🤝｜tech-support

[ornate elk]

Can't check the metada from the first image
Can you reupload it?

12gb ram?
Then yes

[deft hemlock]

alright

there is a tutorial on reddit or yt right?

[ornate elk]

So they are different in the sampler used and in the cfg scale

And fp8 is enabled for sdxl in auto

And fp16 lora cache

[ornate elk]

Should be or here:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html

[deft hemlock]

Cfg shouldn't affect it that much

Sampler too

Right?

Unless something isn't visible in the Metadata

[zenith seal]

is it possible?

[ornate elk]

Yea Normaly not

But you can't generate the same images with different webuis

They work different in the background

[ornate elk]

I don't think so. Maybe in the config_ui.json

[night brook]

Hey all, this is probably classic. I downloaded flux.1 and get images like this
Details: I have an rx6700 12 gb card.
I close and re-open sd maybe 10 times and 1 out of 10 times, it renders properly.
It's the FLUX.1-dev q4 model I'm running. I can check the error messages when I get home, but is this a known issue?

[vocal burrow]

probably trying to render at a wrong resolution

possible too but more likely wrong resolution when it looks like that.

[night brook]

I understand, I always render at 512x512. Do I need a different res?

Also, thanks!

[vocal burrow]

hmmm the image you sent is 256x256 tho

512x512 should work iirc (havent messed with flux in months)

what settings did you use ?

[night brook]

Yeah, I copy-pasted from the internet, since I'm at work now. The very strange part, I see the render, it looks like that, then I get a 512x512 grey png as an output

I use default settings out of the box.

I up the image count to 20 though.

[vocal burrow]

"grey png as an output", what s your gpu ?

[night brook]

AMD Radeon 6700xt / linux / ryzen 2700 / 64 gb ram

[vocal burrow]

and what tool are you using to generate ?

[night brook]

Stable Diffusion, flux.1 model. What other info is needed?

[vocal burrow]

"Stable Diffusion" is not a tool. You re using auto1111, forge, comfyui, etc ?

[night brook]

Auto1111 thanks for the clarification.

[vocal burrow]

do you use directml or zluda for it ? what command line args are you using ?

[night brook]

I am using rocm + pytorch. No zluda afaik. Would pytorch use dml?

Sorry for the ignorance, but thanks, I am learning now.

[vocal burrow]

windows or linux ?

[night brook]

Linux

Yeah, I realized what I said now

[vocal burrow]

ok sounds good then. Probably gonna need the command line args you re using tho to go any further.
(there should be a line for them in webui-user.sh)

and a complete log of one generation

[night brook]

I will check in 8 hours. thnaks.

[ornate elk]

Auto1111 doesn't support flux

[vocal burrow]

welp that answers it then :p

[deft hemlock]

because its weird, a1111 suddently made my loras worse

[deft hemlock]

hey csio

do u think this is enough

i run stablediffuion on a portable drive (F)

should i do it to that too?

[ornate elk]

set it to 16000 min and 24000 max

and dont enable it for any other drive than C

then reboot

[deft hemlock]

It crashes now

A lot

I get a low memory error on my browser

[ornate elk]

In your Browser?

[deft hemlock]

yes

[ornate elk]

Where

[deft hemlock]

opera gx

[ornate elk]

Hmm okay, then try set it back to system managed for C and restart

[deft hemlock]

Ok

its back

i dont get the error now

but the generations are still messed up

it isnt suppose to be this detailed

with and without the masterpiece and source tags

[ornate elk]

Maybe because you have set them to be cached as fp16

Maybe revert that setting

[deft hemlock]

whats that

[ornate elk]

A setting

[deft hemlock]

isnt that a scheuler?

[ornate elk]

Nope

[deft hemlock]

does it affect generation that much?

[ornate elk]

Haven't tested it

[deft hemlock]

ok ill change the setting

cs1o im curious

what does ur username stand for

[deft hemlock]

anything here?

i would have ditched a1111 a long time ago and moved to forge if it wasnt for a specific extention that isnt on any other web ui, except for a111, its the queue extention and its a life saver

[ornate elk]

What's your GPU again?

[deft hemlock]

laptop mobile rtx 2050 with 4gb vram

i got 12gb of ram and the batfile has lowvram

[ornate elk]

Reforge maybe works too with that extension

[deft hemlock]

u should get it

itsa lifesaver

lets u queue different prompts with different settings

has support for the scripts, resolution batch count etc

[deft hemlock]

does it ?

do i need to delete the venv folder even for forge ?

cuz i just updated it

man....

now i cant see my models in forge

ok so basically

it adds this button

to add your current prompts seeds negatives resolution etc, into a queue

u can add and add

a1 a2 and a3 are presets for style loras, same settings and prompt all u have to do is add the lora u want

im not that tech savy

thats why i use this extension

plus

its easier to edit since it just adds a textbox below

and u can edit it if u have a change of mine about a specific thing within the queue

ye

very

thats why i didnt ditch a1111 for that spefic extention

very handy for lora testing

no flux

which is good

do u know how scary flux is ?

its good

really good

and i dont like that

how realistic the images can be

now imagine if bad people got their hands on it

especially scammers

[midnight kiln]

I got about half-way through converting this for Forge - I'll share if manage it all. It's actually a fairly trivial update in the sense that it just needed to have a few sections modified to conform to new standards in the more updated packages (pydantic ect.) I already finished that part. It's getting everything else down the dependency rabbithole to also play ball that is actually the challange. 5-6 dependencies deep into the friggin UI framework I'm not sure if I am qualified to follow the code anymore lol. It may be solvable by what I have already done + tweaking one of the package versions versions slightly so it doesn't break other stuff down the line.

I agree that life is painful without it - but right now Forge is the only one that even supports Flux for playing around with the new stuff

[reef swallow]

hello, I have a question stable diffusion forge or comfy ui works at good capacity with two gpu d (rtx 2080 ti)??

[lucid yoke]

swarmUI can utilize two gpu's by running two generations at once but you can't use two cards for one image yet unless someone developed it recently

[fossil stream]

Hey I know 99% of stable diffusion used nvida cuda cores, and I have an amd gpu. I had to install a special version of webui a couple years ago for amd, and I can no longer find it

https://github.com/LykosAI/StabilityMatrix

a friend sent me that one but idk if it will work on my amd gpu

[ruby yoke]

go to the pinned posts in this channel. look for @ornate elk AMD GPU guides. everything you need is in them

[night brook]

I am back, #export COMMANDLINE_ARGS="" So nothing. In fact the whole file is commented out!

[fossil stream]

Ty

[spare crow]

Hi guys, ive followed the guide to install sd on amd, added

--use-directml --skip-ort --medvram --opt-sub-quad-attention --opt-split-attention --no-half-vae --upcast-sampling

to the webui bat file but still getting the screenshot error

anyone know what to try now?

I tried adding --skip-torch-cuda-test into the webui aswell but no luck

[ornate elk]

Hey, What's your GPU?

[spare crow]

6950xt, I've uninstalled it all and followed your zluda guide instead so ill hang fire and see if this works!

[ornate elk]

But my zluda guide doesn't list the args you put in. You used --directml for example

[spare crow]

nope I think i followed a different guide, ive completely undone all of that and done the zluda instead

[ornate elk]

Ahhh

Okay

Did it worked?

Let me know if not. But with that card it should work easily

[spare crow]

It gave me a few errors but its finally loaded up the webui 🙂 do I just add models into the folder as normal with this then? its given me stable diffusion model failed to load in the cmd though

[ornate elk]

Can you show the full Cmd log?

[ornate elk]

I'm AFK, driving brb

[spare crow]

ok

[vocal burrow]

Do you have a log tho ? Meanwhile you can try uncommenting the line and setting --no-half argument in it. We seem to live in very different timezones.

[ornate elk]

Can you delete the model from the stable-diffusion folder and download an other model to test?

[spare crow]

I had no models downloaded, it seemed to try downloading something for me to use, I'll try that in a mo tho thanks 🙂

[spare crow]

@ornate elk downloaded a model and replaced it and still have this error

[ornate elk]

Zluda failed at start

Something is not correct with the path settings maybe

And did you download any gfx files?

