did environmental for it, just replaced files

[ornate elk]

sometimes it works when the exe is placed in the same folder

[analog finch]

Im just gonna follow the guide, environmental is cleaner I think

doesnt matter if system or user path right?

just has to do with perms on pc

[ornate elk]

yea but better in system path

[analog finch]

weird guide says user, both work or clash?

[ornate elk]

both work yea

[analog finch]

no need for hip path like zluda or w/e it was

[ornate elk]

its the same as hip path or zluda folder

an new entry in path

[analog finch]

rocm or something i needed had to be in path and hip path environmental

but it should be working

cmd command loaded

but it gen this moving around for 2 sec 💀

gpt says the vae2.1 with wan 2.2 is a mismatch 🤔 but can also be settings issue, ill follow his lead for a sec

[ornate elk]

nope vae 2.1 is needed for wan 2.2 to work

chatgpt knows nothing about image generation webuis

[analog finch]

Yeah i figured that out when he gave me the same 4 cmd commands over and over ending in the same 4 unique problems 😂

Is maximum seconds 2 on video gen?

[ornate elk]

nope

it should be more but has to do with the frames amount

[analog finch]

This means its running during gen right?

Utilisation 1-2% tho

I am running q8 now also, 1080x1080 60fps, tweaked settings

pls gen 🤞

[ornate elk]

2.3gb vram usage doesnt lookr right

[analog finch]

CPU Swap Loaded (blocked method) ?

[ornate elk]

hmm strange

[analog finch]

So its producing from cpu?

[ornate elk]

seems so

it didnt had enough vram it seems

[analog finch]

okay, I tone down more and check I guess

im on q8 now so it is limit testing

q6 was same

ValueError: Input and output must have the same number of spatial dimensions, but got input with spatial dimensions of [17, 135, 135] and output size of (270, 270). Please provide input tensor in (N, C, d1, d2, ...,dK) format and output size in (o1, o2, ...,oK) format.

It stopped* at 48% with this

[ornate elk]

hmm

[analog finch]

maybe I need q4?

[green plank]

I have a rx 6750xt 12gb

I will try and download it now haha

[ornate elk]

Maybe, I will test q6 tomorrow

[mystic basin]

so yestersday I had ram issue, after switching of model to stop having those issue, I've got a new issue that I'm not sure what is the cause of it. I think it could be a triton error (even if the triton test when opening comfy work fine)

this error seem to happen when it at the VAE decode node (also happen when using tiled VAE)

[mystic basin]

managed to fix the issue by adding : dtype = torch.float16 # ADD THIS LINE
to sd.py (the same issue happen if instead you add --fp16-vae in the bat file to launch comfy (both of those do the same thing))

but the image that are generated seem to have color issue or something

[analog finch]

lmk how it goes, q2 also gens noise

[plush vine]

hello,
I configured my SD stable-diffusion-webui-amdgpu-forge for AMD, the computer automatically updated to Windows 11, and now the good old freezing has started when changing chepoints and launching the SD. Does anyone know if I need to add something to the ROCM folder to download the missing files for Windows 11?

[ornate elk]

Hey how much RAM do you have?
You probably need to increase the Windows Pagefile.

[mystic basin]

tried to fix the issue, but I can't seem to be able to fix it. either it work but the image are black when I force fp16 or if I don't force fp16, it just tells me this error. I'm guessing the error is most likely caused by triton acceleration thing, probably that my gpu is not supported (rx6850mxt (gfx1031)) or it's just me who did something wrong when I did the install

the log

[ornate elk]

Which model type do you use?

[mystic basin]

tried using illustrious

[ornate elk]

Ah then there is an easy fix, use this vae:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main
The sdxl.vae

Its made in fp16

[mystic basin]

doesn't seem to fix the issue, tried with multiple model, if I force fp16-vae in the comfyui-n.bat, it doesn't give me the error but the result seem to have no colors and if I don't force fp16-vae in the comfyui-n.bat, well it gives me the same error : VAEDecode GET was unable to find an engine to execute this computation

the wworkflow (in case it could be needed)

[plush vine]

16 GB. I configured the SD in Windows 10 fine, but when my computer automatically updated last week, the same lags I'd experienced with the non-AMD-compatible SD reappeared.

im using this guid for amd:
https://github.com/brknsoul/ROCmLibs/wiki/Installing-Forge‐Zluda

[ornate elk]

Yea you need to increase the windows pagefile then. Also the guide you followed is outdated. So you installed an outdated hip SDK version too which is worse for performance and tends to crash.

Please reinstall Forge with Zluda with the guide from here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

And then you still need to increase your Windows Pagefile.
But first set the webui up.

