🤝｜tech-support

[ornate elk]

hey, on an igpu?

[astral oracle]

Yes

Is it possible

[ornate elk]

yea, but whats the cpu and how much vram doies the igpu has ?

[astral oracle]

ryzen 5 5500u, it utilizes system ram

I remember running it before

[ornate elk]

okay, you can try follow my AMD Automatic1111 with DirectML Guide.
you can find the guide in the link from the pinned messages.

if you want to run it on cpu only mode i can help you for that too

directml is faster because it can use the shared ram

but more limited on it too

[astral oracle]

This?

I'm on linux tho

[ornate elk]

ohhh

then i cant help sry.

[astral oracle]

i also run windows

[barren monolith]

sir . its loaded the web ui . but this pop up when cmd running .

but its still load the web ui at the end

[ornate elk]

make sure you have done this:
And then you have to replace the rocm library files again with the one from my guide for your card. (gfx1031) then restart the PC again. Then delete the venv folder and the .zluda folder inside the stable-diffusion-webui-amdgpu folder.

[astral oracle]

if i want to utilize system ram, do i use lowvram option?

[tepid crypt]

Hope someone can help me with this. Im trying to use ip adapter in a1111, im using it in a colab, Sagemaker, following a tutorial. There it shows that they have three different preprocessors that doesnt appear for me, someone knows how to install those?

[ornate elk]

Yes

[astral oracle]

ok

[ornate elk]

Im off now.
Can help tomorrow again.
@barren monolith we will get it to work.

[barren monolith]

yes , btw as info i already uninstall windows in my pc and clean everything and start from new . and do all steps
but the logs on cmd now like this

[lethal kiln]

Dear all,I have a question. I'm currently using Controlnet for semantic segmentation with color recognition, but the generated results always look strange. What could be causing this issue?

It's great to have access to experts!

[gaunt crystal]

is there any way to transfer the style of a prexisting image to another prexisting image using ipadapter?
all the tutorials i see are about generating new images, i want to take an image i have, and put it in the style of another image (in this case a 1970's photograph) basically im talking about making some sort of filter, i've already tried img to img and that didnt seem to work, i have ipadapter installed, but all the tutorials i can find on youtube seem to be about making new images, not adding a style to an existing image, if anyone could point me in the right direction i would appreciate it.

[ruby yoke]

it's called style transfer...

[gaunt crystal]

yes.

[ruby yoke]

https://www.youtube.com/watch?v=gmwZGC8UVHE start here.

[gaunt crystal]

i forgot to mention, im using auto1111

[ruby yoke]

okay well, i'm not sure if auto1111 can use either ipadapter or style transfer. maybe install comfy as well - you can share the models between them

[gaunt crystal]

ok, i have been meaning to do that anyways

[astral oracle]

C:\Users\vaelmar\stable-diffusion-webui-directml\venv\lib\site-packages\pytorch_lightning\utilities\distributed.py:258: LightningDeprecationWarning: pytorch_lightning.utilities.distributed.rank_zero_only has been deprecated in v1.8.1 and will be removed in v2.0.0. You can import it from pytorch_lightning.utilities instead. rank_zero_deprecation( Launching Web UI with arguments: --use-directml --skip-ort --lowvram --opt-sub-quad-attention --opt-split-attention --no-half-vae --upcast-sampling Downloading: "https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors" to C:\Users\vaelmar\stable-diffusion-webui-directml\models\Stable-diffusion\v1-5-pruned-emaonly.safetensors 100%|████████████████████████████████████████████████████████████████████████████▉| 3.97G/3.97G [45:02<00:00, 1.58MB/s]

WebUI freezes once model is downloaded, any workarounds?

enough time to finish

i left it for like 4 hours

i'm going to try to download the model then add it to the folder

[astral oracle]

i fixed it, it works normally now

[misty marten]

my reActor do not works with zluda

[ornate elk]

You have to set reactor to CPU mode

[wispy pond]

Anyone in here have ultrapixel up and running and familiar with this error?

"When loading the graph, the following node types were not found:

UltraPixelProcess
UltraPixelLoad

..."

I'm trying to use it in comfyui

[pallid trench]

is this an expected runtime to train a lora (160 pics, 10 batch size, 5 epochs, 1024,1024 res, 4070 gpu(12gb vram))

[lone wolf]

Hello, Im purchasing a new PC cause of a flooding... RIP,. Do you recommend getting the Top Nvidia GPU or Top AMD GPU for SD3? 🅰️ 📗

[ornate elk]

hey, the top Nvidia GPU performs better with AI stuff right now

[ornate elk]

how does your environment path entry look like?

[dense sandal]

I dont know why my dw open pose processor isnt working.

Can someone help me?

[hushed oracle]

anybody would know if working with an AMD 6800 Xt is decent for a model please ? trying to make Flux working with directml evertime im getting out of GPU memory ...

[half rain]

Hello the StabilityAI/TripoSR is down on HuggingFace any feedback from the dev for the problem ?

https://huggingface.co/spaces/stabilityai/TripoSR

[ornate elk]

dont use directml
you should use Zluda to get a better performance and with that you can use sdxl and upscaling without problems too

You find my AMD Guides here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides

also SD3 may be possible with it

[pastel coral]

Hello! I’m new! Can I generate images on stable diffusion on a MacBook Air m2? Is a graphics card absolutely necessary? If not what is the subscription cost? I can’t seem to find a legit website

[hushed oracle]

thanks a lot @ornate elk

[ornate elk]

hey yes you can use SD on a M2 MAcbook air.
Here is an easy install script for MAC to get it running:
https://github.com/viking1304/a1111-setup

[pastel coral]

No issues on the performance right? I have the 16/256 config , I’m going to make web novels (sfw and nsfw) thus trying to confirm. Also I need to purchase the premium right?

[ornate elk]

Its free if you use it localy

Performance is okay

16gb ram is a bit rough

but still works okay i guess

[pastel coral]

Even for nsfw? Also I read somewhere if I use the generated images for commercial purpose I need to pay something? And also if I don’t use it locally but on the web then it won’t be intensive on my device right? I just bought it a week ago I don’t want to stress it from the get go

I’m up for paying $30ish monthly if that’s what it costs (some api site had $27 listed)

[dense sandal]

my IPadapter unified is throwing me errors about models not found

But I think I got all the models already

ALso, I installed comfyUI from stability-matrix interface, and I don't know how to access the terminal to install dependencies

If I do cmd in the folder it doesn't work

[hushed oracle]

@ornate elk ok i did the whole install but when i launch again the start bat its telling me that i dont have an Nvidia GPU ...:(

[plain bronze]

I have 8gb vram and 32gb ram. I need more vram for faster sd, right?. Can I increase it by using my RAM?

[ruby yoke]

you can try

[ornate elk]

did you tried my comfyui or auto guide?

[hushed oracle]

so it was my mistake wrong path 🙂 now i was able to launch it im trying the dreamshaper_8 model by default but it seems to take also quite a while ? is there a model or something i should pick to start with ?

[ornate elk]

the first image will take 15-40 minutes

[hushed oracle]

ahh ok 🙂 thats true it was written

[ornate elk]

in the cmd should be no error meanwhile

did you installed comfyui or auto1111 ?

[hushed oracle]

ComfyUi

oh i have an error 😦

RuntimeError: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)

[ornate elk]

you have to do step 5

from the guide

[hushed oracle]

but i have a 6800xt ?

i have to copy the dll you mean right ?

[ornate elk]

yea still step 5 is this:
and its required for all cards:
Go into the C:\ZLUDA\ Folder. There copy the cublas.dll and the cusparse.dll and repaste them inside the folder again. Rename the copys to cublas64_11.dll and cusparse64_11.dll Copy these two files to ComfyUI\venv\Lib\site-packages\torch\lib and overwrite if asked.

[hushed oracle]

now when i do that ...he says that he doesnt recognise these files ...

is it normal that they are doing like 180k when the original are doing 200mo or something ?

@ornate elk OSError: [WinError 126] The specified module could not be found. Error loading "C:\Users\david\Documents\Comfy+Zluda\ComfyUI\venv\Lib\site-packages\torch\lib\cublas64_11.dll" or one of its dependencies.