[spare crow]

Gfx?

[spare crow]

Downloaded literally just the things in the guide

[karmic venture]

any idea why the quality is bad? like it's extremely bad, its blurry foggy, idk why

[ornate elk]

Can you show me the path settings?

[ornate elk]

Try an other pony model. The original is bad

[karmic venture]

i am open for suggestions :3

[spare crow]

where its all downloaded you mean ?
C:\SD-Zluda

[ornate elk]

No I mean the environment variable's path settings

[ornate elk]

Nova anime xl is good

[spare crow]

This?

[shut geyser]

hello where do i put ultralytics yolo v8 model?

no ultralytics folder in my comfyui

i do have the comfyui impact pack

[ornate elk]

Yep looks correct

Can you try delete the venv ans the .zluda folder from the stable-diffusion-webui folder?

Then relaunch

[shut geyser]

hey do you know how to get the ultralytics folder? i tried installing comfyui impact pact again in the manager but it doesnt let me uninstall

[ornate elk]

Delete the old impact pac and then reinstall

[spare crow]

@ornate elk reinstalled python and made sure it was add to path, reinstalled zluda as this was C:\Zluda/Zluda as id added a folder, going to see if this fixes the issues and try

[shut geyser]

installing it gives an error like this, still says its installed tho
but the folders arent there like ultralytics

[ornate elk]

Hmm idk then

Maybe the Ultralytics gets downloaded when needed?

[ornate elk]

Have you tried updating comfyui?

[shut geyser]

idk why does it think

my email is that

[ornate elk]

Hmm

[ornate elk]

Found this on the impact PAC github

[shut geyser]

i did in the comfymanager

[ornate elk]

HM okay

[shut geyser]

but still

comfyui doesnt wanna update

it thinks me email is THAT

but its not

idk how to make it know my real email

[ornate elk]

It doesn't need an email

Its a bug

Sometimes it can occur when you have nodes that got removed fron github like the Reactor extension

[shut geyser]

yeah i have reactor node

[ornate elk]

What you can try is a seperate Comfyui install in the SD-Zluda Folder

[shut geyser]

"comfyui-reactor-node"

is that it? do i just delete it?

[spare crow]

@ornate elk reinstalling all, or changing the file path to correctly seems to have fixed it mostly, it loads in and lets me get started now and seems to generate, although it only generates a grey square ?

[ornate elk]

Okay good but, the model is not a normal model

Please use any model that is between 2-7gb

For example juggernaut XL or Nova Anime XL

[lucid yoke]

you cant use hunyuan in auto-reforge right?

[shut geyser]

ima do it from your guide, with the zluda fork
its gonna be the newest version?

[ornate elk]

Yep updated a few days ago to the latest comfyui release

[spare crow]

❤️ you're a legend, seems to work perfectly now, thank you!

Failed to create model quickly; will retry using slow method. is that message normal though? still lets me do images once its done the slow method but idk if id need to do that each time?

[ornate elk]

Yep that message can be ignored. Doesn't affect anything

[ornate elk]

No problem 🙂

[spare crow]

perfect 🙂 thanks for all your help!

[ornate elk]

NP, let me know if you have any other questions

[spare crow]

I've got loads but ill leave you to help other people for now xD

[shut geyser]

i did it and

i can tell i had an outdated version

things look so different

but still, no folders
i have this

i just created the instand id model folder and it works but not for ultralytics

[ornate elk]

Hmm idk. Will check how its on mine later

[shut geyser]

its okay i fixed it

i cant believe i just created the folders

and it worked

🤣

InstantIDFaceAnalysis error 😦 amd

[terse hull]

this happens when i start a1111 with zluda

[ornate elk]

Have you tried it like 10-15 times like I suggested yesterday?

[terse hull]

i tried it like 3 more times and then i went to sleep

[ornate elk]

Having reactor and instantID won't work together is my guess. As I had massive issues when tried to use face id while having reactor installed

[ornate elk]

Try it more

[terse hull]

but it doesnt even say its compiling anymore

[ornate elk]

Its no joke 3 users with a rx580 reported it to work

[terse hull]

im trying to get it to work😭🙏🙏

[ornate elk]

How much ram do you have?

[terse hull]

i could have a possible source for it not to work tho

[terse hull]

16 or 32

idk tbh

[shut geyser]

i do not have it, remember? i moved uis
i made a new folder and installed every node from scratch
just moved my models folder

[terse hull]

maybe its because i use it on my secondary drive and my amd driver is installed on c:

[shut geyser]

would you like an error report?

[ornate elk]

No thats not an issue

Drivers always should be installed on C

[terse hull]

it says its compiling now

it seems to work

ill try it until it works

[ornate elk]

Hmmm okay, can you give me a full cmd output and the workflow?

[shut geyser]

(this is a very nice workflow)

[ornate elk]

Yep exactly. When it crashes, just start it again. Until its done.
After the compile is done it should work normaly

[shut geyser]

i updated pip, pytorch and some other things
now this comes up
what kind of styles.csv does it need?

[spare crow]

Is there a guide anywhere on any of the settings within sd that might be worth changing? and things such as what dampling method to use and when and so on? Trying to generate pixel art assets for a game and each time im trying with different settings and not noticing what is different

Hopefully to potentially stop me from getting errors like this when trying to make larger images also
OutOfMemoryError: CUDA out of memory. Tried to allocate 17.80 GiB. GPU
Time taken: 7 min. 6.3 sec.

[ornate elk]

Which settings did you used?

[ornate elk]

Ah well thats a to large resolution

Sdxl models are trained on 1024x1024

You normaly generate at half the resolution and then Upscale by 2 with hires fix to get Wallpapers

[spare crow]

Ahh so as long as I stick to under 1024x1024 I should be fine? What about sampling methods and everything else, should I stick them all as default?

[ornate elk]

Nope it should be around 1024x1024 but a bit lower and higher works too

Like 720x1280

[lucid yoke]

Aspect ratios i use a lot with good results:
1:1 1024x1024
4:3 1152x864
3:2 1248x832
16:9 1344x768
You can also reverse them to get portrait/landscape image

[spare crow]

Id just made a 960x540 so I can upscale it to the exact size, but the image seems to change when i click create an upscaled version, is that normal ?

[lucid yoke]

Depends on the hiresfix settings/refiner

Did you lock the seed?

[spare crow]

ahhh dammit, didnt even finish the upscale

OutOfMemoryError: CUDA out of memory. Tried to allocate 17.80 GiB. GPU
Time taken: 3 min. 35.1 sec.

id literally just clicked this ✨

[lucid yoke]

Ah i dont use that. I sort of forgot on how to use hiresfix in that UI but with the right settings it will just upscale without changing and will be a lot faster

[spare crow]

there a guide anywhere for what those settings are or what else I should be changing?

[ornate elk]

The denois needs to be 0.5 or 0.4 and you need the tiled diffusion extension.
That prevents it from going out of memory

[spare crow]

OutOfMemoryError: CUDA out of memory. Tried to allocate 17.72 GiB. GPU
Time taken: 3 min. 26.5 sec. still fails, assuming I added the extension correctly, i reloaded the webui with it but still fails

[ornate elk]

After you installed it you need to check the tiled VAE option

Also don't run anything in the background like games or Wallpaper engine or twitch

[spare crow]

it works but only if I turn the resize by scale from 2 to 1.5 which obviously means the image is smaller, not a massive drama tho

[lucid yoke]

i didn't see it earlier but you are on a nvidia gpu right?

[spare crow]

amd unfortunately

6950xt though so not a bad card id say

[ornate elk]

yep its fine

if it works with 1.5 thats good. if you use a smaller res it will work with x2

[lucid yoke]

just wondering why its flipping out over a 2x upscale with thise resolutions though with that much vram

[spare crow]

ive got literally nothing running in the background either

[ornate elk]

is your driver updated?