[plush vine]

thx, i try it

[plush vine]

ROCM and Zluda – do I really need the latest versions? I have an AMD Radeon RX 6700 XT.

[ornate elk]

yea 5.7 is not recommended at all. 6.2 and 6.4 are better

would go for the 6.4 as its the latest

[plush vine]

i need uninstal previous zluda folder or can I leave it?

[ornate elk]

you need to uninstall hip sdk

and if the zluda is older than 3.9.5 then that too

[plush vine]

you mean uninstall this too?

[ornate elk]

uninstall everything from hip sdk, and then install only 6.4

then reboot the PC

[plush vine]

for me it's 6.2

[ornate elk]

oh thanks for pointing that out. as 6.4 was pretty new i have to change the guide on that part. but you can also just use 6.2

if you want to use 6.4 then here are the gfx files for ypur gpu: (gfx1031)
https://github.com/likelovewant/ROCmLibs-for-gfx1103-AMD780M-APU/releases/tag/v0.6.4.2

then the other parts are the same

[plush vine]

already starting downloading SD, if they don't running generate image I'll delete everything and repeat the process or delete HIP 5.7 and 6.4.
Also Hip deleting there?
Settings > Apps > Installed apps

[ornate elk]

yea exactly

[plush vine]

"if you want to use 6.4 then here are the gfx files for ypur gpu: (gfx1031)" ok will check it

[tawny ore]

@ornate elk guys, do u have some advices, i wanna make chibi-emotes for my twitch channel, i find good lora's and checkpoints, but one thing that worries me its: my emotes clothes change every render, how can i control it and make only change expressions and poses

[green plank]

Can somebody help please? I'm trying to get xformers to work. It told me it couldnt load python because it was built for a different version python + pytorch so i instled the versions it said and now a1111 is giving me a error saying it was built for python 3.10.6. What?

[tawny ore]

just install https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#nvidia-forge-neo

[tawny ore]

SD from AUTOMATIC is too old, havve the same issues couple of days ago

u cant fix it

just download this version and u can just copy ur loras and checkpoints

[green plank]

?? nvidia-forge-neo?

[tawny ore]

if u have nvidia GPU then yes

[green plank]

I wanted to use a1111 because of the extentions it has to offer

[tawny ore]

this SD is the same, but newer

[green plank]

Ohhh okay

[tawny ore]

with more options

[green plank]

Thank you ill give it a go then!

[tawny ore]

np bro, u can just try it

[green plank]

Yeah! Sounds cool! Does it generate faster too?

[tawny ore]

yeah coz xformer is working and all

The last time my A1111 was still working, the generations took about 10 minutes, but now it takes no more than 30 seconds. and its with Adetailer

[green plank]

Hmmm... The version of A1111 that im on right now, without xformers, can generate in 30 seconds haha

[tawny ore]

mb ur GPU is better

i have 3070ti

when I installed the a1111 and it still worked fine, I also generated quickly in 30 sec - 2 min max

[tawny ore]

guys help if someone is pro at it, coz im totally noob

[ornate elk]

try fix the seed, or use controlnet IP-Adapter to get the same style or color

[plush vine]

The checkpoint change takes about 20-30 seconds. Is this normal?

[ornate elk]

is your webui or model stored on an HDD?

[plush vine]

on SSD m2

[ornate elk]

okay then you need to do increase the windows pagefile now

Here is the Guide how to increase it:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Enable it only for C and make sure its not enabled for any other drive.
Set it to 16000 Min and 24000 Max for C
Then apply and reboot the PC.

Then it should be faster

[plush vine]

also controlnet downloading in this part of guide?

[ornate elk]

no thats just model linking.

Controlnet is integrated in Forge by default

[plush vine]

worked

[plush vine]

i doesnt have at default

[ornate elk]

which webui have you installed ? normaly its at the bottom

[ornate elk]

well thats not Forge

so it doesnt have controlnet by default

but there is an extension you can install

[plush vine]

have link?

[ornate elk]

click on the extensions tab, then click on Available, then on "load from"
then in the list search for
sd-webui-controlnet and there click Install

then relaunch

[tawny ore]

I don't understand this at all, I tried ControlNet a couple of times but nothing worked.

[ornate elk]

yea its pretty difficult

even with ip-adapter

[green plank]

ERROR: Could not install packages due to an OSError: [WinError 5] Access is denied: 'C:\AI\sd-webui-forge-neo\venv\Lib\site-packages\cv2\cv2.pyd'Check the permissions.

[ornate elk]

whats in your webui-user.bat?