[ornate elk]

how do your system environment paths look like?

[hushed oracle]

which one ?

just the path ?

for system ?

[ornate elk]

no the guide states that you add the c zluda folder and %hip_path% to the system environment path

[hushed oracle]

C:\Program Files\AMD\ROCm\6.1\ i havbe that for HIP_PATH

this is what i have for the HIP

[ornate elk]

can you screenshot both paths

of the system env

i mean in this window.
then after click Edit, you have to add the 2 paths

[ornate elk]

and then this

[hushed oracle]

oh ok this i dont have in path

[ornate elk]

look at my first screenshot, click the red marked and then add 2 new entrys

from screenshot 2

[solemn moon]

Argh been away for a week or two

Feel like I’ve missed everything

May I ask, does flux allow Control Nets?

[hushed oracle]

seems to work again 🙂 lets see

[ornate elk]

currently there are no flux controlnet models

[ornate elk]

does it work?

[ornate elk]

@barren monolith
uninstall HIP SDK 6.1 and install HIP SDK 5.7 and then restart the PC.
After that you need to again replace the Zluda files with these ones here:
(they are made for 5.7)
https://github.com/lshqqytiger/ZLUDA/releases/download/rel.11cc5844514f93161e0e74387f04e2c537705a82/ZLUDA-windows-amd64.zip
Then you have to add %HIP_PATH%bin again to the system variables.
And then you have to replace the rocm library files again with the one from my guide for your card. (gfx1031)
then restart the PC again.
Then delete the ven folder and the .zluda folder inside the stable-diffusion-webui-amdgpu folder.
Finally launch the webui-user.bat

[hushed oracle]

its working thank you so much for your help 🙂

[ornate elk]

perfect, no problem! 🙂

[frigid junco]

is there a definitive way to run SD on 6800 xt ?

ive tried the fixes like directml and the amd fork, it doesnt work as i still get the "Torch is unable to find the gpu" error

im on windows

[hushed oracle]

the next one 🙂

ahah

[ornate elk]

hey yes, i made a lot of AMD Guides and the best usage/performance has
AMD Automatic1111 with Zluda
You can find it here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides

[frigid junco]

im working on the ZLUDA guide right now, hopefully it works. i gave up on directml

[ornate elk]

i made a few changes a few mins ago, so refresh the site one time

but with a 6800XT it doesnt affect you

[frigid junco]

zluda doesnt work ..

[ornate elk]

ahh this error

is caused by having an AMD igpu

you have to disable it in the device manager

[frigid junco]

i have a 7800xt igpu

ok

[ornate elk]

nono the igpu is from the cpu

its called Radeon TM Graphics

in the device manager

only disable that, and not your 7800xt

[frigid junco]

7800x3d

not 7800xt oops

[ornate elk]

ohh xD

yea

i have the same cpu, its a beast

still the Radeon TM Graphics need to be disabled

then relaunch the webui-user.bat

[frigid junco]

webui-user is downloading huggingface.co this time now

[ornate elk]

it downloads the default model mostly

[ornate elk]

did you got FLux to work?

if not i have a link for the Flux 12gb fp8 model

[frigid junco]

@ornate elk i think it works now, im on the webpage. now i just have to figure out how to use it lol

[ornate elk]

generate 1 image to test, the first image will take 15-40 minutes

after that you can test around

[frigid junco]

do i have to use a checkpoint or lora? do i just enter text into the prompt ?

[ornate elk]

checkpoint should be 1.5 ema pruned at the top left or any other checkpoint

no lora

type Cat into the prompt

hit generate

check the cmd if any error appears

[frigid junco]

it worked 😄 thank you so much

that was the image i needed. i will now uninstall as i have peaked in my ai career .... jk

[ornate elk]

perfect ^^ no problem

[ornate elk]

yep you mastered it

nothing more to teach here xD

Now get some nice community made models from Civitai.com

you can delete the 1.5 ema pruned, its not needed anymore

[ornate elk]

i was wrong, there is alreay a cnet for flux:
https://huggingface.co/XLabs-AI/flux-controlnet-canny

[hushed oracle]

are you guys putting the filters off for example on Flux ? is there a way to do that easily ?

[ornate elk]

localy there are no filters

[hushed oracle]

really i asked to do a pikachu to test but it was all blurry lol 🙂

[ornate elk]

make sure you use the 12gb fp8 dev model

if you used "schnell", you have to use less steps

[hushed oracle]

https://huggingface.co/black-forest-labs/FLUX.1-dev you mean this one?

[ornate elk]

Here is a link for the 12gb model:
https://huggingface.co/Kijai/flux-fp8/tree/main

its faster than the 24gb one

[hushed oracle]

it is this one that i have yes

[ornate elk]

can you show how the output looked?

[tranquil zealot]

Guys, I'm kind of new at this, and I wanted to make a thumbnail using this image and wanted to make it look better. Any tips on how to do that?

[ornate elk]

or the workflow

[hushed oracle]

how an i show you the workflow ?

[ornate elk]

screenshot it

[hushed oracle]

cest celui par defaut

its the default one sorry

dont know why i do start talking french for no reason lol

🙂

[dense sandal]

Does anyone know how to access the CMD terminal if using stability matrix?

I don't know in which folder to write CMD to get it working...

[hushed oracle]

better result 🙂

[ornate elk]

yea thats not the right workflow

[hushed oracle]

do you have a good workflow to share ?

[ornate elk]

Thats the official workflow from comfyui examples . Drag and drop it into Comfyui in Browser.

[ornate elk]

what did you changed?

[hushed oracle]

i have a bunch of red node

with this workflow

[ornate elk]

ohhh

[hushed oracle]

?

[ornate elk]

thats strange

but someone else had the same issue

[ornate elk]

thats a question for #📝｜prompting-help

[hushed oracle]

EmptySD3LatentImage
FluxGuidance i dont have these nodes it seems

[ornate elk]

hmmm

[hushed oracle]

are you on Linux or windows @ornate elk ?

[ornate elk]

windows

[hushed oracle]

ok

with the FP8 its pretty fast to be honest

[ornate elk]

can you share me your workflow?

[hushed oracle]

https://comfyanonymous.github.io/ComfyUI_examples/flux/#simple-to-use-fp8-checkpoint-version

i stole this one

if you take the fp8 image

[ornate elk]

its strange i get an error when trying this

where did you placed the model file? in unet or checkpoints?

[drowsy cipher]

How can I use stable diffusion in my python project?

i have a ubuntu server

[hushed oracle]

in checkpoints

theres a version for unet but its the other workflow

and you need a bunch of other files

[drowsy cipher]

checkpoints?

[hushed oracle]

yes

[drowsy cipher]

is there a setup i can use like ollama uses

[drowsy cipher]

i have just realized this conversation has nothing to do with me

[ornate elk]

i have it there too but get an error. ist really strange

[ornate elk]

the online version or a local version of SD ?

[hushed oracle]

what kind of error do you have ? @ornate elk

[ornate elk]

`Error occurred when executing CheckpointLoaderSimple:

ERROR: Could not detect model type of: D:\Programme\AI-Zeug\ComfyUI\models\checkpoints\flux1-dev-fp8.safetensors

File "D:\Programme\AI-Zeug\ComfyUI\execution.py", line 152, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "D:\Programme\AI-Zeug\ComfyUI\execution.py", line 82, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "D:\Programme\AI-Zeug\ComfyUI\execution.py", line 75, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "D:\Programme\AI-Zeug\ComfyUI\nodes.py", line 518, in load_checkpoint
out = comfy.sd.load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, embedding_directory=folder_paths.get_folder_paths("embeddings"))
File "D:\Programme\AI-Zeug\ComfyUI\comfy\sd.py", line 518, in load_checkpoint_guess_config
raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))`

[hushed oracle]

i had the same error i think youi dont have the correct workflow can you show me ?

[ruby yoke]

comfy or auto1111?