[spare crow]

yup, although even 1.5 scale isnt working now after coming back to it again

OutOfMemoryError: CUDA out of memory. Tried to allocate 9.97 GiB. GPU
Time taken: 2.3 sec. and its trying to use less

[ornate elk]

can you show the settings again?

[spare crow]

sec just restarted the ui

gimme a mo

[lucid yoke]

i feel like it has to be the settings. i ran a 1152x896 into a 2x and a 2x again (turning it into a 4608 x 3584) with only 8gb vram those times of 3min are also quite long

it took me 54 for the final refine and 18 for the first one. any special kind of upscaler model your using?

[spare crow]

Okay the restart seems to have helped, it did 1.5 this time

im just dragging the image i liked into img2img and upscaling it from that, idk if thats how im meant to do it or not though, im a complete noob at this

[ornate elk]

you can use the sd upscale script in img2img for that too

[spare crow]

theres so much to this its actually insane 😂 ill give that a try aswell

yup, failed trying the upscale by 2 again

[spare crow]

54 😮 I feel like if it goes past a min for me its doomed to fail xD

[ornate elk]

you should also use the fp16 sdxl vae

it reduces the load a bit and the gen time

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main
its the sdxl.vae.safetensors

[spare crow]

🤯 I dont even think I understood what you said let alone doing it 😂

is it another extension ?

[ornate elk]

nope its a file xD

first go into settings, User interface, quick settings, and there add sd_vae
then hit apply and reload ui.
download the file from my link (sdxl.vae.safetensors) and drop it into models/vae

[spare crow]

here?

and in here?

[ornate elk]

models/vae

[spare crow]

ive only got the stable-diffusion where i put the models and the VAE, do i just put it into both or am I looking in the wrong place?

[ornate elk]

screenshot of your screenshot xD

[spare crow]

Ohhhh fml my bad 😂😂 I was looking for models and VAE 😂

[lucid yoke]

or use this link C:\SD-Zluda\stabe-diffusion-webui-amdgpu\models\VAE

[spare crow]

I'll try it out in a few mins! Sorry just gotta pop out

[spare crow]

Okay added this, wish me luck!

[spare crow]

eughh I dont get why sometimes it gets to 95% or so before it decides it will fail =.=

idk why it fails last min tho

Surely it can decide, I wont request any more, ill just take longer to do it ?

or am I just being naively optimistic xD?

I thought I had a reasonably good gpu aswell xD didnt think id be having these issues

[ornate elk]

You always need to have tiled VAE enabled

When using highres/upscale

Also in img2img use SD upscale script not resize by

[deft hemlock]

Nan

[ruby yoke]

https://www.quora.com/My-computations-on-my-TensorFlow-Graph-keeps-giving-me-NaN-what-are-good-debugging-advice-to-discover-what-the-cause-of-this-is

[soft nest]

guys, im making lora for a style but im not sure which model to use. should i make it more sd 1.5 or sdxl??

[spare crow]

made all these changes this morning, still getting the same issues

[soft nest]

im basically training 1 art style using 71 images

[shut geyser]

amd zluda a1111 i used the guide in pins

[ornate elk]

Can you show the full cmd log?

[shut geyser]

i just installed the directml a1111

it works for now

[ornate elk]

Ah alright

[shut geyser]

i only am installing it to make captions for images

for a lora dataset

[ornate elk]

OneTrainer can do captions too

In the tools section

[shut geyser]

well then i did it all for nothing xD

[shut geyser]

amd for this ai stuff dude, kill me :D

[ornate elk]

Can you show what you try to do again?

[shut geyser]

comfyui lotus giving an error when trying to be installed

it shouldnt conflict with any other node

same with searge llm
Error message occurred while importing the 'Searge-LLM for ComfyUI v1.0' module.

Traceback (most recent call last):
File "C:\SD-Zluda\newest comfyui\ComfyUI\custom_nodes\ComfyUI_Searge_LLM\Searge_LLM_Node.py", line 13, in <module>
Llama = importlib.import_module("llama_cpp_cuda").Llama
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\Lib\importlib_init_.py", line 90, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "<frozen importlib._bootstrap>", line 1387, in _gcd_import
File "<frozen importlib._bootstrap>", line 1360, in _find_and_load
File "<frozen importlib._bootstrap>", line 1324, in _find_and_load_unlocked
ModuleNotFoundError: No module named 'llama_cpp_cuda'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "C:\SD-Zluda\newest comfyui\ComfyUI\nodes.py", line 2145, in load_custom_node
module_spec.loader.exec_module(module)
File "<frozen importlib._bootstrap_external>", line 995, in exec_module
File "<frozen importlib.bootstrap>", line 488, in call_with_frames_removed
File "C:\SD-Zluda\newest comfyui\ComfyUI\custom_nodes\ComfyUI_Searge_LLM_init.py", line 1, in <module>
from .Searge_LLM_Node import *
File "C:\SD-Zluda\newest comfyui\ComfyUI\custom_nodes\ComfyUI_Searge_LLM\Searge_LLM_Node.py", line 15, in <module>
Llama = importlib.import_module("llama_cpp").Llama
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Python312\Lib\importlib_init.py", line 90, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ModuleNotFoundError: No module named 'llama_cpp'

is searge llm for nvidia only along with lotus?

[ornate elk]

llama.cpp can't work with zluda.

Zluda currently can't "translate" every feature that cuda has

[shut geyser]

what about lotus?

"comfyui - lotus"

[ornate elk]

Idk, haven't tested it. But I know depth extensions work in auto1111

Or controlnet works for comfyui too

Has depth and normal map and more

[thorn badge]

In forgeUI, does anyone know a way to keep PNG info (prompt etc) in upscaled version i.e. I lose it when i upscale in img2img?

I am adding it to imgtoimg from folder, not moving it using "painting icon" after txt2img

[ornate elk]

I think there is no way if you add I'd from folder

[ashen gate]

Does Flux on Forge not allow for negative prompts? The negative prompt section is grayed out.

[lucid yoke]

Cfg on 1 = no negatives

And flux wants cfg of 1

[night brook]

How can i find the full logs? Is there an option from the web-ui. When I launch webui.sh, is it just the output to my terminal?

I am sick so in the correct timezone now.

[ornate elk]

Yep, its not available for flux

[supple orchid]

I need help! Whenever I use even a single LoRA in my prompt, my GPU usage drops significantly. Specifically, CUDA utilization gets stuck at around 10–20%, whereas without LoRA, it runs at nearly 100% full load. As a result, the speed difference is almost 10 times or more. For reference, I’m using a 4090 (24GB) with WebUI. Does anyone know a solution to this issue or if it's a known problem?

[shut geyser]

onetrainer flux training, this happens, i already have gguf installed (14.0) and i already have pytorch 🤔

[fierce moat]

bruh dont think gguf work for training

[shut geyser]

.safetensors also dont

[night brook]

Ok, after many reloads, flux.1-dev-uncensored-q4 loads I don't understand why, but it "works" now. I am sure I will have to re-load until it works again after a shutdown.

[shut geyser]

and i am logged in on chrome

[tawny saddle]

i think u should try Opera GX, the sponcer of todays video

[spare crow]

Why do some images come out like this? Some checkpoints just always seem to do this for me even if they are bigger and proper checkpoints, this was with artsyvibe

[glad jackal]

wrong vae issue

[spare crow]

this?

[glad jackal]

yeah you need ae.safetensors for flux1d checkpoints

[spare crow]

how do you know what you need for different checkpoints? does it say somewhere on each checkpoint? and also where do I go to get the different VAE?

[glad jackal]

actually checkpoints have vae, clip and unet included so you should have it already

[ruby yoke]

flux does not use negative prompts. and it does not use CFG

[spare crow]

thats all the ones i have

[glad jackal]

you actually can use cfg and negative prompts to good effect, it just wasnt designed for that use.

[glad jackal]

try none or automatic bc when you load the checkpoint it should have vae included

[spare crow]

tried automatic and none and still didnt work

[ruby yoke]

regardless of whether you can use it or not, it was designed not to allow negative prompts and bfl uses something different than cfg. that's why it is set with cfg not avaialble and no access to negative prompts becase negative prompts need classifier free guidence - which flux does not use.