[green plank]

@echo off

:: set PYTHON=
:: set GIT=
:: set VENV_DIR=

set COMMANDLINE_ARGS= --cuda-stream --sage --fast-fp16 --xformers --cuda-malloc --model-ref "C:\Users\Brandon\Desktop\SwarmUI\Models"

:: --xformers --sage --uv
:: --pin-shared-memory --cuda-malloc --cuda-stream
:: --skip-python-version-check --skip-torch-cuda-test --skip-version-check --skip-prepare-environment --skip-install

call webui.bat

[plush vine]

thx, work. Move 2 move also doesn't work at amd?

[ornate elk]

yea thats outdated

[plush vine]

Are there any other working video 2 video?

[ornate elk]

looks okay, can you show the full cmd error? also make sure you launched the webui-user.bat NOT as admin

[analog finch]

Did you get to test video gen? Are you also amd btw

[plush vine]

I had some incompatible SD version of Pony that worked with mov 2 mov.
But I don't remember which one.

[ornate elk]

it seems like you tried to install controlnet in forge neo, but forge neo has it already integrated

so it gives that error

delete the controlnet folder from the extensions folder

[green plank]

But the other verion has better inpainting tools, like the ability to erase mask

[ornate elk]

forge neo should have that too

if not then you would need to install an other webui like Auto1111 or Forge

[green plank]

facepalm I just had it workinging on a1111 but deleted it because i installed forgeneo lolol

[ornate elk]

gonna test now Q5_K_M with my 7900XTX

@analog finch worked! But it used 21,8gb vram xD

[analog finch]

Whaaat

how much ram you have

[ornate elk]

32gb RAM, the 7900xtx has 24gb vram

[analog finch]

why rx 9070 xt only get 16

scammed

was good price on it

ye vram*

Hopefully 2050 the free ram sites become real

[ornate elk]

maybe you should take a look at the Wan 2.2 5b models instead of the 14b

the 5b ones are much smaller

or you go for a smaler 14b one

[analog finch]

Whats the downside? lower quality or longer time

Can I do offloading somehow? tiling?

[ornate elk]

lower quality i guess and more fps needed

[lucid yoke]

wan generally needs 81 frames iirc

but yes lower parameters lower quality

both in quality, prompt adherence and complexity

[plush vine]

have this error by forge:
Traceback (most recent call last):
File "F:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\launch.py", line 54, in <module>
main()
File "F:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\launch.py", line 50, in main
start()
File "F:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\modules\launch_utils.py", line 675, in start
import webui
File "F:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\webui.py", line 23, in <module>
initialize.imports()
File "F:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\modules\initialize.py", line 29, in imports
shared_init.initialize()
File "F:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\modules\shared_init.py", line 19, in initialize
from modules import options, shared_options
File "F:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\modules\shared_options.py", line 9, in <module>
from modules.onnx_impl.execution_providers import get_default_execution_provider, available_execution_providers
File "F:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\modules\onnx_impl_init_.py", line 6, in <module>
import optimum.onnxruntime
ModuleNotFoundError: No module named 'optimum.onnxruntime'

[ornate elk]

have you followed my setup guide?, it gets explained at the end what to do, at step 7

[plush vine]

yeah, my bad

[ornate elk]

no problem

[green plank]

Is there a list of neoforge compatable extentions? or a repository llike a1111 has?

[ornate elk]

most work in forge neo. what doesnt work is default controlnet as its build in, and animatediff

forge neo has also the same extension list you can use for quick installing adetailer etc

[green plank]

Ok thanks boss!

This is my third time trying to install the inpaint anythig extension on forgeneo and its installs sucessfully, at least no errors in the console. But when i apply and reload theres no tab for inpaint anything, then when i restart console alltogether its still not there

@ornate elk

[ornate elk]

hmm the extension is outdated, but you can try the fixes from this thread:
https://github.com/Uminosachi/sd-webui-inpaint-anything/issues/154

[green plank]

darn thats not working for me either :/ not having fun with sd here today haha thanks for trying to help

[ornate elk]

oh ok xD no problem

[plush vine]

as I understand, I still need to reinstall torch?
venv\scripts\activate
pip uninstall torch torchvision /y
pip install torch==2.3.1 torchvision --index-url https://download.pytorch.org/whl/cu118

[ornate elk]

normaly not

[plush vine]

i have error. On oldest build im reinstall 2.3.1 torch. In this version sd i need reisntall this torch version?

[ornate elk]

this screenshot is from the forge nvidia guide, but your on amd

dont mix them up

[analog finch]

CS1o what do I need for wan 2.2 5b from here? https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B/tree/main