[ornate elk]

comfyui

[ornate elk]

i moved the image from the examples

[ruby yoke]

first problem is that you shouldn't rename the .sft file to safetensors. that indicates comfy isn't updated correctly. unload comfyUI, then go into the comfyUI/update folder and run the update.py file there. it doens't update correctly with manager far too often. then make sure pytorch is at 2.4

[hushed oracle]

yes

[ornate elk]

i didnt renamed it, i used the fp8 model 12gb variant

[hushed oracle]

thats what i thought it is not the correct workflow

[ornate elk]

lol, can you share me yours?

[ruby yoke]

not sure what that even is. i'm sticking with the stuff that comes from black forest labs - not stuff people have cobbled together

[ornate elk]

i used this workflow:

[ruby yoke]

try my workflow

that would should have both schnell and dev in it -schnell connects to the scheduler, dev to the guider

[ornate elk]

i dont have schnell rn

[ruby yoke]

and lets you use fp8 with the same details you get from fp16

go grab a copy of it

[hushed oracle]

but you need all the other files for the clip

[ornate elk]

i have them

[hushed oracle]

you can also use dev with this workflow

should work

[ornate elk]

i used the dev

did you used schnell for yours?

[hushed oracle]

it weird i have a bunch of node again not there i think i now why its the extra stuff to install no in Confy ?

[ruby yoke]

the reason i'm using both dev and schnell is that dev is too distilled, it makes images that look too good - schnell tempers that - using both in the same workflow gives you the widest range of effects without having to destroy the distillation by trying to merge them

[hushed oracle]

Import times for custom nodes:
0.0 seconds (IMPORT FAILED): C:\Users\david\Documents\Comfy+Zluda\ComfyUI\custom_nodes\websocket_image_save.py

WARNING: some comfy_extras/ nodes did not import correctly. This may be because they are missing some dependencies.

do you have that kind of error @ornate elk

?

[ornate elk]

nope

i make a clean reinstall rn

[hushed oracle]

what is your GPU already ?

[ornate elk]

i have a RX 7900XTX

[hushed oracle]

ok so AMD as well its weird i have an error that says that cuDNN is not supported by ZLUDA when i start then later i have a bunch of error

[ornate elk]

thats normal

cudnn is not supported by zluda

the cudnn patch of comfyui fork disables it

[hushed oracle]

but then it desactivate most of the extra nodes ?

thats why i have a bunch of nodes that are in red

[ornate elk]

normaly not

then you have the same issue as a friend right now, we try to fix it the whole day

[hushed oracle]

Import times for custom nodes:
0.0 seconds (IMPORT FAILED): C:\Users\david\Documents\Comfy+Zluda\ComfyUI\custom_nodes\websocket_image_save.py

WARNING: some comfy_extras/ nodes did not import correctly. This may be because they are missing some dependencies.

IMPORT FAILED: nodes_latent.py
IMPORT FAILED: nodes_hypernetwork.py
IMPORT FAILED: nodes_upscale_model.py
IMPORT FAILED: nodes_post_processing.py
IMPORT FAILED: nodes_mask.py
IMPORT FAILED: nodes_compositing.py
IMPORT FAILED: nodes_rebatch.py
IMPORT FAILED: nodes_model_merging.py
IMPORT FAILED: nodes_tomesd.py
IMPORT FAILED: nodes_clip_sdxl.py
IMPORT FAILED: nodes_canny.py
IMPORT FAILED: nodes_freelunch.py
IMPORT FAILED: nodes_custom_sampler.py
IMPORT FAILED: nodes_hypertile.py
IMPORT FAILED: nodes_model_advanced.py
IMPORT FAILED: nodes_model_downscale.py
IMPORT FAILED: nodes_images.py
IMPORT FAILED: nodes_video_model.py
IMPORT FAILED: nodes_sag.py
IMPORT FAILED: nodes_perpneg.py
IMPORT FAILED: nodes_stable3d.py
IMPORT FAILED: nodes_sdupscale.py
IMPORT FAILED: nodes_photomaker.py
IMPORT FAILED: nodes_cond.py
IMPORT FAILED: nodes_morphology.py
IMPORT FAILED: nodes_stable_cascade.py
IMPORT FAILED: nodes_differential_diffusion.py
IMPORT FAILED: nodes_ip2p.py
IMPORT FAILED: nodes_model_merging_model_specific.py
IMPORT FAILED: nodes_pag.py
IMPORT FAILED: nodes_align_your_steps.py
IMPORT FAILED: nodes_attention_multiply.py
IMPORT FAILED: nodes_advanced_samplers.py
IMPORT FAILED: nodes_webcam.py
IMPORT FAILED: nodes_audio.py
IMPORT FAILED: nodes_sd3.py
IMPORT FAILED: nodes_gits.py
IMPORT FAILED: nodes_controlnet.py
IMPORT FAILED: nodes_hunyuan.py
IMPORT FAILED: nodes_flux.py

[ornate elk]

why did you installed it into the Documents folder?

[hushed oracle]

hm i dont know ? i should not ?

[ornate elk]

nope, to prevent permission and onedrive errors

[valid gorge]

Is it possible to merge turbo with normal sdxl together, or turbo with hyper, or even mix them all up with merging?

[hushed oracle]

ok i can move it then

[plain bronze]

I wanted images and hearing some voices from my pc. It started today. Then I saw sd slowly creating one image and completing it and it is silent. Then it stops like 10 seconds and suddenly creating a new image like 10-20 times faster than the first image and I hear the sound. Then repeating cycle. One silent and slow image and one fast noisy image. Why?

[ornate elk]

can you share me the full error?

[hushed oracle]

im reinstalling clean under C:

i send you that if i have the same error again

few minutes

problem solved 🙂 you were right 🙂

thanks again !

[ornate elk]

perfect! no problem

[hushed oracle]

oh no 😦 and now again this problem looll

RuntimeError: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)

Prompt executed in 35.10 seconds

this was the path again right ?

or something else i dont remember ..

[ornate elk]

nope, you have to replace the zluda files

in torch lib

[hushed oracle]

ah shit thats true thanks 🙂

[plain bronze]

It worries me, can someone help?

I upgraded my RAM today and it started after that.

Is this normal?

[ornate elk]

that high pitched sound is called coilwine and happens when the gpu uses its full power

its normal

you can hear it too when playing minecraft with 500-1000fps

[plain bronze]

Why it didn't happen before upgrading ram?

And is this bad for GPU?

[misty marten]

and how could I change this?

[ornate elk]

open the reactor tab in txt2img, click in settings, check CPU mode

[ornate elk]

nope

[ornate elk]

maybe before was it limited because to less RAM

[plain bronze]

Okay, thank you. I searched about this. Relieved now.

[ornate elk]

no problem 🙂

[ornate elk]

found the cause of my issue:
If what you received is the 11GB model distributed by kijai, that's a diffusion model. If it's the 17GB model distributed by Comfy Org, that's a checkpoint model. Both have the same name.

12gb model goes into unet folder to work. and the 17gb one in checkpoints

[safe inlet]

some load with unet loader some load with checkpoint loader, adds much confusion with the 20 versions out ther

[hushed oracle]

perfect 🙂

[crystal nacelle]

I keep getting stunning portraits even though im traing to make an environment drawing

prompt; (+environment), )(+sunset), clouds, vanishing point, sky, high contrast. dramatic clouds, clouds resceding into the vanishing point. sun on horizon. dramatic. (monochrome), (black and white), (ink), HD, high quality, masterpiece. ultra-detailed, intricate, photorealistic, masterpiece, award-winning, sharp focus, crisp lines, smooth gradients, no noise, 8K resolution.
Artstyle: anime realism, comic art meets anime aesthetic. digital illustration, Strong edge lighting, with cinematic lighting. Core shadow. light from below, portrait shot.

negative prompt; ground. soul, mountains, earth,, unclear, noise, lowres, bad quality, blurry, pixelated, jpeg artifacts, unrendered, noisy, people, humans, figures. face, portrait, face, girl, eyes,

im tryna get something like this with no people and words, in ink greyscale

[ruby yoke]

might suggest you totally rewrite that prompt. which model is this for?