[spare crow]

https://civitai.com/models/1162948/artsy-vibe this is the 1 ive been trying to use, although theres a few that dont seem to work with me, this says checkpoint and base model so should be okay right?

[ruby yoke]

you can also use a teaspoon to dig a hole in your backyard, but that isnt' what it's designed for

[glad jackal]

idk much about forge, i just know it barely does regular flux. and this is also a checkpoint merge

[spare crow]

idk what forge is, sorry im a noob! Im on sd with zluda if that means anything

[glad jackal]

you can definitely increase cfg beyond 1 and use negative prompts, all im saying. it does increase generation times but it definitely has a profound effect on image composition/etc

[ruby yoke]

forge is an interface to run the AI models in. so are auto1111 and comfyUI

[ruby yoke]

regardless of whether you can do that or not, flux does not use it - een if you set it to 100, flux itself does nto use it. comfyUI might be doing stuff to modify things when you do that, but the flux model itself does nto use it

[glad jackal]

lol im telling you it can use cfg

different from flux guidance

[ruby yoke]

and I'm telling you that while comfyUI might modify your out put, it's not the flux model that's using it. Robin Rombach, the author and inventor of both stable diffusion and flux, designed flux to use something other than classifier free guidance (cfg)

so when you turn on cfg, and force negative prompts, you are pretty much BREAKING flux

might as well just use sdxl

[glad jackal]

works great for being 'broken'

[spare crow]

So how do i fix it not working? I tried it with flux checkpoint aswell and that didnt work either

[ruby yoke]

yes, well - this discord isn't the black forest labs flux discord. it's the stable diffusion discord. so why is it even being discussed here?

[ruby yoke]

what is it specifically that you are trying to do?

[glad jackal]

no idea, you keep arguing lol

[spare crow]

Just trying to get pics to load with different checkpoints, just trying a bunch out but half of them give me that pixel artifact look and nothing else loads

[ruby yoke]

this?

[glad jackal]

https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
heres the vae for flux models, unsure if this will help

[spare crow]

yup

[glad jackal]

usually that specific issue happens when you use the wrong vae but youre using a checkpoint which has should have the vae included

[ruby yoke]

yeah. you are using the wrong Variable Auto Encoder (VAE) for the model or checkpoint you are tyring to use

if you're using flux, you have to use the flux VAE

[glad jackal]

hes using a checkpoint though

[ruby yoke]

if you're using sdxl, you have to use a VAE trained for sdxl

[ruby yoke]

so? it's trained for a base model of some sort. what base is it trained for?

and does it have a VAE baked into it

if it does, you don't use a second vae

[spare crow]

so how do i know which vae i would need in general?

[glad jackal]

bc its a checkpoint merge? lol

[ruby yoke]

not in general. it's specific to the model or checkpoint you are using

[ruby yoke]

so? which models were merged? there's a base model at the bottom of that or there's no checkpoint at all

where did you get it from?

[glad jackal]

civit

[ruby yoke]

do you have the link to it?

[ornate elk]

Flux is not supported on Auto1111

[glad jackal]

lol thank you^^

[ornate elk]

You would have to use Forge Webui or Comfyui to use flux

[lucid yoke]

Swarm is great with flux too***

[ruby yoke]

Swarm runs comfy in the background

[spare crow]

Ahhh ok, thats a pain! Can I use auto1111 and forge or comfy on the same pc ?

[ruby yoke]

yes. but why?

[lucid yoke]

Yes but since it has a different front end its nice to mention

[ornate elk]

yep and you can also share the models between them

[spare crow]

Id like to try a bunch of different checkpoints so if I cant do it on 1 id like to try on the other

[glad jackal]

yes definitely, they each have their own usecases

[ruby yoke]

you should probalby take @lucid yoke suggestion, install SWarmUI - and let it do all the technical stuff. and just focus on generating

[spare crow]

damn remembering what checkpoint works on which is a pain, can i not just run everything from 1 interface?

[ruby yoke]

sure. use comfy.

[spare crow]

yeah ill probs try and do that if its easy to do! is there a guide somewhere?

[lucid yoke]

Swarm does, comfy too but its more confusing

[ornate elk]

Forge supports more models than auto1111, and has the same ui

[lucid yoke]

Swarm is drag and drop but once you feel confident enough to mess with comfy you can interact with the backend

[ruby yoke]

but you still have to load the right stuff for the models you are loading. when you download a model, there will be an area on the page that tells you what base model it was trained for and if you need a vae, it'll tell you that too

[ornate elk]

whats your gpu again?

[spare crow]

amd 6950xt

[ornate elk]

okay then you can use any webui from my guide

would still recommend forge/auto above comfyui for beginners

[glad jackal]

beginners need more advanced node support to make things easier

[ruby yoke]

that's why i was agreeing with Euro and suggesting swarm.

[glad jackal]

trying to pickup old ui with no support can be difficult when the last people talking about it are from a year+ or more

[lucid yoke]

Beginners need a drag and drop method where they can simply type a prompt and not worry about anything else tbh

[spare crow]

especially me, im new to all of this and its blowing my mind xD

[glad jackal]

google tools like imagen work well for learning prompts

just strictly prompt 2 image

[ruby yoke]

there's a steep learning curve, just take it one step at a time. you'll get the hang of it

[glad jackal]

local is prob a headache/not necessary for beginners lol

[lucid yoke]

I mean, civit ai image gallery is a great way to learn prompts for specific checkpoints

[spare crow]

Im kinda okay for the prompting side of things, when i find a checkpoint that works and i know how to make it work the images that come out are really cool and end up nice and clean so im happy with that

and like ill work using civit gallery and find images i really like and download the checkpoints and loras to try and see what styles I can make with it which is what ive been trying to do, just half the ones ive dled so far dont work for me

[glad jackal]

see for his case he needs to be able to run the latest things so i feel like a1111 would be a huge step in the wrong direction

[lucid yoke]

I wish A1111's page got a big ol warning saying "I dont get updates, see a list below for alternatives" to prevent many such cases

[spare crow]

so comfy or swarm is your recommendations then?

[lucid yoke]

SwarmUI since that way you get both

[spare crow]

with zluda or directML?

[lucid yoke]

And most ||all|| options got a ? Next to it showing you what it means

I recommend following the guide of Cs1o in pinned msg

I can't troubleshoot amd for the life of me

[spare crow]

yeah im on the guide, its just there is with zluda or directml and I have no idea what the difference would be

[lucid yoke]

Zluda is better, directml is slower i heard

[ruby yoke]

you honestly don't want to use auto1111 - it's not really being supported, it's not a very good ui. since you have to learn this from the ground up, please don't spend your time learning something you wont' really be able to use for much.

[spare crow]

ill try that then 🙂 will this cause me huge issues with the auto1111 still on my drive or should I delete all of that?

[ruby yoke]

just leave it where it is. install comfyUI, and then fill out the extra path's file with the paths to where your models and stuff you downlaoded for auto1111 are at

and watch scott's tutorials https://www.youtube.com/watch?v=AbB33AxrcZo&list=PLIF38owJLhR1EGDY4kOnsEnMyolZgza1x

[spare crow]

Thanks 🙂 Ill look into it soon! You guys are stars ❤️ thanks a bunch!

[ruby yoke]

and read @ornate elk 's AMD gpu guides...

[ornate elk]

always choose zluda

[spare crow]

Yup will do:) that vid is going to probs need to be a favorite to set up that each time xD

[flat solstice]

tbf.. I only switched to comfy today as I've heard trying it all out ( controlnet wise on sdxl ) on webui / forge ( which I prefer ngl. ) is a bad idea or simply doesn't work as expected ( yep, same stick figure baking or whatever you wanna call it )

[old widget]

Does anyone know if Nvidia NV link is supported by stable diffusion? Is there any point in installing NVLink?

[fossil stream]

I run an amd system

[fossil stream]

I have 12 gb of vram

[deft hemlock]

Try adding
--skip-torch-cuda-test

[fossil stream]

yeah that worked

ty

[ashen gate]

Do you know if StabilityMatrix automatically updates packages for Forge and Automatic1111?

[deft hemlock]

Depending on what you have in your bat file I think

Usually they have a sperate bat file that u run to update

[ashen gate]

not sure where that'd be located.

[knotty turret]

@ornate elk can we try tmr start the conversation and finish it we always dont finish man😭

[misty marten]

When you use Zluda, this error appears, with Directml it generates images normally, can someone help me?

[ornate elk]

What's your GPU?

[ornate elk]

Can you show the full cmd log?

Also what's your GPU?

[zenith sinew]

I feel stupid for asking this. I have tried to move over to comfy a million times but I never feel like I'm getting the same results I get using WebUI for "hiresfix". I want to use an ESRGAN model as part of that, so decode into image space, then scale up using a loaded ESRGAN model. I think with the nodes I have I also have to scale the image back down if I don't want to use the default multiplier of the ESRGAN model.

Then, I pass the image back into latent space and feed it back into another k-sampler with a low denoise and the same text conditioning.

Am I missing something? Is that really all hiresfix is in WebUI?

[near isle]

Hi, can you tell me how to fix artifacts when generating images? Adjusting the image with the parameters

[ornate elk]

Hey this is a VAE issue

[near isle]

At the generation stage, the image looks in the preview without artifacts.

[ornate elk]

1.5, sdxl, flux etc have their own vae

[near isle]

I used klF8Anime2VAE_klF8Anime2VAE

[ornate elk]

Thats for 1.5 based models

What's the model you use?

[near isle]

zukiCuteILL_v50

How can I check the compatibility of the model and vae?

[ornate elk]

Ah thats an illustrous model.
Pony and illustrious are based on sdxl

[near isle]

Which vae are suitable for this model?

[ornate elk]

So you need an sdxl VAE like this:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors

Thats the one I use mainly. Works for every sdxl, pony, illustrious models

[near isle]

Okay, thanks, I'll try with her.

What settings can be used to speed up generation? rx 580 2048 sp on 8gb vram, when installing medvram, there is not enough video memory even for the minimum resolution, and with lowraw it generates for a very long time even with the minimum resolution. I put a 30gb swap file.
Webui

[ornate elk]

On windows?

Do you use DirectML or Zluda?

[near isle]

Windows/DirectML

[ornate elk]

Yea that won't be fast with any larger model.
Would recommend using my install guide with zluda

That is much faster

[near isle]

How is zluda better than webui?

[cold kernel]

Is there a way to turn a non vpred checkpoint into a vpred one?

[ornate elk]

Zluda is better than directml. Both use the same webui.
Zluda uses less vram and is faster as it "emulates" cuda

[ornate elk]

I don't think so

[cold kernel]

😭😭😭😭!!!!

[near isle]

For the rx 580 2048sp, you only need to install ROCm HIP SDK 5.7?

[ornate elk]

2048 native wouldnt work (the models are trained on 1024x1024)

but you can generate at lower res and upscale

[near isle]

Can't Zluda be set to 2048sp?

[ornate elk]

it can but your vram is not enough for that native resolution

[near isle]

Is 8 GB not enough?

[ornate elk]

not enough for 2048x2048 yes

[near isle]

1024x124 is quite enough, but if I put that much, it will take an extremely long time to generate

2048 is the name of the graphics card, I'm not talking about resolution, maybe there was a misunderstanding.

[ornate elk]

yes because you use directml. its really slow

[ornate elk]

what is a 2048 ?

[near isle]

I'm trying to figure out how to install zluda.

[ornate elk]

and yes i tought you talk about resolution

[near isle]

Radeon rx 580 2048sp

[ornate elk]

ohh this

https://www.techpowerup.com/gpu-specs/radeon-rx-580-2048sp.c3321

4gb vram

that is not easy to work with even with zluda

[near isle]

I have an 8 vram graphics card. rx 580 2048sp (8gb vram)

[ornate elk]

ahh okay

then yep go for zluda and it will work with the illustrious models

[near isle]

Does the processor have a strong effect on the generation rate? Or maybe RAM?

[near isle]

Now I'm trying to figure out how to install it, there's too much in the instructions.

[ornate elk]

cpu nope, but RAM is important for sdxl based stuff.
If you have only 16gb or less you have to increase the windows pagefile to 16000min and 24000max for the C drive

[shut geyser]

hey does any of you know can you run fluxgym on amd and train a flux lora there?

[near isle]

I have 16 GB of RAM and I have installed a 30 GB swap file.

[ornate elk]

thats good then

[fossil stream]

6750 xt, I fixed it tho

[ornate elk]

ah okay, i hope you use zluda and not directml as backend. You would miss out a lot of performance.

[near isle]

Where should I put this folder? to the ZLUDA folder?

[ornate elk]

nope

[near isle]

Just to a new empty folder?

[ornate elk]

can you reload the page, there are words missing in your screenshot

thats how step 3 should look

[ornate elk]

yep now it looks correct

[near isle]

Okay

[ornate elk]

you need the files for step 3

there you see what to do with them

[fossil stream]

zluda isn't in the options

only cpu, cuda, DirectMl, Rocm, and mps

[ornate elk]

yea its not in options, its an install method you have to do to get the best performance with AMD on Windows

directml instead is really slow and runs out of vram constantly

[fossil stream]

yeah would you mind helping me find it then?

[ornate elk]

sure but its not on pinokio

i made AMD Guides for Zluda and DirectML. You can find them here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[fossil stream]

I'm using stability matrix, idk what pinokio is

[ornate elk]

oh right, its nearly the same

both dont have the correct amd setups

only the slow and old directml install

[fossil stream]

is there any way to patch it in? cuz everything else is really simple

[ornate elk]

its not recommended, you could do it but it could break more easily

My setup Guide is a bit more complex but stability matrix gives AMD Users no advantage anyway

so better remove it and then do a clean install

[near isle]

He doesn't want to ship models to zuda. The built-in 1.5 couldn't boot at all

[ornate elk]

its still compiling... xD

you have to wait

give it 15-40 minutes

[near isle]

There was no such thing in directml, is this how it should be?

[ornate elk]

yes

zluda needs to build a cache file

[near isle]

Okay, But the built-in 1.5 model just gave an error, and mine is still loading.

[fossil stream]

which webui do you personally recommend for amd users

[ornate elk]

with 12gb vram i would recommend Forge Webui or Auto1111 with Zluda

you get the best performance and usage on them

you can still install comfyui later too if you want. but as start its not recommended as its more complex

[fossil stream]

how complex

I installed the og webui a couple years ago, can't be as complicated as that process

[ornate elk]

its a different ui system (node based)

[ornate elk]

Forge and auto1111 have the same webui you already know

[ornate elk]

try restart the webui and then let it load until its done, also add --medvram to the webui-user.bat

[fossil stream]

is there any benefit to comfyui other than personal pref

[near isle]

Another model is being loaded

I have set up such settings according to the manual

[shut geyser]

i think more complex posibilities the workflows can give you

[near isle]

error(

[ornate elk]

forge and auto1111 are better to optimize for the vram usage than comfyui.
you can easily upscale on them as they support some extensions to lower vram usage.
Comfyui supports more complex workflows and more models and you can change anything you want in the process. but that also gives you unlimited possibilitys of errors that can appear xD

[fossil stream]

goooootcha

[ornate elk]

on rx580, the webui needs a few restarts like 15-20 everytime you get that error

at the end it will launch

[fossil stream]

I'll install it in a bit, currently cracking my adobe installs cuz adobe wants 60 bucks a month for their shitty apps

genp my beloved

[ornate elk]

they are crazy with their pricing wtf

[near isle]

Do I need to restart 15-20 times?

[fossil stream]

I'll update you on what happens when I install it

[ornate elk]

can you first show me a full cmd log?

[near isle]

Do I need to select a model at startup so that it starts loading it?