[ruby yoke]

positive prompt: pen and ink anime-style, chromatic cell shaded, drone flyover footage high in the clouds. clouds below, in the distance a small castle no negative prompt. not weighting in the positive prompt, just posistion of tokens in the prompt

[fair oxide]

to avoid confusion

[drowsy cipher]

local version

[ornate elk]

You can install a Stable-diffusion webui like Automatic1111. It has an API which you could use in your python project

[restive sparrow]

Is it better to use A1111 on any certain browser?

[solemn moon]

I use chrome

[restive sparrow]

same

With tiled diffusion/VAE, from 1280x720, I can upscale up to 2.5x but beyond that nothing is possible, it either halts or the time is beyond 10min. Is that simple VRAM limitation?

[wispy pond]

When using ultrapixel through comfyui-ultrapixel I get

'NoneType' object is not subscriptable

File "C:\Users\Toniko\Documents\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 152, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Toniko\Documents\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 82, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Toniko\Documents\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 75, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Toniko\Documents\ComfyUI\ComfyUI_windows_portable\ComfyUI\nodes.py", line 1442, in save_images

Not sure if this makes too much difference because it still gives me a final image, wondering if it's important?

[crude trellis]

is there any way I can save stable diffusion settings so I don't need to manually setup everytime? like sampling methods, sampling steps, cfg scale

[fossil swan]

heyhey jjust joined cause im getting a little frustrated following a tutorial

ive done everything it wants, having to take an extra step to specify a path to my python installation

then i get this message when running my webui.user batch file

ive tried two different versions of python, not sure whats going on

https://www.dexerto.com/tech/how-to-install-stable-diffusion-2124809/

using this tutorial btw

if that helps

[ornate elk]

Hey, you have the wrong python version installed

You need 3.10.11 64bit

[fossil swan]

oh right right

so fickle

[ornate elk]

Best is to follow my install guide from the pinned messages of this channel

It covers everything easily

[fossil swan]

i feel like im gonna finish off the one i was working off first

just so i might not have to go through the hassle of redoing everyhting

[ornate elk]

You don't have to. My guide must explains it simpler and also covers the needed performance improvements you can do depending on your GPU vram

You can also do your guide and then add the performance args after that

[fossil swan]

okay yeah i give up on the original now, it says theres no python at the path i gave it and theres DEFINITELY a python

[south vigil]

hey guys so i got a checkpoint that needs a VAE, ive putted that inside the vae stable diffusion folder but dont know how to selct it in the settings

[ornate elk]

Did you got it to work with mine? Or do you need help?

[ornate elk]

Hey, go into the Settings - User interface - Quicksettings
There add sd_vae
And then hit apply and reload ui.
Then you will get a nice VAE dropdown menu in txt2img

[pastel coral]

is stable2go . ai good?

[fossil swan]

i just went with comfyui

seems to be working, images are very very not great but im running at down to either lame checkpoint or hardware, cause im running it on practically a brick

im not used to the ui but it seems to be a little simpler

[ornate elk]

Oh okay.
I would say comfyui is a bit harder learn but offers a lot of freedom on how to get something working

[fossil swan]

i like the visual nodes and stuff

feels like ill be able to understand how it all works eventually

just grabbing all my checkpoints and go-to lora atm

will definitely be asking if i run into any troubles i cant solve with goog, thankyou very much 🙏

[ornate elk]

Alright 🙂 have fun creating!

[fallen blaze]

ı get errors ı dont know how to solve can someone help me ?

[ornate elk]

Yea if you give more information ^^

GPU, the full cmd error log, which webui do you use etc

[fallen blaze]

first that 2

[ornate elk]

Full cmd log would be better and info on GPU and Programm you try

[fallen blaze]

how ı can do that ?

ı am new to this

[ornate elk]

Check what's your GPU.
Then tell what Programm you tried to install or use.
Then copy the whole cmd text as text and paste it here

[fallen blaze]

https://www.youtube.com/watch?v=eO88i8o-BoY&list=WL&index=1&t=20s ı use this video and my gpu is 6650xt

ı couldnt dowload python version

it give me error

so ı used latest one

also sorry for eng its not my mother langue

[ornate elk]

Okay thanks now I have all info I need.

That video guide is not recommended.
If you want the best performance with your GPU, you need to follow my AMD Automatic1111 with ZLUDA
Guide.
You can find it here:
https://github.com/CS1o/Stable-Diffusion-Info

[fallen blaze]

should ı delete everything ı dowload

[ornate elk]

Yea start clean

[fallen blaze]

do ı need visual studio ?

should ı delete that too ?

[ornate elk]

You don't need it

[fallen blaze]

oke

[plucky maple]

Anyone available to check my work? I feel like my images are coming out jumbled and blurry, maybe due to improper VAE use?

This is the first time I've used a model that requires a VAE, but I looks right in the settings, yet there is no change when it is removed during generation.

Generation settings:

model = ponyDiffusionv6XL_v6STartWithThisOne
pos = (((score_9, score_8_up, score_7_up, score_6_up,))), style_anime, flat colors, flower vase
neg = score_5, score_4, painting
sample = DPM++ 2M
Schedule = Auto
w/h = 1024x1024
CFG = 7
seed = 2704000228

-edit- file name for the VAE is actually sdxl_vae.safetensors, I just updated that

[ornate elk]

Hey, thats not a VAE issue.
The base pony model is just bad.
Try any other pony based model like
Perfect Pony
Also add source_pony to the negatives if you want anime

[fallen blaze]

which one ı need to choose ?

[ornate elk]

The one for 6600xt

Gfx1032

[fallen blaze]

oke

[fallen blaze]

what ı need to there ?

ı have that 2

not the bin version

[ornate elk]

You need to click on the Path entry

Not hip path

Its below them

Then click edit

[fallen blaze]

the os one ?

that one ?

[fallen blaze]

okey

[ornate elk]

And then make 2 new entrys

For the folder and the bin

[fallen blaze]

like that ?

where ı need to dowload stable diffusion

[south vigil]

found it, it was all the time on the left side visible lmao, thanks for the info!

[ornate elk]

You mean the models?

[fallen blaze]

no the onbe u need to oppen

ı deleted that too

[ornate elk]

Did you followed the guide? Then you should have an Ai folder on C

[fallen blaze]

ı did

ı dont think ı have tho ?

that one ?

[ornate elk]

Thats the zluda files

Thats not the webui

[fallen blaze]

what ı need to search for it

[ornate elk]

The first step of my guide was to create an extra folder

[fallen blaze]

ı did

[ornate elk]

And what's in there?

[fallen blaze]

ıdk

that right

[ornate elk]

Did you forget where you created it?

[ornate elk]

Yep

That part is the webui setup

[fallen blaze]

ı created that on documents

will that be problem ?

[ornate elk]

Yes thats why my guide says to not do that xD

[fallen blaze]

if ı delete that one and create anıther

will problme solve

[ornate elk]

If you do the steps again there is no problem

You can delete that one from documents

Just don't move the folder

[fallen blaze]

okey

do ı need to dowload git ?

[ornate elk]

Yes

And Python 3.10.11 64bit

[fallen blaze]

can u give me the links ?

[ornate elk]

They are in the guide

Click on Preparation

[fallen blaze]

ı did everthing again

so what ı need to do

which one ı need to oppen

[ornate elk]

If you done everything, run the webui-user.bat

[fallen blaze]

that whats coming ong

[ornate elk]

Can you show the full cmd text?

[ornate elk]

You installed the wrong hip version.
But it can be fixed by downloading the right zluda files.

[fallen blaze]

okey

then what ı need to do

[plain bronze]

Is Opera GX browser good for A1111?

[ornate elk]

Download these file. Unzip and drag an drop them into the C:Programms\AMD\Rocm\6.1\bin\rocblas\library folder.
Then restart the PC:
https://github.com/likelovewant/ROCmLibs-for-gfx1103-AMD780M-APU/releases/download/v0.6.1.2/rocm.gfx1032.for.hip.sdk.6.1.2.7z

[ornate elk]

Not really. Would just use Firefox

[plain bronze]

Firefox is better than Brave for SD?

[ornate elk]

Brave is okay too

It won't matter for performance

[plain bronze]

But GX matter?

[fallen blaze]

the dll one too right ?

[ornate elk]

Nope

[plain bronze]

Then why it is not good for A1111?

[ornate elk]

The DLL goes into the rocblas folder

Not in library

[fallen blaze]

ı did like tgat

[ornate elk]

Because opera VPN can break SD in the browser

[ornate elk]

Looks normal

Dont forget to restart the PC

[fallen blaze]

ı did

what ı need do from now

[ornate elk]

Now relaunch the webui-user.bat and try to generate the first image

[fallen blaze]

will it take long time ?

[ornate elk]

The first image will take 15-40 minutes. One time only.
The cmd shouldnt show any error

[fallen blaze]

the image generator one didnt open

yet

[ornate elk]

What does the cmd shows?

[ornate elk]

Okay let it stay like that

[fallen blaze]

then ?

[ornate elk]

Check back in 30 mins if its open.
Zluda needs 15-40 mins one time to initialise

[fallen blaze]

okey

[ornate elk]

Dont interrupt or close the cmd

[fallen blaze]

will it change something if ı delete them ?

[ornate elk]

Dont delete these folders they are required

[fallen blaze]

okeyge

how ı can add my model now ?

[ornate elk]

Did it opened ? You have to gen 1 image first

[fallen blaze]

ı am waiting it to generate first

[ornate elk]

Looks good. Just wait

The cmd shouldn't show any error

And then you can download new models from Civitai.com

[ornate elk]

Perfect

It will work

[fallen blaze]

Thanks G

[ornate elk]

No problem 🙂

[fallen blaze]

it did

what ı need to do for new models now ?

[ornate elk]

Now you can download new models from Civitai.com

Models are always 2-6 GB files

They go into models/stable-diffusion folder

[fallen blaze]

why are they like that ?

[ornate elk]

Xl and pony models need 1024x1024 resolution

[fallen blaze]

how ı can change thaT ?

[ornate elk]

With the height and width sliders

[fallen blaze]

no not that when ı try to generate something ıt generate random things

[ornate elk]

To generate specific anime chars you need a so called Lora, of this character

[fallen blaze]

how ı can add 2 at the same time ?

[ornate elk]

With an extension called Regional prompter

[fallen blaze]

https://www.youtube.com/watch?v=WwATqUp5gBc is it good for thaT ?

[ornate elk]

Can't check the video rn

But you can try

[sharp ridge]

i cant use flux like this, any solution and help?

it said error

[ornate elk]

Is it the 12gb or the 16gb model?

The 12gb model goes into models/unet
The 16gb into models/checkpoint

[fallen blaze]

ı cant add it

[sharp ridge]

where can i get that 16gb model? can you share the link

[ornate elk]

Here:
https://huggingface.co/Comfy-Org/flux1-dev/blob/main/flux1-dev-fp8.safetensors

[sharp ridge]

thanks!!

[strong mango]

Is that 16gb model the smallest flux version? I have a 1660ti, do i have any hope of using flux?

[south vigil]

where can i find the correct vae i need for this: https://civitai.com/models/14373/nyan-mix?modelVersionId=18151 ?
i found a site where i downloaded a nyan mix vae but my picutres get way too blanc and I really think that i got the wrong one though i cant find it trough the insruction sadly :/
maybe someone can help me

[unborn dagger]

did you manage to get it working?

setting the HF_TOKEN env var to your HF token should fix it

it also seems like you were trying to run it on an AMD GPU. It won't run out of the box because slangtorch has some bugs that doesn't HIPIFY the code properly. We're working on getting rid of the slangtorch dependency altogether, so it should allow you to compile it. However, for AMD, it'll only run on Linux. For Windows, you need pytorch support on Windows, which you won't have with the current HIP SDK on Windows. Sadly you're out of luck for the time being until AMD releases native pytorch support on Windows

[plain bronze]

Can I use Flux in A1111?
https://civitai.com/models/618692/flux

[karmic crown]

I have never heard of it, but generation time in a video on their page takes too long for a cloud-based service. It would be OK for someone's computer, but not for cloud-based service

[ocean hinge]

Hello! I'm completely new to Stable Diffusion, but I believe I have everything set up (I got to the page to generate pictures). I want to generate something similar to this one in terms of art style—it looks perfect, like an AI image I wouldn’t believe was AI-generated. Is there a guide on how to use Stable Diffusion that I should look into, or any videos I should watch? I tried YouTube guides, but I’m getting nowhere with them.

[plush spire]

How easy is it to install SD on a Windows PC and then access it from a Mac web browser? Does it involve setting up an Apache web server on the PC or is it easier than that? Ok never mind, I asked AI and it says: The webui.py script that comes with Stable Diffusion usually sets up its own web server, which you can access directly from your Mac.

[plush spire]

I recommend you looking at the images on civitai.com and seeing what prompts and other settings they use to generate the image. Just copy everything to start with and then experiment yourself.

[ornate elk]

There is a 12gb variant too

[ornate elk]

Currently not

[plain bronze]

Can I use fp32 FLUX with 8gb vram and 32gb ram?

In ComfyUI

[plain bronze]

https://civitai.com/models/618692/flux

What is the difference then?

Between f32, 16 and 8 flux.

I don't get it. I don't know what decimal place is.

What fp version do you use and why?

[fast loom]

I did manage to get it working. I'm on Win11 with a 3090. So it wasn't that actually.

The solution was actually:
python -X utf8 gradio_app.py

It needed that -X and utf8 parameters to launch properly.

https://github.com/Stability-AI/stable-fast-3d/issues/15

[plain bronze]

Oh, I got it, thank you. So fp16 is enough for me.

Who are using fp32 and 64 and why?

[safe inlet]

dont think there are fp32 out there, the one you linked and the one they released is fp16

[safe inlet]

huh

[plain bronze]

Do you know where can I download fp16?

[safe inlet]

ya even though that says fp32 i dont think it is, looks the same size as their released model

[plain bronze]

I think Black Forest Labs in Civitai but there is no fp16 or fp8 version.

https://civitai.com/models/618692/flux

[unborn dagger]

oh cool. glad you got it working. My bad, thought you were running AMD 🙂

[plain bronze]

What is the difference?

[eager raptor]

How do i get all models ID's?

[strong mango]

Can 6 vram handle that?

[eager raptor]

Like this one: stable-diffusion-xl-1024-v1-0

[ornate elk]

i dont think so

[ornate elk]

thats a model name

do you mean model hash?

[ruby yoke]

read the page, it tells you what the difference is

[eager raptor]

Yeah, the model name, but in that format

Where do I find those

[ornate elk]

on Huggingface or civitai.com

there is no list to scrap

you cant find "all" models

[eager raptor]

Idea is that I need the exact names from stability. It's for API

[ornate elk]

then you need the names from their models published on huggingface

[eager raptor]

Do you guys have a list?

Or a way to search them?

[plain bronze]

"Generated outputs can be used for personal, scientific, and commercial purposes"

This is the difference?

[ornate elk]

https://huggingface.co/stabilityai here under models

[plain bronze]

So schnell version have better results?

[eager raptor]

Thank you

[ornate elk]

dev has better results, schnell is just faster

[plain bronze]

You can't use schnell for commercial and you can use dev for it?

For example, you are a Youtuber, which one can you use in your vids?

Or which one you can sell and make money legally?

[ornate elk]

i think the dev one

[dim fossil]

do i need all of those?

[ornate elk]

Go for the SD3 medium incl clips

You only need that

[dim fossil]

ok thanks

[austere pine]

Hey! Is it normal for A1111 to take 60-120 seconds to load a model or switch models (switching takes especially long) on a SATA SSD?

[plain bronze]

How to add batch count to ComfyUI. There is Queue size for it but previous images disappear when I use that feature. Need batch countlike thing for ComfyUI.

[pale salmon]

sup chat, why upscale by are locked/grey ? cant use

[pale salmon]

https://tenor.com/view/nervous-scared-hands-shaking-shivering-gif-18714091

[quick forge]

does anyone know what I need, or need to do to access stable diffusion/A1111 from any web browser?