[ornate elk]

sure, let me know if you have any questions about the steps

[ornate elk]

in the dropdown on the left

[near isle]

I've already closed it, and I'll delete it next time I make a mistake. Should I attach screenshots?

[ornate elk]

yes please

then i can see whats causing the error

[near isle]

It's closed again

[ornate elk]

that is not correct, but could still have worked

[near isle]

Can you show me how to do it? I added what was in the manual.

[ornate elk]

you can remove the part you added

then relaunch

[near isle]

Right?

[ornate elk]

remove the -sdxl

so that its nonly --medvram

then save and close

then do this:
Additional Step ONLY for RX570/580 Users:
Go into the main folder of the webui and click in the explorer bar (not searchbar)
There Type cmd and hit enter and then run these three commands one by one:
venv\scripts\activate
pip uninstall torch torchvision torchaudio -y
pip install torch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 --index-url https://download.pytorch.org/whl/cu118

[near isle]

I didn't really understand where to enter these commands. To the console in the folder?

Ah, yeah

[ornate elk]

in the cmd you open by typing cmd in the explorer path bar and then hit enter

[near isle]

Yes, I saw it

"the system cannot find the specified path"

its okay?

[ornate elk]

now the next 2 commands

one by one

[near isle]

I'll do it now

[mental tide]

Hey all, I've recently formatted c:, but I have all my stuff on a P: drive (programs) for SD. I did a quick reinstall of python, git, and downloaded SD via: "git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git" but I can't get GIT to be reconised by my path:

[near isle]

Is this a problem?

[ornate elk]

no

run the last command to install the torch stuff

[near isle]

Okay

[ornate elk]

hey, try reinstalling git and then do a reboot, then check your environment variables path setting

[near isle]

I'm trying to launch

Run it over and over again?

[ornate elk]

yep

everything is setup correctly

now you need to relaunch it a few times until the compile in the bckground is done

[near isle]

Okay, I'll try.

[ornate elk]

all rx580 users that followed my guide relaunched it like 15-20 times then it worked

the first image will also need to compile again, but after that your fine

[near isle]

15-20 times😱

)

[ornate elk]

yep 😄

[mental tide]

I have GIT in my path, but I keep getting:

[ornate elk]

hmm thats strange and you reinstalled it too and rebooted the PC ?

[mental tide]

Yes sir. If I had hair, I'd pull it out at this point.

[ornate elk]

can you show the git in path?

[mental tide]

sure one moment!

[ornate elk]

ahh

you installed it on P

[mental tide]

correct

[ornate elk]

you have to uninstall it and reinstall it on its default path on C

stuff like python and git should always be on the OS drive

to not get wierd errors

[mental tide]

gotcha, Python worked fine, just Git bein' complicated

[ornate elk]

then also create a folder and name it like Ai or soemthing and then git clone into that

that mitigates permission issues

[mental tide]

seems like permissions issues now... sonofa

[ornate elk]

While installing on C or what?

[mental tide]

trying to run SD

owned by: inconvertable ...

[ornate elk]

Oh

Did you created the Program Files folder by yourself?

If its a system created folder then it wont work

[mental tide]

yeah system created...

I could move everything over to a backup drive, and then slick it and move it all back...

[ornate elk]

Thats why I said to create a folder called ai

[ornate elk]

Yep you can backup the drive and reformat it

[near isle]

It's been loading without errors for about 30 minutes, is everything okay?

[ornate elk]

yes looks good

its nearly done

[mental tide]

@ornate elk formatted P: coppied data back, loading no problems.

Thnk you!

[near isle]

@ornate elk Oh, you really earned it. And the speed has become much faster, it also now works with medvram, thank you very much.

[fossil stream]

@ornate elk my sleep deprived brain thought I did steps 1 and 2 and skipped straight to 3, where would automatic1111 go if my dumbass possibly forgot to make a folder for it

[tawny saddle]

happy valentines day everyone :3 💜

[ornate elk]

perfect no problem 🙂

[ornate elk]

You can still make a new folder

[fossil stream]

I figured, just didn't know where it installed otherwise so I can delete it

[torn blaze]

anybody had problems wioth rtx 5080?

Ryzen 5800x3D
RTX 5080 16gb
32GB RAM

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]

If you opened the Cmd without not in the folder its probably in system32

Or on C directly

[ornate elk]

Hey, make sure your webui is updated.
Then delete the venv folder and relaunch

[fossil stream]

ty

[ornate elk]

If its there delete it. Then start with step 1

[fossil stream]

so I'm on this portion.

found my gpu here

but what exactly do I do here?

do I just get this one?

[fossil stream]

it wasn't in system32 or the c drive so I just installed the regular one on my nvme drive

[ornate elk]

Yes

Gfx1031 is for your gpu

[fossil stream]

ty

[timid dome]

Hello everyone,
I’m sorry if this has been brought up before, but I’m experiencing an issue with the upscaler in the extras section
They just stay on uploading forever, can someone land a hand to this lost soul?

[ornate elk]

Hey make sure the webui is whitelisted in any browser adblocker

[ornate elk]

Np, let me know if it worked

[timid dome]

I did it but same thing happened

thank you for helping btw

[fossil stream]

do I only need the first one

and which of these do I click for it, I tried the rocm6 zip

I'm assuming the zip but I just wanna make sure

[ornate elk]

Leave all default

You need all but not the bottom (pro driver)

[ornate elk]

Yes only rocm6

Not nightly

[fossil stream]

alr

[ornate elk]

What does the cmd shows when you upload them?

[spare crow]

Okayy so I installed comfy ui as per instructions and everything seemed to work fine, except it gets to 29% of an image, closes the cmd window and just says reconnecting in the web ui

[fossil stream]

@ornate elk

haven't gotten to generate anything yet

[fossil stream]

just keeps saying compilation in progress

[dry depot]

hey guys

having some difficulty with connecting stable diffusion to owui

also both are running on docker containers.

[near isle]

After several generations, such an error occurs and the generation does not work until it is restarted

[ornate elk]

Can you show the workflow you tried when it closes?

[ornate elk]

Should work after its done

[spare crow]

Will as soon as I get home:) CMD closes though so I have no logs for that

[fossil stream]

Does it just take that long?

[ornate elk]

Yep first time only

[fossil stream]

Gotcha, thought I did something wrong

[ornate elk]

Did it worked by now?

[little bough]

Hi does anyone know if its possible to manually add Samplers to Forge? I like the 'Gradient estimate' sampler that is in Reforge but it does not come with Forge, is it possible somehow to copy into the folder or install it somehow into Forge? Thanks

[spare crow]

@ornate elk this is the log before I try doing anything in the webui

Ive tried this with a bunch of different models, it just does this each time

[ornate elk]

Numpy error:
Numpy fix:
If you get a NUMPY Error do the following:
Install the Comfyui-Manager from below, then Click on Manager > Click on Install PIP Packages > type numpy==1.26.4 > Click OK and restart the server as prompted.
If you get a security warning you have to go into ComfyUI\user\default\ComfyUI-Manager and there edit the config.ini.
Change the Security Level to Weak, then save and try install numpy again.
After that set the Security Level back to normal.

[spare crow]

Ahh god dammit, I thought that error would pop up in the webui, didnt even see it there! Sorry my dude 😫

[ornate elk]

no problem, yea its only in the cmd

[spare crow]

Im never too sure what im looking for in the cmd so I kinda avoid it as its all jargon to me xD

@ornate elk this when I type in numpy==1.26.4

ERROR: To use this feature, you must either set '--listen' to a local IP and set the security level to 'normal-' or lower, or set the security level to 'middle' or 'weak'. Please contact the administrator.
Reference: https://github.com/ltdrdata/ComfyUI-Manager#security-policy

How do I do that?

[ornate elk]

ready my message from above again (the fix)

[spare crow]

doh! Sorry dude! Feel free to tell me to piss off xD

Still closes at 29% even after the pip install thing and rebooted through cmd and everyhing

Same reconnecting error even without it saying numpy error

[ornate elk]

ohh it tries to use your cpus igpu instead of your dedicated gpu

[spare crow]

god dammit xD how do I tell it to use my other?

[ornate elk]

you can disable the Radeon TM Graphics in the Device Manager of the PC.
or you can try edit the launch bat and at the end behind --auto-launch add --cuda-device 1

[spare crow]

Changed that now, still stuck at 29% but CMD hasnt closed so im assuming this counts as the first time where it can take like 50 mins or so ?

[ornate elk]

yea the compile will take a bit again. but it wont show it in the cmd

so if nothing crashed, let it load

[spare crow]

Alright 🙂 Ill leave it running and hope for the best! Ill let you know how it gets on, thanks again dude you're the goat for real!

[ornate elk]

which option did you took?

disabling the igpu or the edit of the bat?

[spare crow]

--cuda-device 1

[ornate elk]

alright

[spare crow]

everything else on my pc is set to use the dedicated so i doubt it would make much difference

but thought this seemed easier

[ornate elk]

yep

[spare crow]

crashes out again =[

log before the crash

[ornate elk]

how much RAM do you have?

[spare crow]

32gb

[ornate elk]

okay thats enough

and when does it crash?

after you hit generate or just by launching?

[spare crow]

at 29% always, its when its green on the bottom negative prompt

so assuming its when its loading that section

[ornate elk]

is it empty?

[spare crow]

Yeah, it wasnt yday but I cleared it to see if that was the issue

[ornate elk]

hmm should work no matter what

[spare crow]

just tried it with some info in it and still crashes out

should I try making a new workflow and starting again with it all? or is that unlikely to make a difference?

[ornate elk]

try this workflow (drag and drop into comfyui)

[spare crow]

What do i do with that?

[ornate elk]

download and then drag and drop it into the comfyui webui

in browser

[spare crow]

is that because I dont have the vae?

[ornate elk]

yes

[ornate elk]

here is the sdxl vae fp16:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors

the model doesnt matter, choose any sdxl you have

[spare crow]

where do i save the vae?

models/vae?

crashes out, this time at 38% but idk if thats just because theres more or less steps than the other

but crashes out at negative prompt again

[ornate elk]

hmm and the cmd doesnt show anything?

is comfyui whitelisted in any browser adblocker?

[spare crow]

i disabled adblock for this yday when I was in the search someone else had a similar issue

current log, ill try and copy or do a screen recording before it crashes when i try

[ornate elk]

can you try an other browser without any extensions?

[spare crow]

it does a bunch of python stuff before it crashes

[ornate elk]

where do you see that?

the cmd desnt get the prompt to even start genning

something might be blocking between browser and cmd

[spare crow]

i filmed it when i clicked start

hangon ill find it

that was until it crashed

ive got no idea if thats normal or not though

[ornate elk]

ahh okay

nope not normal

whats your python version?

[spare crow]

was the 1 in the link, 3.10.11 64bit I can try and find out exactly incase its wrong

Should I try reinstalling it?

[ornate elk]

yea maybe something wasnt downloaded correctly

or wait

[ornate elk]

this is always before it crashes?

[spare crow]

that was just before yeah, after i click queue and tab into that quickly thats what comes up each time, just tested it again

[ornate elk]

okay and your not running anything heavy in the background ?

like games or wallpaper engine?

[spare crow]

literally just discord

[ornate elk]

ok

how much free space do you have on the C drive?

[spare crow]

334gb

[ornate elk]

open up a cmd and type
python --version
and then
where python

[spare crow]

Microsoft Windows [Version 10.0.22631.4890]
(c) Microsoft Corporation. All rights reserved.

C:\Users\seanb>python --version
Python 3.10.11

C:\Users\seanb>where python
C:\Users\seanb\AppData\Local\Programs\Python\Python310\python.exe
C:\Users\seanb\AppData\Local\Microsoft\WindowsApps\python.exe

C:\Users\seanb>

[ornate elk]

ok open up a cmd as admin and run sfc /scannow
that fixes system bugs.

[spare crow]

running it now

[ornate elk]

and then disable the Radeon TM Graphics in the Device Manager.
then relaunch comfyui

[spare crow]

Microsoft Windows [Version 10.0.22631.4890]
(c) Microsoft Corporation. All rights reserved.

C:\Windows\System32>sfc /scannow

Beginning system scan. This process will take some time.

Beginning verification phase of system scan.
Verification 100% complete.

Windows Resource Protection found corrupt files and successfully repaired them.
For online repairs, details are included in the CBS log file located at
windir\Logs\CBS\CBS.log. For example C:\Windows\Logs\CBS\CBS.log. For offline
repairs, details are included in the log file provided by the /OFFLOGFILE flag.

C:\Windows\System32>

[ornate elk]

okay it found an issue and fixed it

[spare crow]

gunnu restart quickly because its not opening the webui at all atm

brb

okay with AMD Radeon(TM) Graphics disabled it doesnt open the webui at all

[ornate elk]

ahh okay

[spare crow]

do i need to edit the .bat again?

[ornate elk]

then remove the --device-id

[spare crow]

--cuda-device 1 this bit?

done and disabled but crashes out again at the same spot

enabled it again and tried again and crashes still

[ornate elk]

okay then do a reinstall of comfyui

also redownload the launch bat, i updated it rn

[spare crow]

okay 2 secs

just back to this part, be a few mins

Go into the C:\ZLUDA\ Folder.
There make copy of the cublas.dll and the cusparse.dll and nvrtc.dll repaste them inside the folder.
Rename the copies to cublas64_11.dll and cusparse64_11.dll and nvrtc64_112_0.dll
Copy these three files into ComfyUI\venv\Lib\site-packages\torch\lib and overwrite if asked.

okay all done, the launch bat is this right?

https://raw.githubusercontent.com/CS1o/Stable-Diffusion-Info/main/Resources/Start-ComfyUI.bat

okay still crashing out, i tried disabling the driver again and still no joy :S

[ornate elk]

can you check your windows pagefile setting?

[spare crow]

how do i do that?

if thats correct ill follow that

[ornate elk]

looks correct

also try a restart of your PC

[spare crow]

okay 2 mins

same issue

[ornate elk]

okay, can you try with an 1.5 model?

[spare crow]

just tried uninstalling python and reinstalling and it wont even let me reinstall that now

fml its doomed 😬

[ornate elk]

whats the error?

[spare crow]

which i didnt cancel it at all -.-

with this huge log

[ornate elk]

okay then check in system control if any python is still there listed

and uninstall

also uninstall stuff thats called conda

or anaconda

[spare crow]

did that, hang on i got it

ran it as admin and it let me install it

okay whats a 1.5 model ?

[ornate elk]

you then need to delete the venv of comfyui

[spare crow]

how do i do that?

[ornate elk]

this for example
https://civitai.com/models/4384/dreamshaper

[ornate elk]

in the comfyui folder

delete venv

[spare crow]

just literally delete the folder then launch it again

aahh okay

[ornate elk]

you then need to replace the 3 zluda dlls again in there

after the installation

[spare crow]

those 3 zluda files do I leave them in the zluda folder or delete them once i copy them over?

[ornate elk]

you leave them there

[spare crow]

okay I have been so just wanted to make sure

just waiting on the cmd to install it all again

tried with 1.5 and still crashes

[ornate elk]

hmmm and which webuis do you use that work?

[spare crow]

id only used the auto1111 a few days ago which worked fine for a few days then i deleted it all to use this as only half the models I wanted to try worked

ahhh dw dude, ill try 1 of the other uis or go back to auto1111, ive taken up way too much of your time already!

[ornate elk]

hmm okay, yea comfyui is the hardest to get working from my guides.
But the issue you got is fixable.
can you show me your environment path settings?

[spare crow]

where do i find that?

[ornate elk]

my guide listed how to get there and how to edit it

here is a quick way

https://www.wikihow.com/Change-the-PATH-Environment-Variable-on-Windows

and thats how it should look like:

[spare crow]

ohhh rhat

that

sorry

[ornate elk]

remove the C:sd-zluda and the C sd zluda comfyui from there

and then recheck if your zluda folder on C is called ZLUDA or Zluda

[spare crow]

removed

[ornate elk]

i dont see python in path

thats an issue

[spare crow]

i ticked add to path when installing

[ornate elk]

rerun the python installer
click modify,
click next
select "add python to environment variables"

[spare crow]

is that it at the top ? and yeah okay

was already ticked, idk if that means it was already active or not?