I have it running on my linux machin ebut would really like to be able to access it from any browser wherever I am

[ornate elk]

try set hires steps to 10 and denois to 0.5 and check if its still gray.
if yes then set the resolution sliders both to 0

[pale salmon]

where is res sliders ?

[wispy pond]

Trying to work with ultrapixel and stable cascade, getting a few errors:

On ultrapixel I was getting this at first. The images seemed to be working fine though:
_
Error occurred when executing PreviewImage:

'NoneType' object is not subscriptable
_
When I couldn't figure out how to add VAE to the comfyui-ultrapixel installation though, I tried to go use stable cascade. When using that I got this error:
_
Error occurred when executing CheckpointLoaderSimple:

ERROR: Could not detect model type of: C:\Users\Toniko\Documents\ComfyUI\ComfyUI_windows_portable\ComfyUI\models\checkpoints\stage_b.safetensors
_
So I went back to ultrapixel, now I'm getting this error:

_
Error occurred when executing UltraPixelProcess:

Trying to set a tensor of shape torch.Size([2560]) in "bias" (which has shape torch.Size([2304])), this looks incorrect.
_

[pale salmon]

dang

[ornate elk]

only when the two sliders are on 0 you can use upscale by

[pale salmon]

didnt know my bad ty

[ornate elk]

no problem at all

[pale salmon]

i also tryed medram, but it crash in 2min while prompting, with a error who said that my vram is full lol

lets wait i guess

[ornate elk]

What's your GPU?

[pale salmon]

GTX 1080 NVIDIA

but I didn't put thermal paste for years, so my pc is kinda burning and slow

will put some this week

[ornate elk]

Then make sure you have --xformers --medvram-sdxl --no-half-vae
In the webui-user.bat

And only these 3

With that you get the best performance on your GPU.

[pale salmon]

kk ty

bruv

@ornate elkits maybe bc i watch youtube videos idk

[ornate elk]

What are your txt2img settings?

And what's the model

Make sure you close other Programms like games and Wallpaper engine and multiple Browser tabs

[pale salmon]

i close opera gx ?

[ornate elk]

Also make sure operas VPN is disabled

That can cause issues too

Also If you use hires fix always set the hires steps to 10

If not it can run out of vram

[pale salmon]

okay, and btw i use autismmixsdxl, and i use nsfw txt2img, dont think i can send it here lol

[ornate elk]

Okay np. Sdxl models are very resource hungry.
To upscale them without running out of vram you also have to install the tiled vae extension.

I'm off now. Its very late here.

[pale salmon]

tiled vae ? btw i trying again and same, pc crash with discord etc and cmd close

[ornate elk]

The whol PC crashes?

How much ram do you have?

[pale salmon]

yeah, and work fine after secondes

got 16GO

[ornate elk]

Ah okay, you need to increase your Windows Pagefile then

Can help with that tomorrow

[cosmic furnace]

@ornate elk Is forge good & fine to run on AMD now? or nah?

[pale salmon]

bruv in mid prompt

[pale salmon]

fine lol i will ask u tmr gn

[lethal kiln]

Why does this error occur?

[arctic hatch]

i'm facing this issue when using animatediff & SD will auto disconnected

[fathom rose]

how do i get the dualcliploader type as flux? can somebody help me?

worked thanks

[late tree]

Does anyone have a clue what to do about this? Both ADetailer and Controlnet are saying:
AttributeError: 'OptionInfo' object has no attribute 'needs_reload_ui'
Webui and extensions are updated, but I cant use either extension because they keep giving that error.

[ornate elk]

Have you clicked Reload ui in the settings?

[late tree]

That's coming up in the log while starting the UI.

[ornate elk]

Maybe delete and reinstall them

But make sure your controlnet models are not in the extensions folder

[crude trellis]

is there any way I can save stable diffusion settings so I don't need to manually setup everytime? like sampling methods, sampling steps, cfg scale

[ornate elk]

There is an extension for that but I forgot its name

[honest quail]

what are poses and how do I use them

[elder mauve]

can i use rocm on windows using WSL?

[ornate elk]

Yes thats possible

What's your GPU?

[elder mauve]

7800xt

are there any other options?

rn i run SD with directml , and its very slow

[ornate elk]

Oh then you should switch over to use Zluda

[elder mauve]

zluda recently died?

[ornate elk]

Its the best option we have right now for AMD

Its still working

[elder mauve]

how do i make it work?

[ornate elk]

Follow my AMD Automatic1111 with ZLUDA guide from here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides

[elder mauve]

the first one?

[ornate elk]

Yes

[elder mauve]

thanks! checking now

which gpu do u have?

what speed should i expect from my 7800xt?

because on directml it was very slow

[ornate elk]

NP, for any question feel free to ask me.
Im running a 7900xtx

[elder mauve]

and idk what speed its supposed to be at

what speed do u get on your gpu?

[ornate elk]

For 1.5 models
With directml I get 6 it/s
With zluda I get 16 it/s

[elder mauve]

and XL models?

with mine , i get 1 iteration / 16 seconds

and the entire pc freezes when its starting

[ornate elk]

On SDXL like 4-6 it/s

[elder mauve]

how much system RAM do u have?

[ornate elk]

Thats normal for directml. Its just painfully slow and uses the whole vram.

[ornate elk]

32gb

[elder mauve]

same , i got 64GB to fix it , it kinda fixed the freeze issue

but then my ram wasnt doing XMP

cuz different kits

so switched back to 32GB

[ornate elk]

Me to haha

[elder mauve]

it was doing 2933 Mhz 64GB

and now its 3200 on 32GB

[ornate elk]

I tried to get 64 working with 6000mhz but nope not even at 4500 it worked

So I stick to 32gb expo 6000mhz

[elder mauve]

idk if it makes a big diff , but people in pc servers told me to use 32 3200 instead

[elder mauve]

ddr5?

oo nice!

mine is ddr4

[ornate elk]

If you have an AM4 AMD CPU the sweetspot is 3600mhz for DDR4

[elder mauve]

i do have an am4 cpu in my second pc... but

its old

[ornate elk]

For Intel its 3200mhz

[elder mauve]

😂 i am not sure if its even am4 , it fits in am4 socket tho

[ornate elk]

Then it is ^^

[elder mauve]

ancient lol

on my main pc i have intel 13500

[ornate elk]

True

Ah ok

[elder mauve]

in the guide it says i have uncheck the pro driver option while downloading HIP SDK

i dont see an option to do that

the download file says pro

[ornate elk]

Thats fine

Just don't tick pro when installing

Its off by default

[elder mauve]

got it

so i have done the install

now trying to run

i get this error

@ornate elk

[ornate elk]

You installed the wrong python version

[elder mauve]

ah

[ornate elk]

You need 3.10.11 64bit

Also make sure you don't have any third party antivirus Programm active as they break the python install

[elder mauve]

do i remove the current one

or install both

[ornate elk]

Yes

Remove it

[elder mauve]

gotcha

no error now 😄

[ornate elk]

Easy ^^

[elder mauve]

except something broke while installing

says cant find cmake in path

when i ran webui bat

[ornate elk]

Full cmd log please

Also never run the webui.bat
Always and only run the webui-user.bat

[elder mauve]

i re ran it , no error so far

i think what happened before , it was still using py 3.12

from anaconda

i uninstalled normal python earlier but forgot to remove anaconda

[ornate elk]

Ahh

[elder mauve]

now i removed anaconda

🤔 and even after that it was using 3.12

cuz a temp venv folder was there in the sd directory

😂

now i removed that and it says

Skipping onnxruntime installation. You are up to date with the most recent release. no module 'xformers'. Processing without... no module 'xformers'. Processing without... No module 'xformers'. Proceeding without it. D:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\pytorch_lightning\utilities\distributed.py:258: LightningDeprecationWarning: pytorch_lightning.utilities.distributed.rank_zero_onlyhas been deprecated in v1.8.1 and will be removed in v2.0.0. You can import it frompytorch_lightning.utilities instead. rank_zero_deprecation( Launching Web UI with arguments: --use-zluda --update-check --skip-ort

[ornate elk]

Looks good

[elder mauve]

these "no module" is fine?