[ornate elk]

click install then check the environment path again

[spare crow]

still not there

should i add it manually?

[ornate elk]

yep

[spare crow]

added both, that correct?

[ornate elk]

looks good

now you have to reboot the PC again

[spare crow]

okay 2 mins

still the same issue

[ornate elk]

can you show me the full cmd log?

[spare crow]

yup

ill show you it before i queue anything up

then ill send you the vid of it after i queue

file size too large to upload dammit

[ornate elk]

can you go into comfyui/users folder and there copy the comfyui.log file?

[ornate elk]

hmm there is no python error

[spare crow]

managed to cut the vid down

that vid doesnt show any python things like before

[ornate elk]

yep but the error exception code 0xc0000005 is acces violation

[spare crow]

so like my pc doesnt let me have permission to run it?

[ornate elk]

it can access some files hat need to be run

as its windows 11 maybe it has a problem with the C:\ZLUDA

try moving it into SD-Zluda and then alter the path in the environment path variables

[spare crow]

moving the launcher into there?

ohh wait i misread that

move all of zluda into sd-zluda?

[ornate elk]

as folder please

so its C:\SD-Zluda\ZLUDA

[ornate elk]

remove the other path

to C:\ZLUDA

[spare crow]

done

[ornate elk]

also have you any antivirus programm installed?

[spare crow]

nope

still crashes, lemme restart again

ima install this windows update real quick just incase this somehow magically fixes it

[spare crow]

That took ages holy, right ima try again!

Seen this just now in the cmd, think it was there before but does it matter ? Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended.

should I try adding all those things to variables for me?

instead of system

[ornate elk]

Nope

[ornate elk]

And nope torch 2.3.1 is needed for AMD with ZLUDA

@spare crow
I'm out of ideas. Other than a windows reinstall (which fixes mostly issues by 100%) I don't know what else to check.
Its either a permission or RAM issue.
Would suggest using either Auto1111 or Forge Webui.

[spare crow]

No worries dude, ill try going on forge and hopefully that lets me try a bunch of other models and works for me

I really appreciate how long you've spent on this trying to help me dude ❤️ you're a legend

[ornate elk]

Forge let you use flux as addition.

And no problem 🙂

[spare crow]

Is flux like a different model type or just a standalone model? Cause I downloaded a bunch and flux was just 1 of many that wouldn't work

[ornate elk]

Flux is a model type

Like sdxl, 1.5, pony, etc

Auto1111 can use
1.5, 2.0, SDXL (pony, illustrious, both are based on sdxl), SD3.0,

and forge Flux in addition

[spare crow]

I installed swarm, worked straight away 🥲 all that time spent

works fine but feels slower to generate

ahh okay I need to make myself a list of model types so I can keep track of it all then and what can use what

[lucid yoke]

swarm is just that good tbh. but the "feel slower" can be a thing of first loading in the model and some zluda/directML wonkyness

[spare crow]

I read back at your message and because of the comfy vid id saved I thought it was comfy you said to try so I went to that first 🙃

[lucid yoke]

i couldave labeled it a bit better probably

[spare crow]

Naah that was just me trying to do 20 things at once, not reading things properly as usual and watching the vid for comfy while I was trying to do it xD

[topaz escarp]

Hi! sorry guys I´m having trouble to mimic the configuration of flux Dev of Civitai in my Forge Console, anyone knows wich sample methods and schedulle type they use? I´m using same Loras and a flux1-dev-fp8.safetensors checkpoint . I´m breaking my horns trying to get the right configuration, but i keep having weird results.... thx!!!

[ornate elk]

Because it uses directml and thats really slow

And vram limited

[spare crow]

Ill try getting forge setup and working aswell later, hopefully that works because id like multiple options!

[spare crow]

Just realised I have a 1080ti sat doing nothing in my attic, I could probs add this to my pc and use both for this and save myself some headache right?

[ornate elk]

yes

[spare crow]

With most of these, like forge comfy and swarm, can I use the nvidia card as a 2nd card, set it all up with that and use both cards to work at the same time? I know swarm can because I just watched a vid on it

[tight rapids]

which loras are you using ? I've been using beta for some time and its working pretty good.

[cosmic furnace]

oh CS1o h1)

it seemed like 1 guy managed to use hi res fix on flux. to make from 960-1280 a double 1920-2560

with that 10-12 gig of vram... but ive ended up failed with always too low vram message. advising to low vram used. which ive tried up until 16gb but still nothing. Is it like with fp16 and need of 48-64gb of usual ram of wtf ?

yeah flux very bad optimised and eats alot of vram when trying to upscale. but he managed to do it on 3080 =\

dont tell me that amd is still that bad optimised compared to nvidia... shudnt be a thing as of now... pretty sure... espec on forge

[topaz escarp]

im using one of my own and 2 dowloades from civitai that worked ok in the site. Now im testing Euler and Simple (quite Vanilla) and worked on de distilled cfg and looks good! thank u!!!

[chrome rampart]

hello all. Hope somebody can help. I am getting error after i want to generate image. Last week i reinstalled comfy but now again error After restarting ComfyUI, please refresh the browser.
FETCH ComfyRegistry Data: 25/33
FETCH ComfyRegistry Data: 30/33
FETCH ComfyRegistry Data [DONE]
[ComfyUI-Manager] default cache updated: https://api.comfy.org/nodes
nightly_channel: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/remote
FETCH DATA from: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/custom-node-list.json [DONE]
[ComfyUI-Manager] All startup tasks have been completed.
got prompt
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16

F:\comfyui\ComfyUI_windows_portable>pause
Press any key to continue . . .

after press any key the CMD screen just closes

[shut geyser]

why is this a thing in zluda a1111

[ornate elk]

Would need the full cmd log + what you tried as settings

[shut geyser]

nvm it was in directml, im trying to use zluda a1111 now, wrong thing turned on
too many uis

[ornate elk]

Ah well xD

[spare crow]

swarm is using my other gpu by default atm, I dont know why but its using the 1080ti, which is fine because its letting me upscale and do more than the 6950xt does, but how can i get both working at the same time to increase the limit and speed? I tried adding another ComfyUI self starting to the backend with GPU_ID to 1 but then nothing seemed to load at all, not sure if I did it wrong or it doesnt work the way i thought it would?

[ornate elk]

You can't use 2 GPUs for one image output

[lucid yoke]

https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Using More GPUs.md

you can use both gpu's at once to generate/do different things but you cant combine the vram

i wish though

[spare crow]

Ahh that sucks! that would be so handy! I feel cucked, 16gb of vram available to me and I cant do shit with it, but my old 11gb vram card works better xD

[cosmic furnace]

Yyyyy. with zluda bad amd vram optimization shudnt be a thing anymore but im still not sure it is...

Csss111ooooooooooo

[ornate elk]

With Auto1111 and forge you can use 16gb really good but yea xD its much easier with nvidia

[spare crow]

Gunnu set forge up in a little bit so let's hope so! I feel like my amd card just decides to use it all straight away regardless of what it's doing

[cosmic furnace]

it feels more less as if its uses all vram legit. but calculates wrong...

also some vram is always used up even if u close everything

[spare crow]

I couldn't even upscale an image on swarm 2x earlier with 16gb but with my nvidia I've done 2x 3 times to each upscaled image already lmao

[cosmic furnace]

ive read some ppl even connect their monitor into motherboard so it wont use 1-2 gb of extra vram 😄

[ornate elk]

That sounds like Directml

[spare crow]

I think on windows you can go as low as roughly 500mb used doing nothing, from what I looked up earlier

[cosmic furnace]

@ornate elk ? Can say smthing about this?

[cosmic furnace]

This

[spare crow]

I did set up swarm with zluda thought so didn't think it would still use it