[ornate elk]

Yep, xformers is for nvidia

[elder mauve]

Oh

u know so much

[ornate elk]

Thx ^^ im using SD since the beginning and also made the switch from Nvidia to AMD

[elder mauve]

i got this

[ornate elk]

Have you restarted the PC after Installing HIP SDK?

[elder mauve]

no

let me do that now

restarted and ran again

venv "D:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\Scripts\Python.exe" WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next. Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr 5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)] Version: v1.10.1-amd-2-g395ce8dc Commit hash: 395ce8dc2cb01282d48074a89a5e6cb3da4b59ab Using ZLUDA in D:\SD-Zluda\stable-diffusion-webui-amdgpu\.zluda Skipping onnxruntime installation. You are up to date with the most recent release. no module 'xformers'. Processing without... no module 'xformers'. Processing without... No module 'xformers'. Proceeding without it. D:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\pytorch_lightning\utilities\distributed.py:258: LightningDeprecationWarning: pytorch_lightning.utilities.distributed.rank_zero_onlyhas been deprecated in v1.8.1 and will be removed in v2.0.0. You can import it frompytorch_lightning.utilities instead. rank_zero_deprecation( Launching Web UI with arguments: --use-zluda --update-check --skip-ort Press any key to continue . . .

[ornate elk]

Is your AMD Driver updated?

[elder mauve]

its latest

adrenalin driver right?

24.7.1

[ornate elk]

Okay good

Then how do your environment paths settings look like?

[late tree]

I've tried that too. Might have something to do with the .git folder will have to see after work

[elder mauve]

system path

[ornate elk]

Looks correct

[elder mauve]

user path

[ornate elk]

Ohhh now I see

You installed hip 5.7

But you need 6.1

[elder mauve]

i installed the latest one 🤔

[ornate elk]

Did you had 5.7 installed before?

[elder mauve]

yes

[ornate elk]

Because here its listed

[elder mauve]

i had 5.7 before , now i installed the latest one

i have to remove the old one?

[ornate elk]

Yes

[elder mauve]

or just remove the old path variable?

[ornate elk]

The old path should be enough

If not remove the old version fully

[elder mauve]

done

let me try

this looks bad

forgot to clear venv

or is it unrelated to that?

[ornate elk]

You may need to delete the .zluda folder from the stable-diffusion-webui folder
Ans then relaunch

[elder mauve]

gotcha

just .zluda or venv too?

[ornate elk]

Both wouldn't hurt

[elder mauve]

tmp too?

[ornate elk]

Not needed. But you can open up a cmd and type
Pip cache purge before relaunching

[elder mauve]

😮

btw have u tried fooocus?

what would u say about it with comparison to automatic1111 and comfyui

[ornate elk]

Sure, i ve tried any webui that ive made a guide for

Fooocus isnt that great. It only supports sdxl models. So for the low end users or directml its a nogo.

But its good for people that just want to have some good images and dont want to dive deep into SD

[elder mauve]

because with it u dont need to focus on prompting , thats why its good for new users?

otherwise , how do u decide which prompts to use?

[ornate elk]

Yep. It has pre made styles you can just select

After you got auto1111 to work you can easily install fooocus too and test if you want

[elder mauve]

i was using fooocus with directml before

[ornate elk]

Oh sounds painful

[elder mauve]

never tried automatic1111

[ornate elk]

It would have been a much better experience

[elder mauve]

asrock has smth called AIquickset

[ornate elk]

Even on directml

[ornate elk]

Hmm you should uninstall rocm 5.7 and 6.1 then.
Restart the PC. Then install 6.1 again.
Then restart the PC again.
Then delete the .zluda folder again and relaunch the webui-user.bat

[elder mauve]

gotcha

[ornate elk]

Strange error

[elder mauve]

🤔

what to do now

[ornate elk]

Wait

You launched it without the launch args

You have to run the webui-user.bat

With the --use-zluda arg in there

[elder mauve]

that would be a stupid mistake

cleared zluda and venv folders and tried to run it again , now i got this

https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu/issues/514

is this related?

[ornate elk]

No as your GPU is newer

[elder mauve]

😮

[ornate elk]

Try open up a cmd as admin and run
SFC /scannow

[elder mauve]

`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 did not find any integrity violations.`

[ornate elk]

Okay

[ornate elk]

And you downloaded the latest zluda files 3.8.1 ?

[elder mauve]

yes from the guide link

[ornate elk]

Hmm

[elder mauve]

wait

i had another zluda before

did i have it in path tho

[ornate elk]

Yea please recheck the paths again

[elder mauve]

yeah there is only 1 zluda in path

just checked

and its the one i downloaded now

[ornate elk]

Make sure the hip path is still there

[elder mauve]

yeah its there

[ornate elk]

Can you show me the system variables again?

[elder mauve]

sure

just noticed this btw

[ornate elk]

Remove that 5.7

[elder mauve]

these are my system PATH variables

these are rest of the variables

[ornate elk]

Looks okay.

Restart the PC again

[elder mauve]

alr

restarted

what next?

[ornate elk]

Delete the .zluda folder again and relaunch

[elder mauve]

gotcha

same

[ornate elk]

:/

[elder mauve]

😦

[ornate elk]

Can you open an issue on the stable-diffusion-webui-amdgpu github ?

[elder mauve]

sure

[ornate elk]

The dev mostly can tell what's missing here.

When I'm at home I will recheck how it looks for me and then we can compare again

[elder mauve]

thanks!

[ornate elk]

Your GPU was a 7800xt right?

[elder mauve]

yes

[ornate elk]

That should be supported by hip 6.1 natively

[elder mauve]

so why do u think its happening

there is one thing i can do

boot into a fresh install of windows and check there

🤔

what u think?

i have a nvme that i can use for this

[ornate elk]

Thats an Option you could try yes.
Then only install the latest AMD Adrenalin Driver and then follow the Guide again.

[ornate elk]

Something can't access the hip DLL correctly. Or is not compatible

[elder mauve]

my SD-Zluda folder is in another drive than C

i dont think thats causing the issue?

[ornate elk]

That should be no issue

I have it on D too

[elder mauve]

hmmmmmm

there is another ZLUDA folder in that D drive

named ZLUDA

i dont think that should cause issues either?

[ornate elk]

Are you on windows 10 or 11 ?

[ornate elk]

I dont think so

[elder mauve]

11

[ornate elk]

Okay

Driving home now

[dense sandal]

I got a new computer with a 3090 and everything runs fine except when I go to vid2vid at max resolution (high res script with 1.3x) and then after a minute or two the sreen monitors go black and the GPU continues to work (I can hear the fan at full speed) I have to shut down the computer because there is no way to turn the monitors back on.

I don't know what can cause this, the PSU is 1000 W corsair, and the GPU is turbo which as I understand it means that it has only one powerful fan. I also have 32GB RAM which may not be sufficient as before the high res scrip it uses like 31 GB, but it should go and use the pagefile if not enough right?

I did a test with a software to demand everything and it run fine, BUT, I noticed it only does clock speed at maximum but it doesn't load anything on the Vram, which is very much used when using comfyUI

So it's not a reliable test.

[ornate elk]

how does your AMD ROCm folder looks like?

[elder mauve]

which one?

[ornate elk]

C programm files amd

rocm

[ornate elk]

i mena one folder back

is there only 6.1 ?

[elder mauve]

yeah

[ornate elk]

good

how big is that folder?

[elder mauve]

5.14GB

[elder mauve]

same files

[ornate elk]

dont use hires fix while using vid2vid. its just to much. the vram gets maxed out and then you get a black screen

[ornate elk]

your Zluda folder is in C right?

[elder mauve]

yes

[south vigil]

yoo, so Ive made 75 highres prompts over night, the thing is that every time i give stable diffusion more than 1 batch to work on (WITH HIGHRES - without it always works normal) then the picture one (top left) is coming out good but every pic after that is getting worse and worse until there is only a pixelated picture. (the last one was done today again) what can I do to fix this?

[ornate elk]

im starting to think its a bug. so the best you can do is to open up an issue here.
https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu/issues
And then maybe reinstall Windows on your nvme ssd + adrenalin driver and follw the guide again.

[ornate elk]

hey, what txt2img settings did you used ?

[south vigil]

i added Adetailer for better faces (width 967,height 544) DPM++3M SDE Karras 25 steps CFG 7
Highres: latent nearest 15 steps 0.7 denois upscale by 2

[ornate elk]

okay and you used batch count or batch size?

[south vigil]

batch count

[ornate elk]

and whats your gpu and whats inside your webui-user.bat?

[south vigil]

my gpu is a rtx 3060ti
my .bat:
@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --medvram

call webui.bat

[ornate elk]

change --medvram to --medvram-sdxl and also add --xformers --no-half-vae

then save and relaunch

also try first without adetailer

[south vigil]

alright will do thanks a lot

maybe whats also worth telling is, that i use the lobehub extension on stable diffusion

[ornate elk]

you can try use an other Sampler like DPM++ 2M, some samplers have issues with latent upscalers

also 0.7 isnt a good denois for latent

better set it to 0.5

[south vigil]

alright thanks for the tipps, I will now experiment around and try to find out if it has something to do with the method and also change the denois to 0.5

[south vigil]

so I tried it now out again with 5 different sampling methods on a batch count of 10 sadly it still is the same and gets worse after every picture, Though I found out, that i f i go on a batch size of 8 all 8 of the pictures are coming out good, would you maybe have something else in mind what could cause this problem?

[ornate elk]

Is your webui updated?
Are the extensions updated?

[coarse swan]

Hello there, I'm getting the 'Torch is not able to use GPU' error, not sure how I can fix it.

tried deleting venv, updating my NVIDIA Driver to the latest with GRD, the torch-directml trick, nothing's working so far

my PC is windows 10 with GeForce GTX970 if that helps

[ornate elk]

hey make sure your on the latest gpu driver update for your gpu

then open up a cmd and type
pip cache purge
then also delete the venv folder and rerun the wbeui-user.bat

[coarse swan]

it is

[coarse swan]

I'll try that out once my latest image is done rendering 👍

[plain bronze]

How to add batch count to ComfyUI. There is Queue size for it but previous images disappear when I use that feature. Need batch countlike thing for ComfyUI.

So I can't see them unless open the folder.

But I can see them in browser when I use batch size.

Just not when I use batch count. For batch count, they are disappearing.

It wasn't like that in A1111.

There isn't any solution for this?

I can see them in queue history. Problem solved.

@ornate elk How can I activate xformers in ComfyUI?

[ornate elk]

it already should use xformers by default

check the cmd if you see it

[plain bronze]

How

[ornate elk]

when launching and it states this, then its not using xformers:
Using pytorch cross attention

[ornate elk]

you have auto1111 installed too right?

[plain bronze]

Yes.

[ornate elk]

then you can copy over xformers from it

[plain bronze]

How

[ornate elk]

go into:
stable-diffusion-webui\venv\Lib\site-packages
there copy all xformers folders

then put them into:
ComfyUI_windows_portable\python_embeded\Lib\site-packages

[plain bronze]

Here?

@ornate elk only these two folders and not the torch ones?

[elder mauve]

when i put wsl and rocm , it says my gpu is unrecognized type

[dense sandal]

Thanks but the vram went to 15gb and I have 24. Screen monitors just go black and I have to restart the pc. This happens when its almost finished the entire 11 steps. Whch may indicate that at the end there is a spike of.memory usage or something.

[elder mauve]

[dense sandal]

Also, not using h
The high res fix means I cant make high res videos which is why I bought the pc

[plain bronze]

@ornate elk It still says "using pytorch cross attention".

[ornate elk]

at the end is the vae step. it needs more vram than normaly. if your using sdxl models there is a fp16 vae that fixes that spike. Also there is an extension called Tiled DIffusion. Its Tiled Vae function can fix the issue too

[plain bronze]

pytorch version: 2.3.1+cu121
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
PyTorch 2.1.2+cu121 with CUDA 1201 (you have 2.3.1+cu121)
Python 3.10.11 (you have 3.11.9)
Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
Memory-efficient attention, SwiGLU, sparse and more won't be available.
Set XFORMERS_MORE_DETAILS=1 for more details
xformers version: 0.0.23.post1
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4070 Laptop GPU : cudaMallocAsync
Using pytorch cross attention

[dense sandal]

I am using 1.5 lcm for animate diff vid 2 vid

In comfyUI

[ornate elk]

oh your using animatediff

[dense sandal]

Yes

[ornate elk]

then lower the context batch size

that can help

[plain bronze]

From cmd

[ornate elk]

oh yea that xformers version is to new for the torch of comfyui

i thought comfyui uses 2.3.0

[plain bronze]

What should I do?

[dense sandal]

Even if it only happens at last step of the inference?

[ornate elk]

hmmm. so it fails when creating the video?

[dense sandal]

Yes at the last step

The efficient sampler has an option for tiled vae. Maybe that helps?

[ornate elk]

you can try that

[plain bronze]

https://tenor.com/view/quinzha-gif-11008937155546098498

[dense sandal]

Didnt help.. still monitors shut down and gpu continued with the fan really loud indefinetely

I have to shut down the pc manually

[coarse swan]

doesn't work lol. been using this one my webui-user.bat if it helps

[ornate elk]

what error do you get?

[coarse swan]

currently this

also I ran the pip purge on my script folder

[plain bronze]

Please someone help, I have low vram

[ornate elk]

do you have a venv folder?

also please show your whole webui-user.bat file

[ornate elk]

you could try copy the whole venv folder from auto1111 to comfyui

and i mean delete comfyui venv and copy then the auto1111 venv in there

[pale salmon]

@ornate elk sup bro, maybe i can try this like u said

[ornate elk]

ahhh hey

[coarse swan]

I can paste the bat file here if needed. Also I do have a venv folder.

[ornate elk]

follow this guide until your on the pagefile screen:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Then make sure its only enabled for the C drive and disabled for any other drive.
Then set it to custom sized and set 16000 Min and 24000 Max
Then save and restart your PC

[ornate elk]

from the bat remove all -- args
then only add --medvram --no-half-vae
then save.
Then open up a cmd and type
pip cache purge
then relaunch the webui-user.bat

[pale salmon]

@ornate elki did that already sadly

[ornate elk]

okay what was your issue again?

sry xD

[coarse swan]

the pipcache purge should happen in the script folder's cmd right

[ornate elk]

doesnt matter where

[coarse swan]

then I'm bavk to square one lol

[pale salmon]

my whole pc crash after prompting for some seconds, and stable stop in the command line, yesterday my image loaded but i needed towait like 45min for it, and my stable crashed mid time when prompting lol

[coarse swan]

did purged and it said 753 files removed.

[ornate elk]

this was before you have run pip cache purge?

also delete the venv folder

then relaunch the webui-user.bat

[coarse swan]

before and after

I'll try purging and deleting venv again

[ornate elk]

oh this. it happens onyl when using hires fix right?
have you already installed the tiled diffusion extension?

[pale salmon]

nah i never heard of this lol, and yeah probably my pc doesnt support when i upscale with hires fix maybe

but it work fine on comfyui

[ornate elk]

okay then install that extension via the extensions tab

[pale salmon]

https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111

this ?

[ornate elk]

yes

[pale salmon]

kk ty

[ornate elk]

@coarse swan did it worked? i have a command which maybe can fix it

[coarse swan]

gimme a min, I'm still installing

[ornate elk]

after the install restart the whole webui and cmd.
then only enable Tiled VAE (not tiled diffusion or anything)
set the hires steps to 10 ! and then try to generate

[pale salmon]

do i need to stay with --xformers --medvram-sdxl --no-half-vae too ?

[ornate elk]

yes

[pale salmon]

ok i installed tile i restart cmd and webui

i guess its here somewhere

[plain bronze]

Can this cause an error?

[plain bronze]

And there is no venv folder in comfy

Is there another way to do that?

[ornate elk]

oh true i forgot it doesnt use one nativly

[plain bronze]

https://tenor.com/view/shame-gif-26321140